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/ValueTracking.h"
46 #include "llvm/Target/TargetData.h"
47 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
48 #include "llvm/Transforms/Utils/Local.h"
49 #include "llvm/Support/CallSite.h"
50 #include "llvm/Support/ConstantRange.h"
51 #include "llvm/Support/Debug.h"
52 #include "llvm/Support/ErrorHandling.h"
53 #include "llvm/Support/GetElementPtrTypeIterator.h"
54 #include "llvm/Support/InstVisitor.h"
55 #include "llvm/Support/IRBuilder.h"
56 #include "llvm/Support/MathExtras.h"
57 #include "llvm/Support/PatternMatch.h"
58 #include "llvm/Support/Compiler.h"
59 #include "llvm/Support/raw_ostream.h"
60 #include "llvm/ADT/DenseMap.h"
61 #include "llvm/ADT/SmallVector.h"
62 #include "llvm/ADT/SmallPtrSet.h"
63 #include "llvm/ADT/Statistic.h"
64 #include "llvm/ADT/STLExtras.h"
68 using namespace llvm::PatternMatch;
70 STATISTIC(NumCombined , "Number of insts combined");
71 STATISTIC(NumConstProp, "Number of constant folds");
72 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
73 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
74 STATISTIC(NumSunkInst , "Number of instructions sunk");
77 /// InstCombineWorklist - This is the worklist management logic for
79 class InstCombineWorklist {
80 SmallVector<Instruction*, 256> Worklist;
81 DenseMap<Instruction*, unsigned> WorklistMap;
83 void operator=(const InstCombineWorklist&RHS); // DO NOT IMPLEMENT
84 InstCombineWorklist(const InstCombineWorklist&); // DO NOT IMPLEMENT
86 InstCombineWorklist() {}
88 bool isEmpty() const { return Worklist.empty(); }
90 /// Add - Add the specified instruction to the worklist if it isn't already
92 void Add(Instruction *I) {
93 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
94 Worklist.push_back(I);
97 void AddValue(Value *V) {
98 if (Instruction *I = dyn_cast<Instruction>(V))
102 // Remove - remove I from the worklist if it exists.
103 void Remove(Instruction *I) {
104 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
105 if (It == WorklistMap.end()) return; // Not in worklist.
107 // Don't bother moving everything down, just null out the slot.
108 Worklist[It->second] = 0;
110 WorklistMap.erase(It);
113 Instruction *RemoveOne() {
114 Instruction *I = Worklist.back();
116 WorklistMap.erase(I);
120 /// AddUsersToWorkList - When an instruction is simplified, add all users of
121 /// the instruction to the work lists because they might get more simplified
124 void AddUsersToWorkList(Instruction &I) {
125 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
127 Add(cast<Instruction>(*UI));
131 /// Zap - check that the worklist is empty and nuke the backing store for
132 /// the map if it is large.
134 assert(WorklistMap.empty() && "Worklist empty, but map not?");
136 // Do an explicit clear, this shrinks the map if needed.
140 } // end anonymous namespace.
144 /// InstCombineIRInserter - This is an IRBuilder insertion helper that works
145 /// just like the normal insertion helper, but also adds any new instructions
146 /// to the instcombine worklist.
147 class InstCombineIRInserter : public IRBuilderDefaultInserter<true> {
148 InstCombineWorklist &Worklist;
150 InstCombineIRInserter(InstCombineWorklist &WL) : Worklist(WL) {}
152 void InsertHelper(Instruction *I, const Twine &Name,
153 BasicBlock *BB, BasicBlock::iterator InsertPt) const {
154 IRBuilderDefaultInserter<true>::InsertHelper(I, Name, BB, InsertPt);
158 } // end anonymous namespace
162 class VISIBILITY_HIDDEN InstCombiner
163 : public FunctionPass,
164 public InstVisitor<InstCombiner, Instruction*> {
166 bool MustPreserveLCSSA;
168 /// Worklist - All of the instructions that need to be simplified.
169 InstCombineWorklist Worklist;
171 /// Builder - This is an IRBuilder that automatically inserts new
172 /// instructions into the worklist when they are created.
173 typedef IRBuilder<true, ConstantFolder, InstCombineIRInserter> BuilderTy;
176 static char ID; // Pass identification, replacement for typeid
177 InstCombiner() : FunctionPass(&ID), TD(0), Builder(0) {}
179 LLVMContext *Context;
180 LLVMContext *getContext() const { return Context; }
183 virtual bool runOnFunction(Function &F);
185 bool DoOneIteration(Function &F, unsigned ItNum);
187 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
188 AU.addPreservedID(LCSSAID);
189 AU.setPreservesCFG();
192 TargetData *getTargetData() const { return TD; }
194 // Visitation implementation - Implement instruction combining for different
195 // instruction types. The semantics are as follows:
197 // null - No change was made
198 // I - Change was made, I is still valid, I may be dead though
199 // otherwise - Change was made, replace I with returned instruction
201 Instruction *visitAdd(BinaryOperator &I);
202 Instruction *visitFAdd(BinaryOperator &I);
203 Instruction *visitSub(BinaryOperator &I);
204 Instruction *visitFSub(BinaryOperator &I);
205 Instruction *visitMul(BinaryOperator &I);
206 Instruction *visitFMul(BinaryOperator &I);
207 Instruction *visitURem(BinaryOperator &I);
208 Instruction *visitSRem(BinaryOperator &I);
209 Instruction *visitFRem(BinaryOperator &I);
210 bool SimplifyDivRemOfSelect(BinaryOperator &I);
211 Instruction *commonRemTransforms(BinaryOperator &I);
212 Instruction *commonIRemTransforms(BinaryOperator &I);
213 Instruction *commonDivTransforms(BinaryOperator &I);
214 Instruction *commonIDivTransforms(BinaryOperator &I);
215 Instruction *visitUDiv(BinaryOperator &I);
216 Instruction *visitSDiv(BinaryOperator &I);
217 Instruction *visitFDiv(BinaryOperator &I);
218 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
219 Instruction *FoldAndOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
220 Instruction *visitAnd(BinaryOperator &I);
221 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
222 Instruction *FoldOrOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
223 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
224 Value *A, Value *B, Value *C);
225 Instruction *visitOr (BinaryOperator &I);
226 Instruction *visitXor(BinaryOperator &I);
227 Instruction *visitShl(BinaryOperator &I);
228 Instruction *visitAShr(BinaryOperator &I);
229 Instruction *visitLShr(BinaryOperator &I);
230 Instruction *commonShiftTransforms(BinaryOperator &I);
231 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
233 Instruction *visitFCmpInst(FCmpInst &I);
234 Instruction *visitICmpInst(ICmpInst &I);
235 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
236 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
239 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
240 ConstantInt *DivRHS);
242 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
243 ICmpInst::Predicate Cond, Instruction &I);
244 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
246 Instruction *commonCastTransforms(CastInst &CI);
247 Instruction *commonIntCastTransforms(CastInst &CI);
248 Instruction *commonPointerCastTransforms(CastInst &CI);
249 Instruction *visitTrunc(TruncInst &CI);
250 Instruction *visitZExt(ZExtInst &CI);
251 Instruction *visitSExt(SExtInst &CI);
252 Instruction *visitFPTrunc(FPTruncInst &CI);
253 Instruction *visitFPExt(CastInst &CI);
254 Instruction *visitFPToUI(FPToUIInst &FI);
255 Instruction *visitFPToSI(FPToSIInst &FI);
256 Instruction *visitUIToFP(CastInst &CI);
257 Instruction *visitSIToFP(CastInst &CI);
258 Instruction *visitPtrToInt(PtrToIntInst &CI);
259 Instruction *visitIntToPtr(IntToPtrInst &CI);
260 Instruction *visitBitCast(BitCastInst &CI);
261 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
263 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
264 Instruction *visitSelectInst(SelectInst &SI);
265 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
266 Instruction *visitCallInst(CallInst &CI);
267 Instruction *visitInvokeInst(InvokeInst &II);
268 Instruction *visitPHINode(PHINode &PN);
269 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
270 Instruction *visitAllocationInst(AllocationInst &AI);
271 Instruction *visitFreeInst(FreeInst &FI);
272 Instruction *visitLoadInst(LoadInst &LI);
273 Instruction *visitStoreInst(StoreInst &SI);
274 Instruction *visitBranchInst(BranchInst &BI);
275 Instruction *visitSwitchInst(SwitchInst &SI);
276 Instruction *visitInsertElementInst(InsertElementInst &IE);
277 Instruction *visitExtractElementInst(ExtractElementInst &EI);
278 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
279 Instruction *visitExtractValueInst(ExtractValueInst &EV);
281 // visitInstruction - Specify what to return for unhandled instructions...
282 Instruction *visitInstruction(Instruction &I) { return 0; }
285 Instruction *visitCallSite(CallSite CS);
286 bool transformConstExprCastCall(CallSite CS);
287 Instruction *transformCallThroughTrampoline(CallSite CS);
288 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
289 bool DoXform = true);
290 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
291 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
295 // InsertNewInstBefore - insert an instruction New before instruction Old
296 // in the program. Add the new instruction to the worklist.
298 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
299 assert(New && New->getParent() == 0 &&
300 "New instruction already inserted into a basic block!");
301 BasicBlock *BB = Old.getParent();
302 BB->getInstList().insert(&Old, New); // Insert inst
307 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
308 /// This also adds the cast to the worklist. Finally, this returns the
310 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
312 if (V->getType() == Ty) return V;
314 if (Constant *CV = dyn_cast<Constant>(V))
315 return ConstantExpr::getCast(opc, CV, Ty);
317 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
322 // ReplaceInstUsesWith - This method is to be used when an instruction is
323 // found to be dead, replacable with another preexisting expression. Here
324 // we add all uses of I to the worklist, replace all uses of I with the new
325 // value, then return I, so that the inst combiner will know that I was
328 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
329 Worklist.AddUsersToWorkList(I); // Add all modified instrs to worklist.
331 // If we are replacing the instruction with itself, this must be in a
332 // segment of unreachable code, so just clobber the instruction.
334 V = UndefValue::get(I.getType());
336 I.replaceAllUsesWith(V);
340 // EraseInstFromFunction - When dealing with an instruction that has side
341 // effects or produces a void value, we can't rely on DCE to delete the
342 // instruction. Instead, visit methods should return the value returned by
344 Instruction *EraseInstFromFunction(Instruction &I) {
345 assert(I.use_empty() && "Cannot erase instruction that is used!");
346 // Make sure that we reprocess all operands now that we reduced their
348 if (I.getNumOperands() < 8) {
349 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
350 if (Instruction *Op = dyn_cast<Instruction>(*i))
355 return 0; // Don't do anything with FI
358 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
359 APInt &KnownOne, unsigned Depth = 0) const {
360 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
363 bool MaskedValueIsZero(Value *V, const APInt &Mask,
364 unsigned Depth = 0) const {
365 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
367 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
368 return llvm::ComputeNumSignBits(Op, TD, Depth);
373 /// SimplifyCommutative - This performs a few simplifications for
374 /// commutative operators.
375 bool SimplifyCommutative(BinaryOperator &I);
377 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
378 /// most-complex to least-complex order.
379 bool SimplifyCompare(CmpInst &I);
381 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
382 /// based on the demanded bits.
383 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
384 APInt& KnownZero, APInt& KnownOne,
386 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
387 APInt& KnownZero, APInt& KnownOne,
390 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
391 /// SimplifyDemandedBits knows about. See if the instruction has any
392 /// properties that allow us to simplify its operands.
393 bool SimplifyDemandedInstructionBits(Instruction &Inst);
395 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
396 APInt& UndefElts, unsigned Depth = 0);
398 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
399 // PHI node as operand #0, see if we can fold the instruction into the PHI
400 // (which is only possible if all operands to the PHI are constants).
401 Instruction *FoldOpIntoPhi(Instruction &I);
403 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
404 // operator and they all are only used by the PHI, PHI together their
405 // inputs, and do the operation once, to the result of the PHI.
406 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
407 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
408 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
411 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
412 ConstantInt *AndRHS, BinaryOperator &TheAnd);
414 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
415 bool isSub, Instruction &I);
416 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
417 bool isSigned, bool Inside, Instruction &IB);
418 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
419 Instruction *MatchBSwap(BinaryOperator &I);
420 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
421 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
422 Instruction *SimplifyMemSet(MemSetInst *MI);
425 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
427 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
428 unsigned CastOpc, int &NumCastsRemoved);
429 unsigned GetOrEnforceKnownAlignment(Value *V,
430 unsigned PrefAlign = 0);
433 } // end anonymous namespace
435 char InstCombiner::ID = 0;
436 static RegisterPass<InstCombiner>
437 X("instcombine", "Combine redundant instructions");
439 // getComplexity: Assign a complexity or rank value to LLVM Values...
440 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
441 static unsigned getComplexity(Value *V) {
442 if (isa<Instruction>(V)) {
443 if (BinaryOperator::isNeg(V) ||
444 BinaryOperator::isFNeg(V) ||
445 BinaryOperator::isNot(V))
449 if (isa<Argument>(V)) return 3;
450 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
453 // isOnlyUse - Return true if this instruction will be deleted if we stop using
455 static bool isOnlyUse(Value *V) {
456 return V->hasOneUse() || isa<Constant>(V);
459 // getPromotedType - Return the specified type promoted as it would be to pass
460 // though a va_arg area...
461 static const Type *getPromotedType(const Type *Ty) {
462 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
463 if (ITy->getBitWidth() < 32)
464 return Type::getInt32Ty(Ty->getContext());
469 /// getBitCastOperand - If the specified operand is a CastInst, a constant
470 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
471 /// operand value, otherwise return null.
472 static Value *getBitCastOperand(Value *V) {
473 if (Operator *O = dyn_cast<Operator>(V)) {
474 if (O->getOpcode() == Instruction::BitCast)
475 return O->getOperand(0);
476 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
477 if (GEP->hasAllZeroIndices())
478 return GEP->getPointerOperand();
483 /// This function is a wrapper around CastInst::isEliminableCastPair. It
484 /// simply extracts arguments and returns what that function returns.
485 static Instruction::CastOps
486 isEliminableCastPair(
487 const CastInst *CI, ///< The first cast instruction
488 unsigned opcode, ///< The opcode of the second cast instruction
489 const Type *DstTy, ///< The target type for the second cast instruction
490 TargetData *TD ///< The target data for pointer size
493 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
494 const Type *MidTy = CI->getType(); // B from above
496 // Get the opcodes of the two Cast instructions
497 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
498 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
500 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
502 TD ? TD->getIntPtrType(CI->getContext()) : 0);
504 // We don't want to form an inttoptr or ptrtoint that converts to an integer
505 // type that differs from the pointer size.
506 if ((Res == Instruction::IntToPtr &&
507 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
508 (Res == Instruction::PtrToInt &&
509 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
512 return Instruction::CastOps(Res);
515 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
516 /// in any code being generated. It does not require codegen if V is simple
517 /// enough or if the cast can be folded into other casts.
518 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
519 const Type *Ty, TargetData *TD) {
520 if (V->getType() == Ty || isa<Constant>(V)) return false;
522 // If this is another cast that can be eliminated, it isn't codegen either.
523 if (const CastInst *CI = dyn_cast<CastInst>(V))
524 if (isEliminableCastPair(CI, opcode, Ty, TD))
529 // SimplifyCommutative - This performs a few simplifications for commutative
532 // 1. Order operands such that they are listed from right (least complex) to
533 // left (most complex). This puts constants before unary operators before
536 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
537 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
539 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
540 bool Changed = false;
541 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
542 Changed = !I.swapOperands();
544 if (!I.isAssociative()) return Changed;
545 Instruction::BinaryOps Opcode = I.getOpcode();
546 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
547 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
548 if (isa<Constant>(I.getOperand(1))) {
549 Constant *Folded = ConstantExpr::get(I.getOpcode(),
550 cast<Constant>(I.getOperand(1)),
551 cast<Constant>(Op->getOperand(1)));
552 I.setOperand(0, Op->getOperand(0));
553 I.setOperand(1, Folded);
555 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
556 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
557 isOnlyUse(Op) && isOnlyUse(Op1)) {
558 Constant *C1 = cast<Constant>(Op->getOperand(1));
559 Constant *C2 = cast<Constant>(Op1->getOperand(1));
561 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
562 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
563 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
567 I.setOperand(0, New);
568 I.setOperand(1, Folded);
575 /// SimplifyCompare - For a CmpInst this function just orders the operands
576 /// so that theyare listed from right (least complex) to left (most complex).
577 /// This puts constants before unary operators before binary operators.
578 bool InstCombiner::SimplifyCompare(CmpInst &I) {
579 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
582 // Compare instructions are not associative so there's nothing else we can do.
586 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
587 // if the LHS is a constant zero (which is the 'negate' form).
589 static inline Value *dyn_castNegVal(Value *V) {
590 if (BinaryOperator::isNeg(V))
591 return BinaryOperator::getNegArgument(V);
593 // Constants can be considered to be negated values if they can be folded.
594 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
595 return ConstantExpr::getNeg(C);
597 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
598 if (C->getType()->getElementType()->isInteger())
599 return ConstantExpr::getNeg(C);
604 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
605 // instruction if the LHS is a constant negative zero (which is the 'negate'
608 static inline Value *dyn_castFNegVal(Value *V) {
609 if (BinaryOperator::isFNeg(V))
610 return BinaryOperator::getFNegArgument(V);
612 // Constants can be considered to be negated values if they can be folded.
613 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
614 return ConstantExpr::getFNeg(C);
616 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
617 if (C->getType()->getElementType()->isFloatingPoint())
618 return ConstantExpr::getFNeg(C);
623 static inline Value *dyn_castNotVal(Value *V) {
624 if (BinaryOperator::isNot(V))
625 return BinaryOperator::getNotArgument(V);
627 // Constants can be considered to be not'ed values...
628 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
629 return ConstantInt::get(C->getType(), ~C->getValue());
633 // dyn_castFoldableMul - If this value is a multiply that can be folded into
634 // other computations (because it has a constant operand), return the
635 // non-constant operand of the multiply, and set CST to point to the multiplier.
636 // Otherwise, return null.
638 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
639 if (V->hasOneUse() && V->getType()->isInteger())
640 if (Instruction *I = dyn_cast<Instruction>(V)) {
641 if (I->getOpcode() == Instruction::Mul)
642 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
643 return I->getOperand(0);
644 if (I->getOpcode() == Instruction::Shl)
645 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
646 // The multiplier is really 1 << CST.
647 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
648 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
649 CST = ConstantInt::get(V->getType()->getContext(),
650 APInt(BitWidth, 1).shl(CSTVal));
651 return I->getOperand(0);
657 /// AddOne - Add one to a ConstantInt
658 static Constant *AddOne(Constant *C) {
659 return ConstantExpr::getAdd(C,
660 ConstantInt::get(C->getType(), 1));
662 /// SubOne - Subtract one from a ConstantInt
663 static Constant *SubOne(ConstantInt *C) {
664 return ConstantExpr::getSub(C,
665 ConstantInt::get(C->getType(), 1));
667 /// MultiplyOverflows - True if the multiply can not be expressed in an int
669 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
670 uint32_t W = C1->getBitWidth();
671 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
680 APInt MulExt = LHSExt * RHSExt;
683 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
684 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
685 return MulExt.slt(Min) || MulExt.sgt(Max);
687 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
691 /// ShrinkDemandedConstant - Check to see if the specified operand of the
692 /// specified instruction is a constant integer. If so, check to see if there
693 /// are any bits set in the constant that are not demanded. If so, shrink the
694 /// constant and return true.
695 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
697 assert(I && "No instruction?");
698 assert(OpNo < I->getNumOperands() && "Operand index too large");
700 // If the operand is not a constant integer, nothing to do.
701 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
702 if (!OpC) return false;
704 // If there are no bits set that aren't demanded, nothing to do.
705 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
706 if ((~Demanded & OpC->getValue()) == 0)
709 // This instruction is producing bits that are not demanded. Shrink the RHS.
710 Demanded &= OpC->getValue();
711 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Demanded));
715 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
716 // set of known zero and one bits, compute the maximum and minimum values that
717 // could have the specified known zero and known one bits, returning them in
719 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
720 const APInt& KnownOne,
721 APInt& Min, APInt& Max) {
722 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
723 KnownZero.getBitWidth() == Min.getBitWidth() &&
724 KnownZero.getBitWidth() == Max.getBitWidth() &&
725 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
726 APInt UnknownBits = ~(KnownZero|KnownOne);
728 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
729 // bit if it is unknown.
731 Max = KnownOne|UnknownBits;
733 if (UnknownBits.isNegative()) { // Sign bit is unknown
734 Min.set(Min.getBitWidth()-1);
735 Max.clear(Max.getBitWidth()-1);
739 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
740 // a set of known zero and one bits, compute the maximum and minimum values that
741 // could have the specified known zero and known one bits, returning them in
743 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
744 const APInt &KnownOne,
745 APInt &Min, APInt &Max) {
746 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
747 KnownZero.getBitWidth() == Min.getBitWidth() &&
748 KnownZero.getBitWidth() == Max.getBitWidth() &&
749 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
750 APInt UnknownBits = ~(KnownZero|KnownOne);
752 // The minimum value is when the unknown bits are all zeros.
754 // The maximum value is when the unknown bits are all ones.
755 Max = KnownOne|UnknownBits;
758 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
759 /// SimplifyDemandedBits knows about. See if the instruction has any
760 /// properties that allow us to simplify its operands.
761 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
762 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
763 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
764 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
766 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
767 KnownZero, KnownOne, 0);
768 if (V == 0) return false;
769 if (V == &Inst) return true;
770 ReplaceInstUsesWith(Inst, V);
774 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
775 /// specified instruction operand if possible, updating it in place. It returns
776 /// true if it made any change and false otherwise.
777 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
778 APInt &KnownZero, APInt &KnownOne,
780 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
781 KnownZero, KnownOne, Depth);
782 if (NewVal == 0) return false;
788 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
789 /// value based on the demanded bits. When this function is called, it is known
790 /// that only the bits set in DemandedMask of the result of V are ever used
791 /// downstream. Consequently, depending on the mask and V, it may be possible
792 /// to replace V with a constant or one of its operands. In such cases, this
793 /// function does the replacement and returns true. In all other cases, it
794 /// returns false after analyzing the expression and setting KnownOne and known
795 /// to be one in the expression. KnownZero contains all the bits that are known
796 /// to be zero in the expression. These are provided to potentially allow the
797 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
798 /// the expression. KnownOne and KnownZero always follow the invariant that
799 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
800 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
801 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
802 /// and KnownOne must all be the same.
804 /// This returns null if it did not change anything and it permits no
805 /// simplification. This returns V itself if it did some simplification of V's
806 /// operands based on the information about what bits are demanded. This returns
807 /// some other non-null value if it found out that V is equal to another value
808 /// in the context where the specified bits are demanded, but not for all users.
809 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
810 APInt &KnownZero, APInt &KnownOne,
812 assert(V != 0 && "Null pointer of Value???");
813 assert(Depth <= 6 && "Limit Search Depth");
814 uint32_t BitWidth = DemandedMask.getBitWidth();
815 const Type *VTy = V->getType();
816 assert((TD || !isa<PointerType>(VTy)) &&
817 "SimplifyDemandedBits needs to know bit widths!");
818 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
819 (!VTy->isIntOrIntVector() ||
820 VTy->getScalarSizeInBits() == BitWidth) &&
821 KnownZero.getBitWidth() == BitWidth &&
822 KnownOne.getBitWidth() == BitWidth &&
823 "Value *V, DemandedMask, KnownZero and KnownOne "
824 "must have same BitWidth");
825 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
826 // We know all of the bits for a constant!
827 KnownOne = CI->getValue() & DemandedMask;
828 KnownZero = ~KnownOne & DemandedMask;
831 if (isa<ConstantPointerNull>(V)) {
832 // We know all of the bits for a constant!
834 KnownZero = DemandedMask;
840 if (DemandedMask == 0) { // Not demanding any bits from V.
841 if (isa<UndefValue>(V))
843 return UndefValue::get(VTy);
846 if (Depth == 6) // Limit search depth.
849 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
850 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
852 Instruction *I = dyn_cast<Instruction>(V);
854 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
855 return 0; // Only analyze instructions.
858 // If there are multiple uses of this value and we aren't at the root, then
859 // we can't do any simplifications of the operands, because DemandedMask
860 // only reflects the bits demanded by *one* of the users.
861 if (Depth != 0 && !I->hasOneUse()) {
862 // Despite the fact that we can't simplify this instruction in all User's
863 // context, we can at least compute the knownzero/knownone bits, and we can
864 // do simplifications that apply to *just* the one user if we know that
865 // this instruction has a simpler value in that context.
866 if (I->getOpcode() == Instruction::And) {
867 // If either the LHS or the RHS are Zero, the result is zero.
868 ComputeMaskedBits(I->getOperand(1), DemandedMask,
869 RHSKnownZero, RHSKnownOne, Depth+1);
870 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
871 LHSKnownZero, LHSKnownOne, Depth+1);
873 // If all of the demanded bits are known 1 on one side, return the other.
874 // These bits cannot contribute to the result of the 'and' in this
876 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
877 (DemandedMask & ~LHSKnownZero))
878 return I->getOperand(0);
879 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
880 (DemandedMask & ~RHSKnownZero))
881 return I->getOperand(1);
883 // If all of the demanded bits in the inputs are known zeros, return zero.
884 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
885 return Constant::getNullValue(VTy);
887 } else if (I->getOpcode() == Instruction::Or) {
888 // We can simplify (X|Y) -> X or Y in the user's context if we know that
889 // only bits from X or Y are demanded.
891 // If either the LHS or the RHS are One, the result is One.
892 ComputeMaskedBits(I->getOperand(1), DemandedMask,
893 RHSKnownZero, RHSKnownOne, Depth+1);
894 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
895 LHSKnownZero, LHSKnownOne, Depth+1);
897 // If all of the demanded bits are known zero on one side, return the
898 // other. These bits cannot contribute to the result of the 'or' in this
900 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
901 (DemandedMask & ~LHSKnownOne))
902 return I->getOperand(0);
903 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
904 (DemandedMask & ~RHSKnownOne))
905 return I->getOperand(1);
907 // If all of the potentially set bits on one side are known to be set on
908 // the other side, just use the 'other' side.
909 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
910 (DemandedMask & (~RHSKnownZero)))
911 return I->getOperand(0);
912 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
913 (DemandedMask & (~LHSKnownZero)))
914 return I->getOperand(1);
917 // Compute the KnownZero/KnownOne bits to simplify things downstream.
918 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
922 // If this is the root being simplified, allow it to have multiple uses,
923 // just set the DemandedMask to all bits so that we can try to simplify the
924 // operands. This allows visitTruncInst (for example) to simplify the
925 // operand of a trunc without duplicating all the logic below.
926 if (Depth == 0 && !V->hasOneUse())
927 DemandedMask = APInt::getAllOnesValue(BitWidth);
929 switch (I->getOpcode()) {
931 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
933 case Instruction::And:
934 // If either the LHS or the RHS are Zero, the result is zero.
935 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
936 RHSKnownZero, RHSKnownOne, Depth+1) ||
937 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
938 LHSKnownZero, LHSKnownOne, Depth+1))
940 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
941 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
943 // If all of the demanded bits are known 1 on one side, return the other.
944 // These bits cannot contribute to the result of the 'and'.
945 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
946 (DemandedMask & ~LHSKnownZero))
947 return I->getOperand(0);
948 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
949 (DemandedMask & ~RHSKnownZero))
950 return I->getOperand(1);
952 // If all of the demanded bits in the inputs are known zeros, return zero.
953 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
954 return Constant::getNullValue(VTy);
956 // If the RHS is a constant, see if we can simplify it.
957 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
960 // Output known-1 bits are only known if set in both the LHS & RHS.
961 RHSKnownOne &= LHSKnownOne;
962 // Output known-0 are known to be clear if zero in either the LHS | RHS.
963 RHSKnownZero |= LHSKnownZero;
965 case Instruction::Or:
966 // If either the LHS or the RHS are One, the result is One.
967 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
968 RHSKnownZero, RHSKnownOne, Depth+1) ||
969 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
970 LHSKnownZero, LHSKnownOne, Depth+1))
972 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
973 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
975 // If all of the demanded bits are known zero on one side, return the other.
976 // These bits cannot contribute to the result of the 'or'.
977 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
978 (DemandedMask & ~LHSKnownOne))
979 return I->getOperand(0);
980 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
981 (DemandedMask & ~RHSKnownOne))
982 return I->getOperand(1);
984 // If all of the potentially set bits on one side are known to be set on
985 // the other side, just use the 'other' side.
986 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
987 (DemandedMask & (~RHSKnownZero)))
988 return I->getOperand(0);
989 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
990 (DemandedMask & (~LHSKnownZero)))
991 return I->getOperand(1);
993 // If the RHS is a constant, see if we can simplify it.
994 if (ShrinkDemandedConstant(I, 1, DemandedMask))
997 // Output known-0 bits are only known if clear in both the LHS & RHS.
998 RHSKnownZero &= LHSKnownZero;
999 // Output known-1 are known to be set if set in either the LHS | RHS.
1000 RHSKnownOne |= LHSKnownOne;
1002 case Instruction::Xor: {
1003 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1004 RHSKnownZero, RHSKnownOne, Depth+1) ||
1005 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1006 LHSKnownZero, LHSKnownOne, Depth+1))
1008 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1009 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1011 // If all of the demanded bits are known zero on one side, return the other.
1012 // These bits cannot contribute to the result of the 'xor'.
1013 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1014 return I->getOperand(0);
1015 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1016 return I->getOperand(1);
1018 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1019 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1020 (RHSKnownOne & LHSKnownOne);
1021 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1022 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1023 (RHSKnownOne & LHSKnownZero);
1025 // If all of the demanded bits are known to be zero on one side or the
1026 // other, turn this into an *inclusive* or.
1027 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1028 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0)
1029 return Builder->CreateOr(I->getOperand(0), I->getOperand(1),I->getName());
1031 // If all of the demanded bits on one side are known, and all of the set
1032 // bits on that side are also known to be set on the other side, turn this
1033 // into an AND, as we know the bits will be cleared.
1034 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1035 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1037 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1038 Constant *AndC = Constant::getIntegerValue(VTy,
1039 ~RHSKnownOne & DemandedMask);
1041 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1042 return InsertNewInstBefore(And, *I);
1046 // If the RHS is a constant, see if we can simplify it.
1047 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1048 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1051 RHSKnownZero = KnownZeroOut;
1052 RHSKnownOne = KnownOneOut;
1055 case Instruction::Select:
1056 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1057 RHSKnownZero, RHSKnownOne, Depth+1) ||
1058 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1059 LHSKnownZero, LHSKnownOne, Depth+1))
1061 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1062 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1064 // If the operands are constants, see if we can simplify them.
1065 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1066 ShrinkDemandedConstant(I, 2, DemandedMask))
1069 // Only known if known in both the LHS and RHS.
1070 RHSKnownOne &= LHSKnownOne;
1071 RHSKnownZero &= LHSKnownZero;
1073 case Instruction::Trunc: {
1074 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1075 DemandedMask.zext(truncBf);
1076 RHSKnownZero.zext(truncBf);
1077 RHSKnownOne.zext(truncBf);
1078 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1079 RHSKnownZero, RHSKnownOne, Depth+1))
1081 DemandedMask.trunc(BitWidth);
1082 RHSKnownZero.trunc(BitWidth);
1083 RHSKnownOne.trunc(BitWidth);
1084 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1087 case Instruction::BitCast:
1088 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1089 return false; // vector->int or fp->int?
1091 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1092 if (const VectorType *SrcVTy =
1093 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1094 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1095 // Don't touch a bitcast between vectors of different element counts.
1098 // Don't touch a scalar-to-vector bitcast.
1100 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1101 // Don't touch a vector-to-scalar bitcast.
1104 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1105 RHSKnownZero, RHSKnownOne, Depth+1))
1107 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1109 case Instruction::ZExt: {
1110 // Compute the bits in the result that are not present in the input.
1111 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1113 DemandedMask.trunc(SrcBitWidth);
1114 RHSKnownZero.trunc(SrcBitWidth);
1115 RHSKnownOne.trunc(SrcBitWidth);
1116 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1117 RHSKnownZero, RHSKnownOne, Depth+1))
1119 DemandedMask.zext(BitWidth);
1120 RHSKnownZero.zext(BitWidth);
1121 RHSKnownOne.zext(BitWidth);
1122 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1123 // The top bits are known to be zero.
1124 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1127 case Instruction::SExt: {
1128 // Compute the bits in the result that are not present in the input.
1129 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1131 APInt InputDemandedBits = DemandedMask &
1132 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1134 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1135 // If any of the sign extended bits are demanded, we know that the sign
1137 if ((NewBits & DemandedMask) != 0)
1138 InputDemandedBits.set(SrcBitWidth-1);
1140 InputDemandedBits.trunc(SrcBitWidth);
1141 RHSKnownZero.trunc(SrcBitWidth);
1142 RHSKnownOne.trunc(SrcBitWidth);
1143 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1144 RHSKnownZero, RHSKnownOne, Depth+1))
1146 InputDemandedBits.zext(BitWidth);
1147 RHSKnownZero.zext(BitWidth);
1148 RHSKnownOne.zext(BitWidth);
1149 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1151 // If the sign bit of the input is known set or clear, then we know the
1152 // top bits of the result.
1154 // If the input sign bit is known zero, or if the NewBits are not demanded
1155 // convert this into a zero extension.
1156 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1157 // Convert to ZExt cast
1158 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1159 return InsertNewInstBefore(NewCast, *I);
1160 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1161 RHSKnownOne |= NewBits;
1165 case Instruction::Add: {
1166 // Figure out what the input bits are. If the top bits of the and result
1167 // are not demanded, then the add doesn't demand them from its input
1169 unsigned NLZ = DemandedMask.countLeadingZeros();
1171 // If there is a constant on the RHS, there are a variety of xformations
1173 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1174 // If null, this should be simplified elsewhere. Some of the xforms here
1175 // won't work if the RHS is zero.
1179 // If the top bit of the output is demanded, demand everything from the
1180 // input. Otherwise, we demand all the input bits except NLZ top bits.
1181 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1183 // Find information about known zero/one bits in the input.
1184 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1185 LHSKnownZero, LHSKnownOne, Depth+1))
1188 // If the RHS of the add has bits set that can't affect the input, reduce
1190 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1193 // Avoid excess work.
1194 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1197 // Turn it into OR if input bits are zero.
1198 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1200 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1202 return InsertNewInstBefore(Or, *I);
1205 // We can say something about the output known-zero and known-one bits,
1206 // depending on potential carries from the input constant and the
1207 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1208 // bits set and the RHS constant is 0x01001, then we know we have a known
1209 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1211 // To compute this, we first compute the potential carry bits. These are
1212 // the bits which may be modified. I'm not aware of a better way to do
1214 const APInt &RHSVal = RHS->getValue();
1215 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1217 // Now that we know which bits have carries, compute the known-1/0 sets.
1219 // Bits are known one if they are known zero in one operand and one in the
1220 // other, and there is no input carry.
1221 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1222 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1224 // Bits are known zero if they are known zero in both operands and there
1225 // is no input carry.
1226 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1228 // If the high-bits of this ADD are not demanded, then it does not demand
1229 // the high bits of its LHS or RHS.
1230 if (DemandedMask[BitWidth-1] == 0) {
1231 // Right fill the mask of bits for this ADD to demand the most
1232 // significant bit and all those below it.
1233 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1234 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1235 LHSKnownZero, LHSKnownOne, Depth+1) ||
1236 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1237 LHSKnownZero, LHSKnownOne, Depth+1))
1243 case Instruction::Sub:
1244 // If the high-bits of this SUB are not demanded, then it does not demand
1245 // the high bits of its LHS or RHS.
1246 if (DemandedMask[BitWidth-1] == 0) {
1247 // Right fill the mask of bits for this SUB to demand the most
1248 // significant bit and all those below it.
1249 uint32_t NLZ = DemandedMask.countLeadingZeros();
1250 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1251 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1252 LHSKnownZero, LHSKnownOne, Depth+1) ||
1253 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1254 LHSKnownZero, LHSKnownOne, Depth+1))
1257 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1258 // the known zeros and ones.
1259 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1261 case Instruction::Shl:
1262 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1263 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1264 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1265 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1266 RHSKnownZero, RHSKnownOne, Depth+1))
1268 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1269 RHSKnownZero <<= ShiftAmt;
1270 RHSKnownOne <<= ShiftAmt;
1271 // low bits known zero.
1273 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1276 case Instruction::LShr:
1277 // For a logical shift right
1278 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1279 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1281 // Unsigned shift right.
1282 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1283 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1284 RHSKnownZero, RHSKnownOne, Depth+1))
1286 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1287 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1288 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1290 // Compute the new bits that are at the top now.
1291 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1292 RHSKnownZero |= HighBits; // high bits known zero.
1296 case Instruction::AShr:
1297 // If this is an arithmetic shift right and only the low-bit is set, we can
1298 // always convert this into a logical shr, even if the shift amount is
1299 // variable. The low bit of the shift cannot be an input sign bit unless
1300 // the shift amount is >= the size of the datatype, which is undefined.
1301 if (DemandedMask == 1) {
1302 // Perform the logical shift right.
1303 Instruction *NewVal = BinaryOperator::CreateLShr(
1304 I->getOperand(0), I->getOperand(1), I->getName());
1305 return InsertNewInstBefore(NewVal, *I);
1308 // If the sign bit is the only bit demanded by this ashr, then there is no
1309 // need to do it, the shift doesn't change the high bit.
1310 if (DemandedMask.isSignBit())
1311 return I->getOperand(0);
1313 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1314 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1316 // Signed shift right.
1317 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1318 // If any of the "high bits" are demanded, we should set the sign bit as
1320 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1321 DemandedMaskIn.set(BitWidth-1);
1322 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1323 RHSKnownZero, RHSKnownOne, Depth+1))
1325 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1326 // Compute the new bits that are at the top now.
1327 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1328 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1329 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1331 // Handle the sign bits.
1332 APInt SignBit(APInt::getSignBit(BitWidth));
1333 // Adjust to where it is now in the mask.
1334 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1336 // If the input sign bit is known to be zero, or if none of the top bits
1337 // are demanded, turn this into an unsigned shift right.
1338 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1339 (HighBits & ~DemandedMask) == HighBits) {
1340 // Perform the logical shift right.
1341 Instruction *NewVal = BinaryOperator::CreateLShr(
1342 I->getOperand(0), SA, I->getName());
1343 return InsertNewInstBefore(NewVal, *I);
1344 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1345 RHSKnownOne |= HighBits;
1349 case Instruction::SRem:
1350 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1351 APInt RA = Rem->getValue().abs();
1352 if (RA.isPowerOf2()) {
1353 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1354 return I->getOperand(0);
1356 APInt LowBits = RA - 1;
1357 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1358 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1359 LHSKnownZero, LHSKnownOne, Depth+1))
1362 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1363 LHSKnownZero |= ~LowBits;
1365 KnownZero |= LHSKnownZero & DemandedMask;
1367 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1371 case Instruction::URem: {
1372 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1373 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1374 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1375 KnownZero2, KnownOne2, Depth+1) ||
1376 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1377 KnownZero2, KnownOne2, Depth+1))
1380 unsigned Leaders = KnownZero2.countLeadingOnes();
1381 Leaders = std::max(Leaders,
1382 KnownZero2.countLeadingOnes());
1383 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1386 case Instruction::Call:
1387 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1388 switch (II->getIntrinsicID()) {
1390 case Intrinsic::bswap: {
1391 // If the only bits demanded come from one byte of the bswap result,
1392 // just shift the input byte into position to eliminate the bswap.
1393 unsigned NLZ = DemandedMask.countLeadingZeros();
1394 unsigned NTZ = DemandedMask.countTrailingZeros();
1396 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1397 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1398 // have 14 leading zeros, round to 8.
1401 // If we need exactly one byte, we can do this transformation.
1402 if (BitWidth-NLZ-NTZ == 8) {
1403 unsigned ResultBit = NTZ;
1404 unsigned InputBit = BitWidth-NTZ-8;
1406 // Replace this with either a left or right shift to get the byte into
1408 Instruction *NewVal;
1409 if (InputBit > ResultBit)
1410 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1411 ConstantInt::get(I->getType(), InputBit-ResultBit));
1413 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1414 ConstantInt::get(I->getType(), ResultBit-InputBit));
1415 NewVal->takeName(I);
1416 return InsertNewInstBefore(NewVal, *I);
1419 // TODO: Could compute known zero/one bits based on the input.
1424 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1428 // If the client is only demanding bits that we know, return the known
1430 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1431 return Constant::getIntegerValue(VTy, RHSKnownOne);
1436 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1437 /// any number of elements. DemandedElts contains the set of elements that are
1438 /// actually used by the caller. This method analyzes which elements of the
1439 /// operand are undef and returns that information in UndefElts.
1441 /// If the information about demanded elements can be used to simplify the
1442 /// operation, the operation is simplified, then the resultant value is
1443 /// returned. This returns null if no change was made.
1444 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1447 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1448 APInt EltMask(APInt::getAllOnesValue(VWidth));
1449 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1451 if (isa<UndefValue>(V)) {
1452 // If the entire vector is undefined, just return this info.
1453 UndefElts = EltMask;
1455 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1456 UndefElts = EltMask;
1457 return UndefValue::get(V->getType());
1461 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1462 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1463 Constant *Undef = UndefValue::get(EltTy);
1465 std::vector<Constant*> Elts;
1466 for (unsigned i = 0; i != VWidth; ++i)
1467 if (!DemandedElts[i]) { // If not demanded, set to undef.
1468 Elts.push_back(Undef);
1470 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1471 Elts.push_back(Undef);
1473 } else { // Otherwise, defined.
1474 Elts.push_back(CP->getOperand(i));
1477 // If we changed the constant, return it.
1478 Constant *NewCP = ConstantVector::get(Elts);
1479 return NewCP != CP ? NewCP : 0;
1480 } else if (isa<ConstantAggregateZero>(V)) {
1481 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1484 // Check if this is identity. If so, return 0 since we are not simplifying
1486 if (DemandedElts == ((1ULL << VWidth) -1))
1489 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1490 Constant *Zero = Constant::getNullValue(EltTy);
1491 Constant *Undef = UndefValue::get(EltTy);
1492 std::vector<Constant*> Elts;
1493 for (unsigned i = 0; i != VWidth; ++i) {
1494 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1495 Elts.push_back(Elt);
1497 UndefElts = DemandedElts ^ EltMask;
1498 return ConstantVector::get(Elts);
1501 // Limit search depth.
1505 // If multiple users are using the root value, procede with
1506 // simplification conservatively assuming that all elements
1508 if (!V->hasOneUse()) {
1509 // Quit if we find multiple users of a non-root value though.
1510 // They'll be handled when it's their turn to be visited by
1511 // the main instcombine process.
1513 // TODO: Just compute the UndefElts information recursively.
1516 // Conservatively assume that all elements are needed.
1517 DemandedElts = EltMask;
1520 Instruction *I = dyn_cast<Instruction>(V);
1521 if (!I) return 0; // Only analyze instructions.
1523 bool MadeChange = false;
1524 APInt UndefElts2(VWidth, 0);
1526 switch (I->getOpcode()) {
1529 case Instruction::InsertElement: {
1530 // If this is a variable index, we don't know which element it overwrites.
1531 // demand exactly the same input as we produce.
1532 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1534 // Note that we can't propagate undef elt info, because we don't know
1535 // which elt is getting updated.
1536 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1537 UndefElts2, Depth+1);
1538 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1542 // If this is inserting an element that isn't demanded, remove this
1544 unsigned IdxNo = Idx->getZExtValue();
1545 if (IdxNo >= VWidth || !DemandedElts[IdxNo]) {
1547 return I->getOperand(0);
1550 // Otherwise, the element inserted overwrites whatever was there, so the
1551 // input demanded set is simpler than the output set.
1552 APInt DemandedElts2 = DemandedElts;
1553 DemandedElts2.clear(IdxNo);
1554 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1555 UndefElts, Depth+1);
1556 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1558 // The inserted element is defined.
1559 UndefElts.clear(IdxNo);
1562 case Instruction::ShuffleVector: {
1563 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1564 uint64_t LHSVWidth =
1565 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1566 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1567 for (unsigned i = 0; i < VWidth; i++) {
1568 if (DemandedElts[i]) {
1569 unsigned MaskVal = Shuffle->getMaskValue(i);
1570 if (MaskVal != -1u) {
1571 assert(MaskVal < LHSVWidth * 2 &&
1572 "shufflevector mask index out of range!");
1573 if (MaskVal < LHSVWidth)
1574 LeftDemanded.set(MaskVal);
1576 RightDemanded.set(MaskVal - LHSVWidth);
1581 APInt UndefElts4(LHSVWidth, 0);
1582 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1583 UndefElts4, Depth+1);
1584 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1586 APInt UndefElts3(LHSVWidth, 0);
1587 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1588 UndefElts3, Depth+1);
1589 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1591 bool NewUndefElts = false;
1592 for (unsigned i = 0; i < VWidth; i++) {
1593 unsigned MaskVal = Shuffle->getMaskValue(i);
1594 if (MaskVal == -1u) {
1596 } else if (MaskVal < LHSVWidth) {
1597 if (UndefElts4[MaskVal]) {
1598 NewUndefElts = true;
1602 if (UndefElts3[MaskVal - LHSVWidth]) {
1603 NewUndefElts = true;
1610 // Add additional discovered undefs.
1611 std::vector<Constant*> Elts;
1612 for (unsigned i = 0; i < VWidth; ++i) {
1614 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
1616 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
1617 Shuffle->getMaskValue(i)));
1619 I->setOperand(2, ConstantVector::get(Elts));
1624 case Instruction::BitCast: {
1625 // Vector->vector casts only.
1626 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1628 unsigned InVWidth = VTy->getNumElements();
1629 APInt InputDemandedElts(InVWidth, 0);
1632 if (VWidth == InVWidth) {
1633 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1634 // elements as are demanded of us.
1636 InputDemandedElts = DemandedElts;
1637 } else if (VWidth > InVWidth) {
1641 // If there are more elements in the result than there are in the source,
1642 // then an input element is live if any of the corresponding output
1643 // elements are live.
1644 Ratio = VWidth/InVWidth;
1645 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1646 if (DemandedElts[OutIdx])
1647 InputDemandedElts.set(OutIdx/Ratio);
1653 // If there are more elements in the source than there are in the result,
1654 // then an input element is live if the corresponding output element is
1656 Ratio = InVWidth/VWidth;
1657 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1658 if (DemandedElts[InIdx/Ratio])
1659 InputDemandedElts.set(InIdx);
1662 // div/rem demand all inputs, because they don't want divide by zero.
1663 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1664 UndefElts2, Depth+1);
1666 I->setOperand(0, TmpV);
1670 UndefElts = UndefElts2;
1671 if (VWidth > InVWidth) {
1672 llvm_unreachable("Unimp");
1673 // If there are more elements in the result than there are in the source,
1674 // then an output element is undef if the corresponding input element is
1676 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1677 if (UndefElts2[OutIdx/Ratio])
1678 UndefElts.set(OutIdx);
1679 } else if (VWidth < InVWidth) {
1680 llvm_unreachable("Unimp");
1681 // If there are more elements in the source than there are in the result,
1682 // then a result element is undef if all of the corresponding input
1683 // elements are undef.
1684 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1685 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1686 if (!UndefElts2[InIdx]) // Not undef?
1687 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1691 case Instruction::And:
1692 case Instruction::Or:
1693 case Instruction::Xor:
1694 case Instruction::Add:
1695 case Instruction::Sub:
1696 case Instruction::Mul:
1697 // div/rem demand all inputs, because they don't want divide by zero.
1698 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1699 UndefElts, Depth+1);
1700 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1701 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1702 UndefElts2, Depth+1);
1703 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1705 // Output elements are undefined if both are undefined. Consider things
1706 // like undef&0. The result is known zero, not undef.
1707 UndefElts &= UndefElts2;
1710 case Instruction::Call: {
1711 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1713 switch (II->getIntrinsicID()) {
1716 // Binary vector operations that work column-wise. A dest element is a
1717 // function of the corresponding input elements from the two inputs.
1718 case Intrinsic::x86_sse_sub_ss:
1719 case Intrinsic::x86_sse_mul_ss:
1720 case Intrinsic::x86_sse_min_ss:
1721 case Intrinsic::x86_sse_max_ss:
1722 case Intrinsic::x86_sse2_sub_sd:
1723 case Intrinsic::x86_sse2_mul_sd:
1724 case Intrinsic::x86_sse2_min_sd:
1725 case Intrinsic::x86_sse2_max_sd:
1726 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1727 UndefElts, Depth+1);
1728 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1729 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1730 UndefElts2, Depth+1);
1731 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1733 // If only the low elt is demanded and this is a scalarizable intrinsic,
1734 // scalarize it now.
1735 if (DemandedElts == 1) {
1736 switch (II->getIntrinsicID()) {
1738 case Intrinsic::x86_sse_sub_ss:
1739 case Intrinsic::x86_sse_mul_ss:
1740 case Intrinsic::x86_sse2_sub_sd:
1741 case Intrinsic::x86_sse2_mul_sd:
1742 // TODO: Lower MIN/MAX/ABS/etc
1743 Value *LHS = II->getOperand(1);
1744 Value *RHS = II->getOperand(2);
1745 // Extract the element as scalars.
1746 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1747 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1748 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1749 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1751 switch (II->getIntrinsicID()) {
1752 default: llvm_unreachable("Case stmts out of sync!");
1753 case Intrinsic::x86_sse_sub_ss:
1754 case Intrinsic::x86_sse2_sub_sd:
1755 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1756 II->getName()), *II);
1758 case Intrinsic::x86_sse_mul_ss:
1759 case Intrinsic::x86_sse2_mul_sd:
1760 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1761 II->getName()), *II);
1766 InsertElementInst::Create(
1767 UndefValue::get(II->getType()), TmpV,
1768 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), II->getName());
1769 InsertNewInstBefore(New, *II);
1774 // Output elements are undefined if both are undefined. Consider things
1775 // like undef&0. The result is known zero, not undef.
1776 UndefElts &= UndefElts2;
1782 return MadeChange ? I : 0;
1786 /// AssociativeOpt - Perform an optimization on an associative operator. This
1787 /// function is designed to check a chain of associative operators for a
1788 /// potential to apply a certain optimization. Since the optimization may be
1789 /// applicable if the expression was reassociated, this checks the chain, then
1790 /// reassociates the expression as necessary to expose the optimization
1791 /// opportunity. This makes use of a special Functor, which must define
1792 /// 'shouldApply' and 'apply' methods.
1794 template<typename Functor>
1795 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1796 unsigned Opcode = Root.getOpcode();
1797 Value *LHS = Root.getOperand(0);
1799 // Quick check, see if the immediate LHS matches...
1800 if (F.shouldApply(LHS))
1801 return F.apply(Root);
1803 // Otherwise, if the LHS is not of the same opcode as the root, return.
1804 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1805 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1806 // Should we apply this transform to the RHS?
1807 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1809 // If not to the RHS, check to see if we should apply to the LHS...
1810 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1811 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1815 // If the functor wants to apply the optimization to the RHS of LHSI,
1816 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1818 // Now all of the instructions are in the current basic block, go ahead
1819 // and perform the reassociation.
1820 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1822 // First move the selected RHS to the LHS of the root...
1823 Root.setOperand(0, LHSI->getOperand(1));
1825 // Make what used to be the LHS of the root be the user of the root...
1826 Value *ExtraOperand = TmpLHSI->getOperand(1);
1827 if (&Root == TmpLHSI) {
1828 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1831 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1832 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1833 BasicBlock::iterator ARI = &Root; ++ARI;
1834 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1837 // Now propagate the ExtraOperand down the chain of instructions until we
1839 while (TmpLHSI != LHSI) {
1840 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1841 // Move the instruction to immediately before the chain we are
1842 // constructing to avoid breaking dominance properties.
1843 NextLHSI->moveBefore(ARI);
1846 Value *NextOp = NextLHSI->getOperand(1);
1847 NextLHSI->setOperand(1, ExtraOperand);
1849 ExtraOperand = NextOp;
1852 // Now that the instructions are reassociated, have the functor perform
1853 // the transformation...
1854 return F.apply(Root);
1857 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1864 // AddRHS - Implements: X + X --> X << 1
1867 explicit AddRHS(Value *rhs) : RHS(rhs) {}
1868 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1869 Instruction *apply(BinaryOperator &Add) const {
1870 return BinaryOperator::CreateShl(Add.getOperand(0),
1871 ConstantInt::get(Add.getType(), 1));
1875 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1877 struct AddMaskingAnd {
1879 explicit AddMaskingAnd(Constant *c) : C2(c) {}
1880 bool shouldApply(Value *LHS) const {
1882 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1883 ConstantExpr::getAnd(C1, C2)->isNullValue();
1885 Instruction *apply(BinaryOperator &Add) const {
1886 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1892 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1894 if (CastInst *CI = dyn_cast<CastInst>(&I))
1895 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1897 // Figure out if the constant is the left or the right argument.
1898 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1899 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1901 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1903 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1904 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1907 Value *Op0 = SO, *Op1 = ConstOperand;
1909 std::swap(Op0, Op1);
1911 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1912 return IC->Builder->CreateBinOp(BO->getOpcode(), Op0, Op1,
1913 SO->getName()+".op");
1914 if (ICmpInst *CI = dyn_cast<ICmpInst>(&I))
1915 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
1916 SO->getName()+".cmp");
1917 if (FCmpInst *CI = dyn_cast<FCmpInst>(&I))
1918 return IC->Builder->CreateICmp(CI->getPredicate(), Op0, Op1,
1919 SO->getName()+".cmp");
1920 llvm_unreachable("Unknown binary instruction type!");
1923 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1924 // constant as the other operand, try to fold the binary operator into the
1925 // select arguments. This also works for Cast instructions, which obviously do
1926 // not have a second operand.
1927 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1929 // Don't modify shared select instructions
1930 if (!SI->hasOneUse()) return 0;
1931 Value *TV = SI->getOperand(1);
1932 Value *FV = SI->getOperand(2);
1934 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1935 // Bool selects with constant operands can be folded to logical ops.
1936 if (SI->getType() == Type::getInt1Ty(*IC->getContext())) return 0;
1938 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1939 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1941 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1948 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1949 /// node as operand #0, see if we can fold the instruction into the PHI (which
1950 /// is only possible if all operands to the PHI are constants).
1951 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1952 PHINode *PN = cast<PHINode>(I.getOperand(0));
1953 unsigned NumPHIValues = PN->getNumIncomingValues();
1954 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1956 // Check to see if all of the operands of the PHI are constants. If there is
1957 // one non-constant value, remember the BB it is. If there is more than one
1958 // or if *it* is a PHI, bail out.
1959 BasicBlock *NonConstBB = 0;
1960 for (unsigned i = 0; i != NumPHIValues; ++i)
1961 if (!isa<Constant>(PN->getIncomingValue(i))) {
1962 if (NonConstBB) return 0; // More than one non-const value.
1963 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1964 NonConstBB = PN->getIncomingBlock(i);
1966 // If the incoming non-constant value is in I's block, we have an infinite
1968 if (NonConstBB == I.getParent())
1972 // If there is exactly one non-constant value, we can insert a copy of the
1973 // operation in that block. However, if this is a critical edge, we would be
1974 // inserting the computation one some other paths (e.g. inside a loop). Only
1975 // do this if the pred block is unconditionally branching into the phi block.
1977 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1978 if (!BI || !BI->isUnconditional()) return 0;
1981 // Okay, we can do the transformation: create the new PHI node.
1982 PHINode *NewPN = PHINode::Create(I.getType(), "");
1983 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1984 InsertNewInstBefore(NewPN, *PN);
1985 NewPN->takeName(PN);
1987 // Next, add all of the operands to the PHI.
1988 if (I.getNumOperands() == 2) {
1989 Constant *C = cast<Constant>(I.getOperand(1));
1990 for (unsigned i = 0; i != NumPHIValues; ++i) {
1992 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1993 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1994 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1996 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1998 assert(PN->getIncomingBlock(i) == NonConstBB);
1999 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2000 InV = BinaryOperator::Create(BO->getOpcode(),
2001 PN->getIncomingValue(i), C, "phitmp",
2002 NonConstBB->getTerminator());
2003 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2004 InV = CmpInst::Create(CI->getOpcode(),
2006 PN->getIncomingValue(i), C, "phitmp",
2007 NonConstBB->getTerminator());
2009 llvm_unreachable("Unknown binop!");
2011 Worklist.Add(cast<Instruction>(InV));
2013 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2016 CastInst *CI = cast<CastInst>(&I);
2017 const Type *RetTy = CI->getType();
2018 for (unsigned i = 0; i != NumPHIValues; ++i) {
2020 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2021 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2023 assert(PN->getIncomingBlock(i) == NonConstBB);
2024 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2025 I.getType(), "phitmp",
2026 NonConstBB->getTerminator());
2027 Worklist.Add(cast<Instruction>(InV));
2029 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2032 return ReplaceInstUsesWith(I, NewPN);
2036 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2037 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2038 /// This basically requires proving that the add in the original type would not
2039 /// overflow to change the sign bit or have a carry out.
2040 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2041 // There are different heuristics we can use for this. Here are some simple
2044 // Add has the property that adding any two 2's complement numbers can only
2045 // have one carry bit which can change a sign. As such, if LHS and RHS each
2046 // have at least two sign bits, we know that the addition of the two values will
2047 // sign extend fine.
2048 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2052 // If one of the operands only has one non-zero bit, and if the other operand
2053 // has a known-zero bit in a more significant place than it (not including the
2054 // sign bit) the ripple may go up to and fill the zero, but won't change the
2055 // sign. For example, (X & ~4) + 1.
2063 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2064 bool Changed = SimplifyCommutative(I);
2065 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2067 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2068 // X + undef -> undef
2069 if (isa<UndefValue>(RHS))
2070 return ReplaceInstUsesWith(I, RHS);
2073 if (RHSC->isNullValue())
2074 return ReplaceInstUsesWith(I, LHS);
2076 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2077 // X + (signbit) --> X ^ signbit
2078 const APInt& Val = CI->getValue();
2079 uint32_t BitWidth = Val.getBitWidth();
2080 if (Val == APInt::getSignBit(BitWidth))
2081 return BinaryOperator::CreateXor(LHS, RHS);
2083 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2084 // (X & 254)+1 -> (X&254)|1
2085 if (SimplifyDemandedInstructionBits(I))
2088 // zext(bool) + C -> bool ? C + 1 : C
2089 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2090 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2091 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2094 if (isa<PHINode>(LHS))
2095 if (Instruction *NV = FoldOpIntoPhi(I))
2098 ConstantInt *XorRHS = 0;
2100 if (isa<ConstantInt>(RHSC) &&
2101 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2102 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2103 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2105 uint32_t Size = TySizeBits / 2;
2106 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2107 APInt CFF80Val(-C0080Val);
2109 if (TySizeBits > Size) {
2110 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2111 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2112 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2113 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2114 // This is a sign extend if the top bits are known zero.
2115 if (!MaskedValueIsZero(XorLHS,
2116 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2117 Size = 0; // Not a sign ext, but can't be any others either.
2122 C0080Val = APIntOps::lshr(C0080Val, Size);
2123 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2124 } while (Size >= 1);
2126 // FIXME: This shouldn't be necessary. When the backends can handle types
2127 // with funny bit widths then this switch statement should be removed. It
2128 // is just here to get the size of the "middle" type back up to something
2129 // that the back ends can handle.
2130 const Type *MiddleType = 0;
2133 case 32: MiddleType = Type::getInt32Ty(*Context); break;
2134 case 16: MiddleType = Type::getInt16Ty(*Context); break;
2135 case 8: MiddleType = Type::getInt8Ty(*Context); break;
2138 Value *NewTrunc = Builder->CreateTrunc(XorLHS, MiddleType, "sext");
2139 return new SExtInst(NewTrunc, I.getType(), I.getName());
2144 if (I.getType() == Type::getInt1Ty(*Context))
2145 return BinaryOperator::CreateXor(LHS, RHS);
2148 if (I.getType()->isInteger()) {
2149 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2152 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2153 if (RHSI->getOpcode() == Instruction::Sub)
2154 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2155 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2157 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2158 if (LHSI->getOpcode() == Instruction::Sub)
2159 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2160 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2165 // -A + -B --> -(A + B)
2166 if (Value *LHSV = dyn_castNegVal(LHS)) {
2167 if (LHS->getType()->isIntOrIntVector()) {
2168 if (Value *RHSV = dyn_castNegVal(RHS)) {
2169 Value *NewAdd = Builder->CreateAdd(LHSV, RHSV, "sum");
2170 return BinaryOperator::CreateNeg(NewAdd);
2174 return BinaryOperator::CreateSub(RHS, LHSV);
2178 if (!isa<Constant>(RHS))
2179 if (Value *V = dyn_castNegVal(RHS))
2180 return BinaryOperator::CreateSub(LHS, V);
2184 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2185 if (X == RHS) // X*C + X --> X * (C+1)
2186 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2188 // X*C1 + X*C2 --> X * (C1+C2)
2190 if (X == dyn_castFoldableMul(RHS, C1))
2191 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2194 // X + X*C --> X * (C+1)
2195 if (dyn_castFoldableMul(RHS, C2) == LHS)
2196 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2198 // X + ~X --> -1 since ~X = -X-1
2199 if (dyn_castNotVal(LHS) == RHS ||
2200 dyn_castNotVal(RHS) == LHS)
2201 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2204 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2205 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2206 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2209 // A+B --> A|B iff A and B have no bits set in common.
2210 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2211 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2212 APInt LHSKnownOne(IT->getBitWidth(), 0);
2213 APInt LHSKnownZero(IT->getBitWidth(), 0);
2214 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2215 if (LHSKnownZero != 0) {
2216 APInt RHSKnownOne(IT->getBitWidth(), 0);
2217 APInt RHSKnownZero(IT->getBitWidth(), 0);
2218 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2220 // No bits in common -> bitwise or.
2221 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2222 return BinaryOperator::CreateOr(LHS, RHS);
2226 // W*X + Y*Z --> W * (X+Z) iff W == Y
2227 if (I.getType()->isIntOrIntVector()) {
2228 Value *W, *X, *Y, *Z;
2229 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2230 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2234 } else if (Y == X) {
2236 } else if (X == Z) {
2243 Value *NewAdd = Builder->CreateAdd(X, Z, LHS->getName());
2244 return BinaryOperator::CreateMul(W, NewAdd);
2249 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2251 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2252 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2254 // (X & FF00) + xx00 -> (X+xx00) & FF00
2255 if (LHS->hasOneUse() &&
2256 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2257 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2258 if (Anded == CRHS) {
2259 // See if all bits from the first bit set in the Add RHS up are included
2260 // in the mask. First, get the rightmost bit.
2261 const APInt& AddRHSV = CRHS->getValue();
2263 // Form a mask of all bits from the lowest bit added through the top.
2264 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2266 // See if the and mask includes all of these bits.
2267 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2269 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2270 // Okay, the xform is safe. Insert the new add pronto.
2271 Value *NewAdd = Builder->CreateAdd(X, CRHS, LHS->getName());
2272 return BinaryOperator::CreateAnd(NewAdd, C2);
2277 // Try to fold constant add into select arguments.
2278 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2279 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2283 // add (select X 0 (sub n A)) A --> select X A n
2285 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2288 SI = dyn_cast<SelectInst>(RHS);
2291 if (SI && SI->hasOneUse()) {
2292 Value *TV = SI->getTrueValue();
2293 Value *FV = SI->getFalseValue();
2296 // Can we fold the add into the argument of the select?
2297 // We check both true and false select arguments for a matching subtract.
2298 if (match(FV, m_Zero()) &&
2299 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2300 // Fold the add into the true select value.
2301 return SelectInst::Create(SI->getCondition(), N, A);
2302 if (match(TV, m_Zero()) &&
2303 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2304 // Fold the add into the false select value.
2305 return SelectInst::Create(SI->getCondition(), A, N);
2309 // Check for (add (sext x), y), see if we can merge this into an
2310 // integer add followed by a sext.
2311 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2312 // (add (sext x), cst) --> (sext (add x, cst'))
2313 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2315 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2316 if (LHSConv->hasOneUse() &&
2317 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2318 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2319 // Insert the new, smaller add.
2320 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2322 return new SExtInst(NewAdd, I.getType());
2326 // (add (sext x), (sext y)) --> (sext (add int x, y))
2327 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2328 // Only do this if x/y have the same type, if at last one of them has a
2329 // single use (so we don't increase the number of sexts), and if the
2330 // integer add will not overflow.
2331 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2332 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2333 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2334 RHSConv->getOperand(0))) {
2335 // Insert the new integer add.
2336 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2337 RHSConv->getOperand(0), "addconv");
2338 return new SExtInst(NewAdd, I.getType());
2343 return Changed ? &I : 0;
2346 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2347 bool Changed = SimplifyCommutative(I);
2348 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2350 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2352 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2353 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2354 (I.getType())->getValueAPF()))
2355 return ReplaceInstUsesWith(I, LHS);
2358 if (isa<PHINode>(LHS))
2359 if (Instruction *NV = FoldOpIntoPhi(I))
2364 // -A + -B --> -(A + B)
2365 if (Value *LHSV = dyn_castFNegVal(LHS))
2366 return BinaryOperator::CreateFSub(RHS, LHSV);
2369 if (!isa<Constant>(RHS))
2370 if (Value *V = dyn_castFNegVal(RHS))
2371 return BinaryOperator::CreateFSub(LHS, V);
2373 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2374 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2375 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2376 return ReplaceInstUsesWith(I, LHS);
2378 // Check for (add double (sitofp x), y), see if we can merge this into an
2379 // integer add followed by a promotion.
2380 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2381 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2382 // ... if the constant fits in the integer value. This is useful for things
2383 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2384 // requires a constant pool load, and generally allows the add to be better
2386 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2388 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2389 if (LHSConv->hasOneUse() &&
2390 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2391 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2392 // Insert the new integer add.
2393 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2395 return new SIToFPInst(NewAdd, I.getType());
2399 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2400 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2401 // Only do this if x/y have the same type, if at last one of them has a
2402 // single use (so we don't increase the number of int->fp conversions),
2403 // and if the integer add will not overflow.
2404 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2405 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2406 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2407 RHSConv->getOperand(0))) {
2408 // Insert the new integer add.
2409 Value *NewAdd = Builder->CreateAdd(LHSConv->getOperand(0),
2410 RHSConv->getOperand(0), "addconv");
2411 return new SIToFPInst(NewAdd, I.getType());
2416 return Changed ? &I : 0;
2419 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2420 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2422 if (Op0 == Op1) // sub X, X -> 0
2423 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2425 // If this is a 'B = x-(-A)', change to B = x+A...
2426 if (Value *V = dyn_castNegVal(Op1))
2427 return BinaryOperator::CreateAdd(Op0, V);
2429 if (isa<UndefValue>(Op0))
2430 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2431 if (isa<UndefValue>(Op1))
2432 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2434 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2435 // Replace (-1 - A) with (~A)...
2436 if (C->isAllOnesValue())
2437 return BinaryOperator::CreateNot(Op1);
2439 // C - ~X == X + (1+C)
2441 if (match(Op1, m_Not(m_Value(X))))
2442 return BinaryOperator::CreateAdd(X, AddOne(C));
2444 // -(X >>u 31) -> (X >>s 31)
2445 // -(X >>s 31) -> (X >>u 31)
2447 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2448 if (SI->getOpcode() == Instruction::LShr) {
2449 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2450 // Check to see if we are shifting out everything but the sign bit.
2451 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2452 SI->getType()->getPrimitiveSizeInBits()-1) {
2453 // Ok, the transformation is safe. Insert AShr.
2454 return BinaryOperator::Create(Instruction::AShr,
2455 SI->getOperand(0), CU, SI->getName());
2459 else if (SI->getOpcode() == Instruction::AShr) {
2460 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2461 // Check to see if we are shifting out everything but the sign bit.
2462 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2463 SI->getType()->getPrimitiveSizeInBits()-1) {
2464 // Ok, the transformation is safe. Insert LShr.
2465 return BinaryOperator::CreateLShr(
2466 SI->getOperand(0), CU, SI->getName());
2473 // Try to fold constant sub into select arguments.
2474 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2475 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2478 // C - zext(bool) -> bool ? C - 1 : C
2479 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2480 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2481 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2484 if (I.getType() == Type::getInt1Ty(*Context))
2485 return BinaryOperator::CreateXor(Op0, Op1);
2487 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2488 if (Op1I->getOpcode() == Instruction::Add) {
2489 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2490 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2492 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2493 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2495 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2496 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2497 // C1-(X+C2) --> (C1-C2)-X
2498 return BinaryOperator::CreateSub(
2499 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2503 if (Op1I->hasOneUse()) {
2504 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2505 // is not used by anyone else...
2507 if (Op1I->getOpcode() == Instruction::Sub) {
2508 // Swap the two operands of the subexpr...
2509 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2510 Op1I->setOperand(0, IIOp1);
2511 Op1I->setOperand(1, IIOp0);
2513 // Create the new top level add instruction...
2514 return BinaryOperator::CreateAdd(Op0, Op1);
2517 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2519 if (Op1I->getOpcode() == Instruction::And &&
2520 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2521 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2523 Value *NewNot = Builder->CreateNot(OtherOp, "B.not");
2524 return BinaryOperator::CreateAnd(Op0, NewNot);
2527 // 0 - (X sdiv C) -> (X sdiv -C)
2528 if (Op1I->getOpcode() == Instruction::SDiv)
2529 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2531 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2532 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2533 ConstantExpr::getNeg(DivRHS));
2535 // X - X*C --> X * (1-C)
2536 ConstantInt *C2 = 0;
2537 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2539 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2541 return BinaryOperator::CreateMul(Op0, CP1);
2546 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2547 if (Op0I->getOpcode() == Instruction::Add) {
2548 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2549 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2550 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2551 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2552 } else if (Op0I->getOpcode() == Instruction::Sub) {
2553 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2554 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2560 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2561 if (X == Op1) // X*C - X --> X * (C-1)
2562 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2564 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2565 if (X == dyn_castFoldableMul(Op1, C2))
2566 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2571 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2572 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2574 // If this is a 'B = x-(-A)', change to B = x+A...
2575 if (Value *V = dyn_castFNegVal(Op1))
2576 return BinaryOperator::CreateFAdd(Op0, V);
2578 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2579 if (Op1I->getOpcode() == Instruction::FAdd) {
2580 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2581 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
2583 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2584 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
2592 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2593 /// comparison only checks the sign bit. If it only checks the sign bit, set
2594 /// TrueIfSigned if the result of the comparison is true when the input value is
2596 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2597 bool &TrueIfSigned) {
2599 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2600 TrueIfSigned = true;
2601 return RHS->isZero();
2602 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2603 TrueIfSigned = true;
2604 return RHS->isAllOnesValue();
2605 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2606 TrueIfSigned = false;
2607 return RHS->isAllOnesValue();
2608 case ICmpInst::ICMP_UGT:
2609 // True if LHS u> RHS and RHS == high-bit-mask - 1
2610 TrueIfSigned = true;
2611 return RHS->getValue() ==
2612 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2613 case ICmpInst::ICMP_UGE:
2614 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2615 TrueIfSigned = true;
2616 return RHS->getValue().isSignBit();
2622 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2623 bool Changed = SimplifyCommutative(I);
2624 Value *Op0 = I.getOperand(0);
2626 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2627 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2629 // Simplify mul instructions with a constant RHS...
2630 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2631 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2633 // ((X << C1)*C2) == (X * (C2 << C1))
2634 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2635 if (SI->getOpcode() == Instruction::Shl)
2636 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2637 return BinaryOperator::CreateMul(SI->getOperand(0),
2638 ConstantExpr::getShl(CI, ShOp));
2641 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2642 if (CI->equalsInt(1)) // X * 1 == X
2643 return ReplaceInstUsesWith(I, Op0);
2644 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2645 return BinaryOperator::CreateNeg(Op0, I.getName());
2647 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2648 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2649 return BinaryOperator::CreateShl(Op0,
2650 ConstantInt::get(Op0->getType(), Val.logBase2()));
2652 } else if (isa<VectorType>(Op1->getType())) {
2653 if (Op1->isNullValue())
2654 return ReplaceInstUsesWith(I, Op1);
2656 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2657 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2658 return BinaryOperator::CreateNeg(Op0, I.getName());
2660 // As above, vector X*splat(1.0) -> X in all defined cases.
2661 if (Constant *Splat = Op1V->getSplatValue()) {
2662 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2663 if (CI->equalsInt(1))
2664 return ReplaceInstUsesWith(I, Op0);
2669 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2670 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2671 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2672 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2673 Value *Add = Builder->CreateMul(Op0I->getOperand(0), Op1, "tmp");
2674 Value *C1C2 = Builder->CreateMul(Op1, Op0I->getOperand(1));
2675 return BinaryOperator::CreateAdd(Add, C1C2);
2679 // Try to fold constant mul into select arguments.
2680 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2681 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2684 if (isa<PHINode>(Op0))
2685 if (Instruction *NV = FoldOpIntoPhi(I))
2689 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2690 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2691 return BinaryOperator::CreateMul(Op0v, Op1v);
2693 // (X / Y) * Y = X - (X % Y)
2694 // (X / Y) * -Y = (X % Y) - X
2696 Value *Op1 = I.getOperand(1);
2697 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2699 (BO->getOpcode() != Instruction::UDiv &&
2700 BO->getOpcode() != Instruction::SDiv)) {
2702 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2704 Value *Neg = dyn_castNegVal(Op1);
2705 if (BO && BO->hasOneUse() &&
2706 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2707 (BO->getOpcode() == Instruction::UDiv ||
2708 BO->getOpcode() == Instruction::SDiv)) {
2709 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2711 // If the division is exact, X % Y is zero.
2712 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
2713 if (SDiv->isExact()) {
2715 return ReplaceInstUsesWith(I, Op0BO);
2717 return BinaryOperator::CreateNeg(Op0BO);
2721 if (BO->getOpcode() == Instruction::UDiv)
2722 Rem = Builder->CreateURem(Op0BO, Op1BO);
2724 Rem = Builder->CreateSRem(Op0BO, Op1BO);
2728 return BinaryOperator::CreateSub(Op0BO, Rem);
2729 return BinaryOperator::CreateSub(Rem, Op0BO);
2733 if (I.getType() == Type::getInt1Ty(*Context))
2734 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2736 // If one of the operands of the multiply is a cast from a boolean value, then
2737 // we know the bool is either zero or one, so this is a 'masking' multiply.
2738 // See if we can simplify things based on how the boolean was originally
2740 CastInst *BoolCast = 0;
2741 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2742 if (CI->getOperand(0)->getType() == Type::getInt1Ty(*Context))
2745 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2746 if (CI->getOperand(0)->getType() == Type::getInt1Ty(*Context))
2749 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2750 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2751 const Type *SCOpTy = SCIOp0->getType();
2754 // If the icmp is true iff the sign bit of X is set, then convert this
2755 // multiply into a shift/and combination.
2756 if (isa<ConstantInt>(SCIOp1) &&
2757 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2759 // Shift the X value right to turn it into "all signbits".
2760 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2761 SCOpTy->getPrimitiveSizeInBits()-1);
2762 Value *V = Builder->CreateAShr(SCIOp0, Amt,
2763 BoolCast->getOperand(0)->getName()+".mask");
2765 // If the multiply type is not the same as the source type, sign extend
2766 // or truncate to the multiply type.
2767 if (I.getType() != V->getType()) {
2768 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2769 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2770 Instruction::CastOps opcode =
2771 (SrcBits == DstBits ? Instruction::BitCast :
2772 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2773 V = InsertCastBefore(opcode, V, I.getType(), I);
2776 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2777 return BinaryOperator::CreateAnd(V, OtherOp);
2782 return Changed ? &I : 0;
2785 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2786 bool Changed = SimplifyCommutative(I);
2787 Value *Op0 = I.getOperand(0);
2789 // Simplify mul instructions with a constant RHS...
2790 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2791 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2792 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2793 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2794 if (Op1F->isExactlyValue(1.0))
2795 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2796 } else if (isa<VectorType>(Op1->getType())) {
2797 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2798 // As above, vector X*splat(1.0) -> X in all defined cases.
2799 if (Constant *Splat = Op1V->getSplatValue()) {
2800 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2801 if (F->isExactlyValue(1.0))
2802 return ReplaceInstUsesWith(I, Op0);
2807 // Try to fold constant mul into select arguments.
2808 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2809 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2812 if (isa<PHINode>(Op0))
2813 if (Instruction *NV = FoldOpIntoPhi(I))
2817 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
2818 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1)))
2819 return BinaryOperator::CreateFMul(Op0v, Op1v);
2821 return Changed ? &I : 0;
2824 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2826 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2827 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2829 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2830 int NonNullOperand = -1;
2831 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2832 if (ST->isNullValue())
2834 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2835 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2836 if (ST->isNullValue())
2839 if (NonNullOperand == -1)
2842 Value *SelectCond = SI->getOperand(0);
2844 // Change the div/rem to use 'Y' instead of the select.
2845 I.setOperand(1, SI->getOperand(NonNullOperand));
2847 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2848 // problem. However, the select, or the condition of the select may have
2849 // multiple uses. Based on our knowledge that the operand must be non-zero,
2850 // propagate the known value for the select into other uses of it, and
2851 // propagate a known value of the condition into its other users.
2853 // If the select and condition only have a single use, don't bother with this,
2855 if (SI->use_empty() && SelectCond->hasOneUse())
2858 // Scan the current block backward, looking for other uses of SI.
2859 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2861 while (BBI != BBFront) {
2863 // If we found a call to a function, we can't assume it will return, so
2864 // information from below it cannot be propagated above it.
2865 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2868 // Replace uses of the select or its condition with the known values.
2869 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2872 *I = SI->getOperand(NonNullOperand);
2874 } else if (*I == SelectCond) {
2875 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
2876 ConstantInt::getFalse(*Context);
2881 // If we past the instruction, quit looking for it.
2884 if (&*BBI == SelectCond)
2887 // If we ran out of things to eliminate, break out of the loop.
2888 if (SelectCond == 0 && SI == 0)
2896 /// This function implements the transforms on div instructions that work
2897 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2898 /// used by the visitors to those instructions.
2899 /// @brief Transforms common to all three div instructions
2900 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2901 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2903 // undef / X -> 0 for integer.
2904 // undef / X -> undef for FP (the undef could be a snan).
2905 if (isa<UndefValue>(Op0)) {
2906 if (Op0->getType()->isFPOrFPVector())
2907 return ReplaceInstUsesWith(I, Op0);
2908 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2911 // X / undef -> undef
2912 if (isa<UndefValue>(Op1))
2913 return ReplaceInstUsesWith(I, Op1);
2918 /// This function implements the transforms common to both integer division
2919 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2920 /// division instructions.
2921 /// @brief Common integer divide transforms
2922 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2923 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2925 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2927 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2928 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
2929 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2930 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2933 Constant *CI = ConstantInt::get(I.getType(), 1);
2934 return ReplaceInstUsesWith(I, CI);
2937 if (Instruction *Common = commonDivTransforms(I))
2940 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2941 // This does not apply for fdiv.
2942 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2945 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2947 if (RHS->equalsInt(1))
2948 return ReplaceInstUsesWith(I, Op0);
2950 // (X / C1) / C2 -> X / (C1*C2)
2951 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2952 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2953 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2954 if (MultiplyOverflows(RHS, LHSRHS,
2955 I.getOpcode()==Instruction::SDiv))
2956 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2958 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2959 ConstantExpr::getMul(RHS, LHSRHS));
2962 if (!RHS->isZero()) { // avoid X udiv 0
2963 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2964 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2966 if (isa<PHINode>(Op0))
2967 if (Instruction *NV = FoldOpIntoPhi(I))
2972 // 0 / X == 0, we don't need to preserve faults!
2973 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2974 if (LHS->equalsInt(0))
2975 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2977 // It can't be division by zero, hence it must be division by one.
2978 if (I.getType() == Type::getInt1Ty(*Context))
2979 return ReplaceInstUsesWith(I, Op0);
2981 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2982 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2985 return ReplaceInstUsesWith(I, Op0);
2991 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2992 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2994 // Handle the integer div common cases
2995 if (Instruction *Common = commonIDivTransforms(I))
2998 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2999 // X udiv C^2 -> X >> C
3000 // Check to see if this is an unsigned division with an exact power of 2,
3001 // if so, convert to a right shift.
3002 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3003 return BinaryOperator::CreateLShr(Op0,
3004 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3006 // X udiv C, where C >= signbit
3007 if (C->getValue().isNegative()) {
3008 Value *IC = Builder->CreateICmpULT( Op0, C);
3009 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3010 ConstantInt::get(I.getType(), 1));
3014 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3015 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3016 if (RHSI->getOpcode() == Instruction::Shl &&
3017 isa<ConstantInt>(RHSI->getOperand(0))) {
3018 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3019 if (C1.isPowerOf2()) {
3020 Value *N = RHSI->getOperand(1);
3021 const Type *NTy = N->getType();
3022 if (uint32_t C2 = C1.logBase2())
3023 N = Builder->CreateAdd(N, ConstantInt::get(NTy, C2), "tmp");
3024 return BinaryOperator::CreateLShr(Op0, N);
3029 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3030 // where C1&C2 are powers of two.
3031 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3032 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3033 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3034 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3035 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3036 // Compute the shift amounts
3037 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3038 // Construct the "on true" case of the select
3039 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3040 Value *TSI = Builder->CreateLShr(Op0, TC, SI->getName()+".t");
3042 // Construct the "on false" case of the select
3043 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3044 Value *FSI = Builder->CreateLShr(Op0, FC, SI->getName()+".f");
3046 // construct the select instruction and return it.
3047 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3053 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3054 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3056 // Handle the integer div common cases
3057 if (Instruction *Common = commonIDivTransforms(I))
3060 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3062 if (RHS->isAllOnesValue())
3063 return BinaryOperator::CreateNeg(Op0);
3065 // sdiv X, C --> ashr X, log2(C)
3066 if (cast<SDivOperator>(&I)->isExact() &&
3067 RHS->getValue().isNonNegative() &&
3068 RHS->getValue().isPowerOf2()) {
3069 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3070 RHS->getValue().exactLogBase2());
3071 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3074 // -X/C --> X/-C provided the negation doesn't overflow.
3075 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
3076 if (isa<Constant>(Sub->getOperand(0)) &&
3077 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
3078 Sub->hasNoSignedWrap())
3079 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
3080 ConstantExpr::getNeg(RHS));
3083 // If the sign bits of both operands are zero (i.e. we can prove they are
3084 // unsigned inputs), turn this into a udiv.
3085 if (I.getType()->isInteger()) {
3086 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3087 if (MaskedValueIsZero(Op0, Mask)) {
3088 if (MaskedValueIsZero(Op1, Mask)) {
3089 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3090 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3092 ConstantInt *ShiftedInt;
3093 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3094 ShiftedInt->getValue().isPowerOf2()) {
3095 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3096 // Safe because the only negative value (1 << Y) can take on is
3097 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3098 // the sign bit set.
3099 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3107 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3108 return commonDivTransforms(I);
3111 /// This function implements the transforms on rem instructions that work
3112 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3113 /// is used by the visitors to those instructions.
3114 /// @brief Transforms common to all three rem instructions
3115 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3116 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3118 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3119 if (I.getType()->isFPOrFPVector())
3120 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3121 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3123 if (isa<UndefValue>(Op1))
3124 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3126 // Handle cases involving: rem X, (select Cond, Y, Z)
3127 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3133 /// This function implements the transforms common to both integer remainder
3134 /// instructions (urem and srem). It is called by the visitors to those integer
3135 /// remainder instructions.
3136 /// @brief Common integer remainder transforms
3137 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3138 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3140 if (Instruction *common = commonRemTransforms(I))
3143 // 0 % X == 0 for integer, we don't need to preserve faults!
3144 if (Constant *LHS = dyn_cast<Constant>(Op0))
3145 if (LHS->isNullValue())
3146 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3148 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3149 // X % 0 == undef, we don't need to preserve faults!
3150 if (RHS->equalsInt(0))
3151 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3153 if (RHS->equalsInt(1)) // X % 1 == 0
3154 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3156 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3157 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3158 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3160 } else if (isa<PHINode>(Op0I)) {
3161 if (Instruction *NV = FoldOpIntoPhi(I))
3165 // See if we can fold away this rem instruction.
3166 if (SimplifyDemandedInstructionBits(I))
3174 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3175 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3177 if (Instruction *common = commonIRemTransforms(I))
3180 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3181 // X urem C^2 -> X and C
3182 // Check to see if this is an unsigned remainder with an exact power of 2,
3183 // if so, convert to a bitwise and.
3184 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3185 if (C->getValue().isPowerOf2())
3186 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3189 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3190 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3191 if (RHSI->getOpcode() == Instruction::Shl &&
3192 isa<ConstantInt>(RHSI->getOperand(0))) {
3193 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3194 Constant *N1 = Constant::getAllOnesValue(I.getType());
3195 Value *Add = Builder->CreateAdd(RHSI, N1, "tmp");
3196 return BinaryOperator::CreateAnd(Op0, Add);
3201 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3202 // where C1&C2 are powers of two.
3203 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3204 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3205 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3206 // STO == 0 and SFO == 0 handled above.
3207 if ((STO->getValue().isPowerOf2()) &&
3208 (SFO->getValue().isPowerOf2())) {
3209 Value *TrueAnd = Builder->CreateAnd(Op0, SubOne(STO),
3210 SI->getName()+".t");
3211 Value *FalseAnd = Builder->CreateAnd(Op0, SubOne(SFO),
3212 SI->getName()+".f");
3213 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3221 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3222 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3224 // Handle the integer rem common cases
3225 if (Instruction *Common = commonIRemTransforms(I))
3228 if (Value *RHSNeg = dyn_castNegVal(Op1))
3229 if (!isa<Constant>(RHSNeg) ||
3230 (isa<ConstantInt>(RHSNeg) &&
3231 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3233 Worklist.AddValue(I.getOperand(1));
3234 I.setOperand(1, RHSNeg);
3238 // If the sign bits of both operands are zero (i.e. we can prove they are
3239 // unsigned inputs), turn this into a urem.
3240 if (I.getType()->isInteger()) {
3241 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3242 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3243 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3244 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3248 // If it's a constant vector, flip any negative values positive.
3249 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3250 unsigned VWidth = RHSV->getNumOperands();
3252 bool hasNegative = false;
3253 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3254 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3255 if (RHS->getValue().isNegative())
3259 std::vector<Constant *> Elts(VWidth);
3260 for (unsigned i = 0; i != VWidth; ++i) {
3261 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3262 if (RHS->getValue().isNegative())
3263 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3269 Constant *NewRHSV = ConstantVector::get(Elts);
3270 if (NewRHSV != RHSV) {
3271 Worklist.AddValue(I.getOperand(1));
3272 I.setOperand(1, NewRHSV);
3281 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3282 return commonRemTransforms(I);
3285 // isOneBitSet - Return true if there is exactly one bit set in the specified
3287 static bool isOneBitSet(const ConstantInt *CI) {
3288 return CI->getValue().isPowerOf2();
3291 // isHighOnes - Return true if the constant is of the form 1+0+.
3292 // This is the same as lowones(~X).
3293 static bool isHighOnes(const ConstantInt *CI) {
3294 return (~CI->getValue() + 1).isPowerOf2();
3297 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3298 /// are carefully arranged to allow folding of expressions such as:
3300 /// (A < B) | (A > B) --> (A != B)
3302 /// Note that this is only valid if the first and second predicates have the
3303 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3305 /// Three bits are used to represent the condition, as follows:
3310 /// <=> Value Definition
3311 /// 000 0 Always false
3318 /// 111 7 Always true
3320 static unsigned getICmpCode(const ICmpInst *ICI) {
3321 switch (ICI->getPredicate()) {
3323 case ICmpInst::ICMP_UGT: return 1; // 001
3324 case ICmpInst::ICMP_SGT: return 1; // 001
3325 case ICmpInst::ICMP_EQ: return 2; // 010
3326 case ICmpInst::ICMP_UGE: return 3; // 011
3327 case ICmpInst::ICMP_SGE: return 3; // 011
3328 case ICmpInst::ICMP_ULT: return 4; // 100
3329 case ICmpInst::ICMP_SLT: return 4; // 100
3330 case ICmpInst::ICMP_NE: return 5; // 101
3331 case ICmpInst::ICMP_ULE: return 6; // 110
3332 case ICmpInst::ICMP_SLE: return 6; // 110
3335 llvm_unreachable("Invalid ICmp predicate!");
3340 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3341 /// predicate into a three bit mask. It also returns whether it is an ordered
3342 /// predicate by reference.
3343 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3346 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3347 case FCmpInst::FCMP_UNO: return 0; // 000
3348 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3349 case FCmpInst::FCMP_UGT: return 1; // 001
3350 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3351 case FCmpInst::FCMP_UEQ: return 2; // 010
3352 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3353 case FCmpInst::FCMP_UGE: return 3; // 011
3354 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3355 case FCmpInst::FCMP_ULT: return 4; // 100
3356 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3357 case FCmpInst::FCMP_UNE: return 5; // 101
3358 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3359 case FCmpInst::FCMP_ULE: return 6; // 110
3362 // Not expecting FCMP_FALSE and FCMP_TRUE;
3363 llvm_unreachable("Unexpected FCmp predicate!");
3368 /// getICmpValue - This is the complement of getICmpCode, which turns an
3369 /// opcode and two operands into either a constant true or false, or a brand
3370 /// new ICmp instruction. The sign is passed in to determine which kind
3371 /// of predicate to use in the new icmp instruction.
3372 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3373 LLVMContext *Context) {
3375 default: llvm_unreachable("Illegal ICmp code!");
3376 case 0: return ConstantInt::getFalse(*Context);
3379 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3381 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3382 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3385 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3387 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3390 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3392 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3393 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3396 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3398 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3399 case 7: return ConstantInt::getTrue(*Context);
3403 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3404 /// opcode and two operands into either a FCmp instruction. isordered is passed
3405 /// in to determine which kind of predicate to use in the new fcmp instruction.
3406 static Value *getFCmpValue(bool isordered, unsigned code,
3407 Value *LHS, Value *RHS, LLVMContext *Context) {
3409 default: llvm_unreachable("Illegal FCmp code!");
3412 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3414 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3417 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3419 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3422 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3424 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3427 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3429 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3432 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3434 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3437 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3439 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3442 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3444 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3445 case 7: return ConstantInt::getTrue(*Context);
3449 /// PredicatesFoldable - Return true if both predicates match sign or if at
3450 /// least one of them is an equality comparison (which is signless).
3451 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3452 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3453 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3454 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3458 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3459 struct FoldICmpLogical {
3462 ICmpInst::Predicate pred;
3463 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3464 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3465 pred(ICI->getPredicate()) {}
3466 bool shouldApply(Value *V) const {
3467 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3468 if (PredicatesFoldable(pred, ICI->getPredicate()))
3469 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3470 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3473 Instruction *apply(Instruction &Log) const {
3474 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3475 if (ICI->getOperand(0) != LHS) {
3476 assert(ICI->getOperand(1) == LHS);
3477 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3480 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3481 unsigned LHSCode = getICmpCode(ICI);
3482 unsigned RHSCode = getICmpCode(RHSICI);
3484 switch (Log.getOpcode()) {
3485 case Instruction::And: Code = LHSCode & RHSCode; break;
3486 case Instruction::Or: Code = LHSCode | RHSCode; break;
3487 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3488 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3491 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3492 ICmpInst::isSignedPredicate(ICI->getPredicate());
3494 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3495 if (Instruction *I = dyn_cast<Instruction>(RV))
3497 // Otherwise, it's a constant boolean value...
3498 return IC.ReplaceInstUsesWith(Log, RV);
3501 } // end anonymous namespace
3503 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3504 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3505 // guaranteed to be a binary operator.
3506 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3508 ConstantInt *AndRHS,
3509 BinaryOperator &TheAnd) {
3510 Value *X = Op->getOperand(0);
3511 Constant *Together = 0;
3513 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3515 switch (Op->getOpcode()) {
3516 case Instruction::Xor:
3517 if (Op->hasOneUse()) {
3518 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3519 Value *And = Builder->CreateAnd(X, AndRHS);
3521 return BinaryOperator::CreateXor(And, Together);
3524 case Instruction::Or:
3525 if (Together == AndRHS) // (X | C) & C --> C
3526 return ReplaceInstUsesWith(TheAnd, AndRHS);
3528 if (Op->hasOneUse() && Together != OpRHS) {
3529 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3530 Value *Or = Builder->CreateOr(X, Together);
3532 return BinaryOperator::CreateAnd(Or, AndRHS);
3535 case Instruction::Add:
3536 if (Op->hasOneUse()) {
3537 // Adding a one to a single bit bit-field should be turned into an XOR
3538 // of the bit. First thing to check is to see if this AND is with a
3539 // single bit constant.
3540 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3542 // If there is only one bit set...
3543 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3544 // Ok, at this point, we know that we are masking the result of the
3545 // ADD down to exactly one bit. If the constant we are adding has
3546 // no bits set below this bit, then we can eliminate the ADD.
3547 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3549 // Check to see if any bits below the one bit set in AndRHSV are set.
3550 if ((AddRHS & (AndRHSV-1)) == 0) {
3551 // If not, the only thing that can effect the output of the AND is
3552 // the bit specified by AndRHSV. If that bit is set, the effect of
3553 // the XOR is to toggle the bit. If it is clear, then the ADD has
3555 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3556 TheAnd.setOperand(0, X);
3559 // Pull the XOR out of the AND.
3560 Value *NewAnd = Builder->CreateAnd(X, AndRHS);
3561 NewAnd->takeName(Op);
3562 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3569 case Instruction::Shl: {
3570 // We know that the AND will not produce any of the bits shifted in, so if
3571 // the anded constant includes them, clear them now!
3573 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3574 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3575 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3576 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3578 if (CI->getValue() == ShlMask) {
3579 // Masking out bits that the shift already masks
3580 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3581 } else if (CI != AndRHS) { // Reducing bits set in and.
3582 TheAnd.setOperand(1, CI);
3587 case Instruction::LShr:
3589 // We know that the AND will not produce any of the bits shifted in, so if
3590 // the anded constant includes them, clear them now! This only applies to
3591 // unsigned shifts, because a signed shr may bring in set bits!
3593 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3594 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3595 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3596 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3598 if (CI->getValue() == ShrMask) {
3599 // Masking out bits that the shift already masks.
3600 return ReplaceInstUsesWith(TheAnd, Op);
3601 } else if (CI != AndRHS) {
3602 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3607 case Instruction::AShr:
3609 // See if this is shifting in some sign extension, then masking it out
3611 if (Op->hasOneUse()) {
3612 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3613 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3614 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3615 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3616 if (C == AndRHS) { // Masking out bits shifted in.
3617 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3618 // Make the argument unsigned.
3619 Value *ShVal = Op->getOperand(0);
3620 ShVal = Builder->CreateLShr(ShVal, OpRHS, Op->getName());
3621 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3630 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3631 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3632 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3633 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3634 /// insert new instructions.
3635 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3636 bool isSigned, bool Inside,
3638 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3639 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3640 "Lo is not <= Hi in range emission code!");
3643 if (Lo == Hi) // Trivially false.
3644 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3646 // V >= Min && V < Hi --> V < Hi
3647 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3648 ICmpInst::Predicate pred = (isSigned ?
3649 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3650 return new ICmpInst(pred, V, Hi);
3653 // Emit V-Lo <u Hi-Lo
3654 Constant *NegLo = ConstantExpr::getNeg(Lo);
3655 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
3656 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3657 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3660 if (Lo == Hi) // Trivially true.
3661 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3663 // V < Min || V >= Hi -> V > Hi-1
3664 Hi = SubOne(cast<ConstantInt>(Hi));
3665 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3666 ICmpInst::Predicate pred = (isSigned ?
3667 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3668 return new ICmpInst(pred, V, Hi);
3671 // Emit V-Lo >u Hi-1-Lo
3672 // Note that Hi has already had one subtracted from it, above.
3673 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3674 Value *Add = Builder->CreateAdd(V, NegLo, V->getName()+".off");
3675 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3676 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3679 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3680 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3681 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3682 // not, since all 1s are not contiguous.
3683 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3684 const APInt& V = Val->getValue();
3685 uint32_t BitWidth = Val->getType()->getBitWidth();
3686 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3688 // look for the first zero bit after the run of ones
3689 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3690 // look for the first non-zero bit
3691 ME = V.getActiveBits();
3695 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3696 /// where isSub determines whether the operator is a sub. If we can fold one of
3697 /// the following xforms:
3699 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3700 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3701 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3703 /// return (A +/- B).
3705 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3706 ConstantInt *Mask, bool isSub,
3708 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3709 if (!LHSI || LHSI->getNumOperands() != 2 ||
3710 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3712 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3714 switch (LHSI->getOpcode()) {
3716 case Instruction::And:
3717 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3718 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3719 if ((Mask->getValue().countLeadingZeros() +
3720 Mask->getValue().countPopulation()) ==
3721 Mask->getValue().getBitWidth())
3724 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3725 // part, we don't need any explicit masks to take them out of A. If that
3726 // is all N is, ignore it.
3727 uint32_t MB = 0, ME = 0;
3728 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3729 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3730 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3731 if (MaskedValueIsZero(RHS, Mask))
3736 case Instruction::Or:
3737 case Instruction::Xor:
3738 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3739 if ((Mask->getValue().countLeadingZeros() +
3740 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3741 && ConstantExpr::getAnd(N, Mask)->isNullValue())
3747 return Builder->CreateSub(LHSI->getOperand(0), RHS, "fold");
3748 return Builder->CreateAdd(LHSI->getOperand(0), RHS, "fold");
3751 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3752 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3753 ICmpInst *LHS, ICmpInst *RHS) {
3755 ConstantInt *LHSCst, *RHSCst;
3756 ICmpInst::Predicate LHSCC, RHSCC;
3758 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3759 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3760 m_ConstantInt(LHSCst))) ||
3761 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3762 m_ConstantInt(RHSCst))))
3765 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3766 // where C is a power of 2
3767 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3768 LHSCst->getValue().isPowerOf2()) {
3769 Value *NewOr = Builder->CreateOr(Val, Val2);
3770 return new ICmpInst(LHSCC, NewOr, LHSCst);
3773 // From here on, we only handle:
3774 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3775 if (Val != Val2) return 0;
3777 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3778 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3779 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3780 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3781 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3784 // We can't fold (ugt x, C) & (sgt x, C2).
3785 if (!PredicatesFoldable(LHSCC, RHSCC))
3788 // Ensure that the larger constant is on the RHS.
3790 if (ICmpInst::isSignedPredicate(LHSCC) ||
3791 (ICmpInst::isEquality(LHSCC) &&
3792 ICmpInst::isSignedPredicate(RHSCC)))
3793 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3795 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3798 std::swap(LHS, RHS);
3799 std::swap(LHSCst, RHSCst);
3800 std::swap(LHSCC, RHSCC);
3803 // At this point, we know we have have two icmp instructions
3804 // comparing a value against two constants and and'ing the result
3805 // together. Because of the above check, we know that we only have
3806 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3807 // (from the FoldICmpLogical check above), that the two constants
3808 // are not equal and that the larger constant is on the RHS
3809 assert(LHSCst != RHSCst && "Compares not folded above?");
3812 default: llvm_unreachable("Unknown integer condition code!");
3813 case ICmpInst::ICMP_EQ:
3815 default: llvm_unreachable("Unknown integer condition code!");
3816 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3817 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3818 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3819 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3820 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3821 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3822 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3823 return ReplaceInstUsesWith(I, LHS);
3825 case ICmpInst::ICMP_NE:
3827 default: llvm_unreachable("Unknown integer condition code!");
3828 case ICmpInst::ICMP_ULT:
3829 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3830 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3831 break; // (X != 13 & X u< 15) -> no change
3832 case ICmpInst::ICMP_SLT:
3833 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3834 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3835 break; // (X != 13 & X s< 15) -> no change
3836 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3837 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3838 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3839 return ReplaceInstUsesWith(I, RHS);
3840 case ICmpInst::ICMP_NE:
3841 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3842 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3843 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
3844 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3845 ConstantInt::get(Add->getType(), 1));
3847 break; // (X != 13 & X != 15) -> no change
3850 case ICmpInst::ICMP_ULT:
3852 default: llvm_unreachable("Unknown integer condition code!");
3853 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3854 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3855 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3856 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3858 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3859 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3860 return ReplaceInstUsesWith(I, LHS);
3861 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3865 case ICmpInst::ICMP_SLT:
3867 default: llvm_unreachable("Unknown integer condition code!");
3868 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3869 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3870 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3871 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3873 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3874 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3875 return ReplaceInstUsesWith(I, LHS);
3876 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3880 case ICmpInst::ICMP_UGT:
3882 default: llvm_unreachable("Unknown integer condition code!");
3883 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3884 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3885 return ReplaceInstUsesWith(I, RHS);
3886 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3888 case ICmpInst::ICMP_NE:
3889 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3890 return new ICmpInst(LHSCC, Val, RHSCst);
3891 break; // (X u> 13 & X != 15) -> no change
3892 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3893 return InsertRangeTest(Val, AddOne(LHSCst),
3894 RHSCst, false, true, I);
3895 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3899 case ICmpInst::ICMP_SGT:
3901 default: llvm_unreachable("Unknown integer condition code!");
3902 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3903 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3904 return ReplaceInstUsesWith(I, RHS);
3905 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3907 case ICmpInst::ICMP_NE:
3908 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3909 return new ICmpInst(LHSCC, Val, RHSCst);
3910 break; // (X s> 13 & X != 15) -> no change
3911 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3912 return InsertRangeTest(Val, AddOne(LHSCst),
3913 RHSCst, true, true, I);
3914 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3923 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
3926 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3927 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3928 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3929 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3930 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3931 // If either of the constants are nans, then the whole thing returns
3933 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3934 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3935 return new FCmpInst(FCmpInst::FCMP_ORD,
3936 LHS->getOperand(0), RHS->getOperand(0));
3939 // Handle vector zeros. This occurs because the canonical form of
3940 // "fcmp ord x,x" is "fcmp ord x, 0".
3941 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
3942 isa<ConstantAggregateZero>(RHS->getOperand(1)))
3943 return new FCmpInst(FCmpInst::FCMP_ORD,
3944 LHS->getOperand(0), RHS->getOperand(0));
3948 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
3949 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
3950 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
3953 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
3954 // Swap RHS operands to match LHS.
3955 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
3956 std::swap(Op1LHS, Op1RHS);
3959 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
3960 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
3962 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
3964 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
3965 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3966 if (Op0CC == FCmpInst::FCMP_TRUE)
3967 return ReplaceInstUsesWith(I, RHS);
3968 if (Op1CC == FCmpInst::FCMP_TRUE)
3969 return ReplaceInstUsesWith(I, LHS);
3973 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
3974 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
3976 std::swap(LHS, RHS);
3977 std::swap(Op0Pred, Op1Pred);
3978 std::swap(Op0Ordered, Op1Ordered);
3981 // uno && ueq -> uno && (uno || eq) -> ueq
3982 // ord && olt -> ord && (ord && lt) -> olt
3983 if (Op0Ordered == Op1Ordered)
3984 return ReplaceInstUsesWith(I, RHS);
3986 // uno && oeq -> uno && (ord && eq) -> false
3987 // uno && ord -> false
3989 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3990 // ord && ueq -> ord && (uno || eq) -> oeq
3991 return cast<Instruction>(getFCmpValue(true, Op1Pred,
3992 Op0LHS, Op0RHS, Context));
4000 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4001 bool Changed = SimplifyCommutative(I);
4002 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4004 if (isa<UndefValue>(Op1)) // X & undef -> 0
4005 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4009 return ReplaceInstUsesWith(I, Op1);
4011 // See if we can simplify any instructions used by the instruction whose sole
4012 // purpose is to compute bits we don't care about.
4013 if (SimplifyDemandedInstructionBits(I))
4015 if (isa<VectorType>(I.getType())) {
4016 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4017 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4018 return ReplaceInstUsesWith(I, I.getOperand(0));
4019 } else if (isa<ConstantAggregateZero>(Op1)) {
4020 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4024 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4025 const APInt& AndRHSMask = AndRHS->getValue();
4026 APInt NotAndRHS(~AndRHSMask);
4028 // Optimize a variety of ((val OP C1) & C2) combinations...
4029 if (isa<BinaryOperator>(Op0)) {
4030 Instruction *Op0I = cast<Instruction>(Op0);
4031 Value *Op0LHS = Op0I->getOperand(0);
4032 Value *Op0RHS = Op0I->getOperand(1);
4033 switch (Op0I->getOpcode()) {
4034 case Instruction::Xor:
4035 case Instruction::Or:
4036 // If the mask is only needed on one incoming arm, push it up.
4037 if (Op0I->hasOneUse()) {
4038 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4039 // Not masking anything out for the LHS, move to RHS.
4040 Value *NewRHS = Builder->CreateAnd(Op0RHS, AndRHS,
4041 Op0RHS->getName()+".masked");
4042 return BinaryOperator::Create(
4043 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4045 if (!isa<Constant>(Op0RHS) &&
4046 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4047 // Not masking anything out for the RHS, move to LHS.
4048 Value *NewLHS = Builder->CreateAnd(Op0LHS, AndRHS,
4049 Op0LHS->getName()+".masked");
4050 return BinaryOperator::Create(
4051 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4056 case Instruction::Add:
4057 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4058 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4059 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4060 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4061 return BinaryOperator::CreateAnd(V, AndRHS);
4062 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4063 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4066 case Instruction::Sub:
4067 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4068 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4069 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4070 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4071 return BinaryOperator::CreateAnd(V, AndRHS);
4073 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4074 // has 1's for all bits that the subtraction with A might affect.
4075 if (Op0I->hasOneUse()) {
4076 uint32_t BitWidth = AndRHSMask.getBitWidth();
4077 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4078 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4080 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4081 if (!(A && A->isZero()) && // avoid infinite recursion.
4082 MaskedValueIsZero(Op0LHS, Mask)) {
4083 Value *NewNeg = Builder->CreateNeg(Op0RHS);
4084 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4089 case Instruction::Shl:
4090 case Instruction::LShr:
4091 // (1 << x) & 1 --> zext(x == 0)
4092 // (1 >> x) & 1 --> zext(x == 0)
4093 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4095 Builder->CreateICmpEQ(Op0RHS, Constant::getNullValue(I.getType()));
4096 return new ZExtInst(NewICmp, I.getType());
4101 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4102 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4104 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4105 // If this is an integer truncation or change from signed-to-unsigned, and
4106 // if the source is an and/or with immediate, transform it. This
4107 // frequently occurs for bitfield accesses.
4108 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4109 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4110 CastOp->getNumOperands() == 2)
4111 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4112 if (CastOp->getOpcode() == Instruction::And) {
4113 // Change: and (cast (and X, C1) to T), C2
4114 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4115 // This will fold the two constants together, which may allow
4116 // other simplifications.
4117 Value *NewCast = Builder->CreateTruncOrBitCast(
4118 CastOp->getOperand(0), I.getType(),
4119 CastOp->getName()+".shrunk");
4120 // trunc_or_bitcast(C1)&C2
4121 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4122 C3 = ConstantExpr::getAnd(C3, AndRHS);
4123 return BinaryOperator::CreateAnd(NewCast, C3);
4124 } else if (CastOp->getOpcode() == Instruction::Or) {
4125 // Change: and (cast (or X, C1) to T), C2
4126 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4127 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4128 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4130 return ReplaceInstUsesWith(I, AndRHS);
4136 // Try to fold constant and into select arguments.
4137 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4138 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4140 if (isa<PHINode>(Op0))
4141 if (Instruction *NV = FoldOpIntoPhi(I))
4145 Value *Op0NotVal = dyn_castNotVal(Op0);
4146 Value *Op1NotVal = dyn_castNotVal(Op1);
4148 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4149 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4151 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4152 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4153 Value *Or = Builder->CreateOr(Op0NotVal, Op1NotVal,
4154 I.getName()+".demorgan");
4155 return BinaryOperator::CreateNot(Or);
4159 Value *A = 0, *B = 0, *C = 0, *D = 0;
4160 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4161 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4162 return ReplaceInstUsesWith(I, Op1);
4164 // (A|B) & ~(A&B) -> A^B
4165 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4166 if ((A == C && B == D) || (A == D && B == C))
4167 return BinaryOperator::CreateXor(A, B);
4171 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4172 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4173 return ReplaceInstUsesWith(I, Op0);
4175 // ~(A&B) & (A|B) -> A^B
4176 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4177 if ((A == C && B == D) || (A == D && B == C))
4178 return BinaryOperator::CreateXor(A, B);
4182 if (Op0->hasOneUse() &&
4183 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4184 if (A == Op1) { // (A^B)&A -> A&(A^B)
4185 I.swapOperands(); // Simplify below
4186 std::swap(Op0, Op1);
4187 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4188 cast<BinaryOperator>(Op0)->swapOperands();
4189 I.swapOperands(); // Simplify below
4190 std::swap(Op0, Op1);
4194 if (Op1->hasOneUse() &&
4195 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4196 if (B == Op0) { // B&(A^B) -> B&(B^A)
4197 cast<BinaryOperator>(Op1)->swapOperands();
4200 if (A == Op0) // A&(A^B) -> A & ~B
4201 return BinaryOperator::CreateAnd(A, Builder->CreateNot(B, "tmp"));
4204 // (A&((~A)|B)) -> A&B
4205 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4206 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4207 return BinaryOperator::CreateAnd(A, Op1);
4208 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4209 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4210 return BinaryOperator::CreateAnd(A, Op0);
4213 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4214 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4215 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4218 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4219 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4223 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4224 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4225 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4226 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4227 const Type *SrcTy = Op0C->getOperand(0)->getType();
4228 if (SrcTy == Op1C->getOperand(0)->getType() &&
4229 SrcTy->isIntOrIntVector() &&
4230 // Only do this if the casts both really cause code to be generated.
4231 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4233 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4235 Value *NewOp = Builder->CreateAnd(Op0C->getOperand(0),
4236 Op1C->getOperand(0), I.getName());
4237 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4241 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4242 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4243 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4244 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4245 SI0->getOperand(1) == SI1->getOperand(1) &&
4246 (SI0->hasOneUse() || SI1->hasOneUse())) {
4248 Builder->CreateAnd(SI0->getOperand(0), SI1->getOperand(0),
4250 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4251 SI1->getOperand(1));
4255 // If and'ing two fcmp, try combine them into one.
4256 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4257 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4258 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4262 return Changed ? &I : 0;
4265 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4266 /// capable of providing pieces of a bswap. The subexpression provides pieces
4267 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4268 /// the expression came from the corresponding "byte swapped" byte in some other
4269 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4270 /// we know that the expression deposits the low byte of %X into the high byte
4271 /// of the bswap result and that all other bytes are zero. This expression is
4272 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4275 /// This function returns true if the match was unsuccessful and false if so.
4276 /// On entry to the function the "OverallLeftShift" is a signed integer value
4277 /// indicating the number of bytes that the subexpression is later shifted. For
4278 /// example, if the expression is later right shifted by 16 bits, the
4279 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4280 /// byte of ByteValues is actually being set.
4282 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4283 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4284 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4285 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4286 /// always in the local (OverallLeftShift) coordinate space.
4288 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4289 SmallVector<Value*, 8> &ByteValues) {
4290 if (Instruction *I = dyn_cast<Instruction>(V)) {
4291 // If this is an or instruction, it may be an inner node of the bswap.
4292 if (I->getOpcode() == Instruction::Or) {
4293 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4295 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4299 // If this is a logical shift by a constant multiple of 8, recurse with
4300 // OverallLeftShift and ByteMask adjusted.
4301 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4303 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4304 // Ensure the shift amount is defined and of a byte value.
4305 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4308 unsigned ByteShift = ShAmt >> 3;
4309 if (I->getOpcode() == Instruction::Shl) {
4310 // X << 2 -> collect(X, +2)
4311 OverallLeftShift += ByteShift;
4312 ByteMask >>= ByteShift;
4314 // X >>u 2 -> collect(X, -2)
4315 OverallLeftShift -= ByteShift;
4316 ByteMask <<= ByteShift;
4317 ByteMask &= (~0U >> (32-ByteValues.size()));
4320 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4321 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4323 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4327 // If this is a logical 'and' with a mask that clears bytes, clear the
4328 // corresponding bytes in ByteMask.
4329 if (I->getOpcode() == Instruction::And &&
4330 isa<ConstantInt>(I->getOperand(1))) {
4331 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4332 unsigned NumBytes = ByteValues.size();
4333 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4334 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4336 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4337 // If this byte is masked out by a later operation, we don't care what
4339 if ((ByteMask & (1 << i)) == 0)
4342 // If the AndMask is all zeros for this byte, clear the bit.
4343 APInt MaskB = AndMask & Byte;
4345 ByteMask &= ~(1U << i);
4349 // If the AndMask is not all ones for this byte, it's not a bytezap.
4353 // Otherwise, this byte is kept.
4356 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4361 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4362 // the input value to the bswap. Some observations: 1) if more than one byte
4363 // is demanded from this input, then it could not be successfully assembled
4364 // into a byteswap. At least one of the two bytes would not be aligned with
4365 // their ultimate destination.
4366 if (!isPowerOf2_32(ByteMask)) return true;
4367 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4369 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4370 // is demanded, it needs to go into byte 0 of the result. This means that the
4371 // byte needs to be shifted until it lands in the right byte bucket. The
4372 // shift amount depends on the position: if the byte is coming from the high
4373 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4374 // low part, it must be shifted left.
4375 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4376 if (InputByteNo < ByteValues.size()/2) {
4377 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4380 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4384 // If the destination byte value is already defined, the values are or'd
4385 // together, which isn't a bswap (unless it's an or of the same bits).
4386 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4388 ByteValues[DestByteNo] = V;
4392 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4393 /// If so, insert the new bswap intrinsic and return it.
4394 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4395 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4396 if (!ITy || ITy->getBitWidth() % 16 ||
4397 // ByteMask only allows up to 32-byte values.
4398 ITy->getBitWidth() > 32*8)
4399 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4401 /// ByteValues - For each byte of the result, we keep track of which value
4402 /// defines each byte.
4403 SmallVector<Value*, 8> ByteValues;
4404 ByteValues.resize(ITy->getBitWidth()/8);
4406 // Try to find all the pieces corresponding to the bswap.
4407 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4408 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4411 // Check to see if all of the bytes come from the same value.
4412 Value *V = ByteValues[0];
4413 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4415 // Check to make sure that all of the bytes come from the same value.
4416 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4417 if (ByteValues[i] != V)
4419 const Type *Tys[] = { ITy };
4420 Module *M = I.getParent()->getParent()->getParent();
4421 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4422 return CallInst::Create(F, V);
4425 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4426 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4427 /// we can simplify this expression to "cond ? C : D or B".
4428 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4430 LLVMContext *Context) {
4431 // If A is not a select of -1/0, this cannot match.
4433 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4436 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4437 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4438 return SelectInst::Create(Cond, C, B);
4439 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4440 return SelectInst::Create(Cond, C, B);
4441 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4442 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4443 return SelectInst::Create(Cond, C, D);
4444 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4445 return SelectInst::Create(Cond, C, D);
4449 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4450 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4451 ICmpInst *LHS, ICmpInst *RHS) {
4453 ConstantInt *LHSCst, *RHSCst;
4454 ICmpInst::Predicate LHSCC, RHSCC;
4456 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4457 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4458 m_ConstantInt(LHSCst))) ||
4459 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4460 m_ConstantInt(RHSCst))))
4463 // From here on, we only handle:
4464 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4465 if (Val != Val2) return 0;
4467 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4468 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4469 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4470 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4471 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4474 // We can't fold (ugt x, C) | (sgt x, C2).
4475 if (!PredicatesFoldable(LHSCC, RHSCC))
4478 // Ensure that the larger constant is on the RHS.
4480 if (ICmpInst::isSignedPredicate(LHSCC) ||
4481 (ICmpInst::isEquality(LHSCC) &&
4482 ICmpInst::isSignedPredicate(RHSCC)))
4483 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4485 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4488 std::swap(LHS, RHS);
4489 std::swap(LHSCst, RHSCst);
4490 std::swap(LHSCC, RHSCC);
4493 // At this point, we know we have have two icmp instructions
4494 // comparing a value against two constants and or'ing the result
4495 // together. Because of the above check, we know that we only have
4496 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4497 // FoldICmpLogical check above), that the two constants are not
4499 assert(LHSCst != RHSCst && "Compares not folded above?");
4502 default: llvm_unreachable("Unknown integer condition code!");
4503 case ICmpInst::ICMP_EQ:
4505 default: llvm_unreachable("Unknown integer condition code!");
4506 case ICmpInst::ICMP_EQ:
4507 if (LHSCst == SubOne(RHSCst)) {
4508 // (X == 13 | X == 14) -> X-13 <u 2
4509 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4510 Value *Add = Builder->CreateAdd(Val, AddCST, Val->getName()+".off");
4511 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4512 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4514 break; // (X == 13 | X == 15) -> no change
4515 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4516 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4518 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4519 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4520 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4521 return ReplaceInstUsesWith(I, RHS);
4524 case ICmpInst::ICMP_NE:
4526 default: llvm_unreachable("Unknown integer condition code!");
4527 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4528 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4529 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4530 return ReplaceInstUsesWith(I, LHS);
4531 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4532 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4533 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4534 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4537 case ICmpInst::ICMP_ULT:
4539 default: llvm_unreachable("Unknown integer condition code!");
4540 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4542 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4543 // If RHSCst is [us]MAXINT, it is always false. Not handling
4544 // this can cause overflow.
4545 if (RHSCst->isMaxValue(false))
4546 return ReplaceInstUsesWith(I, LHS);
4547 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4549 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4551 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4552 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4553 return ReplaceInstUsesWith(I, RHS);
4554 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4558 case ICmpInst::ICMP_SLT:
4560 default: llvm_unreachable("Unknown integer condition code!");
4561 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4563 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4564 // If RHSCst is [us]MAXINT, it is always false. Not handling
4565 // this can cause overflow.
4566 if (RHSCst->isMaxValue(true))
4567 return ReplaceInstUsesWith(I, LHS);
4568 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4570 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4572 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4573 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4574 return ReplaceInstUsesWith(I, RHS);
4575 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4579 case ICmpInst::ICMP_UGT:
4581 default: llvm_unreachable("Unknown integer condition code!");
4582 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4583 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4584 return ReplaceInstUsesWith(I, LHS);
4585 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4587 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4588 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4589 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4590 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4594 case ICmpInst::ICMP_SGT:
4596 default: llvm_unreachable("Unknown integer condition code!");
4597 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4598 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4599 return ReplaceInstUsesWith(I, LHS);
4600 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4602 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4603 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4604 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4605 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4613 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
4615 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4616 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4617 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4618 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4619 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4620 // If either of the constants are nans, then the whole thing returns
4622 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4623 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4625 // Otherwise, no need to compare the two constants, compare the
4627 return new FCmpInst(FCmpInst::FCMP_UNO,
4628 LHS->getOperand(0), RHS->getOperand(0));
4631 // Handle vector zeros. This occurs because the canonical form of
4632 // "fcmp uno x,x" is "fcmp uno x, 0".
4633 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4634 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4635 return new FCmpInst(FCmpInst::FCMP_UNO,
4636 LHS->getOperand(0), RHS->getOperand(0));
4641 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4642 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4643 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4645 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4646 // Swap RHS operands to match LHS.
4647 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4648 std::swap(Op1LHS, Op1RHS);
4650 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4651 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4653 return new FCmpInst((FCmpInst::Predicate)Op0CC,
4655 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
4656 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4657 if (Op0CC == FCmpInst::FCMP_FALSE)
4658 return ReplaceInstUsesWith(I, RHS);
4659 if (Op1CC == FCmpInst::FCMP_FALSE)
4660 return ReplaceInstUsesWith(I, LHS);
4663 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4664 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4665 if (Op0Ordered == Op1Ordered) {
4666 // If both are ordered or unordered, return a new fcmp with
4667 // or'ed predicates.
4668 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4669 Op0LHS, Op0RHS, Context);
4670 if (Instruction *I = dyn_cast<Instruction>(RV))
4672 // Otherwise, it's a constant boolean value...
4673 return ReplaceInstUsesWith(I, RV);
4679 /// FoldOrWithConstants - This helper function folds:
4681 /// ((A | B) & C1) | (B & C2)
4687 /// when the XOR of the two constants is "all ones" (-1).
4688 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4689 Value *A, Value *B, Value *C) {
4690 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4694 ConstantInt *CI2 = 0;
4695 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4697 APInt Xor = CI1->getValue() ^ CI2->getValue();
4698 if (!Xor.isAllOnesValue()) return 0;
4700 if (V1 == A || V1 == B) {
4701 Value *NewOp = Builder->CreateAnd((V1 == A) ? B : A, CI1);
4702 return BinaryOperator::CreateOr(NewOp, V1);
4708 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4709 bool Changed = SimplifyCommutative(I);
4710 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4712 if (isa<UndefValue>(Op1)) // X | undef -> -1
4713 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4717 return ReplaceInstUsesWith(I, Op0);
4719 // See if we can simplify any instructions used by the instruction whose sole
4720 // purpose is to compute bits we don't care about.
4721 if (SimplifyDemandedInstructionBits(I))
4723 if (isa<VectorType>(I.getType())) {
4724 if (isa<ConstantAggregateZero>(Op1)) {
4725 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4726 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4727 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4728 return ReplaceInstUsesWith(I, I.getOperand(1));
4733 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4734 ConstantInt *C1 = 0; Value *X = 0;
4735 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4736 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
4738 Value *Or = Builder->CreateOr(X, RHS);
4740 return BinaryOperator::CreateAnd(Or,
4741 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
4744 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4745 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
4747 Value *Or = Builder->CreateOr(X, RHS);
4749 return BinaryOperator::CreateXor(Or,
4750 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
4753 // Try to fold constant and into select arguments.
4754 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4755 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4757 if (isa<PHINode>(Op0))
4758 if (Instruction *NV = FoldOpIntoPhi(I))
4762 Value *A = 0, *B = 0;
4763 ConstantInt *C1 = 0, *C2 = 0;
4765 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4766 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4767 return ReplaceInstUsesWith(I, Op1);
4768 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4769 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4770 return ReplaceInstUsesWith(I, Op0);
4772 // (A | B) | C and A | (B | C) -> bswap if possible.
4773 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4774 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4775 match(Op1, m_Or(m_Value(), m_Value())) ||
4776 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4777 match(Op1, m_Shift(m_Value(), m_Value())))) {
4778 if (Instruction *BSwap = MatchBSwap(I))
4782 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4783 if (Op0->hasOneUse() &&
4784 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4785 MaskedValueIsZero(Op1, C1->getValue())) {
4786 Value *NOr = Builder->CreateOr(A, Op1);
4788 return BinaryOperator::CreateXor(NOr, C1);
4791 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4792 if (Op1->hasOneUse() &&
4793 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4794 MaskedValueIsZero(Op0, C1->getValue())) {
4795 Value *NOr = Builder->CreateOr(A, Op0);
4797 return BinaryOperator::CreateXor(NOr, C1);
4801 Value *C = 0, *D = 0;
4802 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4803 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4804 Value *V1 = 0, *V2 = 0, *V3 = 0;
4805 C1 = dyn_cast<ConstantInt>(C);
4806 C2 = dyn_cast<ConstantInt>(D);
4807 if (C1 && C2) { // (A & C1)|(B & C2)
4808 // If we have: ((V + N) & C1) | (V & C2)
4809 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4810 // replace with V+N.
4811 if (C1->getValue() == ~C2->getValue()) {
4812 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4813 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4814 // Add commutes, try both ways.
4815 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4816 return ReplaceInstUsesWith(I, A);
4817 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4818 return ReplaceInstUsesWith(I, A);
4820 // Or commutes, try both ways.
4821 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4822 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4823 // Add commutes, try both ways.
4824 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4825 return ReplaceInstUsesWith(I, B);
4826 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4827 return ReplaceInstUsesWith(I, B);
4830 V1 = 0; V2 = 0; V3 = 0;
4833 // Check to see if we have any common things being and'ed. If so, find the
4834 // terms for V1 & (V2|V3).
4835 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4836 if (A == B) // (A & C)|(A & D) == A & (C|D)
4837 V1 = A, V2 = C, V3 = D;
4838 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4839 V1 = A, V2 = B, V3 = C;
4840 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4841 V1 = C, V2 = A, V3 = D;
4842 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4843 V1 = C, V2 = A, V3 = B;
4846 Value *Or = Builder->CreateOr(V2, V3, "tmp");
4847 return BinaryOperator::CreateAnd(V1, Or);
4851 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4852 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4854 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4856 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4858 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4861 // ((A&~B)|(~A&B)) -> A^B
4862 if ((match(C, m_Not(m_Specific(D))) &&
4863 match(B, m_Not(m_Specific(A)))))
4864 return BinaryOperator::CreateXor(A, D);
4865 // ((~B&A)|(~A&B)) -> A^B
4866 if ((match(A, m_Not(m_Specific(D))) &&
4867 match(B, m_Not(m_Specific(C)))))
4868 return BinaryOperator::CreateXor(C, D);
4869 // ((A&~B)|(B&~A)) -> A^B
4870 if ((match(C, m_Not(m_Specific(B))) &&
4871 match(D, m_Not(m_Specific(A)))))
4872 return BinaryOperator::CreateXor(A, B);
4873 // ((~B&A)|(B&~A)) -> A^B
4874 if ((match(A, m_Not(m_Specific(B))) &&
4875 match(D, m_Not(m_Specific(C)))))
4876 return BinaryOperator::CreateXor(C, B);
4879 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4880 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4881 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4882 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4883 SI0->getOperand(1) == SI1->getOperand(1) &&
4884 (SI0->hasOneUse() || SI1->hasOneUse())) {
4885 Value *NewOp = Builder->CreateOr(SI0->getOperand(0), SI1->getOperand(0),
4887 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4888 SI1->getOperand(1));
4892 // ((A|B)&1)|(B&-2) -> (A&1) | B
4893 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4894 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4895 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4896 if (Ret) return Ret;
4898 // (B&-2)|((A|B)&1) -> (A&1) | B
4899 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4900 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4901 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4902 if (Ret) return Ret;
4905 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4906 if (A == Op1) // ~A | A == -1
4907 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4911 // Note, A is still live here!
4912 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4914 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4916 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4917 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4918 Value *And = Builder->CreateAnd(A, B, I.getName()+".demorgan");
4919 return BinaryOperator::CreateNot(And);
4923 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4924 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4925 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4928 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4929 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4933 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4934 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4935 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4936 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4937 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4938 !isa<ICmpInst>(Op1C->getOperand(0))) {
4939 const Type *SrcTy = Op0C->getOperand(0)->getType();
4940 if (SrcTy == Op1C->getOperand(0)->getType() &&
4941 SrcTy->isIntOrIntVector() &&
4942 // Only do this if the casts both really cause code to be
4944 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4946 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4948 Value *NewOp = Builder->CreateOr(Op0C->getOperand(0),
4949 Op1C->getOperand(0), I.getName());
4950 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4957 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4958 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4959 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4960 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
4964 return Changed ? &I : 0;
4969 // XorSelf - Implements: X ^ X --> 0
4972 XorSelf(Value *rhs) : RHS(rhs) {}
4973 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4974 Instruction *apply(BinaryOperator &Xor) const {
4981 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4982 bool Changed = SimplifyCommutative(I);
4983 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4985 if (isa<UndefValue>(Op1)) {
4986 if (isa<UndefValue>(Op0))
4987 // Handle undef ^ undef -> 0 special case. This is a common
4989 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4990 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4993 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4994 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4995 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4996 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4999 // See if we can simplify any instructions used by the instruction whose sole
5000 // purpose is to compute bits we don't care about.
5001 if (SimplifyDemandedInstructionBits(I))
5003 if (isa<VectorType>(I.getType()))
5004 if (isa<ConstantAggregateZero>(Op1))
5005 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5007 // Is this a ~ operation?
5008 if (Value *NotOp = dyn_castNotVal(&I)) {
5009 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5010 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5011 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5012 if (Op0I->getOpcode() == Instruction::And ||
5013 Op0I->getOpcode() == Instruction::Or) {
5014 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
5015 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5017 Builder->CreateNot(Op0I->getOperand(1),
5018 Op0I->getOperand(1)->getName()+".not");
5019 if (Op0I->getOpcode() == Instruction::And)
5020 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5021 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5028 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5029 if (RHS == ConstantInt::getTrue(*Context) && Op0->hasOneUse()) {
5030 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5031 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5032 return new ICmpInst(ICI->getInversePredicate(),
5033 ICI->getOperand(0), ICI->getOperand(1));
5035 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5036 return new FCmpInst(FCI->getInversePredicate(),
5037 FCI->getOperand(0), FCI->getOperand(1));
5040 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5041 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5042 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5043 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5044 Instruction::CastOps Opcode = Op0C->getOpcode();
5045 if ((Opcode == Instruction::ZExt || Opcode == Instruction::SExt) &&
5046 (RHS == ConstantExpr::getCast(Opcode,
5047 ConstantInt::getTrue(*Context),
5048 Op0C->getDestTy()))) {
5049 CI->setPredicate(CI->getInversePredicate());
5050 return CastInst::Create(Opcode, CI, Op0C->getType());
5056 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5057 // ~(c-X) == X-c-1 == X+(-c-1)
5058 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5059 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5060 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5061 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5062 ConstantInt::get(I.getType(), 1));
5063 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5066 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5067 if (Op0I->getOpcode() == Instruction::Add) {
5068 // ~(X-c) --> (-c-1)-X
5069 if (RHS->isAllOnesValue()) {
5070 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5071 return BinaryOperator::CreateSub(
5072 ConstantExpr::getSub(NegOp0CI,
5073 ConstantInt::get(I.getType(), 1)),
5074 Op0I->getOperand(0));
5075 } else if (RHS->getValue().isSignBit()) {
5076 // (X + C) ^ signbit -> (X + C + signbit)
5077 Constant *C = ConstantInt::get(*Context,
5078 RHS->getValue() + Op0CI->getValue());
5079 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5082 } else if (Op0I->getOpcode() == Instruction::Or) {
5083 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5084 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5085 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5086 // Anything in both C1 and C2 is known to be zero, remove it from
5088 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5089 NewRHS = ConstantExpr::getAnd(NewRHS,
5090 ConstantExpr::getNot(CommonBits));
5092 I.setOperand(0, Op0I->getOperand(0));
5093 I.setOperand(1, NewRHS);
5100 // Try to fold constant and into select arguments.
5101 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5102 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5104 if (isa<PHINode>(Op0))
5105 if (Instruction *NV = FoldOpIntoPhi(I))
5109 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5111 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5113 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5115 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5118 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5121 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5122 if (A == Op0) { // B^(B|A) == (A|B)^B
5123 Op1I->swapOperands();
5125 std::swap(Op0, Op1);
5126 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5127 I.swapOperands(); // Simplified below.
5128 std::swap(Op0, Op1);
5130 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5131 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5132 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5133 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5134 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5136 if (A == Op0) { // A^(A&B) -> A^(B&A)
5137 Op1I->swapOperands();
5140 if (B == Op0) { // A^(B&A) -> (B&A)^A
5141 I.swapOperands(); // Simplified below.
5142 std::swap(Op0, Op1);
5147 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5150 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5151 Op0I->hasOneUse()) {
5152 if (A == Op1) // (B|A)^B == (A|B)^B
5154 if (B == Op1) // (A|B)^B == A & ~B
5155 return BinaryOperator::CreateAnd(A, Builder->CreateNot(Op1, "tmp"));
5156 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5157 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5158 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5159 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5160 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5162 if (A == Op1) // (A&B)^A -> (B&A)^A
5164 if (B == Op1 && // (B&A)^A == ~B & A
5165 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5166 return BinaryOperator::CreateAnd(Builder->CreateNot(A, "tmp"), Op1);
5171 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5172 if (Op0I && Op1I && Op0I->isShift() &&
5173 Op0I->getOpcode() == Op1I->getOpcode() &&
5174 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5175 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5177 Builder->CreateXor(Op0I->getOperand(0), Op1I->getOperand(0),
5179 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5180 Op1I->getOperand(1));
5184 Value *A, *B, *C, *D;
5185 // (A & B)^(A | B) -> A ^ B
5186 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5187 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5188 if ((A == C && B == D) || (A == D && B == C))
5189 return BinaryOperator::CreateXor(A, B);
5191 // (A | B)^(A & B) -> A ^ B
5192 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5193 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5194 if ((A == C && B == D) || (A == D && B == C))
5195 return BinaryOperator::CreateXor(A, B);
5199 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5200 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5201 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5202 // (X & Y)^(X & Y) -> (Y^Z) & X
5203 Value *X = 0, *Y = 0, *Z = 0;
5205 X = A, Y = B, Z = D;
5207 X = A, Y = B, Z = C;
5209 X = B, Y = A, Z = D;
5211 X = B, Y = A, Z = C;
5214 Value *NewOp = Builder->CreateXor(Y, Z, Op0->getName());
5215 return BinaryOperator::CreateAnd(NewOp, X);
5220 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5221 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5222 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5225 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5226 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5227 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5228 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5229 const Type *SrcTy = Op0C->getOperand(0)->getType();
5230 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5231 // Only do this if the casts both really cause code to be generated.
5232 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5234 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5236 Value *NewOp = Builder->CreateXor(Op0C->getOperand(0),
5237 Op1C->getOperand(0), I.getName());
5238 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5243 return Changed ? &I : 0;
5246 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5247 LLVMContext *Context) {
5248 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5251 static bool HasAddOverflow(ConstantInt *Result,
5252 ConstantInt *In1, ConstantInt *In2,
5255 if (In2->getValue().isNegative())
5256 return Result->getValue().sgt(In1->getValue());
5258 return Result->getValue().slt(In1->getValue());
5260 return Result->getValue().ult(In1->getValue());
5263 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5264 /// overflowed for this type.
5265 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5266 Constant *In2, LLVMContext *Context,
5267 bool IsSigned = false) {
5268 Result = ConstantExpr::getAdd(In1, In2);
5270 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5271 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5272 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5273 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5274 ExtractElement(In1, Idx, Context),
5275 ExtractElement(In2, Idx, Context),
5282 return HasAddOverflow(cast<ConstantInt>(Result),
5283 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5287 static bool HasSubOverflow(ConstantInt *Result,
5288 ConstantInt *In1, ConstantInt *In2,
5291 if (In2->getValue().isNegative())
5292 return Result->getValue().slt(In1->getValue());
5294 return Result->getValue().sgt(In1->getValue());
5296 return Result->getValue().ugt(In1->getValue());
5299 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5300 /// overflowed for this type.
5301 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5302 Constant *In2, LLVMContext *Context,
5303 bool IsSigned = false) {
5304 Result = ConstantExpr::getSub(In1, In2);
5306 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5307 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5308 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5309 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5310 ExtractElement(In1, Idx, Context),
5311 ExtractElement(In2, Idx, Context),
5318 return HasSubOverflow(cast<ConstantInt>(Result),
5319 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5323 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5324 /// code necessary to compute the offset from the base pointer (without adding
5325 /// in the base pointer). Return the result as a signed integer of intptr size.
5326 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5327 TargetData &TD = *IC.getTargetData();
5328 gep_type_iterator GTI = gep_type_begin(GEP);
5329 const Type *IntPtrTy = TD.getIntPtrType(I.getContext());
5330 Value *Result = Constant::getNullValue(IntPtrTy);
5332 // Build a mask for high order bits.
5333 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5334 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5336 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5339 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5340 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5341 if (OpC->isZero()) continue;
5343 // Handle a struct index, which adds its field offset to the pointer.
5344 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5345 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5347 Result = IC.Builder->CreateAdd(Result,
5348 ConstantInt::get(IntPtrTy, Size),
5349 GEP->getName()+".offs");
5353 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5355 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5356 Scale = ConstantExpr::getMul(OC, Scale);
5357 // Emit an add instruction.
5358 Result = IC.Builder->CreateAdd(Result, Scale, GEP->getName()+".offs");
5361 // Convert to correct type.
5362 if (Op->getType() != IntPtrTy)
5363 Op = IC.Builder->CreateIntCast(Op, IntPtrTy, true, Op->getName()+".c");
5365 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5366 // We'll let instcombine(mul) convert this to a shl if possible.
5367 Op = IC.Builder->CreateMul(Op, Scale, GEP->getName()+".idx");
5370 // Emit an add instruction.
5371 Result = IC.Builder->CreateAdd(Op, Result, GEP->getName()+".offs");
5377 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5378 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5379 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5380 /// be complex, and scales are involved. The above expression would also be
5381 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5382 /// This later form is less amenable to optimization though, and we are allowed
5383 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5385 /// If we can't emit an optimized form for this expression, this returns null.
5387 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5389 TargetData &TD = *IC.getTargetData();
5390 gep_type_iterator GTI = gep_type_begin(GEP);
5392 // Check to see if this gep only has a single variable index. If so, and if
5393 // any constant indices are a multiple of its scale, then we can compute this
5394 // in terms of the scale of the variable index. For example, if the GEP
5395 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5396 // because the expression will cross zero at the same point.
5397 unsigned i, e = GEP->getNumOperands();
5399 for (i = 1; i != e; ++i, ++GTI) {
5400 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5401 // Compute the aggregate offset of constant indices.
5402 if (CI->isZero()) continue;
5404 // Handle a struct index, which adds its field offset to the pointer.
5405 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5406 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5408 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5409 Offset += Size*CI->getSExtValue();
5412 // Found our variable index.
5417 // If there are no variable indices, we must have a constant offset, just
5418 // evaluate it the general way.
5419 if (i == e) return 0;
5421 Value *VariableIdx = GEP->getOperand(i);
5422 // Determine the scale factor of the variable element. For example, this is
5423 // 4 if the variable index is into an array of i32.
5424 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5426 // Verify that there are no other variable indices. If so, emit the hard way.
5427 for (++i, ++GTI; i != e; ++i, ++GTI) {
5428 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5431 // Compute the aggregate offset of constant indices.
5432 if (CI->isZero()) continue;
5434 // Handle a struct index, which adds its field offset to the pointer.
5435 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5436 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5438 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5439 Offset += Size*CI->getSExtValue();
5443 // Okay, we know we have a single variable index, which must be a
5444 // pointer/array/vector index. If there is no offset, life is simple, return
5446 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5448 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5449 // we don't need to bother extending: the extension won't affect where the
5450 // computation crosses zero.
5451 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5452 VariableIdx = new TruncInst(VariableIdx,
5453 TD.getIntPtrType(VariableIdx->getContext()),
5454 VariableIdx->getName(), &I);
5458 // Otherwise, there is an index. The computation we will do will be modulo
5459 // the pointer size, so get it.
5460 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5462 Offset &= PtrSizeMask;
5463 VariableScale &= PtrSizeMask;
5465 // To do this transformation, any constant index must be a multiple of the
5466 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5467 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5468 // multiple of the variable scale.
5469 int64_t NewOffs = Offset / (int64_t)VariableScale;
5470 if (Offset != NewOffs*(int64_t)VariableScale)
5473 // Okay, we can do this evaluation. Start by converting the index to intptr.
5474 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
5475 if (VariableIdx->getType() != IntPtrTy)
5476 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5478 VariableIdx->getName(), &I);
5479 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5480 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5484 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5485 /// else. At this point we know that the GEP is on the LHS of the comparison.
5486 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5487 ICmpInst::Predicate Cond,
5489 // Look through bitcasts.
5490 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5491 RHS = BCI->getOperand(0);
5493 Value *PtrBase = GEPLHS->getOperand(0);
5494 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5495 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5496 // This transformation (ignoring the base and scales) is valid because we
5497 // know pointers can't overflow since the gep is inbounds. See if we can
5498 // output an optimized form.
5499 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5501 // If not, synthesize the offset the hard way.
5503 Offset = EmitGEPOffset(GEPLHS, I, *this);
5504 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5505 Constant::getNullValue(Offset->getType()));
5506 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5507 // If the base pointers are different, but the indices are the same, just
5508 // compare the base pointer.
5509 if (PtrBase != GEPRHS->getOperand(0)) {
5510 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5511 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5512 GEPRHS->getOperand(0)->getType();
5514 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5515 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5516 IndicesTheSame = false;
5520 // If all indices are the same, just compare the base pointers.
5522 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5523 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5525 // Otherwise, the base pointers are different and the indices are
5526 // different, bail out.
5530 // If one of the GEPs has all zero indices, recurse.
5531 bool AllZeros = true;
5532 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5533 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5534 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5539 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5540 ICmpInst::getSwappedPredicate(Cond), I);
5542 // If the other GEP has all zero indices, recurse.
5544 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5545 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5546 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5551 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5553 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5554 // If the GEPs only differ by one index, compare it.
5555 unsigned NumDifferences = 0; // Keep track of # differences.
5556 unsigned DiffOperand = 0; // The operand that differs.
5557 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5558 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5559 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5560 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5561 // Irreconcilable differences.
5565 if (NumDifferences++) break;
5570 if (NumDifferences == 0) // SAME GEP?
5571 return ReplaceInstUsesWith(I, // No comparison is needed here.
5572 ConstantInt::get(Type::getInt1Ty(*Context),
5573 ICmpInst::isTrueWhenEqual(Cond)));
5575 else if (NumDifferences == 1) {
5576 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5577 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5578 // Make sure we do a signed comparison here.
5579 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5583 // Only lower this if the icmp is the only user of the GEP or if we expect
5584 // the result to fold to a constant!
5586 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5587 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5588 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5589 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5590 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5591 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5597 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5599 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5602 if (!isa<ConstantFP>(RHSC)) return 0;
5603 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5605 // Get the width of the mantissa. We don't want to hack on conversions that
5606 // might lose information from the integer, e.g. "i64 -> float"
5607 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5608 if (MantissaWidth == -1) return 0; // Unknown.
5610 // Check to see that the input is converted from an integer type that is small
5611 // enough that preserves all bits. TODO: check here for "known" sign bits.
5612 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5613 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5615 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5616 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5620 // If the conversion would lose info, don't hack on this.
5621 if ((int)InputSize > MantissaWidth)
5624 // Otherwise, we can potentially simplify the comparison. We know that it
5625 // will always come through as an integer value and we know the constant is
5626 // not a NAN (it would have been previously simplified).
5627 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5629 ICmpInst::Predicate Pred;
5630 switch (I.getPredicate()) {
5631 default: llvm_unreachable("Unexpected predicate!");
5632 case FCmpInst::FCMP_UEQ:
5633 case FCmpInst::FCMP_OEQ:
5634 Pred = ICmpInst::ICMP_EQ;
5636 case FCmpInst::FCMP_UGT:
5637 case FCmpInst::FCMP_OGT:
5638 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5640 case FCmpInst::FCMP_UGE:
5641 case FCmpInst::FCMP_OGE:
5642 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5644 case FCmpInst::FCMP_ULT:
5645 case FCmpInst::FCMP_OLT:
5646 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5648 case FCmpInst::FCMP_ULE:
5649 case FCmpInst::FCMP_OLE:
5650 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5652 case FCmpInst::FCMP_UNE:
5653 case FCmpInst::FCMP_ONE:
5654 Pred = ICmpInst::ICMP_NE;
5656 case FCmpInst::FCMP_ORD:
5657 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5658 case FCmpInst::FCMP_UNO:
5659 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5662 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5664 // Now we know that the APFloat is a normal number, zero or inf.
5666 // See if the FP constant is too large for the integer. For example,
5667 // comparing an i8 to 300.0.
5668 unsigned IntWidth = IntTy->getScalarSizeInBits();
5671 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5672 // and large values.
5673 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5674 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5675 APFloat::rmNearestTiesToEven);
5676 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5677 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5678 Pred == ICmpInst::ICMP_SLE)
5679 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5680 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5683 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5684 // +INF and large values.
5685 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5686 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5687 APFloat::rmNearestTiesToEven);
5688 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5689 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5690 Pred == ICmpInst::ICMP_ULE)
5691 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5692 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5697 // See if the RHS value is < SignedMin.
5698 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5699 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5700 APFloat::rmNearestTiesToEven);
5701 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5702 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5703 Pred == ICmpInst::ICMP_SGE)
5704 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5705 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5709 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5710 // [0, UMAX], but it may still be fractional. See if it is fractional by
5711 // casting the FP value to the integer value and back, checking for equality.
5712 // Don't do this for zero, because -0.0 is not fractional.
5713 Constant *RHSInt = LHSUnsigned
5714 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5715 : ConstantExpr::getFPToSI(RHSC, IntTy);
5716 if (!RHS.isZero()) {
5717 bool Equal = LHSUnsigned
5718 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5719 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5721 // If we had a comparison against a fractional value, we have to adjust
5722 // the compare predicate and sometimes the value. RHSC is rounded towards
5723 // zero at this point.
5725 default: llvm_unreachable("Unexpected integer comparison!");
5726 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5727 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5728 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5729 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5730 case ICmpInst::ICMP_ULE:
5731 // (float)int <= 4.4 --> int <= 4
5732 // (float)int <= -4.4 --> false
5733 if (RHS.isNegative())
5734 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5736 case ICmpInst::ICMP_SLE:
5737 // (float)int <= 4.4 --> int <= 4
5738 // (float)int <= -4.4 --> int < -4
5739 if (RHS.isNegative())
5740 Pred = ICmpInst::ICMP_SLT;
5742 case ICmpInst::ICMP_ULT:
5743 // (float)int < -4.4 --> false
5744 // (float)int < 4.4 --> int <= 4
5745 if (RHS.isNegative())
5746 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5747 Pred = ICmpInst::ICMP_ULE;
5749 case ICmpInst::ICMP_SLT:
5750 // (float)int < -4.4 --> int < -4
5751 // (float)int < 4.4 --> int <= 4
5752 if (!RHS.isNegative())
5753 Pred = ICmpInst::ICMP_SLE;
5755 case ICmpInst::ICMP_UGT:
5756 // (float)int > 4.4 --> int > 4
5757 // (float)int > -4.4 --> true
5758 if (RHS.isNegative())
5759 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5761 case ICmpInst::ICMP_SGT:
5762 // (float)int > 4.4 --> int > 4
5763 // (float)int > -4.4 --> int >= -4
5764 if (RHS.isNegative())
5765 Pred = ICmpInst::ICMP_SGE;
5767 case ICmpInst::ICMP_UGE:
5768 // (float)int >= -4.4 --> true
5769 // (float)int >= 4.4 --> int > 4
5770 if (!RHS.isNegative())
5771 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5772 Pred = ICmpInst::ICMP_UGT;
5774 case ICmpInst::ICMP_SGE:
5775 // (float)int >= -4.4 --> int >= -4
5776 // (float)int >= 4.4 --> int > 4
5777 if (!RHS.isNegative())
5778 Pred = ICmpInst::ICMP_SGT;
5784 // Lower this FP comparison into an appropriate integer version of the
5786 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5789 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5790 bool Changed = SimplifyCompare(I);
5791 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5793 // Fold trivial predicates.
5794 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5795 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5796 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5797 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5799 // Simplify 'fcmp pred X, X'
5801 switch (I.getPredicate()) {
5802 default: llvm_unreachable("Unknown predicate!");
5803 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5804 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5805 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5806 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5807 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5808 case FCmpInst::FCMP_OLT: // True if ordered and less than
5809 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5810 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5812 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5813 case FCmpInst::FCMP_ULT: // True if unordered or less than
5814 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5815 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5816 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5817 I.setPredicate(FCmpInst::FCMP_UNO);
5818 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5821 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5822 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5823 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5824 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5825 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5826 I.setPredicate(FCmpInst::FCMP_ORD);
5827 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5832 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5833 return ReplaceInstUsesWith(I, UndefValue::get(Type::getInt1Ty(*Context)));
5835 // Handle fcmp with constant RHS
5836 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5837 // If the constant is a nan, see if we can fold the comparison based on it.
5838 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5839 if (CFP->getValueAPF().isNaN()) {
5840 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5841 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5842 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5843 "Comparison must be either ordered or unordered!");
5844 // True if unordered.
5845 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5849 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5850 switch (LHSI->getOpcode()) {
5851 case Instruction::PHI:
5852 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5853 // block. If in the same block, we're encouraging jump threading. If
5854 // not, we are just pessimizing the code by making an i1 phi.
5855 if (LHSI->getParent() == I.getParent())
5856 if (Instruction *NV = FoldOpIntoPhi(I))
5859 case Instruction::SIToFP:
5860 case Instruction::UIToFP:
5861 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5864 case Instruction::Select:
5865 // If either operand of the select is a constant, we can fold the
5866 // comparison into the select arms, which will cause one to be
5867 // constant folded and the select turned into a bitwise or.
5868 Value *Op1 = 0, *Op2 = 0;
5869 if (LHSI->hasOneUse()) {
5870 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5871 // Fold the known value into the constant operand.
5872 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5873 // Insert a new FCmp of the other select operand.
5874 Op2 = Builder->CreateFCmp(I.getPredicate(),
5875 LHSI->getOperand(2), RHSC, I.getName());
5876 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5877 // Fold the known value into the constant operand.
5878 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5879 // Insert a new FCmp of the other select operand.
5880 Op1 = Builder->CreateFCmp(I.getPredicate(), LHSI->getOperand(1),
5886 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5891 return Changed ? &I : 0;
5894 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5895 bool Changed = SimplifyCompare(I);
5896 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5897 const Type *Ty = Op0->getType();
5901 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
5902 I.isTrueWhenEqual()));
5904 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5905 return ReplaceInstUsesWith(I, UndefValue::get(Type::getInt1Ty(*Context)));
5907 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5908 // addresses never equal each other! We already know that Op0 != Op1.
5909 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5910 isa<ConstantPointerNull>(Op0)) &&
5911 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5912 isa<ConstantPointerNull>(Op1)))
5913 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
5914 !I.isTrueWhenEqual()));
5916 // icmp's with boolean values can always be turned into bitwise operations
5917 if (Ty == Type::getInt1Ty(*Context)) {
5918 switch (I.getPredicate()) {
5919 default: llvm_unreachable("Invalid icmp instruction!");
5920 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5921 Value *Xor = Builder->CreateXor(Op0, Op1, I.getName()+"tmp");
5922 return BinaryOperator::CreateNot(Xor);
5924 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5925 return BinaryOperator::CreateXor(Op0, Op1);
5927 case ICmpInst::ICMP_UGT:
5928 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
5930 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
5931 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
5932 return BinaryOperator::CreateAnd(Not, Op1);
5934 case ICmpInst::ICMP_SGT:
5935 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
5937 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
5938 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
5939 return BinaryOperator::CreateAnd(Not, Op0);
5941 case ICmpInst::ICMP_UGE:
5942 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
5944 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
5945 Value *Not = Builder->CreateNot(Op0, I.getName()+"tmp");
5946 return BinaryOperator::CreateOr(Not, Op1);
5948 case ICmpInst::ICMP_SGE:
5949 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
5951 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
5952 Value *Not = Builder->CreateNot(Op1, I.getName()+"tmp");
5953 return BinaryOperator::CreateOr(Not, Op0);
5958 unsigned BitWidth = 0;
5960 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
5961 else if (Ty->isIntOrIntVector())
5962 BitWidth = Ty->getScalarSizeInBits();
5964 bool isSignBit = false;
5966 // See if we are doing a comparison with a constant.
5967 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5968 Value *A = 0, *B = 0;
5970 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5971 if (I.isEquality() && CI->isNullValue() &&
5972 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5973 // (icmp cond A B) if cond is equality
5974 return new ICmpInst(I.getPredicate(), A, B);
5977 // If we have an icmp le or icmp ge instruction, turn it into the
5978 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
5979 // them being folded in the code below.
5980 switch (I.getPredicate()) {
5982 case ICmpInst::ICMP_ULE:
5983 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5984 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5985 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
5987 case ICmpInst::ICMP_SLE:
5988 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5989 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5990 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5992 case ICmpInst::ICMP_UGE:
5993 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5994 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5995 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
5997 case ICmpInst::ICMP_SGE:
5998 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5999 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6000 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6004 // If this comparison is a normal comparison, it demands all
6005 // bits, if it is a sign bit comparison, it only demands the sign bit.
6007 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6010 // See if we can fold the comparison based on range information we can get
6011 // by checking whether bits are known to be zero or one in the input.
6012 if (BitWidth != 0) {
6013 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6014 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6016 if (SimplifyDemandedBits(I.getOperandUse(0),
6017 isSignBit ? APInt::getSignBit(BitWidth)
6018 : APInt::getAllOnesValue(BitWidth),
6019 Op0KnownZero, Op0KnownOne, 0))
6021 if (SimplifyDemandedBits(I.getOperandUse(1),
6022 APInt::getAllOnesValue(BitWidth),
6023 Op1KnownZero, Op1KnownOne, 0))
6026 // Given the known and unknown bits, compute a range that the LHS could be
6027 // in. Compute the Min, Max and RHS values based on the known bits. For the
6028 // EQ and NE we use unsigned values.
6029 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6030 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6031 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6032 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6034 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6037 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6039 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6043 // If Min and Max are known to be the same, then SimplifyDemandedBits
6044 // figured out that the LHS is a constant. Just constant fold this now so
6045 // that code below can assume that Min != Max.
6046 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6047 return new ICmpInst(I.getPredicate(),
6048 ConstantInt::get(*Context, Op0Min), Op1);
6049 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6050 return new ICmpInst(I.getPredicate(), Op0,
6051 ConstantInt::get(*Context, Op1Min));
6053 // Based on the range information we know about the LHS, see if we can
6054 // simplify this comparison. For example, (x&4) < 8 is always true.
6055 switch (I.getPredicate()) {
6056 default: llvm_unreachable("Unknown icmp opcode!");
6057 case ICmpInst::ICMP_EQ:
6058 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6059 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6061 case ICmpInst::ICMP_NE:
6062 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6063 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6065 case ICmpInst::ICMP_ULT:
6066 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6067 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6068 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6069 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6070 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6071 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6072 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6073 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6074 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6077 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6078 if (CI->isMinValue(true))
6079 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6080 Constant::getAllOnesValue(Op0->getType()));
6083 case ICmpInst::ICMP_UGT:
6084 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6085 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6086 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6087 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6089 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6090 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6091 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6092 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6093 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6096 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6097 if (CI->isMaxValue(true))
6098 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6099 Constant::getNullValue(Op0->getType()));
6102 case ICmpInst::ICMP_SLT:
6103 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6104 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6105 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6106 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6107 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6108 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6109 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6110 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6111 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6115 case ICmpInst::ICMP_SGT:
6116 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6117 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6118 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6119 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6121 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6122 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6123 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6124 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6125 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6129 case ICmpInst::ICMP_SGE:
6130 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6131 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6132 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6133 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6134 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6136 case ICmpInst::ICMP_SLE:
6137 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6138 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6139 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6140 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6141 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6143 case ICmpInst::ICMP_UGE:
6144 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6145 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6146 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6147 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6148 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6150 case ICmpInst::ICMP_ULE:
6151 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6152 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6153 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6154 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6155 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6159 // Turn a signed comparison into an unsigned one if both operands
6160 // are known to have the same sign.
6161 if (I.isSignedPredicate() &&
6162 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6163 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6164 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
6167 // Test if the ICmpInst instruction is used exclusively by a select as
6168 // part of a minimum or maximum operation. If so, refrain from doing
6169 // any other folding. This helps out other analyses which understand
6170 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6171 // and CodeGen. And in this case, at least one of the comparison
6172 // operands has at least one user besides the compare (the select),
6173 // which would often largely negate the benefit of folding anyway.
6175 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6176 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6177 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6180 // See if we are doing a comparison between a constant and an instruction that
6181 // can be folded into the comparison.
6182 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6183 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6184 // instruction, see if that instruction also has constants so that the
6185 // instruction can be folded into the icmp
6186 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6187 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6191 // Handle icmp with constant (but not simple integer constant) RHS
6192 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6193 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6194 switch (LHSI->getOpcode()) {
6195 case Instruction::GetElementPtr:
6196 if (RHSC->isNullValue()) {
6197 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6198 bool isAllZeros = true;
6199 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6200 if (!isa<Constant>(LHSI->getOperand(i)) ||
6201 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6206 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6207 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6211 case Instruction::PHI:
6212 // Only fold icmp into the PHI if the phi and fcmp are in the same
6213 // block. If in the same block, we're encouraging jump threading. If
6214 // not, we are just pessimizing the code by making an i1 phi.
6215 if (LHSI->getParent() == I.getParent())
6216 if (Instruction *NV = FoldOpIntoPhi(I))
6219 case Instruction::Select: {
6220 // If either operand of the select is a constant, we can fold the
6221 // comparison into the select arms, which will cause one to be
6222 // constant folded and the select turned into a bitwise or.
6223 Value *Op1 = 0, *Op2 = 0;
6224 if (LHSI->hasOneUse()) {
6225 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6226 // Fold the known value into the constant operand.
6227 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6228 // Insert a new ICmp of the other select operand.
6229 Op2 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(2),
6231 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6232 // Fold the known value into the constant operand.
6233 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6234 // Insert a new ICmp of the other select operand.
6235 Op1 = Builder->CreateICmp(I.getPredicate(), LHSI->getOperand(1),
6241 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6244 case Instruction::Malloc:
6245 // If we have (malloc != null), and if the malloc has a single use, we
6246 // can assume it is successful and remove the malloc.
6247 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6249 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
6250 !I.isTrueWhenEqual()));
6256 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6257 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6258 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6260 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6261 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6262 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6265 // Test to see if the operands of the icmp are casted versions of other
6266 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6268 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6269 if (isa<PointerType>(Op0->getType()) &&
6270 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6271 // We keep moving the cast from the left operand over to the right
6272 // operand, where it can often be eliminated completely.
6273 Op0 = CI->getOperand(0);
6275 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6276 // so eliminate it as well.
6277 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6278 Op1 = CI2->getOperand(0);
6280 // If Op1 is a constant, we can fold the cast into the constant.
6281 if (Op0->getType() != Op1->getType()) {
6282 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6283 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6285 // Otherwise, cast the RHS right before the icmp
6286 Op1 = Builder->CreateBitCast(Op1, Op0->getType());
6289 return new ICmpInst(I.getPredicate(), Op0, Op1);
6293 if (isa<CastInst>(Op0)) {
6294 // Handle the special case of: icmp (cast bool to X), <cst>
6295 // This comes up when you have code like
6298 // For generality, we handle any zero-extension of any operand comparison
6299 // with a constant or another cast from the same type.
6300 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6301 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6305 // See if it's the same type of instruction on the left and right.
6306 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6307 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6308 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6309 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6310 switch (Op0I->getOpcode()) {
6312 case Instruction::Add:
6313 case Instruction::Sub:
6314 case Instruction::Xor:
6315 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6316 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6317 Op1I->getOperand(0));
6318 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6319 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6320 if (CI->getValue().isSignBit()) {
6321 ICmpInst::Predicate Pred = I.isSignedPredicate()
6322 ? I.getUnsignedPredicate()
6323 : I.getSignedPredicate();
6324 return new ICmpInst(Pred, Op0I->getOperand(0),
6325 Op1I->getOperand(0));
6328 if (CI->getValue().isMaxSignedValue()) {
6329 ICmpInst::Predicate Pred = I.isSignedPredicate()
6330 ? I.getUnsignedPredicate()
6331 : I.getSignedPredicate();
6332 Pred = I.getSwappedPredicate(Pred);
6333 return new ICmpInst(Pred, Op0I->getOperand(0),
6334 Op1I->getOperand(0));
6338 case Instruction::Mul:
6339 if (!I.isEquality())
6342 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6343 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6344 // Mask = -1 >> count-trailing-zeros(Cst).
6345 if (!CI->isZero() && !CI->isOne()) {
6346 const APInt &AP = CI->getValue();
6347 ConstantInt *Mask = ConstantInt::get(*Context,
6348 APInt::getLowBitsSet(AP.getBitWidth(),
6350 AP.countTrailingZeros()));
6351 Value *And1 = Builder->CreateAnd(Op0I->getOperand(0), Mask);
6352 Value *And2 = Builder->CreateAnd(Op1I->getOperand(0), Mask);
6353 return new ICmpInst(I.getPredicate(), And1, And2);
6362 // ~x < ~y --> y < x
6364 if (match(Op0, m_Not(m_Value(A))) &&
6365 match(Op1, m_Not(m_Value(B))))
6366 return new ICmpInst(I.getPredicate(), B, A);
6369 if (I.isEquality()) {
6370 Value *A, *B, *C, *D;
6372 // -x == -y --> x == y
6373 if (match(Op0, m_Neg(m_Value(A))) &&
6374 match(Op1, m_Neg(m_Value(B))))
6375 return new ICmpInst(I.getPredicate(), A, B);
6377 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6378 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6379 Value *OtherVal = A == Op1 ? B : A;
6380 return new ICmpInst(I.getPredicate(), OtherVal,
6381 Constant::getNullValue(A->getType()));
6384 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6385 // A^c1 == C^c2 --> A == C^(c1^c2)
6386 ConstantInt *C1, *C2;
6387 if (match(B, m_ConstantInt(C1)) &&
6388 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6390 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6391 Value *Xor = Builder->CreateXor(C, NC, "tmp");
6392 return new ICmpInst(I.getPredicate(), A, Xor);
6395 // A^B == A^D -> B == D
6396 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6397 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6398 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6399 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6403 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6404 (A == Op0 || B == Op0)) {
6405 // A == (A^B) -> B == 0
6406 Value *OtherVal = A == Op0 ? B : A;
6407 return new ICmpInst(I.getPredicate(), OtherVal,
6408 Constant::getNullValue(A->getType()));
6411 // (A-B) == A -> B == 0
6412 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6413 return new ICmpInst(I.getPredicate(), B,
6414 Constant::getNullValue(B->getType()));
6416 // A == (A-B) -> B == 0
6417 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6418 return new ICmpInst(I.getPredicate(), B,
6419 Constant::getNullValue(B->getType()));
6421 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6422 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6423 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6424 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6425 Value *X = 0, *Y = 0, *Z = 0;
6428 X = B; Y = D; Z = A;
6429 } else if (A == D) {
6430 X = B; Y = C; Z = A;
6431 } else if (B == C) {
6432 X = A; Y = D; Z = B;
6433 } else if (B == D) {
6434 X = A; Y = C; Z = B;
6437 if (X) { // Build (X^Y) & Z
6438 Op1 = Builder->CreateXor(X, Y, "tmp");
6439 Op1 = Builder->CreateAnd(Op1, Z, "tmp");
6440 I.setOperand(0, Op1);
6441 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6446 return Changed ? &I : 0;
6450 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6451 /// and CmpRHS are both known to be integer constants.
6452 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6453 ConstantInt *DivRHS) {
6454 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6455 const APInt &CmpRHSV = CmpRHS->getValue();
6457 // FIXME: If the operand types don't match the type of the divide
6458 // then don't attempt this transform. The code below doesn't have the
6459 // logic to deal with a signed divide and an unsigned compare (and
6460 // vice versa). This is because (x /s C1) <s C2 produces different
6461 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6462 // (x /u C1) <u C2. Simply casting the operands and result won't
6463 // work. :( The if statement below tests that condition and bails
6465 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6466 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6468 if (DivRHS->isZero())
6469 return 0; // The ProdOV computation fails on divide by zero.
6470 if (DivIsSigned && DivRHS->isAllOnesValue())
6471 return 0; // The overflow computation also screws up here
6472 if (DivRHS->isOne())
6473 return 0; // Not worth bothering, and eliminates some funny cases
6476 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6477 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6478 // C2 (CI). By solving for X we can turn this into a range check
6479 // instead of computing a divide.
6480 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6482 // Determine if the product overflows by seeing if the product is
6483 // not equal to the divide. Make sure we do the same kind of divide
6484 // as in the LHS instruction that we're folding.
6485 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6486 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6488 // Get the ICmp opcode
6489 ICmpInst::Predicate Pred = ICI.getPredicate();
6491 // Figure out the interval that is being checked. For example, a comparison
6492 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6493 // Compute this interval based on the constants involved and the signedness of
6494 // the compare/divide. This computes a half-open interval, keeping track of
6495 // whether either value in the interval overflows. After analysis each
6496 // overflow variable is set to 0 if it's corresponding bound variable is valid
6497 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6498 int LoOverflow = 0, HiOverflow = 0;
6499 Constant *LoBound = 0, *HiBound = 0;
6501 if (!DivIsSigned) { // udiv
6502 // e.g. X/5 op 3 --> [15, 20)
6504 HiOverflow = LoOverflow = ProdOV;
6506 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6507 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6508 if (CmpRHSV == 0) { // (X / pos) op 0
6509 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6510 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6512 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6513 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6514 HiOverflow = LoOverflow = ProdOV;
6516 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6517 } else { // (X / pos) op neg
6518 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6519 HiBound = AddOne(Prod);
6520 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6522 ConstantInt* DivNeg =
6523 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6524 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6528 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6529 if (CmpRHSV == 0) { // (X / neg) op 0
6530 // e.g. X/-5 op 0 --> [-4, 5)
6531 LoBound = AddOne(DivRHS);
6532 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6533 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6534 HiOverflow = 1; // [INTMIN+1, overflow)
6535 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6537 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6538 // e.g. X/-5 op 3 --> [-19, -14)
6539 HiBound = AddOne(Prod);
6540 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6542 LoOverflow = AddWithOverflow(LoBound, HiBound,
6543 DivRHS, Context, true) ? -1 : 0;
6544 } else { // (X / neg) op neg
6545 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6546 LoOverflow = HiOverflow = ProdOV;
6548 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6551 // Dividing by a negative swaps the condition. LT <-> GT
6552 Pred = ICmpInst::getSwappedPredicate(Pred);
6555 Value *X = DivI->getOperand(0);
6557 default: llvm_unreachable("Unhandled icmp opcode!");
6558 case ICmpInst::ICMP_EQ:
6559 if (LoOverflow && HiOverflow)
6560 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6561 else if (HiOverflow)
6562 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6563 ICmpInst::ICMP_UGE, X, LoBound);
6564 else if (LoOverflow)
6565 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6566 ICmpInst::ICMP_ULT, X, HiBound);
6568 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6569 case ICmpInst::ICMP_NE:
6570 if (LoOverflow && HiOverflow)
6571 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6572 else if (HiOverflow)
6573 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6574 ICmpInst::ICMP_ULT, X, LoBound);
6575 else if (LoOverflow)
6576 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6577 ICmpInst::ICMP_UGE, X, HiBound);
6579 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6580 case ICmpInst::ICMP_ULT:
6581 case ICmpInst::ICMP_SLT:
6582 if (LoOverflow == +1) // Low bound is greater than input range.
6583 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6584 if (LoOverflow == -1) // Low bound is less than input range.
6585 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6586 return new ICmpInst(Pred, X, LoBound);
6587 case ICmpInst::ICMP_UGT:
6588 case ICmpInst::ICMP_SGT:
6589 if (HiOverflow == +1) // High bound greater than input range.
6590 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6591 else if (HiOverflow == -1) // High bound less than input range.
6592 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6593 if (Pred == ICmpInst::ICMP_UGT)
6594 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6596 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6601 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6603 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6606 const APInt &RHSV = RHS->getValue();
6608 switch (LHSI->getOpcode()) {
6609 case Instruction::Trunc:
6610 if (ICI.isEquality() && LHSI->hasOneUse()) {
6611 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6612 // of the high bits truncated out of x are known.
6613 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6614 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6615 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6616 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6617 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6619 // If all the high bits are known, we can do this xform.
6620 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6621 // Pull in the high bits from known-ones set.
6622 APInt NewRHS(RHS->getValue());
6623 NewRHS.zext(SrcBits);
6625 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6626 ConstantInt::get(*Context, NewRHS));
6631 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6632 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6633 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6635 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6636 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6637 Value *CompareVal = LHSI->getOperand(0);
6639 // If the sign bit of the XorCST is not set, there is no change to
6640 // the operation, just stop using the Xor.
6641 if (!XorCST->getValue().isNegative()) {
6642 ICI.setOperand(0, CompareVal);
6647 // Was the old condition true if the operand is positive?
6648 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6650 // If so, the new one isn't.
6651 isTrueIfPositive ^= true;
6653 if (isTrueIfPositive)
6654 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
6657 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
6661 if (LHSI->hasOneUse()) {
6662 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6663 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6664 const APInt &SignBit = XorCST->getValue();
6665 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6666 ? ICI.getUnsignedPredicate()
6667 : ICI.getSignedPredicate();
6668 return new ICmpInst(Pred, LHSI->getOperand(0),
6669 ConstantInt::get(*Context, RHSV ^ SignBit));
6672 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6673 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6674 const APInt &NotSignBit = XorCST->getValue();
6675 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6676 ? ICI.getUnsignedPredicate()
6677 : ICI.getSignedPredicate();
6678 Pred = ICI.getSwappedPredicate(Pred);
6679 return new ICmpInst(Pred, LHSI->getOperand(0),
6680 ConstantInt::get(*Context, RHSV ^ NotSignBit));
6685 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6686 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6687 LHSI->getOperand(0)->hasOneUse()) {
6688 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6690 // If the LHS is an AND of a truncating cast, we can widen the
6691 // and/compare to be the input width without changing the value
6692 // produced, eliminating a cast.
6693 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6694 // We can do this transformation if either the AND constant does not
6695 // have its sign bit set or if it is an equality comparison.
6696 // Extending a relational comparison when we're checking the sign
6697 // bit would not work.
6698 if (Cast->hasOneUse() &&
6699 (ICI.isEquality() ||
6700 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6702 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6703 APInt NewCST = AndCST->getValue();
6704 NewCST.zext(BitWidth);
6706 NewCI.zext(BitWidth);
6708 Builder->CreateAnd(Cast->getOperand(0),
6709 ConstantInt::get(*Context, NewCST), LHSI->getName());
6710 return new ICmpInst(ICI.getPredicate(), NewAnd,
6711 ConstantInt::get(*Context, NewCI));
6715 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6716 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6717 // happens a LOT in code produced by the C front-end, for bitfield
6719 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6720 if (Shift && !Shift->isShift())
6724 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6725 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6726 const Type *AndTy = AndCST->getType(); // Type of the and.
6728 // We can fold this as long as we can't shift unknown bits
6729 // into the mask. This can only happen with signed shift
6730 // rights, as they sign-extend.
6732 bool CanFold = Shift->isLogicalShift();
6734 // To test for the bad case of the signed shr, see if any
6735 // of the bits shifted in could be tested after the mask.
6736 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6737 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6739 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6740 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6741 AndCST->getValue()) == 0)
6747 if (Shift->getOpcode() == Instruction::Shl)
6748 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6750 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6752 // Check to see if we are shifting out any of the bits being
6754 if (ConstantExpr::get(Shift->getOpcode(),
6755 NewCst, ShAmt) != RHS) {
6756 // If we shifted bits out, the fold is not going to work out.
6757 // As a special case, check to see if this means that the
6758 // result is always true or false now.
6759 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6760 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6761 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6762 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6764 ICI.setOperand(1, NewCst);
6765 Constant *NewAndCST;
6766 if (Shift->getOpcode() == Instruction::Shl)
6767 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6769 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6770 LHSI->setOperand(1, NewAndCST);
6771 LHSI->setOperand(0, Shift->getOperand(0));
6772 Worklist.Add(Shift); // Shift is dead.
6778 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6779 // preferable because it allows the C<<Y expression to be hoisted out
6780 // of a loop if Y is invariant and X is not.
6781 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6782 ICI.isEquality() && !Shift->isArithmeticShift() &&
6783 !isa<Constant>(Shift->getOperand(0))) {
6786 if (Shift->getOpcode() == Instruction::LShr) {
6787 NS = Builder->CreateShl(AndCST, Shift->getOperand(1), "tmp");
6789 // Insert a logical shift.
6790 NS = Builder->CreateLShr(AndCST, Shift->getOperand(1), "tmp");
6793 // Compute X & (C << Y).
6795 Builder->CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6797 ICI.setOperand(0, NewAnd);
6803 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6804 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6807 uint32_t TypeBits = RHSV.getBitWidth();
6809 // Check that the shift amount is in range. If not, don't perform
6810 // undefined shifts. When the shift is visited it will be
6812 if (ShAmt->uge(TypeBits))
6815 if (ICI.isEquality()) {
6816 // If we are comparing against bits always shifted out, the
6817 // comparison cannot succeed.
6819 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
6821 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6822 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6823 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6824 return ReplaceInstUsesWith(ICI, Cst);
6827 if (LHSI->hasOneUse()) {
6828 // Otherwise strength reduce the shift into an and.
6829 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6831 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
6832 TypeBits-ShAmtVal));
6835 Builder->CreateAnd(LHSI->getOperand(0),Mask, LHSI->getName()+".mask");
6836 return new ICmpInst(ICI.getPredicate(), And,
6837 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
6841 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6842 bool TrueIfSigned = false;
6843 if (LHSI->hasOneUse() &&
6844 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6845 // (X << 31) <s 0 --> (X&1) != 0
6846 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
6847 (TypeBits-ShAmt->getZExtValue()-1));
6849 Builder->CreateAnd(LHSI->getOperand(0), Mask, LHSI->getName()+".mask");
6850 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6851 And, Constant::getNullValue(And->getType()));
6856 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6857 case Instruction::AShr: {
6858 // Only handle equality comparisons of shift-by-constant.
6859 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6860 if (!ShAmt || !ICI.isEquality()) break;
6862 // Check that the shift amount is in range. If not, don't perform
6863 // undefined shifts. When the shift is visited it will be
6865 uint32_t TypeBits = RHSV.getBitWidth();
6866 if (ShAmt->uge(TypeBits))
6869 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6871 // If we are comparing against bits always shifted out, the
6872 // comparison cannot succeed.
6873 APInt Comp = RHSV << ShAmtVal;
6874 if (LHSI->getOpcode() == Instruction::LShr)
6875 Comp = Comp.lshr(ShAmtVal);
6877 Comp = Comp.ashr(ShAmtVal);
6879 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6880 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6881 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6882 return ReplaceInstUsesWith(ICI, Cst);
6885 // Otherwise, check to see if the bits shifted out are known to be zero.
6886 // If so, we can compare against the unshifted value:
6887 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6888 if (LHSI->hasOneUse() &&
6889 MaskedValueIsZero(LHSI->getOperand(0),
6890 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6891 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6892 ConstantExpr::getShl(RHS, ShAmt));
6895 if (LHSI->hasOneUse()) {
6896 // Otherwise strength reduce the shift into an and.
6897 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6898 Constant *Mask = ConstantInt::get(*Context, Val);
6900 Value *And = Builder->CreateAnd(LHSI->getOperand(0),
6901 Mask, LHSI->getName()+".mask");
6902 return new ICmpInst(ICI.getPredicate(), And,
6903 ConstantExpr::getShl(RHS, ShAmt));
6908 case Instruction::SDiv:
6909 case Instruction::UDiv:
6910 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6911 // Fold this div into the comparison, producing a range check.
6912 // Determine, based on the divide type, what the range is being
6913 // checked. If there is an overflow on the low or high side, remember
6914 // it, otherwise compute the range [low, hi) bounding the new value.
6915 // See: InsertRangeTest above for the kinds of replacements possible.
6916 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6917 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6922 case Instruction::Add:
6923 // Fold: icmp pred (add, X, C1), C2
6925 if (!ICI.isEquality()) {
6926 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6928 const APInt &LHSV = LHSC->getValue();
6930 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6933 if (ICI.isSignedPredicate()) {
6934 if (CR.getLower().isSignBit()) {
6935 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6936 ConstantInt::get(*Context, CR.getUpper()));
6937 } else if (CR.getUpper().isSignBit()) {
6938 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6939 ConstantInt::get(*Context, CR.getLower()));
6942 if (CR.getLower().isMinValue()) {
6943 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6944 ConstantInt::get(*Context, CR.getUpper()));
6945 } else if (CR.getUpper().isMinValue()) {
6946 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6947 ConstantInt::get(*Context, CR.getLower()));
6954 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6955 if (ICI.isEquality()) {
6956 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6958 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6959 // the second operand is a constant, simplify a bit.
6960 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6961 switch (BO->getOpcode()) {
6962 case Instruction::SRem:
6963 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6964 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6965 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6966 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6968 Builder->CreateURem(BO->getOperand(0), BO->getOperand(1),
6970 return new ICmpInst(ICI.getPredicate(), NewRem,
6971 Constant::getNullValue(BO->getType()));
6975 case Instruction::Add:
6976 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6977 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6978 if (BO->hasOneUse())
6979 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6980 ConstantExpr::getSub(RHS, BOp1C));
6981 } else if (RHSV == 0) {
6982 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6983 // efficiently invertible, or if the add has just this one use.
6984 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6986 if (Value *NegVal = dyn_castNegVal(BOp1))
6987 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6988 else if (Value *NegVal = dyn_castNegVal(BOp0))
6989 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6990 else if (BO->hasOneUse()) {
6991 Value *Neg = Builder->CreateNeg(BOp1);
6993 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6997 case Instruction::Xor:
6998 // For the xor case, we can xor two constants together, eliminating
6999 // the explicit xor.
7000 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7001 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7002 ConstantExpr::getXor(RHS, BOC));
7005 case Instruction::Sub:
7006 // Replace (([sub|xor] A, B) != 0) with (A != B)
7008 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7012 case Instruction::Or:
7013 // If bits are being or'd in that are not present in the constant we
7014 // are comparing against, then the comparison could never succeed!
7015 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7016 Constant *NotCI = ConstantExpr::getNot(RHS);
7017 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7018 return ReplaceInstUsesWith(ICI,
7019 ConstantInt::get(Type::getInt1Ty(*Context),
7024 case Instruction::And:
7025 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7026 // If bits are being compared against that are and'd out, then the
7027 // comparison can never succeed!
7028 if ((RHSV & ~BOC->getValue()) != 0)
7029 return ReplaceInstUsesWith(ICI,
7030 ConstantInt::get(Type::getInt1Ty(*Context),
7033 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7034 if (RHS == BOC && RHSV.isPowerOf2())
7035 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
7036 ICmpInst::ICMP_NE, LHSI,
7037 Constant::getNullValue(RHS->getType()));
7039 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7040 if (BOC->getValue().isSignBit()) {
7041 Value *X = BO->getOperand(0);
7042 Constant *Zero = Constant::getNullValue(X->getType());
7043 ICmpInst::Predicate pred = isICMP_NE ?
7044 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7045 return new ICmpInst(pred, X, Zero);
7048 // ((X & ~7) == 0) --> X < 8
7049 if (RHSV == 0 && isHighOnes(BOC)) {
7050 Value *X = BO->getOperand(0);
7051 Constant *NegX = ConstantExpr::getNeg(BOC);
7052 ICmpInst::Predicate pred = isICMP_NE ?
7053 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7054 return new ICmpInst(pred, X, NegX);
7059 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7060 // Handle icmp {eq|ne} <intrinsic>, intcst.
7061 if (II->getIntrinsicID() == Intrinsic::bswap) {
7063 ICI.setOperand(0, II->getOperand(1));
7064 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7072 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7073 /// We only handle extending casts so far.
7075 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7076 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7077 Value *LHSCIOp = LHSCI->getOperand(0);
7078 const Type *SrcTy = LHSCIOp->getType();
7079 const Type *DestTy = LHSCI->getType();
7082 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7083 // integer type is the same size as the pointer type.
7084 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7085 TD->getPointerSizeInBits() ==
7086 cast<IntegerType>(DestTy)->getBitWidth()) {
7088 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7089 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7090 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7091 RHSOp = RHSC->getOperand(0);
7092 // If the pointer types don't match, insert a bitcast.
7093 if (LHSCIOp->getType() != RHSOp->getType())
7094 RHSOp = Builder->CreateBitCast(RHSOp, LHSCIOp->getType());
7098 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
7101 // The code below only handles extension cast instructions, so far.
7103 if (LHSCI->getOpcode() != Instruction::ZExt &&
7104 LHSCI->getOpcode() != Instruction::SExt)
7107 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7108 bool isSignedCmp = ICI.isSignedPredicate();
7110 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7111 // Not an extension from the same type?
7112 RHSCIOp = CI->getOperand(0);
7113 if (RHSCIOp->getType() != LHSCIOp->getType())
7116 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7117 // and the other is a zext), then we can't handle this.
7118 if (CI->getOpcode() != LHSCI->getOpcode())
7121 // Deal with equality cases early.
7122 if (ICI.isEquality())
7123 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7125 // A signed comparison of sign extended values simplifies into a
7126 // signed comparison.
7127 if (isSignedCmp && isSignedExt)
7128 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7130 // The other three cases all fold into an unsigned comparison.
7131 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7134 // If we aren't dealing with a constant on the RHS, exit early
7135 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7139 // Compute the constant that would happen if we truncated to SrcTy then
7140 // reextended to DestTy.
7141 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7142 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7145 // If the re-extended constant didn't change...
7147 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7148 // For example, we might have:
7149 // %A = sext i16 %X to i32
7150 // %B = icmp ugt i32 %A, 1330
7151 // It is incorrect to transform this into
7152 // %B = icmp ugt i16 %X, 1330
7153 // because %A may have negative value.
7155 // However, we allow this when the compare is EQ/NE, because they are
7157 if (isSignedExt == isSignedCmp || ICI.isEquality())
7158 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7162 // The re-extended constant changed so the constant cannot be represented
7163 // in the shorter type. Consequently, we cannot emit a simple comparison.
7165 // First, handle some easy cases. We know the result cannot be equal at this
7166 // point so handle the ICI.isEquality() cases
7167 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7168 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7169 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7170 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7172 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7173 // should have been folded away previously and not enter in here.
7176 // We're performing a signed comparison.
7177 if (cast<ConstantInt>(CI)->getValue().isNegative())
7178 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7180 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7182 // We're performing an unsigned comparison.
7184 // We're performing an unsigned comp with a sign extended value.
7185 // This is true if the input is >= 0. [aka >s -1]
7186 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7187 Result = Builder->CreateICmpSGT(LHSCIOp, NegOne, ICI.getName());
7189 // Unsigned extend & unsigned compare -> always true.
7190 Result = ConstantInt::getTrue(*Context);
7194 // Finally, return the value computed.
7195 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7196 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7197 return ReplaceInstUsesWith(ICI, Result);
7199 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7200 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7201 "ICmp should be folded!");
7202 if (Constant *CI = dyn_cast<Constant>(Result))
7203 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7204 return BinaryOperator::CreateNot(Result);
7207 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7208 return commonShiftTransforms(I);
7211 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7212 return commonShiftTransforms(I);
7215 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7216 if (Instruction *R = commonShiftTransforms(I))
7219 Value *Op0 = I.getOperand(0);
7221 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7222 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7223 if (CSI->isAllOnesValue())
7224 return ReplaceInstUsesWith(I, CSI);
7226 // See if we can turn a signed shr into an unsigned shr.
7227 if (MaskedValueIsZero(Op0,
7228 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7229 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7231 // Arithmetic shifting an all-sign-bit value is a no-op.
7232 unsigned NumSignBits = ComputeNumSignBits(Op0);
7233 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7234 return ReplaceInstUsesWith(I, Op0);
7239 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7240 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7241 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7243 // shl X, 0 == X and shr X, 0 == X
7244 // shl 0, X == 0 and shr 0, X == 0
7245 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7246 Op0 == Constant::getNullValue(Op0->getType()))
7247 return ReplaceInstUsesWith(I, Op0);
7249 if (isa<UndefValue>(Op0)) {
7250 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7251 return ReplaceInstUsesWith(I, Op0);
7252 else // undef << X -> 0, undef >>u X -> 0
7253 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7255 if (isa<UndefValue>(Op1)) {
7256 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7257 return ReplaceInstUsesWith(I, Op0);
7258 else // X << undef, X >>u undef -> 0
7259 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7262 // See if we can fold away this shift.
7263 if (SimplifyDemandedInstructionBits(I))
7266 // Try to fold constant and into select arguments.
7267 if (isa<Constant>(Op0))
7268 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7269 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7272 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7273 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7278 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7279 BinaryOperator &I) {
7280 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7282 // See if we can simplify any instructions used by the instruction whose sole
7283 // purpose is to compute bits we don't care about.
7284 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7286 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7289 if (Op1->uge(TypeBits)) {
7290 if (I.getOpcode() != Instruction::AShr)
7291 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7293 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7298 // ((X*C1) << C2) == (X * (C1 << C2))
7299 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7300 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7301 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7302 return BinaryOperator::CreateMul(BO->getOperand(0),
7303 ConstantExpr::getShl(BOOp, Op1));
7305 // Try to fold constant and into select arguments.
7306 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7307 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7309 if (isa<PHINode>(Op0))
7310 if (Instruction *NV = FoldOpIntoPhi(I))
7313 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7314 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7315 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7316 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7317 // place. Don't try to do this transformation in this case. Also, we
7318 // require that the input operand is a shift-by-constant so that we have
7319 // confidence that the shifts will get folded together. We could do this
7320 // xform in more cases, but it is unlikely to be profitable.
7321 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7322 isa<ConstantInt>(TrOp->getOperand(1))) {
7323 // Okay, we'll do this xform. Make the shift of shift.
7324 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7325 // (shift2 (shift1 & 0x00FF), c2)
7326 Value *NSh = Builder->CreateBinOp(I.getOpcode(), TrOp, ShAmt,I.getName());
7328 // For logical shifts, the truncation has the effect of making the high
7329 // part of the register be zeros. Emulate this by inserting an AND to
7330 // clear the top bits as needed. This 'and' will usually be zapped by
7331 // other xforms later if dead.
7332 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7333 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7334 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7336 // The mask we constructed says what the trunc would do if occurring
7337 // between the shifts. We want to know the effect *after* the second
7338 // shift. We know that it is a logical shift by a constant, so adjust the
7339 // mask as appropriate.
7340 if (I.getOpcode() == Instruction::Shl)
7341 MaskV <<= Op1->getZExtValue();
7343 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7344 MaskV = MaskV.lshr(Op1->getZExtValue());
7348 Value *And = Builder->CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7351 // Return the value truncated to the interesting size.
7352 return new TruncInst(And, I.getType());
7356 if (Op0->hasOneUse()) {
7357 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7358 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7361 switch (Op0BO->getOpcode()) {
7363 case Instruction::Add:
7364 case Instruction::And:
7365 case Instruction::Or:
7366 case Instruction::Xor: {
7367 // These operators commute.
7368 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7369 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7370 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7371 m_Specific(Op1)))) {
7372 Value *YS = // (Y << C)
7373 Builder->CreateShl(Op0BO->getOperand(0), Op1, Op0BO->getName());
7375 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), YS, V1,
7376 Op0BO->getOperand(1)->getName());
7377 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7378 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7379 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7382 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7383 Value *Op0BOOp1 = Op0BO->getOperand(1);
7384 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7386 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7387 m_ConstantInt(CC))) &&
7388 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7389 Value *YS = // (Y << C)
7390 Builder->CreateShl(Op0BO->getOperand(0), Op1,
7393 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7394 V1->getName()+".mask");
7395 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7400 case Instruction::Sub: {
7401 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7402 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7403 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7404 m_Specific(Op1)))) {
7405 Value *YS = // (Y << C)
7406 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7408 Value *X = Builder->CreateBinOp(Op0BO->getOpcode(), V1, YS,
7409 Op0BO->getOperand(0)->getName());
7410 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7411 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7412 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7415 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7416 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7417 match(Op0BO->getOperand(0),
7418 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7419 m_ConstantInt(CC))) && V2 == Op1 &&
7420 cast<BinaryOperator>(Op0BO->getOperand(0))
7421 ->getOperand(0)->hasOneUse()) {
7422 Value *YS = // (Y << C)
7423 Builder->CreateShl(Op0BO->getOperand(1), Op1, Op0BO->getName());
7425 Value *XM = Builder->CreateAnd(V1, ConstantExpr::getShl(CC, Op1),
7426 V1->getName()+".mask");
7428 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7436 // If the operand is an bitwise operator with a constant RHS, and the
7437 // shift is the only use, we can pull it out of the shift.
7438 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7439 bool isValid = true; // Valid only for And, Or, Xor
7440 bool highBitSet = false; // Transform if high bit of constant set?
7442 switch (Op0BO->getOpcode()) {
7443 default: isValid = false; break; // Do not perform transform!
7444 case Instruction::Add:
7445 isValid = isLeftShift;
7447 case Instruction::Or:
7448 case Instruction::Xor:
7451 case Instruction::And:
7456 // If this is a signed shift right, and the high bit is modified
7457 // by the logical operation, do not perform the transformation.
7458 // The highBitSet boolean indicates the value of the high bit of
7459 // the constant which would cause it to be modified for this
7462 if (isValid && I.getOpcode() == Instruction::AShr)
7463 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7466 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7469 Builder->CreateBinOp(I.getOpcode(), Op0BO->getOperand(0), Op1);
7470 NewShift->takeName(Op0BO);
7472 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7479 // Find out if this is a shift of a shift by a constant.
7480 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7481 if (ShiftOp && !ShiftOp->isShift())
7484 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7485 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7486 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7487 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7488 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7489 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7490 Value *X = ShiftOp->getOperand(0);
7492 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7494 const IntegerType *Ty = cast<IntegerType>(I.getType());
7496 // Check for (X << c1) << c2 and (X >> c1) >> c2
7497 if (I.getOpcode() == ShiftOp->getOpcode()) {
7498 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7500 if (AmtSum >= TypeBits) {
7501 if (I.getOpcode() != Instruction::AShr)
7502 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7503 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7506 return BinaryOperator::Create(I.getOpcode(), X,
7507 ConstantInt::get(Ty, AmtSum));
7510 if (ShiftOp->getOpcode() == Instruction::LShr &&
7511 I.getOpcode() == Instruction::AShr) {
7512 if (AmtSum >= TypeBits)
7513 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7515 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7516 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7519 if (ShiftOp->getOpcode() == Instruction::AShr &&
7520 I.getOpcode() == Instruction::LShr) {
7521 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7522 if (AmtSum >= TypeBits)
7523 AmtSum = TypeBits-1;
7525 Value *Shift = Builder->CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7527 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7528 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7531 // Okay, if we get here, one shift must be left, and the other shift must be
7532 // right. See if the amounts are equal.
7533 if (ShiftAmt1 == ShiftAmt2) {
7534 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7535 if (I.getOpcode() == Instruction::Shl) {
7536 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7537 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7539 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7540 if (I.getOpcode() == Instruction::LShr) {
7541 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7542 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7544 // We can simplify ((X << C) >>s C) into a trunc + sext.
7545 // NOTE: we could do this for any C, but that would make 'unusual' integer
7546 // types. For now, just stick to ones well-supported by the code
7548 const Type *SExtType = 0;
7549 switch (Ty->getBitWidth() - ShiftAmt1) {
7556 SExtType = IntegerType::get(*Context, Ty->getBitWidth() - ShiftAmt1);
7561 return new SExtInst(Builder->CreateTrunc(X, SExtType, "sext"), Ty);
7562 // Otherwise, we can't handle it yet.
7563 } else if (ShiftAmt1 < ShiftAmt2) {
7564 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7566 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7567 if (I.getOpcode() == Instruction::Shl) {
7568 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7569 ShiftOp->getOpcode() == Instruction::AShr);
7570 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7572 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7573 return BinaryOperator::CreateAnd(Shift,
7574 ConstantInt::get(*Context, Mask));
7577 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7578 if (I.getOpcode() == Instruction::LShr) {
7579 assert(ShiftOp->getOpcode() == Instruction::Shl);
7580 Value *Shift = Builder->CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7582 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7583 return BinaryOperator::CreateAnd(Shift,
7584 ConstantInt::get(*Context, Mask));
7587 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7589 assert(ShiftAmt2 < ShiftAmt1);
7590 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7592 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7593 if (I.getOpcode() == Instruction::Shl) {
7594 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7595 ShiftOp->getOpcode() == Instruction::AShr);
7596 Value *Shift = Builder->CreateBinOp(ShiftOp->getOpcode(), X,
7597 ConstantInt::get(Ty, ShiftDiff));
7599 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7600 return BinaryOperator::CreateAnd(Shift,
7601 ConstantInt::get(*Context, Mask));
7604 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7605 if (I.getOpcode() == Instruction::LShr) {
7606 assert(ShiftOp->getOpcode() == Instruction::Shl);
7607 Value *Shift = Builder->CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7609 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7610 return BinaryOperator::CreateAnd(Shift,
7611 ConstantInt::get(*Context, Mask));
7614 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7621 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7622 /// expression. If so, decompose it, returning some value X, such that Val is
7625 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7626 int &Offset, LLVMContext *Context) {
7627 assert(Val->getType() == Type::getInt32Ty(*Context) &&
7628 "Unexpected allocation size type!");
7629 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7630 Offset = CI->getZExtValue();
7632 return ConstantInt::get(Type::getInt32Ty(*Context), 0);
7633 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7634 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7635 if (I->getOpcode() == Instruction::Shl) {
7636 // This is a value scaled by '1 << the shift amt'.
7637 Scale = 1U << RHS->getZExtValue();
7639 return I->getOperand(0);
7640 } else if (I->getOpcode() == Instruction::Mul) {
7641 // This value is scaled by 'RHS'.
7642 Scale = RHS->getZExtValue();
7644 return I->getOperand(0);
7645 } else if (I->getOpcode() == Instruction::Add) {
7646 // We have X+C. Check to see if we really have (X*C2)+C1,
7647 // where C1 is divisible by C2.
7650 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7652 Offset += RHS->getZExtValue();
7659 // Otherwise, we can't look past this.
7666 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7667 /// try to eliminate the cast by moving the type information into the alloc.
7668 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7669 AllocationInst &AI) {
7670 const PointerType *PTy = cast<PointerType>(CI.getType());
7672 BuilderTy AllocaBuilder(*Builder);
7673 AllocaBuilder.SetInsertPoint(AI.getParent(), &AI);
7675 // Remove any uses of AI that are dead.
7676 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7678 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7679 Instruction *User = cast<Instruction>(*UI++);
7680 if (isInstructionTriviallyDead(User)) {
7681 while (UI != E && *UI == User)
7682 ++UI; // If this instruction uses AI more than once, don't break UI.
7685 DEBUG(errs() << "IC: DCE: " << *User << '\n');
7686 EraseInstFromFunction(*User);
7690 // This requires TargetData to get the alloca alignment and size information.
7693 // Get the type really allocated and the type casted to.
7694 const Type *AllocElTy = AI.getAllocatedType();
7695 const Type *CastElTy = PTy->getElementType();
7696 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7698 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7699 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7700 if (CastElTyAlign < AllocElTyAlign) return 0;
7702 // If the allocation has multiple uses, only promote it if we are strictly
7703 // increasing the alignment of the resultant allocation. If we keep it the
7704 // same, we open the door to infinite loops of various kinds. (A reference
7705 // from a dbg.declare doesn't count as a use for this purpose.)
7706 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7707 CastElTyAlign == AllocElTyAlign) return 0;
7709 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7710 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7711 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7713 // See if we can satisfy the modulus by pulling a scale out of the array
7715 unsigned ArraySizeScale;
7717 Value *NumElements = // See if the array size is a decomposable linear expr.
7718 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7719 ArrayOffset, Context);
7721 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7723 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7724 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7726 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7731 Amt = ConstantInt::get(Type::getInt32Ty(*Context), Scale);
7732 // Insert before the alloca, not before the cast.
7733 Amt = AllocaBuilder.CreateMul(Amt, NumElements, "tmp");
7736 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7737 Value *Off = ConstantInt::get(Type::getInt32Ty(*Context), Offset, true);
7738 Amt = AllocaBuilder.CreateAdd(Amt, Off, "tmp");
7741 AllocationInst *New;
7742 if (isa<MallocInst>(AI))
7743 New = AllocaBuilder.CreateMalloc(CastElTy, Amt);
7745 New = AllocaBuilder.CreateAlloca(CastElTy, Amt);
7746 New->setAlignment(AI.getAlignment());
7749 // If the allocation has one real use plus a dbg.declare, just remove the
7751 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7752 EraseInstFromFunction(*DI);
7754 // If the allocation has multiple real uses, insert a cast and change all
7755 // things that used it to use the new cast. This will also hack on CI, but it
7757 else if (!AI.hasOneUse()) {
7758 // New is the allocation instruction, pointer typed. AI is the original
7759 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7760 Value *NewCast = AllocaBuilder.CreateBitCast(New, AI.getType(), "tmpcast");
7761 AI.replaceAllUsesWith(NewCast);
7763 return ReplaceInstUsesWith(CI, New);
7766 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7767 /// and return it as type Ty without inserting any new casts and without
7768 /// changing the computed value. This is used by code that tries to decide
7769 /// whether promoting or shrinking integer operations to wider or smaller types
7770 /// will allow us to eliminate a truncate or extend.
7772 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7773 /// extension operation if Ty is larger.
7775 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7776 /// should return true if trunc(V) can be computed by computing V in the smaller
7777 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7778 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7779 /// efficiently truncated.
7781 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7782 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7783 /// the final result.
7784 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7786 int &NumCastsRemoved){
7787 // We can always evaluate constants in another type.
7788 if (isa<Constant>(V))
7791 Instruction *I = dyn_cast<Instruction>(V);
7792 if (!I) return false;
7794 const Type *OrigTy = V->getType();
7796 // If this is an extension or truncate, we can often eliminate it.
7797 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7798 // If this is a cast from the destination type, we can trivially eliminate
7799 // it, and this will remove a cast overall.
7800 if (I->getOperand(0)->getType() == Ty) {
7801 // If the first operand is itself a cast, and is eliminable, do not count
7802 // this as an eliminable cast. We would prefer to eliminate those two
7804 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7810 // We can't extend or shrink something that has multiple uses: doing so would
7811 // require duplicating the instruction in general, which isn't profitable.
7812 if (!I->hasOneUse()) return false;
7814 unsigned Opc = I->getOpcode();
7816 case Instruction::Add:
7817 case Instruction::Sub:
7818 case Instruction::Mul:
7819 case Instruction::And:
7820 case Instruction::Or:
7821 case Instruction::Xor:
7822 // These operators can all arbitrarily be extended or truncated.
7823 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7825 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7828 case Instruction::UDiv:
7829 case Instruction::URem: {
7830 // UDiv and URem can be truncated if all the truncated bits are zero.
7831 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7832 uint32_t BitWidth = Ty->getScalarSizeInBits();
7833 if (BitWidth < OrigBitWidth) {
7834 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7835 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7836 MaskedValueIsZero(I->getOperand(1), Mask)) {
7837 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7839 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7845 case Instruction::Shl:
7846 // If we are truncating the result of this SHL, and if it's a shift of a
7847 // constant amount, we can always perform a SHL in a smaller type.
7848 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7849 uint32_t BitWidth = Ty->getScalarSizeInBits();
7850 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7851 CI->getLimitedValue(BitWidth) < BitWidth)
7852 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7856 case Instruction::LShr:
7857 // If this is a truncate of a logical shr, we can truncate it to a smaller
7858 // lshr iff we know that the bits we would otherwise be shifting in are
7860 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7861 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7862 uint32_t BitWidth = Ty->getScalarSizeInBits();
7863 if (BitWidth < OrigBitWidth &&
7864 MaskedValueIsZero(I->getOperand(0),
7865 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7866 CI->getLimitedValue(BitWidth) < BitWidth) {
7867 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7872 case Instruction::ZExt:
7873 case Instruction::SExt:
7874 case Instruction::Trunc:
7875 // If this is the same kind of case as our original (e.g. zext+zext), we
7876 // can safely replace it. Note that replacing it does not reduce the number
7877 // of casts in the input.
7881 // sext (zext ty1), ty2 -> zext ty2
7882 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
7885 case Instruction::Select: {
7886 SelectInst *SI = cast<SelectInst>(I);
7887 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
7889 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
7892 case Instruction::PHI: {
7893 // We can change a phi if we can change all operands.
7894 PHINode *PN = cast<PHINode>(I);
7895 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
7896 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
7902 // TODO: Can handle more cases here.
7909 /// EvaluateInDifferentType - Given an expression that
7910 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7911 /// evaluate the expression.
7912 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7914 if (Constant *C = dyn_cast<Constant>(V))
7915 return ConstantExpr::getIntegerCast(C, Ty,
7916 isSigned /*Sext or ZExt*/);
7918 // Otherwise, it must be an instruction.
7919 Instruction *I = cast<Instruction>(V);
7920 Instruction *Res = 0;
7921 unsigned Opc = I->getOpcode();
7923 case Instruction::Add:
7924 case Instruction::Sub:
7925 case Instruction::Mul:
7926 case Instruction::And:
7927 case Instruction::Or:
7928 case Instruction::Xor:
7929 case Instruction::AShr:
7930 case Instruction::LShr:
7931 case Instruction::Shl:
7932 case Instruction::UDiv:
7933 case Instruction::URem: {
7934 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7935 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7936 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
7939 case Instruction::Trunc:
7940 case Instruction::ZExt:
7941 case Instruction::SExt:
7942 // If the source type of the cast is the type we're trying for then we can
7943 // just return the source. There's no need to insert it because it is not
7945 if (I->getOperand(0)->getType() == Ty)
7946 return I->getOperand(0);
7948 // Otherwise, must be the same type of cast, so just reinsert a new one.
7949 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7952 case Instruction::Select: {
7953 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7954 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
7955 Res = SelectInst::Create(I->getOperand(0), True, False);
7958 case Instruction::PHI: {
7959 PHINode *OPN = cast<PHINode>(I);
7960 PHINode *NPN = PHINode::Create(Ty);
7961 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
7962 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
7963 NPN->addIncoming(V, OPN->getIncomingBlock(i));
7969 // TODO: Can handle more cases here.
7970 llvm_unreachable("Unreachable!");
7975 return InsertNewInstBefore(Res, *I);
7978 /// @brief Implement the transforms common to all CastInst visitors.
7979 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7980 Value *Src = CI.getOperand(0);
7982 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7983 // eliminate it now.
7984 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7985 if (Instruction::CastOps opc =
7986 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7987 // The first cast (CSrc) is eliminable so we need to fix up or replace
7988 // the second cast (CI). CSrc will then have a good chance of being dead.
7989 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
7993 // If we are casting a select then fold the cast into the select
7994 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7995 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7998 // If we are casting a PHI then fold the cast into the PHI
7999 if (isa<PHINode>(Src))
8000 if (Instruction *NV = FoldOpIntoPhi(CI))
8006 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8007 /// or not there is a sequence of GEP indices into the type that will land us at
8008 /// the specified offset. If so, fill them into NewIndices and return the
8009 /// resultant element type, otherwise return null.
8010 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8011 SmallVectorImpl<Value*> &NewIndices,
8012 const TargetData *TD,
8013 LLVMContext *Context) {
8015 if (!Ty->isSized()) return 0;
8017 // Start with the index over the outer type. Note that the type size
8018 // might be zero (even if the offset isn't zero) if the indexed type
8019 // is something like [0 x {int, int}]
8020 const Type *IntPtrTy = TD->getIntPtrType(*Context);
8021 int64_t FirstIdx = 0;
8022 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8023 FirstIdx = Offset/TySize;
8024 Offset -= FirstIdx*TySize;
8026 // Handle hosts where % returns negative instead of values [0..TySize).
8030 assert(Offset >= 0);
8032 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8035 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8037 // Index into the types. If we fail, set OrigBase to null.
8039 // Indexing into tail padding between struct/array elements.
8040 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8043 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8044 const StructLayout *SL = TD->getStructLayout(STy);
8045 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8046 "Offset must stay within the indexed type");
8048 unsigned Elt = SL->getElementContainingOffset(Offset);
8049 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Elt));
8051 Offset -= SL->getElementOffset(Elt);
8052 Ty = STy->getElementType(Elt);
8053 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8054 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8055 assert(EltSize && "Cannot index into a zero-sized array");
8056 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8058 Ty = AT->getElementType();
8060 // Otherwise, we can't index into the middle of this atomic type, bail.
8068 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8069 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8070 Value *Src = CI.getOperand(0);
8072 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8073 // If casting the result of a getelementptr instruction with no offset, turn
8074 // this into a cast of the original pointer!
8075 if (GEP->hasAllZeroIndices()) {
8076 // Changing the cast operand is usually not a good idea but it is safe
8077 // here because the pointer operand is being replaced with another
8078 // pointer operand so the opcode doesn't need to change.
8080 CI.setOperand(0, GEP->getOperand(0));
8084 // If the GEP has a single use, and the base pointer is a bitcast, and the
8085 // GEP computes a constant offset, see if we can convert these three
8086 // instructions into fewer. This typically happens with unions and other
8087 // non-type-safe code.
8088 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8089 if (GEP->hasAllConstantIndices()) {
8090 // We are guaranteed to get a constant from EmitGEPOffset.
8091 ConstantInt *OffsetV =
8092 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8093 int64_t Offset = OffsetV->getSExtValue();
8095 // Get the base pointer input of the bitcast, and the type it points to.
8096 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8097 const Type *GEPIdxTy =
8098 cast<PointerType>(OrigBase->getType())->getElementType();
8099 SmallVector<Value*, 8> NewIndices;
8100 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8101 // If we were able to index down into an element, create the GEP
8102 // and bitcast the result. This eliminates one bitcast, potentially
8104 Value *NGEP = Builder->CreateGEP(OrigBase, NewIndices.begin(),
8106 NGEP->takeName(GEP);
8107 if (isa<Instruction>(NGEP) && cast<GEPOperator>(GEP)->isInBounds())
8108 cast<GEPOperator>(NGEP)->setIsInBounds(true);
8110 if (isa<BitCastInst>(CI))
8111 return new BitCastInst(NGEP, CI.getType());
8112 assert(isa<PtrToIntInst>(CI));
8113 return new PtrToIntInst(NGEP, CI.getType());
8119 return commonCastTransforms(CI);
8122 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8123 /// type like i42. We don't want to introduce operations on random non-legal
8124 /// integer types where they don't already exist in the code. In the future,
8125 /// we should consider making this based off target-data, so that 32-bit targets
8126 /// won't get i64 operations etc.
8127 static bool isSafeIntegerType(const Type *Ty) {
8128 switch (Ty->getPrimitiveSizeInBits()) {
8139 /// commonIntCastTransforms - This function implements the common transforms
8140 /// for trunc, zext, and sext.
8141 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8142 if (Instruction *Result = commonCastTransforms(CI))
8145 Value *Src = CI.getOperand(0);
8146 const Type *SrcTy = Src->getType();
8147 const Type *DestTy = CI.getType();
8148 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8149 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8151 // See if we can simplify any instructions used by the LHS whose sole
8152 // purpose is to compute bits we don't care about.
8153 if (SimplifyDemandedInstructionBits(CI))
8156 // If the source isn't an instruction or has more than one use then we
8157 // can't do anything more.
8158 Instruction *SrcI = dyn_cast<Instruction>(Src);
8159 if (!SrcI || !Src->hasOneUse())
8162 // Attempt to propagate the cast into the instruction for int->int casts.
8163 int NumCastsRemoved = 0;
8164 // Only do this if the dest type is a simple type, don't convert the
8165 // expression tree to something weird like i93 unless the source is also
8167 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8168 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8169 CanEvaluateInDifferentType(SrcI, DestTy,
8170 CI.getOpcode(), NumCastsRemoved)) {
8171 // If this cast is a truncate, evaluting in a different type always
8172 // eliminates the cast, so it is always a win. If this is a zero-extension,
8173 // we need to do an AND to maintain the clear top-part of the computation,
8174 // so we require that the input have eliminated at least one cast. If this
8175 // is a sign extension, we insert two new casts (to do the extension) so we
8176 // require that two casts have been eliminated.
8177 bool DoXForm = false;
8178 bool JustReplace = false;
8179 switch (CI.getOpcode()) {
8181 // All the others use floating point so we shouldn't actually
8182 // get here because of the check above.
8183 llvm_unreachable("Unknown cast type");
8184 case Instruction::Trunc:
8187 case Instruction::ZExt: {
8188 DoXForm = NumCastsRemoved >= 1;
8189 if (!DoXForm && 0) {
8190 // If it's unnecessary to issue an AND to clear the high bits, it's
8191 // always profitable to do this xform.
8192 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8193 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8194 if (MaskedValueIsZero(TryRes, Mask))
8195 return ReplaceInstUsesWith(CI, TryRes);
8197 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8198 if (TryI->use_empty())
8199 EraseInstFromFunction(*TryI);
8203 case Instruction::SExt: {
8204 DoXForm = NumCastsRemoved >= 2;
8205 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8206 // If we do not have to emit the truncate + sext pair, then it's always
8207 // profitable to do this xform.
8209 // It's not safe to eliminate the trunc + sext pair if one of the
8210 // eliminated cast is a truncate. e.g.
8211 // t2 = trunc i32 t1 to i16
8212 // t3 = sext i16 t2 to i32
8215 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8216 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8217 if (NumSignBits > (DestBitSize - SrcBitSize))
8218 return ReplaceInstUsesWith(CI, TryRes);
8220 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8221 if (TryI->use_empty())
8222 EraseInstFromFunction(*TryI);
8229 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
8230 " to avoid cast: " << CI);
8231 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8232 CI.getOpcode() == Instruction::SExt);
8234 // Just replace this cast with the result.
8235 return ReplaceInstUsesWith(CI, Res);
8237 assert(Res->getType() == DestTy);
8238 switch (CI.getOpcode()) {
8239 default: llvm_unreachable("Unknown cast type!");
8240 case Instruction::Trunc:
8241 // Just replace this cast with the result.
8242 return ReplaceInstUsesWith(CI, Res);
8243 case Instruction::ZExt: {
8244 assert(SrcBitSize < DestBitSize && "Not a zext?");
8246 // If the high bits are already zero, just replace this cast with the
8248 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8249 if (MaskedValueIsZero(Res, Mask))
8250 return ReplaceInstUsesWith(CI, Res);
8252 // We need to emit an AND to clear the high bits.
8253 Constant *C = ConstantInt::get(*Context,
8254 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8255 return BinaryOperator::CreateAnd(Res, C);
8257 case Instruction::SExt: {
8258 // If the high bits are already filled with sign bit, just replace this
8259 // cast with the result.
8260 unsigned NumSignBits = ComputeNumSignBits(Res);
8261 if (NumSignBits > (DestBitSize - SrcBitSize))
8262 return ReplaceInstUsesWith(CI, Res);
8264 // We need to emit a cast to truncate, then a cast to sext.
8265 return CastInst::Create(Instruction::SExt,
8266 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8273 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8274 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8276 switch (SrcI->getOpcode()) {
8277 case Instruction::Add:
8278 case Instruction::Mul:
8279 case Instruction::And:
8280 case Instruction::Or:
8281 case Instruction::Xor:
8282 // If we are discarding information, rewrite.
8283 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8284 // Don't insert two casts unless at least one can be eliminated.
8285 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8286 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8287 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8288 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8289 return BinaryOperator::Create(
8290 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8294 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8295 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8296 SrcI->getOpcode() == Instruction::Xor &&
8297 Op1 == ConstantInt::getTrue(*Context) &&
8298 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8299 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8300 return BinaryOperator::CreateXor(New,
8301 ConstantInt::get(CI.getType(), 1));
8305 case Instruction::Shl: {
8306 // Canonicalize trunc inside shl, if we can.
8307 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8308 if (CI && DestBitSize < SrcBitSize &&
8309 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8310 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8311 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8312 return BinaryOperator::CreateShl(Op0c, Op1c);
8320 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8321 if (Instruction *Result = commonIntCastTransforms(CI))
8324 Value *Src = CI.getOperand(0);
8325 const Type *Ty = CI.getType();
8326 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8327 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8329 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8330 if (DestBitWidth == 1) {
8331 Constant *One = ConstantInt::get(Src->getType(), 1);
8332 Src = Builder->CreateAnd(Src, One, "tmp");
8333 Value *Zero = Constant::getNullValue(Src->getType());
8334 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8337 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8338 ConstantInt *ShAmtV = 0;
8340 if (Src->hasOneUse() &&
8341 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8342 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8344 // Get a mask for the bits shifting in.
8345 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8346 if (MaskedValueIsZero(ShiftOp, Mask)) {
8347 if (ShAmt >= DestBitWidth) // All zeros.
8348 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8350 // Okay, we can shrink this. Truncate the input, then return a new
8352 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8353 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8354 return BinaryOperator::CreateLShr(V1, V2);
8361 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8362 /// in order to eliminate the icmp.
8363 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8365 // If we are just checking for a icmp eq of a single bit and zext'ing it
8366 // to an integer, then shift the bit to the appropriate place and then
8367 // cast to integer to avoid the comparison.
8368 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8369 const APInt &Op1CV = Op1C->getValue();
8371 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8372 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8373 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8374 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8375 if (!DoXform) return ICI;
8377 Value *In = ICI->getOperand(0);
8378 Value *Sh = ConstantInt::get(In->getType(),
8379 In->getType()->getScalarSizeInBits()-1);
8380 In = Builder->CreateLShr(In, Sh, In->getName()+".lobit");
8381 if (In->getType() != CI.getType())
8382 In = Builder->CreateIntCast(In, CI.getType(), false/*ZExt*/, "tmp");
8384 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8385 Constant *One = ConstantInt::get(In->getType(), 1);
8386 In = Builder->CreateXor(In, One, In->getName()+".not");
8389 return ReplaceInstUsesWith(CI, In);
8394 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8395 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8396 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8397 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8398 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8399 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8400 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8401 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8402 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8403 // This only works for EQ and NE
8404 ICI->isEquality()) {
8405 // If Op1C some other power of two, convert:
8406 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8407 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8408 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8409 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8411 APInt KnownZeroMask(~KnownZero);
8412 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8413 if (!DoXform) return ICI;
8415 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8416 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8417 // (X&4) == 2 --> false
8418 // (X&4) != 2 --> true
8419 Constant *Res = ConstantInt::get(Type::getInt1Ty(*Context), isNE);
8420 Res = ConstantExpr::getZExt(Res, CI.getType());
8421 return ReplaceInstUsesWith(CI, Res);
8424 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8425 Value *In = ICI->getOperand(0);
8427 // Perform a logical shr by shiftamt.
8428 // Insert the shift to put the result in the low bit.
8429 In = Builder->CreateLShr(In, ConstantInt::get(In->getType(),ShiftAmt),
8430 In->getName()+".lobit");
8433 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8434 Constant *One = ConstantInt::get(In->getType(), 1);
8435 In = Builder->CreateXor(In, One, "tmp");
8438 if (CI.getType() == In->getType())
8439 return ReplaceInstUsesWith(CI, In);
8441 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8449 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8450 // If one of the common conversion will work ..
8451 if (Instruction *Result = commonIntCastTransforms(CI))
8454 Value *Src = CI.getOperand(0);
8456 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8457 // types and if the sizes are just right we can convert this into a logical
8458 // 'and' which will be much cheaper than the pair of casts.
8459 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8460 // Get the sizes of the types involved. We know that the intermediate type
8461 // will be smaller than A or C, but don't know the relation between A and C.
8462 Value *A = CSrc->getOperand(0);
8463 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8464 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8465 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8466 // If we're actually extending zero bits, then if
8467 // SrcSize < DstSize: zext(a & mask)
8468 // SrcSize == DstSize: a & mask
8469 // SrcSize > DstSize: trunc(a) & mask
8470 if (SrcSize < DstSize) {
8471 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8472 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8473 Value *And = Builder->CreateAnd(A, AndConst, CSrc->getName()+".mask");
8474 return new ZExtInst(And, CI.getType());
8477 if (SrcSize == DstSize) {
8478 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8479 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8482 if (SrcSize > DstSize) {
8483 Value *Trunc = Builder->CreateTrunc(A, CI.getType(), "tmp");
8484 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8485 return BinaryOperator::CreateAnd(Trunc,
8486 ConstantInt::get(Trunc->getType(),
8491 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8492 return transformZExtICmp(ICI, CI);
8494 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8495 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8496 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8497 // of the (zext icmp) will be transformed.
8498 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8499 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8500 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8501 (transformZExtICmp(LHS, CI, false) ||
8502 transformZExtICmp(RHS, CI, false))) {
8503 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8504 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8505 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8509 // zext(trunc(t) & C) -> (t & zext(C)).
8510 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8511 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8512 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8513 Value *TI0 = TI->getOperand(0);
8514 if (TI0->getType() == CI.getType())
8516 BinaryOperator::CreateAnd(TI0,
8517 ConstantExpr::getZExt(C, CI.getType()));
8520 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8521 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8522 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8523 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8524 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8525 And->getOperand(1) == C)
8526 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8527 Value *TI0 = TI->getOperand(0);
8528 if (TI0->getType() == CI.getType()) {
8529 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8530 Value *NewAnd = Builder->CreateAnd(TI0, ZC, "tmp");
8531 return BinaryOperator::CreateXor(NewAnd, ZC);
8538 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8539 if (Instruction *I = commonIntCastTransforms(CI))
8542 Value *Src = CI.getOperand(0);
8544 // Canonicalize sign-extend from i1 to a select.
8545 if (Src->getType() == Type::getInt1Ty(*Context))
8546 return SelectInst::Create(Src,
8547 Constant::getAllOnesValue(CI.getType()),
8548 Constant::getNullValue(CI.getType()));
8550 // See if the value being truncated is already sign extended. If so, just
8551 // eliminate the trunc/sext pair.
8552 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8553 Value *Op = cast<User>(Src)->getOperand(0);
8554 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8555 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8556 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8557 unsigned NumSignBits = ComputeNumSignBits(Op);
8559 if (OpBits == DestBits) {
8560 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8561 // bits, it is already ready.
8562 if (NumSignBits > DestBits-MidBits)
8563 return ReplaceInstUsesWith(CI, Op);
8564 } else if (OpBits < DestBits) {
8565 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8566 // bits, just sext from i32.
8567 if (NumSignBits > OpBits-MidBits)
8568 return new SExtInst(Op, CI.getType(), "tmp");
8570 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8571 // bits, just truncate to i32.
8572 if (NumSignBits > OpBits-MidBits)
8573 return new TruncInst(Op, CI.getType(), "tmp");
8577 // If the input is a shl/ashr pair of a same constant, then this is a sign
8578 // extension from a smaller value. If we could trust arbitrary bitwidth
8579 // integers, we could turn this into a truncate to the smaller bit and then
8580 // use a sext for the whole extension. Since we don't, look deeper and check
8581 // for a truncate. If the source and dest are the same type, eliminate the
8582 // trunc and extend and just do shifts. For example, turn:
8583 // %a = trunc i32 %i to i8
8584 // %b = shl i8 %a, 6
8585 // %c = ashr i8 %b, 6
8586 // %d = sext i8 %c to i32
8588 // %a = shl i32 %i, 30
8589 // %d = ashr i32 %a, 30
8591 ConstantInt *BA = 0, *CA = 0;
8592 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8593 m_ConstantInt(CA))) &&
8594 BA == CA && isa<TruncInst>(A)) {
8595 Value *I = cast<TruncInst>(A)->getOperand(0);
8596 if (I->getType() == CI.getType()) {
8597 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8598 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8599 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8600 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8601 I = Builder->CreateShl(I, ShAmtV, CI.getName());
8602 return BinaryOperator::CreateAShr(I, ShAmtV);
8609 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8610 /// in the specified FP type without changing its value.
8611 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8612 LLVMContext *Context) {
8614 APFloat F = CFP->getValueAPF();
8615 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8617 return ConstantFP::get(*Context, F);
8621 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8622 /// through it until we get the source value.
8623 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8624 if (Instruction *I = dyn_cast<Instruction>(V))
8625 if (I->getOpcode() == Instruction::FPExt)
8626 return LookThroughFPExtensions(I->getOperand(0), Context);
8628 // If this value is a constant, return the constant in the smallest FP type
8629 // that can accurately represent it. This allows us to turn
8630 // (float)((double)X+2.0) into x+2.0f.
8631 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8632 if (CFP->getType() == Type::getPPC_FP128Ty(*Context))
8633 return V; // No constant folding of this.
8634 // See if the value can be truncated to float and then reextended.
8635 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8637 if (CFP->getType() == Type::getDoubleTy(*Context))
8638 return V; // Won't shrink.
8639 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8641 // Don't try to shrink to various long double types.
8647 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8648 if (Instruction *I = commonCastTransforms(CI))
8651 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8652 // smaller than the destination type, we can eliminate the truncate by doing
8653 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8654 // many builtins (sqrt, etc).
8655 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8656 if (OpI && OpI->hasOneUse()) {
8657 switch (OpI->getOpcode()) {
8659 case Instruction::FAdd:
8660 case Instruction::FSub:
8661 case Instruction::FMul:
8662 case Instruction::FDiv:
8663 case Instruction::FRem:
8664 const Type *SrcTy = OpI->getType();
8665 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8666 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8667 if (LHSTrunc->getType() != SrcTy &&
8668 RHSTrunc->getType() != SrcTy) {
8669 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8670 // If the source types were both smaller than the destination type of
8671 // the cast, do this xform.
8672 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8673 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8674 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8676 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8678 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8687 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8688 return commonCastTransforms(CI);
8691 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8692 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8694 return commonCastTransforms(FI);
8696 // fptoui(uitofp(X)) --> X
8697 // fptoui(sitofp(X)) --> X
8698 // This is safe if the intermediate type has enough bits in its mantissa to
8699 // accurately represent all values of X. For example, do not do this with
8700 // i64->float->i64. This is also safe for sitofp case, because any negative
8701 // 'X' value would cause an undefined result for the fptoui.
8702 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8703 OpI->getOperand(0)->getType() == FI.getType() &&
8704 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8705 OpI->getType()->getFPMantissaWidth())
8706 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8708 return commonCastTransforms(FI);
8711 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8712 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8714 return commonCastTransforms(FI);
8716 // fptosi(sitofp(X)) --> X
8717 // fptosi(uitofp(X)) --> X
8718 // This is safe if the intermediate type has enough bits in its mantissa to
8719 // accurately represent all values of X. For example, do not do this with
8720 // i64->float->i64. This is also safe for sitofp case, because any negative
8721 // 'X' value would cause an undefined result for the fptoui.
8722 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8723 OpI->getOperand(0)->getType() == FI.getType() &&
8724 (int)FI.getType()->getScalarSizeInBits() <=
8725 OpI->getType()->getFPMantissaWidth())
8726 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8728 return commonCastTransforms(FI);
8731 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8732 return commonCastTransforms(CI);
8735 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8736 return commonCastTransforms(CI);
8739 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8740 // If the destination integer type is smaller than the intptr_t type for
8741 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8742 // trunc to be exposed to other transforms. Don't do this for extending
8743 // ptrtoint's, because we don't know if the target sign or zero extends its
8746 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8747 Value *P = Builder->CreatePtrToInt(CI.getOperand(0),
8748 TD->getIntPtrType(CI.getContext()),
8750 return new TruncInst(P, CI.getType());
8753 return commonPointerCastTransforms(CI);
8756 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8757 // If the source integer type is larger than the intptr_t type for
8758 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8759 // allows the trunc to be exposed to other transforms. Don't do this for
8760 // extending inttoptr's, because we don't know if the target sign or zero
8761 // extends to pointers.
8762 if (TD && CI.getOperand(0)->getType()->getScalarSizeInBits() >
8763 TD->getPointerSizeInBits()) {
8764 Value *P = Builder->CreateTrunc(CI.getOperand(0),
8765 TD->getIntPtrType(CI.getContext()), "tmp");
8766 return new IntToPtrInst(P, CI.getType());
8769 if (Instruction *I = commonCastTransforms(CI))
8775 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8776 // If the operands are integer typed then apply the integer transforms,
8777 // otherwise just apply the common ones.
8778 Value *Src = CI.getOperand(0);
8779 const Type *SrcTy = Src->getType();
8780 const Type *DestTy = CI.getType();
8782 if (isa<PointerType>(SrcTy)) {
8783 if (Instruction *I = commonPointerCastTransforms(CI))
8786 if (Instruction *Result = commonCastTransforms(CI))
8791 // Get rid of casts from one type to the same type. These are useless and can
8792 // be replaced by the operand.
8793 if (DestTy == Src->getType())
8794 return ReplaceInstUsesWith(CI, Src);
8796 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8797 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8798 const Type *DstElTy = DstPTy->getElementType();
8799 const Type *SrcElTy = SrcPTy->getElementType();
8801 // If the address spaces don't match, don't eliminate the bitcast, which is
8802 // required for changing types.
8803 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8806 // If we are casting a malloc or alloca to a pointer to a type of the same
8807 // size, rewrite the allocation instruction to allocate the "right" type.
8808 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8809 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8812 // If the source and destination are pointers, and this cast is equivalent
8813 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8814 // This can enhance SROA and other transforms that want type-safe pointers.
8815 Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(*Context));
8816 unsigned NumZeros = 0;
8817 while (SrcElTy != DstElTy &&
8818 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8819 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8820 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8824 // If we found a path from the src to dest, create the getelementptr now.
8825 if (SrcElTy == DstElTy) {
8826 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8827 Instruction *GEP = GetElementPtrInst::Create(Src,
8828 Idxs.begin(), Idxs.end(), "",
8829 ((Instruction*) NULL));
8830 cast<GEPOperator>(GEP)->setIsInBounds(true);
8835 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
8836 if (DestVTy->getNumElements() == 1) {
8837 if (!isa<VectorType>(SrcTy)) {
8838 Value *Elem = InsertCastBefore(Instruction::BitCast, Src,
8839 DestVTy->getElementType(), CI);
8840 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
8841 Constant::getNullValue(Type::getInt32Ty(*Context)));
8843 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
8847 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
8848 if (SrcVTy->getNumElements() == 1) {
8849 if (!isa<VectorType>(DestTy)) {
8851 Builder->CreateExtractElement(Src,
8852 Constant::getNullValue(Type::getInt32Ty(*Context)));
8853 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
8858 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
8859 if (SVI->hasOneUse()) {
8860 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
8861 // a bitconvert to a vector with the same # elts.
8862 if (isa<VectorType>(DestTy) &&
8863 cast<VectorType>(DestTy)->getNumElements() ==
8864 SVI->getType()->getNumElements() &&
8865 SVI->getType()->getNumElements() ==
8866 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
8868 // If either of the operands is a cast from CI.getType(), then
8869 // evaluating the shuffle in the casted destination's type will allow
8870 // us to eliminate at least one cast.
8871 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
8872 Tmp->getOperand(0)->getType() == DestTy) ||
8873 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
8874 Tmp->getOperand(0)->getType() == DestTy)) {
8875 Value *LHS = InsertCastBefore(Instruction::BitCast,
8876 SVI->getOperand(0), DestTy, CI);
8877 Value *RHS = InsertCastBefore(Instruction::BitCast,
8878 SVI->getOperand(1), DestTy, CI);
8879 // Return a new shuffle vector. Use the same element ID's, as we
8880 // know the vector types match #elts.
8881 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
8889 /// GetSelectFoldableOperands - We want to turn code that looks like this:
8891 /// %D = select %cond, %C, %A
8893 /// %C = select %cond, %B, 0
8896 /// Assuming that the specified instruction is an operand to the select, return
8897 /// a bitmask indicating which operands of this instruction are foldable if they
8898 /// equal the other incoming value of the select.
8900 static unsigned GetSelectFoldableOperands(Instruction *I) {
8901 switch (I->getOpcode()) {
8902 case Instruction::Add:
8903 case Instruction::Mul:
8904 case Instruction::And:
8905 case Instruction::Or:
8906 case Instruction::Xor:
8907 return 3; // Can fold through either operand.
8908 case Instruction::Sub: // Can only fold on the amount subtracted.
8909 case Instruction::Shl: // Can only fold on the shift amount.
8910 case Instruction::LShr:
8911 case Instruction::AShr:
8914 return 0; // Cannot fold
8918 /// GetSelectFoldableConstant - For the same transformation as the previous
8919 /// function, return the identity constant that goes into the select.
8920 static Constant *GetSelectFoldableConstant(Instruction *I,
8921 LLVMContext *Context) {
8922 switch (I->getOpcode()) {
8923 default: llvm_unreachable("This cannot happen!");
8924 case Instruction::Add:
8925 case Instruction::Sub:
8926 case Instruction::Or:
8927 case Instruction::Xor:
8928 case Instruction::Shl:
8929 case Instruction::LShr:
8930 case Instruction::AShr:
8931 return Constant::getNullValue(I->getType());
8932 case Instruction::And:
8933 return Constant::getAllOnesValue(I->getType());
8934 case Instruction::Mul:
8935 return ConstantInt::get(I->getType(), 1);
8939 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8940 /// have the same opcode and only one use each. Try to simplify this.
8941 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8943 if (TI->getNumOperands() == 1) {
8944 // If this is a non-volatile load or a cast from the same type,
8947 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8950 return 0; // unknown unary op.
8953 // Fold this by inserting a select from the input values.
8954 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8955 FI->getOperand(0), SI.getName()+".v");
8956 InsertNewInstBefore(NewSI, SI);
8957 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
8961 // Only handle binary operators here.
8962 if (!isa<BinaryOperator>(TI))
8965 // Figure out if the operations have any operands in common.
8966 Value *MatchOp, *OtherOpT, *OtherOpF;
8968 if (TI->getOperand(0) == FI->getOperand(0)) {
8969 MatchOp = TI->getOperand(0);
8970 OtherOpT = TI->getOperand(1);
8971 OtherOpF = FI->getOperand(1);
8972 MatchIsOpZero = true;
8973 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8974 MatchOp = TI->getOperand(1);
8975 OtherOpT = TI->getOperand(0);
8976 OtherOpF = FI->getOperand(0);
8977 MatchIsOpZero = false;
8978 } else if (!TI->isCommutative()) {
8980 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8981 MatchOp = TI->getOperand(0);
8982 OtherOpT = TI->getOperand(1);
8983 OtherOpF = FI->getOperand(0);
8984 MatchIsOpZero = true;
8985 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8986 MatchOp = TI->getOperand(1);
8987 OtherOpT = TI->getOperand(0);
8988 OtherOpF = FI->getOperand(1);
8989 MatchIsOpZero = true;
8994 // If we reach here, they do have operations in common.
8995 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8996 OtherOpF, SI.getName()+".v");
8997 InsertNewInstBefore(NewSI, SI);
8999 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9001 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9003 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9005 llvm_unreachable("Shouldn't get here");
9009 static bool isSelect01(Constant *C1, Constant *C2) {
9010 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9013 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9016 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9019 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9020 /// facilitate further optimization.
9021 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9023 // See the comment above GetSelectFoldableOperands for a description of the
9024 // transformation we are doing here.
9025 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9026 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9027 !isa<Constant>(FalseVal)) {
9028 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9029 unsigned OpToFold = 0;
9030 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9032 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9037 Constant *C = GetSelectFoldableConstant(TVI, Context);
9038 Value *OOp = TVI->getOperand(2-OpToFold);
9039 // Avoid creating select between 2 constants unless it's selecting
9041 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9042 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9043 InsertNewInstBefore(NewSel, SI);
9044 NewSel->takeName(TVI);
9045 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9046 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9047 llvm_unreachable("Unknown instruction!!");
9054 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9055 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9056 !isa<Constant>(TrueVal)) {
9057 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9058 unsigned OpToFold = 0;
9059 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9061 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9066 Constant *C = GetSelectFoldableConstant(FVI, Context);
9067 Value *OOp = FVI->getOperand(2-OpToFold);
9068 // Avoid creating select between 2 constants unless it's selecting
9070 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9071 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9072 InsertNewInstBefore(NewSel, SI);
9073 NewSel->takeName(FVI);
9074 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9075 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9076 llvm_unreachable("Unknown instruction!!");
9086 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9087 /// ICmpInst as its first operand.
9089 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9091 bool Changed = false;
9092 ICmpInst::Predicate Pred = ICI->getPredicate();
9093 Value *CmpLHS = ICI->getOperand(0);
9094 Value *CmpRHS = ICI->getOperand(1);
9095 Value *TrueVal = SI.getTrueValue();
9096 Value *FalseVal = SI.getFalseValue();
9098 // Check cases where the comparison is with a constant that
9099 // can be adjusted to fit the min/max idiom. We may edit ICI in
9100 // place here, so make sure the select is the only user.
9101 if (ICI->hasOneUse())
9102 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9105 case ICmpInst::ICMP_ULT:
9106 case ICmpInst::ICMP_SLT: {
9107 // X < MIN ? T : F --> F
9108 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9109 return ReplaceInstUsesWith(SI, FalseVal);
9110 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9111 Constant *AdjustedRHS = SubOne(CI);
9112 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9113 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9114 Pred = ICmpInst::getSwappedPredicate(Pred);
9115 CmpRHS = AdjustedRHS;
9116 std::swap(FalseVal, TrueVal);
9117 ICI->setPredicate(Pred);
9118 ICI->setOperand(1, CmpRHS);
9119 SI.setOperand(1, TrueVal);
9120 SI.setOperand(2, FalseVal);
9125 case ICmpInst::ICMP_UGT:
9126 case ICmpInst::ICMP_SGT: {
9127 // X > MAX ? T : F --> F
9128 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9129 return ReplaceInstUsesWith(SI, FalseVal);
9130 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9131 Constant *AdjustedRHS = AddOne(CI);
9132 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9133 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9134 Pred = ICmpInst::getSwappedPredicate(Pred);
9135 CmpRHS = AdjustedRHS;
9136 std::swap(FalseVal, TrueVal);
9137 ICI->setPredicate(Pred);
9138 ICI->setOperand(1, CmpRHS);
9139 SI.setOperand(1, TrueVal);
9140 SI.setOperand(2, FalseVal);
9147 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9148 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9149 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9150 if (match(TrueVal, m_ConstantInt<-1>()) &&
9151 match(FalseVal, m_ConstantInt<0>()))
9152 Pred = ICI->getPredicate();
9153 else if (match(TrueVal, m_ConstantInt<0>()) &&
9154 match(FalseVal, m_ConstantInt<-1>()))
9155 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9157 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9158 // If we are just checking for a icmp eq of a single bit and zext'ing it
9159 // to an integer, then shift the bit to the appropriate place and then
9160 // cast to integer to avoid the comparison.
9161 const APInt &Op1CV = CI->getValue();
9163 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9164 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9165 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9166 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9167 Value *In = ICI->getOperand(0);
9168 Value *Sh = ConstantInt::get(In->getType(),
9169 In->getType()->getScalarSizeInBits()-1);
9170 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9171 In->getName()+".lobit"),
9173 if (In->getType() != SI.getType())
9174 In = CastInst::CreateIntegerCast(In, SI.getType(),
9175 true/*SExt*/, "tmp", ICI);
9177 if (Pred == ICmpInst::ICMP_SGT)
9178 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9179 In->getName()+".not"), *ICI);
9181 return ReplaceInstUsesWith(SI, In);
9186 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9187 // Transform (X == Y) ? X : Y -> Y
9188 if (Pred == ICmpInst::ICMP_EQ)
9189 return ReplaceInstUsesWith(SI, FalseVal);
9190 // Transform (X != Y) ? X : Y -> X
9191 if (Pred == ICmpInst::ICMP_NE)
9192 return ReplaceInstUsesWith(SI, TrueVal);
9193 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9195 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9196 // Transform (X == Y) ? Y : X -> X
9197 if (Pred == ICmpInst::ICMP_EQ)
9198 return ReplaceInstUsesWith(SI, FalseVal);
9199 // Transform (X != Y) ? Y : X -> Y
9200 if (Pred == ICmpInst::ICMP_NE)
9201 return ReplaceInstUsesWith(SI, TrueVal);
9202 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9205 /// NOTE: if we wanted to, this is where to detect integer ABS
9207 return Changed ? &SI : 0;
9210 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9211 Value *CondVal = SI.getCondition();
9212 Value *TrueVal = SI.getTrueValue();
9213 Value *FalseVal = SI.getFalseValue();
9215 // select true, X, Y -> X
9216 // select false, X, Y -> Y
9217 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9218 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9220 // select C, X, X -> X
9221 if (TrueVal == FalseVal)
9222 return ReplaceInstUsesWith(SI, TrueVal);
9224 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9225 return ReplaceInstUsesWith(SI, FalseVal);
9226 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9227 return ReplaceInstUsesWith(SI, TrueVal);
9228 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9229 if (isa<Constant>(TrueVal))
9230 return ReplaceInstUsesWith(SI, TrueVal);
9232 return ReplaceInstUsesWith(SI, FalseVal);
9235 if (SI.getType() == Type::getInt1Ty(*Context)) {
9236 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9237 if (C->getZExtValue()) {
9238 // Change: A = select B, true, C --> A = or B, C
9239 return BinaryOperator::CreateOr(CondVal, FalseVal);
9241 // Change: A = select B, false, C --> A = and !B, C
9243 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9244 "not."+CondVal->getName()), SI);
9245 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9247 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9248 if (C->getZExtValue() == false) {
9249 // Change: A = select B, C, false --> A = and B, C
9250 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9252 // Change: A = select B, C, true --> A = or !B, C
9254 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9255 "not."+CondVal->getName()), SI);
9256 return BinaryOperator::CreateOr(NotCond, TrueVal);
9260 // select a, b, a -> a&b
9261 // select a, a, b -> a|b
9262 if (CondVal == TrueVal)
9263 return BinaryOperator::CreateOr(CondVal, FalseVal);
9264 else if (CondVal == FalseVal)
9265 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9268 // Selecting between two integer constants?
9269 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9270 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9271 // select C, 1, 0 -> zext C to int
9272 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9273 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9274 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9275 // select C, 0, 1 -> zext !C to int
9277 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9278 "not."+CondVal->getName()), SI);
9279 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9282 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9283 // If one of the constants is zero (we know they can't both be) and we
9284 // have an icmp instruction with zero, and we have an 'and' with the
9285 // non-constant value, eliminate this whole mess. This corresponds to
9286 // cases like this: ((X & 27) ? 27 : 0)
9287 if (TrueValC->isZero() || FalseValC->isZero())
9288 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9289 cast<Constant>(IC->getOperand(1))->isNullValue())
9290 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9291 if (ICA->getOpcode() == Instruction::And &&
9292 isa<ConstantInt>(ICA->getOperand(1)) &&
9293 (ICA->getOperand(1) == TrueValC ||
9294 ICA->getOperand(1) == FalseValC) &&
9295 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9296 // Okay, now we know that everything is set up, we just don't
9297 // know whether we have a icmp_ne or icmp_eq and whether the
9298 // true or false val is the zero.
9299 bool ShouldNotVal = !TrueValC->isZero();
9300 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9303 V = InsertNewInstBefore(BinaryOperator::Create(
9304 Instruction::Xor, V, ICA->getOperand(1)), SI);
9305 return ReplaceInstUsesWith(SI, V);
9310 // See if we are selecting two values based on a comparison of the two values.
9311 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9312 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9313 // Transform (X == Y) ? X : Y -> Y
9314 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9315 // This is not safe in general for floating point:
9316 // consider X== -0, Y== +0.
9317 // It becomes safe if either operand is a nonzero constant.
9318 ConstantFP *CFPt, *CFPf;
9319 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9320 !CFPt->getValueAPF().isZero()) ||
9321 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9322 !CFPf->getValueAPF().isZero()))
9323 return ReplaceInstUsesWith(SI, FalseVal);
9325 // Transform (X != Y) ? X : Y -> X
9326 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9327 return ReplaceInstUsesWith(SI, TrueVal);
9328 // NOTE: if we wanted to, this is where to detect MIN/MAX
9330 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9331 // Transform (X == Y) ? Y : X -> X
9332 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9333 // This is not safe in general for floating point:
9334 // consider X== -0, Y== +0.
9335 // It becomes safe if either operand is a nonzero constant.
9336 ConstantFP *CFPt, *CFPf;
9337 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9338 !CFPt->getValueAPF().isZero()) ||
9339 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9340 !CFPf->getValueAPF().isZero()))
9341 return ReplaceInstUsesWith(SI, FalseVal);
9343 // Transform (X != Y) ? Y : X -> Y
9344 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9345 return ReplaceInstUsesWith(SI, TrueVal);
9346 // NOTE: if we wanted to, this is where to detect MIN/MAX
9348 // NOTE: if we wanted to, this is where to detect ABS
9351 // See if we are selecting two values based on a comparison of the two values.
9352 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9353 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9356 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9357 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9358 if (TI->hasOneUse() && FI->hasOneUse()) {
9359 Instruction *AddOp = 0, *SubOp = 0;
9361 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9362 if (TI->getOpcode() == FI->getOpcode())
9363 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9366 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9367 // even legal for FP.
9368 if ((TI->getOpcode() == Instruction::Sub &&
9369 FI->getOpcode() == Instruction::Add) ||
9370 (TI->getOpcode() == Instruction::FSub &&
9371 FI->getOpcode() == Instruction::FAdd)) {
9372 AddOp = FI; SubOp = TI;
9373 } else if ((FI->getOpcode() == Instruction::Sub &&
9374 TI->getOpcode() == Instruction::Add) ||
9375 (FI->getOpcode() == Instruction::FSub &&
9376 TI->getOpcode() == Instruction::FAdd)) {
9377 AddOp = TI; SubOp = FI;
9381 Value *OtherAddOp = 0;
9382 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9383 OtherAddOp = AddOp->getOperand(1);
9384 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9385 OtherAddOp = AddOp->getOperand(0);
9389 // So at this point we know we have (Y -> OtherAddOp):
9390 // select C, (add X, Y), (sub X, Z)
9391 Value *NegVal; // Compute -Z
9392 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9393 NegVal = ConstantExpr::getNeg(C);
9395 NegVal = InsertNewInstBefore(
9396 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9400 Value *NewTrueOp = OtherAddOp;
9401 Value *NewFalseOp = NegVal;
9403 std::swap(NewTrueOp, NewFalseOp);
9404 Instruction *NewSel =
9405 SelectInst::Create(CondVal, NewTrueOp,
9406 NewFalseOp, SI.getName() + ".p");
9408 NewSel = InsertNewInstBefore(NewSel, SI);
9409 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9414 // See if we can fold the select into one of our operands.
9415 if (SI.getType()->isInteger()) {
9416 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9421 if (BinaryOperator::isNot(CondVal)) {
9422 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9423 SI.setOperand(1, FalseVal);
9424 SI.setOperand(2, TrueVal);
9431 /// EnforceKnownAlignment - If the specified pointer points to an object that
9432 /// we control, modify the object's alignment to PrefAlign. This isn't
9433 /// often possible though. If alignment is important, a more reliable approach
9434 /// is to simply align all global variables and allocation instructions to
9435 /// their preferred alignment from the beginning.
9437 static unsigned EnforceKnownAlignment(Value *V,
9438 unsigned Align, unsigned PrefAlign) {
9440 User *U = dyn_cast<User>(V);
9441 if (!U) return Align;
9443 switch (Operator::getOpcode(U)) {
9445 case Instruction::BitCast:
9446 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9447 case Instruction::GetElementPtr: {
9448 // If all indexes are zero, it is just the alignment of the base pointer.
9449 bool AllZeroOperands = true;
9450 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9451 if (!isa<Constant>(*i) ||
9452 !cast<Constant>(*i)->isNullValue()) {
9453 AllZeroOperands = false;
9457 if (AllZeroOperands) {
9458 // Treat this like a bitcast.
9459 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9465 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9466 // If there is a large requested alignment and we can, bump up the alignment
9468 if (!GV->isDeclaration()) {
9469 if (GV->getAlignment() >= PrefAlign)
9470 Align = GV->getAlignment();
9472 GV->setAlignment(PrefAlign);
9476 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9477 // If there is a requested alignment and if this is an alloca, round up. We
9478 // don't do this for malloc, because some systems can't respect the request.
9479 if (isa<AllocaInst>(AI)) {
9480 if (AI->getAlignment() >= PrefAlign)
9481 Align = AI->getAlignment();
9483 AI->setAlignment(PrefAlign);
9492 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9493 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9494 /// and it is more than the alignment of the ultimate object, see if we can
9495 /// increase the alignment of the ultimate object, making this check succeed.
9496 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9497 unsigned PrefAlign) {
9498 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9499 sizeof(PrefAlign) * CHAR_BIT;
9500 APInt Mask = APInt::getAllOnesValue(BitWidth);
9501 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9502 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9503 unsigned TrailZ = KnownZero.countTrailingOnes();
9504 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9506 if (PrefAlign > Align)
9507 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9509 // We don't need to make any adjustment.
9513 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9514 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9515 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9516 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9517 unsigned CopyAlign = MI->getAlignment();
9519 if (CopyAlign < MinAlign) {
9520 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9525 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9527 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9528 if (MemOpLength == 0) return 0;
9530 // Source and destination pointer types are always "i8*" for intrinsic. See
9531 // if the size is something we can handle with a single primitive load/store.
9532 // A single load+store correctly handles overlapping memory in the memmove
9534 unsigned Size = MemOpLength->getZExtValue();
9535 if (Size == 0) return MI; // Delete this mem transfer.
9537 if (Size > 8 || (Size&(Size-1)))
9538 return 0; // If not 1/2/4/8 bytes, exit.
9540 // Use an integer load+store unless we can find something better.
9542 PointerType::getUnqual(IntegerType::get(*Context, Size<<3));
9544 // Memcpy forces the use of i8* for the source and destination. That means
9545 // that if you're using memcpy to move one double around, you'll get a cast
9546 // from double* to i8*. We'd much rather use a double load+store rather than
9547 // an i64 load+store, here because this improves the odds that the source or
9548 // dest address will be promotable. See if we can find a better type than the
9549 // integer datatype.
9550 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9551 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9552 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9553 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9554 // down through these levels if so.
9555 while (!SrcETy->isSingleValueType()) {
9556 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9557 if (STy->getNumElements() == 1)
9558 SrcETy = STy->getElementType(0);
9561 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9562 if (ATy->getNumElements() == 1)
9563 SrcETy = ATy->getElementType();
9570 if (SrcETy->isSingleValueType())
9571 NewPtrTy = PointerType::getUnqual(SrcETy);
9576 // If the memcpy/memmove provides better alignment info than we can
9578 SrcAlign = std::max(SrcAlign, CopyAlign);
9579 DstAlign = std::max(DstAlign, CopyAlign);
9581 Value *Src = Builder->CreateBitCast(MI->getOperand(2), NewPtrTy);
9582 Value *Dest = Builder->CreateBitCast(MI->getOperand(1), NewPtrTy);
9583 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9584 InsertNewInstBefore(L, *MI);
9585 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9587 // Set the size of the copy to 0, it will be deleted on the next iteration.
9588 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9592 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9593 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9594 if (MI->getAlignment() < Alignment) {
9595 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9600 // Extract the length and alignment and fill if they are constant.
9601 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9602 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9603 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(*Context))
9605 uint64_t Len = LenC->getZExtValue();
9606 Alignment = MI->getAlignment();
9608 // If the length is zero, this is a no-op
9609 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9611 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9612 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9613 const Type *ITy = IntegerType::get(*Context, Len*8); // n=1 -> i8.
9615 Value *Dest = MI->getDest();
9616 Dest = Builder->CreateBitCast(Dest, PointerType::getUnqual(ITy));
9618 // Alignment 0 is identity for alignment 1 for memset, but not store.
9619 if (Alignment == 0) Alignment = 1;
9621 // Extract the fill value and store.
9622 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9623 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
9624 Dest, false, Alignment), *MI);
9626 // Set the size of the copy to 0, it will be deleted on the next iteration.
9627 MI->setLength(Constant::getNullValue(LenC->getType()));
9635 /// visitCallInst - CallInst simplification. This mostly only handles folding
9636 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9637 /// the heavy lifting.
9639 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9640 // If the caller function is nounwind, mark the call as nounwind, even if the
9642 if (CI.getParent()->getParent()->doesNotThrow() &&
9643 !CI.doesNotThrow()) {
9644 CI.setDoesNotThrow();
9648 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9649 if (!II) return visitCallSite(&CI);
9651 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9653 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9654 bool Changed = false;
9656 // memmove/cpy/set of zero bytes is a noop.
9657 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9658 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9660 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9661 if (CI->getZExtValue() == 1) {
9662 // Replace the instruction with just byte operations. We would
9663 // transform other cases to loads/stores, but we don't know if
9664 // alignment is sufficient.
9668 // If we have a memmove and the source operation is a constant global,
9669 // then the source and dest pointers can't alias, so we can change this
9670 // into a call to memcpy.
9671 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9672 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9673 if (GVSrc->isConstant()) {
9674 Module *M = CI.getParent()->getParent()->getParent();
9675 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9677 Tys[0] = CI.getOperand(3)->getType();
9679 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9683 // memmove(x,x,size) -> noop.
9684 if (MMI->getSource() == MMI->getDest())
9685 return EraseInstFromFunction(CI);
9688 // If we can determine a pointer alignment that is bigger than currently
9689 // set, update the alignment.
9690 if (isa<MemTransferInst>(MI)) {
9691 if (Instruction *I = SimplifyMemTransfer(MI))
9693 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9694 if (Instruction *I = SimplifyMemSet(MSI))
9698 if (Changed) return II;
9701 switch (II->getIntrinsicID()) {
9703 case Intrinsic::bswap:
9704 // bswap(bswap(x)) -> x
9705 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9706 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9707 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9709 case Intrinsic::ppc_altivec_lvx:
9710 case Intrinsic::ppc_altivec_lvxl:
9711 case Intrinsic::x86_sse_loadu_ps:
9712 case Intrinsic::x86_sse2_loadu_pd:
9713 case Intrinsic::x86_sse2_loadu_dq:
9714 // Turn PPC lvx -> load if the pointer is known aligned.
9715 // Turn X86 loadups -> load if the pointer is known aligned.
9716 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9717 Value *Ptr = Builder->CreateBitCast(II->getOperand(1),
9718 PointerType::getUnqual(II->getType()));
9719 return new LoadInst(Ptr);
9722 case Intrinsic::ppc_altivec_stvx:
9723 case Intrinsic::ppc_altivec_stvxl:
9724 // Turn stvx -> store if the pointer is known aligned.
9725 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9726 const Type *OpPtrTy =
9727 PointerType::getUnqual(II->getOperand(1)->getType());
9728 Value *Ptr = Builder->CreateBitCast(II->getOperand(2), OpPtrTy);
9729 return new StoreInst(II->getOperand(1), Ptr);
9732 case Intrinsic::x86_sse_storeu_ps:
9733 case Intrinsic::x86_sse2_storeu_pd:
9734 case Intrinsic::x86_sse2_storeu_dq:
9735 // Turn X86 storeu -> store if the pointer is known aligned.
9736 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9737 const Type *OpPtrTy =
9738 PointerType::getUnqual(II->getOperand(2)->getType());
9739 Value *Ptr = Builder->CreateBitCast(II->getOperand(1), OpPtrTy);
9740 return new StoreInst(II->getOperand(2), Ptr);
9744 case Intrinsic::x86_sse_cvttss2si: {
9745 // These intrinsics only demands the 0th element of its input vector. If
9746 // we can simplify the input based on that, do so now.
9748 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9749 APInt DemandedElts(VWidth, 1);
9750 APInt UndefElts(VWidth, 0);
9751 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9753 II->setOperand(1, V);
9759 case Intrinsic::ppc_altivec_vperm:
9760 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9761 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9762 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9764 // Check that all of the elements are integer constants or undefs.
9765 bool AllEltsOk = true;
9766 for (unsigned i = 0; i != 16; ++i) {
9767 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9768 !isa<UndefValue>(Mask->getOperand(i))) {
9775 // Cast the input vectors to byte vectors.
9776 Value *Op0 = Builder->CreateBitCast(II->getOperand(1), Mask->getType());
9777 Value *Op1 = Builder->CreateBitCast(II->getOperand(2), Mask->getType());
9778 Value *Result = UndefValue::get(Op0->getType());
9780 // Only extract each element once.
9781 Value *ExtractedElts[32];
9782 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9784 for (unsigned i = 0; i != 16; ++i) {
9785 if (isa<UndefValue>(Mask->getOperand(i)))
9787 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9788 Idx &= 31; // Match the hardware behavior.
9790 if (ExtractedElts[Idx] == 0) {
9791 ExtractedElts[Idx] =
9792 Builder->CreateExtractElement(Idx < 16 ? Op0 : Op1,
9793 ConstantInt::get(Type::getInt32Ty(*Context), Idx&15, false),
9797 // Insert this value into the result vector.
9798 Result = Builder->CreateInsertElement(Result, ExtractedElts[Idx],
9799 ConstantInt::get(Type::getInt32Ty(*Context), i, false),
9802 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9807 case Intrinsic::stackrestore: {
9808 // If the save is right next to the restore, remove the restore. This can
9809 // happen when variable allocas are DCE'd.
9810 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9811 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9812 BasicBlock::iterator BI = SS;
9814 return EraseInstFromFunction(CI);
9818 // Scan down this block to see if there is another stack restore in the
9819 // same block without an intervening call/alloca.
9820 BasicBlock::iterator BI = II;
9821 TerminatorInst *TI = II->getParent()->getTerminator();
9822 bool CannotRemove = false;
9823 for (++BI; &*BI != TI; ++BI) {
9824 if (isa<AllocaInst>(BI)) {
9825 CannotRemove = true;
9828 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9829 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9830 // If there is a stackrestore below this one, remove this one.
9831 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9832 return EraseInstFromFunction(CI);
9833 // Otherwise, ignore the intrinsic.
9835 // If we found a non-intrinsic call, we can't remove the stack
9837 CannotRemove = true;
9843 // If the stack restore is in a return/unwind block and if there are no
9844 // allocas or calls between the restore and the return, nuke the restore.
9845 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
9846 return EraseInstFromFunction(CI);
9851 return visitCallSite(II);
9854 // InvokeInst simplification
9856 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
9857 return visitCallSite(&II);
9860 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
9861 /// passed through the varargs area, we can eliminate the use of the cast.
9862 static bool isSafeToEliminateVarargsCast(const CallSite CS,
9863 const CastInst * const CI,
9864 const TargetData * const TD,
9866 if (!CI->isLosslessCast())
9869 // The size of ByVal arguments is derived from the type, so we
9870 // can't change to a type with a different size. If the size were
9871 // passed explicitly we could avoid this check.
9872 if (!CS.paramHasAttr(ix, Attribute::ByVal))
9876 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
9877 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
9878 if (!SrcTy->isSized() || !DstTy->isSized())
9880 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
9885 // visitCallSite - Improvements for call and invoke instructions.
9887 Instruction *InstCombiner::visitCallSite(CallSite CS) {
9888 bool Changed = false;
9890 // If the callee is a constexpr cast of a function, attempt to move the cast
9891 // to the arguments of the call/invoke.
9892 if (transformConstExprCastCall(CS)) return 0;
9894 Value *Callee = CS.getCalledValue();
9896 if (Function *CalleeF = dyn_cast<Function>(Callee))
9897 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
9898 Instruction *OldCall = CS.getInstruction();
9899 // If the call and callee calling conventions don't match, this call must
9900 // be unreachable, as the call is undefined.
9901 new StoreInst(ConstantInt::getTrue(*Context),
9902 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))),
9904 if (!OldCall->use_empty())
9905 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
9906 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
9907 return EraseInstFromFunction(*OldCall);
9911 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
9912 // This instruction is not reachable, just remove it. We insert a store to
9913 // undef so that we know that this code is not reachable, despite the fact
9914 // that we can't modify the CFG here.
9915 new StoreInst(ConstantInt::getTrue(*Context),
9916 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))),
9917 CS.getInstruction());
9919 if (!CS.getInstruction()->use_empty())
9920 CS.getInstruction()->
9921 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
9923 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
9924 // Don't break the CFG, insert a dummy cond branch.
9925 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
9926 ConstantInt::getTrue(*Context), II);
9928 return EraseInstFromFunction(*CS.getInstruction());
9931 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
9932 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
9933 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
9934 return transformCallThroughTrampoline(CS);
9936 const PointerType *PTy = cast<PointerType>(Callee->getType());
9937 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9938 if (FTy->isVarArg()) {
9939 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
9940 // See if we can optimize any arguments passed through the varargs area of
9942 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
9943 E = CS.arg_end(); I != E; ++I, ++ix) {
9944 CastInst *CI = dyn_cast<CastInst>(*I);
9945 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
9946 *I = CI->getOperand(0);
9952 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
9953 // Inline asm calls cannot throw - mark them 'nounwind'.
9954 CS.setDoesNotThrow();
9958 return Changed ? CS.getInstruction() : 0;
9961 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
9962 // attempt to move the cast to the arguments of the call/invoke.
9964 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
9965 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
9966 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
9967 if (CE->getOpcode() != Instruction::BitCast ||
9968 !isa<Function>(CE->getOperand(0)))
9970 Function *Callee = cast<Function>(CE->getOperand(0));
9971 Instruction *Caller = CS.getInstruction();
9972 const AttrListPtr &CallerPAL = CS.getAttributes();
9974 // Okay, this is a cast from a function to a different type. Unless doing so
9975 // would cause a type conversion of one of our arguments, change this call to
9976 // be a direct call with arguments casted to the appropriate types.
9978 const FunctionType *FT = Callee->getFunctionType();
9979 const Type *OldRetTy = Caller->getType();
9980 const Type *NewRetTy = FT->getReturnType();
9982 if (isa<StructType>(NewRetTy))
9983 return false; // TODO: Handle multiple return values.
9985 // Check to see if we are changing the return type...
9986 if (OldRetTy != NewRetTy) {
9987 if (Callee->isDeclaration() &&
9988 // Conversion is ok if changing from one pointer type to another or from
9989 // a pointer to an integer of the same size.
9990 !((isa<PointerType>(OldRetTy) || !TD ||
9991 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
9992 (isa<PointerType>(NewRetTy) || !TD ||
9993 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
9994 return false; // Cannot transform this return value.
9996 if (!Caller->use_empty() &&
9997 // void -> non-void is handled specially
9998 NewRetTy != Type::getVoidTy(*Context) && !CastInst::isCastable(NewRetTy, OldRetTy))
9999 return false; // Cannot transform this return value.
10001 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10002 Attributes RAttrs = CallerPAL.getRetAttributes();
10003 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10004 return false; // Attribute not compatible with transformed value.
10007 // If the callsite is an invoke instruction, and the return value is used by
10008 // a PHI node in a successor, we cannot change the return type of the call
10009 // because there is no place to put the cast instruction (without breaking
10010 // the critical edge). Bail out in this case.
10011 if (!Caller->use_empty())
10012 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10013 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10015 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10016 if (PN->getParent() == II->getNormalDest() ||
10017 PN->getParent() == II->getUnwindDest())
10021 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10022 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10024 CallSite::arg_iterator AI = CS.arg_begin();
10025 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10026 const Type *ParamTy = FT->getParamType(i);
10027 const Type *ActTy = (*AI)->getType();
10029 if (!CastInst::isCastable(ActTy, ParamTy))
10030 return false; // Cannot transform this parameter value.
10032 if (CallerPAL.getParamAttributes(i + 1)
10033 & Attribute::typeIncompatible(ParamTy))
10034 return false; // Attribute not compatible with transformed value.
10036 // Converting from one pointer type to another or between a pointer and an
10037 // integer of the same size is safe even if we do not have a body.
10038 bool isConvertible = ActTy == ParamTy ||
10039 (TD && ((isa<PointerType>(ParamTy) ||
10040 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
10041 (isa<PointerType>(ActTy) ||
10042 ActTy == TD->getIntPtrType(Caller->getContext()))));
10043 if (Callee->isDeclaration() && !isConvertible) return false;
10046 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10047 Callee->isDeclaration())
10048 return false; // Do not delete arguments unless we have a function body.
10050 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10051 !CallerPAL.isEmpty())
10052 // In this case we have more arguments than the new function type, but we
10053 // won't be dropping them. Check that these extra arguments have attributes
10054 // that are compatible with being a vararg call argument.
10055 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10056 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10058 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10059 if (PAttrs & Attribute::VarArgsIncompatible)
10063 // Okay, we decided that this is a safe thing to do: go ahead and start
10064 // inserting cast instructions as necessary...
10065 std::vector<Value*> Args;
10066 Args.reserve(NumActualArgs);
10067 SmallVector<AttributeWithIndex, 8> attrVec;
10068 attrVec.reserve(NumCommonArgs);
10070 // Get any return attributes.
10071 Attributes RAttrs = CallerPAL.getRetAttributes();
10073 // If the return value is not being used, the type may not be compatible
10074 // with the existing attributes. Wipe out any problematic attributes.
10075 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10077 // Add the new return attributes.
10079 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10081 AI = CS.arg_begin();
10082 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10083 const Type *ParamTy = FT->getParamType(i);
10084 if ((*AI)->getType() == ParamTy) {
10085 Args.push_back(*AI);
10087 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10088 false, ParamTy, false);
10089 Args.push_back(Builder->CreateCast(opcode, *AI, ParamTy, "tmp"));
10092 // Add any parameter attributes.
10093 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10094 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10097 // If the function takes more arguments than the call was taking, add them
10099 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10100 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10102 // If we are removing arguments to the function, emit an obnoxious warning.
10103 if (FT->getNumParams() < NumActualArgs) {
10104 if (!FT->isVarArg()) {
10105 errs() << "WARNING: While resolving call to function '"
10106 << Callee->getName() << "' arguments were dropped!\n";
10108 // Add all of the arguments in their promoted form to the arg list.
10109 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10110 const Type *PTy = getPromotedType((*AI)->getType());
10111 if (PTy != (*AI)->getType()) {
10112 // Must promote to pass through va_arg area!
10113 Instruction::CastOps opcode =
10114 CastInst::getCastOpcode(*AI, false, PTy, false);
10115 Args.push_back(Builder->CreateCast(opcode, *AI, PTy, "tmp"));
10117 Args.push_back(*AI);
10120 // Add any parameter attributes.
10121 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10122 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10127 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10128 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10130 if (NewRetTy == Type::getVoidTy(*Context))
10131 Caller->setName(""); // Void type should not have a name.
10133 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10137 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10138 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10139 Args.begin(), Args.end(),
10140 Caller->getName(), Caller);
10141 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10142 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10144 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10145 Caller->getName(), Caller);
10146 CallInst *CI = cast<CallInst>(Caller);
10147 if (CI->isTailCall())
10148 cast<CallInst>(NC)->setTailCall();
10149 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10150 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10153 // Insert a cast of the return type as necessary.
10155 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10156 if (NV->getType() != Type::getVoidTy(*Context)) {
10157 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10159 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10161 // If this is an invoke instruction, we should insert it after the first
10162 // non-phi, instruction in the normal successor block.
10163 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10164 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10165 InsertNewInstBefore(NC, *I);
10167 // Otherwise, it's a call, just insert cast right after the call instr
10168 InsertNewInstBefore(NC, *Caller);
10170 Worklist.AddUsersToWorkList(*Caller);
10172 NV = UndefValue::get(Caller->getType());
10176 if (Caller->getType() != Type::getVoidTy(*Context) && !Caller->use_empty())
10177 Caller->replaceAllUsesWith(NV);
10178 Caller->eraseFromParent();
10179 Worklist.Remove(Caller);
10183 // transformCallThroughTrampoline - Turn a call to a function created by the
10184 // init_trampoline intrinsic into a direct call to the underlying function.
10186 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10187 Value *Callee = CS.getCalledValue();
10188 const PointerType *PTy = cast<PointerType>(Callee->getType());
10189 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10190 const AttrListPtr &Attrs = CS.getAttributes();
10192 // If the call already has the 'nest' attribute somewhere then give up -
10193 // otherwise 'nest' would occur twice after splicing in the chain.
10194 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10197 IntrinsicInst *Tramp =
10198 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10200 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10201 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10202 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10204 const AttrListPtr &NestAttrs = NestF->getAttributes();
10205 if (!NestAttrs.isEmpty()) {
10206 unsigned NestIdx = 1;
10207 const Type *NestTy = 0;
10208 Attributes NestAttr = Attribute::None;
10210 // Look for a parameter marked with the 'nest' attribute.
10211 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10212 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10213 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10214 // Record the parameter type and any other attributes.
10216 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10221 Instruction *Caller = CS.getInstruction();
10222 std::vector<Value*> NewArgs;
10223 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10225 SmallVector<AttributeWithIndex, 8> NewAttrs;
10226 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10228 // Insert the nest argument into the call argument list, which may
10229 // mean appending it. Likewise for attributes.
10231 // Add any result attributes.
10232 if (Attributes Attr = Attrs.getRetAttributes())
10233 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10237 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10239 if (Idx == NestIdx) {
10240 // Add the chain argument and attributes.
10241 Value *NestVal = Tramp->getOperand(3);
10242 if (NestVal->getType() != NestTy)
10243 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10244 NewArgs.push_back(NestVal);
10245 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10251 // Add the original argument and attributes.
10252 NewArgs.push_back(*I);
10253 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10255 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10261 // Add any function attributes.
10262 if (Attributes Attr = Attrs.getFnAttributes())
10263 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10265 // The trampoline may have been bitcast to a bogus type (FTy).
10266 // Handle this by synthesizing a new function type, equal to FTy
10267 // with the chain parameter inserted.
10269 std::vector<const Type*> NewTypes;
10270 NewTypes.reserve(FTy->getNumParams()+1);
10272 // Insert the chain's type into the list of parameter types, which may
10273 // mean appending it.
10276 FunctionType::param_iterator I = FTy->param_begin(),
10277 E = FTy->param_end();
10280 if (Idx == NestIdx)
10281 // Add the chain's type.
10282 NewTypes.push_back(NestTy);
10287 // Add the original type.
10288 NewTypes.push_back(*I);
10294 // Replace the trampoline call with a direct call. Let the generic
10295 // code sort out any function type mismatches.
10296 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10298 Constant *NewCallee =
10299 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10300 NestF : ConstantExpr::getBitCast(NestF,
10301 PointerType::getUnqual(NewFTy));
10302 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10305 Instruction *NewCaller;
10306 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10307 NewCaller = InvokeInst::Create(NewCallee,
10308 II->getNormalDest(), II->getUnwindDest(),
10309 NewArgs.begin(), NewArgs.end(),
10310 Caller->getName(), Caller);
10311 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10312 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10314 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10315 Caller->getName(), Caller);
10316 if (cast<CallInst>(Caller)->isTailCall())
10317 cast<CallInst>(NewCaller)->setTailCall();
10318 cast<CallInst>(NewCaller)->
10319 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10320 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10322 if (Caller->getType() != Type::getVoidTy(*Context) && !Caller->use_empty())
10323 Caller->replaceAllUsesWith(NewCaller);
10324 Caller->eraseFromParent();
10325 Worklist.Remove(Caller);
10330 // Replace the trampoline call with a direct call. Since there is no 'nest'
10331 // parameter, there is no need to adjust the argument list. Let the generic
10332 // code sort out any function type mismatches.
10333 Constant *NewCallee =
10334 NestF->getType() == PTy ? NestF :
10335 ConstantExpr::getBitCast(NestF, PTy);
10336 CS.setCalledFunction(NewCallee);
10337 return CS.getInstruction();
10340 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10341 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10342 /// and a single binop.
10343 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10344 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10345 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10346 unsigned Opc = FirstInst->getOpcode();
10347 Value *LHSVal = FirstInst->getOperand(0);
10348 Value *RHSVal = FirstInst->getOperand(1);
10350 const Type *LHSType = LHSVal->getType();
10351 const Type *RHSType = RHSVal->getType();
10353 // Scan to see if all operands are the same opcode, all have one use, and all
10354 // kill their operands (i.e. the operands have one use).
10355 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10356 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10357 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10358 // Verify type of the LHS matches so we don't fold cmp's of different
10359 // types or GEP's with different index types.
10360 I->getOperand(0)->getType() != LHSType ||
10361 I->getOperand(1)->getType() != RHSType)
10364 // If they are CmpInst instructions, check their predicates
10365 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10366 if (cast<CmpInst>(I)->getPredicate() !=
10367 cast<CmpInst>(FirstInst)->getPredicate())
10370 // Keep track of which operand needs a phi node.
10371 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10372 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10375 // Otherwise, this is safe to transform!
10377 Value *InLHS = FirstInst->getOperand(0);
10378 Value *InRHS = FirstInst->getOperand(1);
10379 PHINode *NewLHS = 0, *NewRHS = 0;
10381 NewLHS = PHINode::Create(LHSType,
10382 FirstInst->getOperand(0)->getName() + ".pn");
10383 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10384 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10385 InsertNewInstBefore(NewLHS, PN);
10390 NewRHS = PHINode::Create(RHSType,
10391 FirstInst->getOperand(1)->getName() + ".pn");
10392 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10393 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10394 InsertNewInstBefore(NewRHS, PN);
10398 // Add all operands to the new PHIs.
10399 if (NewLHS || NewRHS) {
10400 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10401 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10403 Value *NewInLHS = InInst->getOperand(0);
10404 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10407 Value *NewInRHS = InInst->getOperand(1);
10408 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10413 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10414 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10415 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10416 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10420 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10421 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10423 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10424 FirstInst->op_end());
10425 // This is true if all GEP bases are allocas and if all indices into them are
10427 bool AllBasePointersAreAllocas = true;
10429 // Scan to see if all operands are the same opcode, all have one use, and all
10430 // kill their operands (i.e. the operands have one use).
10431 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10432 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10433 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10434 GEP->getNumOperands() != FirstInst->getNumOperands())
10437 // Keep track of whether or not all GEPs are of alloca pointers.
10438 if (AllBasePointersAreAllocas &&
10439 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10440 !GEP->hasAllConstantIndices()))
10441 AllBasePointersAreAllocas = false;
10443 // Compare the operand lists.
10444 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10445 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10448 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10449 // if one of the PHIs has a constant for the index. The index may be
10450 // substantially cheaper to compute for the constants, so making it a
10451 // variable index could pessimize the path. This also handles the case
10452 // for struct indices, which must always be constant.
10453 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10454 isa<ConstantInt>(GEP->getOperand(op)))
10457 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10459 FixedOperands[op] = 0; // Needs a PHI.
10463 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10464 // bother doing this transformation. At best, this will just save a bit of
10465 // offset calculation, but all the predecessors will have to materialize the
10466 // stack address into a register anyway. We'd actually rather *clone* the
10467 // load up into the predecessors so that we have a load of a gep of an alloca,
10468 // which can usually all be folded into the load.
10469 if (AllBasePointersAreAllocas)
10472 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10473 // that is variable.
10474 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10476 bool HasAnyPHIs = false;
10477 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10478 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10479 Value *FirstOp = FirstInst->getOperand(i);
10480 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10481 FirstOp->getName()+".pn");
10482 InsertNewInstBefore(NewPN, PN);
10484 NewPN->reserveOperandSpace(e);
10485 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10486 OperandPhis[i] = NewPN;
10487 FixedOperands[i] = NewPN;
10492 // Add all operands to the new PHIs.
10494 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10495 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10496 BasicBlock *InBB = PN.getIncomingBlock(i);
10498 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10499 if (PHINode *OpPhi = OperandPhis[op])
10500 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10504 Value *Base = FixedOperands[0];
10505 GetElementPtrInst *GEP =
10506 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10507 FixedOperands.end());
10508 if (cast<GEPOperator>(FirstInst)->isInBounds())
10509 cast<GEPOperator>(GEP)->setIsInBounds(true);
10514 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10515 /// sink the load out of the block that defines it. This means that it must be
10516 /// obvious the value of the load is not changed from the point of the load to
10517 /// the end of the block it is in.
10519 /// Finally, it is safe, but not profitable, to sink a load targetting a
10520 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10522 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10523 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10525 for (++BBI; BBI != E; ++BBI)
10526 if (BBI->mayWriteToMemory())
10529 // Check for non-address taken alloca. If not address-taken already, it isn't
10530 // profitable to do this xform.
10531 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10532 bool isAddressTaken = false;
10533 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10535 if (isa<LoadInst>(UI)) continue;
10536 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10537 // If storing TO the alloca, then the address isn't taken.
10538 if (SI->getOperand(1) == AI) continue;
10540 isAddressTaken = true;
10544 if (!isAddressTaken && AI->isStaticAlloca())
10548 // If this load is a load from a GEP with a constant offset from an alloca,
10549 // then we don't want to sink it. In its present form, it will be
10550 // load [constant stack offset]. Sinking it will cause us to have to
10551 // materialize the stack addresses in each predecessor in a register only to
10552 // do a shared load from register in the successor.
10553 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10554 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10555 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10562 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10563 // operator and they all are only used by the PHI, PHI together their
10564 // inputs, and do the operation once, to the result of the PHI.
10565 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10566 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10568 // Scan the instruction, looking for input operations that can be folded away.
10569 // If all input operands to the phi are the same instruction (e.g. a cast from
10570 // the same type or "+42") we can pull the operation through the PHI, reducing
10571 // code size and simplifying code.
10572 Constant *ConstantOp = 0;
10573 const Type *CastSrcTy = 0;
10574 bool isVolatile = false;
10575 if (isa<CastInst>(FirstInst)) {
10576 CastSrcTy = FirstInst->getOperand(0)->getType();
10577 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10578 // Can fold binop, compare or shift here if the RHS is a constant,
10579 // otherwise call FoldPHIArgBinOpIntoPHI.
10580 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10581 if (ConstantOp == 0)
10582 return FoldPHIArgBinOpIntoPHI(PN);
10583 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10584 isVolatile = LI->isVolatile();
10585 // We can't sink the load if the loaded value could be modified between the
10586 // load and the PHI.
10587 if (LI->getParent() != PN.getIncomingBlock(0) ||
10588 !isSafeAndProfitableToSinkLoad(LI))
10591 // If the PHI is of volatile loads and the load block has multiple
10592 // successors, sinking it would remove a load of the volatile value from
10593 // the path through the other successor.
10595 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10598 } else if (isa<GetElementPtrInst>(FirstInst)) {
10599 return FoldPHIArgGEPIntoPHI(PN);
10601 return 0; // Cannot fold this operation.
10604 // Check to see if all arguments are the same operation.
10605 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10606 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10607 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10608 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10611 if (I->getOperand(0)->getType() != CastSrcTy)
10612 return 0; // Cast operation must match.
10613 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10614 // We can't sink the load if the loaded value could be modified between
10615 // the load and the PHI.
10616 if (LI->isVolatile() != isVolatile ||
10617 LI->getParent() != PN.getIncomingBlock(i) ||
10618 !isSafeAndProfitableToSinkLoad(LI))
10621 // If the PHI is of volatile loads and the load block has multiple
10622 // successors, sinking it would remove a load of the volatile value from
10623 // the path through the other successor.
10625 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10628 } else if (I->getOperand(1) != ConstantOp) {
10633 // Okay, they are all the same operation. Create a new PHI node of the
10634 // correct type, and PHI together all of the LHS's of the instructions.
10635 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10636 PN.getName()+".in");
10637 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10639 Value *InVal = FirstInst->getOperand(0);
10640 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10642 // Add all operands to the new PHI.
10643 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10644 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10645 if (NewInVal != InVal)
10647 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10652 // The new PHI unions all of the same values together. This is really
10653 // common, so we handle it intelligently here for compile-time speed.
10657 InsertNewInstBefore(NewPN, PN);
10661 // Insert and return the new operation.
10662 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10663 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10664 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10665 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10666 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10667 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10668 PhiVal, ConstantOp);
10669 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10671 // If this was a volatile load that we are merging, make sure to loop through
10672 // and mark all the input loads as non-volatile. If we don't do this, we will
10673 // insert a new volatile load and the old ones will not be deletable.
10675 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10676 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10678 return new LoadInst(PhiVal, "", isVolatile);
10681 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10683 static bool DeadPHICycle(PHINode *PN,
10684 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10685 if (PN->use_empty()) return true;
10686 if (!PN->hasOneUse()) return false;
10688 // Remember this node, and if we find the cycle, return.
10689 if (!PotentiallyDeadPHIs.insert(PN))
10692 // Don't scan crazily complex things.
10693 if (PotentiallyDeadPHIs.size() == 16)
10696 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10697 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10702 /// PHIsEqualValue - Return true if this phi node is always equal to
10703 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10704 /// z = some value; x = phi (y, z); y = phi (x, z)
10705 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10706 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10707 // See if we already saw this PHI node.
10708 if (!ValueEqualPHIs.insert(PN))
10711 // Don't scan crazily complex things.
10712 if (ValueEqualPHIs.size() == 16)
10715 // Scan the operands to see if they are either phi nodes or are equal to
10717 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10718 Value *Op = PN->getIncomingValue(i);
10719 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10720 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10722 } else if (Op != NonPhiInVal)
10730 // PHINode simplification
10732 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10733 // If LCSSA is around, don't mess with Phi nodes
10734 if (MustPreserveLCSSA) return 0;
10736 if (Value *V = PN.hasConstantValue())
10737 return ReplaceInstUsesWith(PN, V);
10739 // If all PHI operands are the same operation, pull them through the PHI,
10740 // reducing code size.
10741 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10742 isa<Instruction>(PN.getIncomingValue(1)) &&
10743 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10744 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10745 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10746 // than themselves more than once.
10747 PN.getIncomingValue(0)->hasOneUse())
10748 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10751 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10752 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10753 // PHI)... break the cycle.
10754 if (PN.hasOneUse()) {
10755 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10756 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10757 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10758 PotentiallyDeadPHIs.insert(&PN);
10759 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10760 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10763 // If this phi has a single use, and if that use just computes a value for
10764 // the next iteration of a loop, delete the phi. This occurs with unused
10765 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10766 // common case here is good because the only other things that catch this
10767 // are induction variable analysis (sometimes) and ADCE, which is only run
10769 if (PHIUser->hasOneUse() &&
10770 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10771 PHIUser->use_back() == &PN) {
10772 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10776 // We sometimes end up with phi cycles that non-obviously end up being the
10777 // same value, for example:
10778 // z = some value; x = phi (y, z); y = phi (x, z)
10779 // where the phi nodes don't necessarily need to be in the same block. Do a
10780 // quick check to see if the PHI node only contains a single non-phi value, if
10781 // so, scan to see if the phi cycle is actually equal to that value.
10783 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10784 // Scan for the first non-phi operand.
10785 while (InValNo != NumOperandVals &&
10786 isa<PHINode>(PN.getIncomingValue(InValNo)))
10789 if (InValNo != NumOperandVals) {
10790 Value *NonPhiInVal = PN.getOperand(InValNo);
10792 // Scan the rest of the operands to see if there are any conflicts, if so
10793 // there is no need to recursively scan other phis.
10794 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10795 Value *OpVal = PN.getIncomingValue(InValNo);
10796 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10800 // If we scanned over all operands, then we have one unique value plus
10801 // phi values. Scan PHI nodes to see if they all merge in each other or
10803 if (InValNo == NumOperandVals) {
10804 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10805 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10806 return ReplaceInstUsesWith(PN, NonPhiInVal);
10813 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10814 Value *PtrOp = GEP.getOperand(0);
10815 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10816 // If so, eliminate the noop.
10817 if (GEP.getNumOperands() == 1)
10818 return ReplaceInstUsesWith(GEP, PtrOp);
10820 if (isa<UndefValue>(GEP.getOperand(0)))
10821 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10823 bool HasZeroPointerIndex = false;
10824 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10825 HasZeroPointerIndex = C->isNullValue();
10827 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
10828 return ReplaceInstUsesWith(GEP, PtrOp);
10830 // Eliminate unneeded casts for indices.
10832 bool MadeChange = false;
10833 unsigned PtrSize = TD->getPointerSizeInBits();
10835 gep_type_iterator GTI = gep_type_begin(GEP);
10836 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
10837 I != E; ++I, ++GTI) {
10838 if (!isa<SequentialType>(*GTI)) continue;
10840 // If we are using a wider index than needed for this platform, shrink it
10841 // to what we need. If narrower, sign-extend it to what we need. This
10842 // explicit cast can make subsequent optimizations more obvious.
10843 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
10845 if (OpBits == PtrSize)
10848 Instruction::CastOps Opc =
10849 OpBits > PtrSize ? Instruction::Trunc : Instruction::SExt;
10850 *I = InsertCastBefore(Opc, *I, TD->getIntPtrType(GEP.getContext()), GEP);
10853 if (MadeChange) return &GEP;
10856 // Combine Indices - If the source pointer to this getelementptr instruction
10857 // is a getelementptr instruction, combine the indices of the two
10858 // getelementptr instructions into a single instruction.
10860 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
10861 // Note that if our source is a gep chain itself that we wait for that
10862 // chain to be resolved before we perform this transformation. This
10863 // avoids us creating a TON of code in some cases.
10865 if (GetElementPtrInst *SrcGEP =
10866 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
10867 if (SrcGEP->getNumOperands() == 2)
10868 return 0; // Wait until our source is folded to completion.
10870 SmallVector<Value*, 8> Indices;
10872 // Find out whether the last index in the source GEP is a sequential idx.
10873 bool EndsWithSequential = false;
10874 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
10876 EndsWithSequential = !isa<StructType>(*I);
10878 // Can we combine the two pointer arithmetics offsets?
10879 if (EndsWithSequential) {
10880 // Replace: gep (gep %P, long B), long A, ...
10881 // With: T = long A+B; gep %P, T, ...
10884 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
10885 Value *GO1 = GEP.getOperand(1);
10886 if (SO1 == Constant::getNullValue(SO1->getType())) {
10888 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
10891 // If they aren't the same type, then the input hasn't been processed
10892 // by the loop above yet (which canonicalizes sequential index types to
10893 // intptr_t). Just avoid transforming this until the input has been
10895 if (SO1->getType() != GO1->getType())
10897 Sum = Builder->CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
10900 // Update the GEP in place if possible.
10901 if (Src->getNumOperands() == 2) {
10902 GEP.setOperand(0, Src->getOperand(0));
10903 GEP.setOperand(1, Sum);
10906 Indices.append(Src->op_begin()+1, Src->op_end()-1);
10907 Indices.push_back(Sum);
10908 Indices.append(GEP.op_begin()+2, GEP.op_end());
10909 } else if (isa<Constant>(*GEP.idx_begin()) &&
10910 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
10911 Src->getNumOperands() != 1) {
10912 // Otherwise we can do the fold if the first index of the GEP is a zero
10913 Indices.append(Src->op_begin()+1, Src->op_end());
10914 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
10917 if (!Indices.empty()) {
10918 GetElementPtrInst *NewGEP =
10919 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
10920 Indices.end(), GEP.getName());
10921 if (cast<GEPOperator>(&GEP)->isInBounds() && Src->isInBounds())
10922 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
10927 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
10928 if (Value *X = getBitCastOperand(PtrOp)) {
10929 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
10931 if (HasZeroPointerIndex) {
10932 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
10933 // into : GEP [10 x i8]* X, i32 0, ...
10935 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
10936 // into : GEP i8* X, ...
10938 // This occurs when the program declares an array extern like "int X[];"
10939 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
10940 const PointerType *XTy = cast<PointerType>(X->getType());
10941 if (const ArrayType *CATy =
10942 dyn_cast<ArrayType>(CPTy->getElementType())) {
10943 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
10944 if (CATy->getElementType() == XTy->getElementType()) {
10945 // -> GEP i8* X, ...
10946 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
10947 GetElementPtrInst *NewGEP =
10948 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
10950 if (cast<GEPOperator>(&GEP)->isInBounds())
10951 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
10953 } else if (const ArrayType *XATy =
10954 dyn_cast<ArrayType>(XTy->getElementType())) {
10955 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
10956 if (CATy->getElementType() == XATy->getElementType()) {
10957 // -> GEP [10 x i8]* X, i32 0, ...
10958 // At this point, we know that the cast source type is a pointer
10959 // to an array of the same type as the destination pointer
10960 // array. Because the array type is never stepped over (there
10961 // is a leading zero) we can fold the cast into this GEP.
10962 GEP.setOperand(0, X);
10967 } else if (GEP.getNumOperands() == 2) {
10968 // Transform things like:
10969 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
10970 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
10971 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
10972 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
10973 if (TD && isa<ArrayType>(SrcElTy) &&
10974 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
10975 TD->getTypeAllocSize(ResElTy)) {
10977 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
10978 Idx[1] = GEP.getOperand(1);
10980 Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
10981 if (cast<GEPOperator>(&GEP)->isInBounds())
10982 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
10983 // V and GEP are both pointer types --> BitCast
10984 return new BitCastInst(NewGEP, GEP.getType());
10987 // Transform things like:
10988 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
10989 // (where tmp = 8*tmp2) into:
10990 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
10992 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::getInt8Ty(*Context)) {
10993 uint64_t ArrayEltSize =
10994 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
10996 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
10997 // allow either a mul, shift, or constant here.
10999 ConstantInt *Scale = 0;
11000 if (ArrayEltSize == 1) {
11001 NewIdx = GEP.getOperand(1);
11002 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11003 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11004 NewIdx = ConstantInt::get(CI->getType(), 1);
11006 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11007 if (Inst->getOpcode() == Instruction::Shl &&
11008 isa<ConstantInt>(Inst->getOperand(1))) {
11009 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11010 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11011 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11013 NewIdx = Inst->getOperand(0);
11014 } else if (Inst->getOpcode() == Instruction::Mul &&
11015 isa<ConstantInt>(Inst->getOperand(1))) {
11016 Scale = cast<ConstantInt>(Inst->getOperand(1));
11017 NewIdx = Inst->getOperand(0);
11021 // If the index will be to exactly the right offset with the scale taken
11022 // out, perform the transformation. Note, we don't know whether Scale is
11023 // signed or not. We'll use unsigned version of division/modulo
11024 // operation after making sure Scale doesn't have the sign bit set.
11025 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11026 Scale->getZExtValue() % ArrayEltSize == 0) {
11027 Scale = ConstantInt::get(Scale->getType(),
11028 Scale->getZExtValue() / ArrayEltSize);
11029 if (Scale->getZExtValue() != 1) {
11030 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11032 NewIdx = Builder->CreateMul(NewIdx, C, "idxscale");
11035 // Insert the new GEP instruction.
11037 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11039 Value *NewGEP = Builder->CreateGEP(X, Idx, Idx + 2, GEP.getName());
11040 if (cast<GEPOperator>(&GEP)->isInBounds())
11041 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11042 // The NewGEP must be pointer typed, so must the old one -> BitCast
11043 return new BitCastInst(NewGEP, GEP.getType());
11049 /// See if we can simplify:
11050 /// X = bitcast A* to B*
11051 /// Y = gep X, <...constant indices...>
11052 /// into a gep of the original struct. This is important for SROA and alias
11053 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11054 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11056 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11057 // Determine how much the GEP moves the pointer. We are guaranteed to get
11058 // a constant back from EmitGEPOffset.
11059 ConstantInt *OffsetV =
11060 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11061 int64_t Offset = OffsetV->getSExtValue();
11063 // If this GEP instruction doesn't move the pointer, just replace the GEP
11064 // with a bitcast of the real input to the dest type.
11066 // If the bitcast is of an allocation, and the allocation will be
11067 // converted to match the type of the cast, don't touch this.
11068 if (isa<AllocationInst>(BCI->getOperand(0))) {
11069 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11070 if (Instruction *I = visitBitCast(*BCI)) {
11073 BCI->getParent()->getInstList().insert(BCI, I);
11074 ReplaceInstUsesWith(*BCI, I);
11079 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11082 // Otherwise, if the offset is non-zero, we need to find out if there is a
11083 // field at Offset in 'A's type. If so, we can pull the cast through the
11085 SmallVector<Value*, 8> NewIndices;
11087 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11088 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11089 Value *NGEP = Builder->CreateGEP(BCI->getOperand(0), NewIndices.begin(),
11091 if (cast<GEPOperator>(&GEP)->isInBounds())
11092 cast<GEPOperator>(NGEP)->setIsInBounds(true);
11094 if (NGEP->getType() == GEP.getType())
11095 return ReplaceInstUsesWith(GEP, NGEP);
11096 NGEP->takeName(&GEP);
11097 return new BitCastInst(NGEP, GEP.getType());
11105 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11106 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11107 if (AI.isArrayAllocation()) { // Check C != 1
11108 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11109 const Type *NewTy =
11110 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11111 AllocationInst *New = 0;
11113 // Create and insert the replacement instruction...
11114 if (isa<MallocInst>(AI))
11115 New = Builder->CreateMalloc(NewTy, 0, AI.getName());
11117 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11118 New = Builder->CreateAlloca(NewTy, 0, AI.getName());
11120 New->setAlignment(AI.getAlignment());
11122 // Scan to the end of the allocation instructions, to skip over a block of
11123 // allocas if possible...also skip interleaved debug info
11125 BasicBlock::iterator It = New;
11126 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11128 // Now that I is pointing to the first non-allocation-inst in the block,
11129 // insert our getelementptr instruction...
11131 Value *NullIdx = Constant::getNullValue(Type::getInt32Ty(*Context));
11135 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11136 New->getName()+".sub", It);
11137 cast<GEPOperator>(V)->setIsInBounds(true);
11139 // Now make everything use the getelementptr instead of the original
11141 return ReplaceInstUsesWith(AI, V);
11142 } else if (isa<UndefValue>(AI.getArraySize())) {
11143 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11147 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11148 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11149 // Note that we only do this for alloca's, because malloc should allocate
11150 // and return a unique pointer, even for a zero byte allocation.
11151 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11152 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11154 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11155 if (AI.getAlignment() == 0)
11156 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11162 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11163 Value *Op = FI.getOperand(0);
11165 // free undef -> unreachable.
11166 if (isa<UndefValue>(Op)) {
11167 // Insert a new store to null because we cannot modify the CFG here.
11168 new StoreInst(ConstantInt::getTrue(*Context),
11169 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))), &FI);
11170 return EraseInstFromFunction(FI);
11173 // If we have 'free null' delete the instruction. This can happen in stl code
11174 // when lots of inlining happens.
11175 if (isa<ConstantPointerNull>(Op))
11176 return EraseInstFromFunction(FI);
11178 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11179 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11180 FI.setOperand(0, CI->getOperand(0));
11184 // Change free (gep X, 0,0,0,0) into free(X)
11185 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11186 if (GEPI->hasAllZeroIndices()) {
11187 Worklist.Add(GEPI);
11188 FI.setOperand(0, GEPI->getOperand(0));
11193 // Change free(malloc) into nothing, if the malloc has a single use.
11194 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11195 if (MI->hasOneUse()) {
11196 EraseInstFromFunction(FI);
11197 return EraseInstFromFunction(*MI);
11204 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11205 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11206 const TargetData *TD) {
11207 User *CI = cast<User>(LI.getOperand(0));
11208 Value *CastOp = CI->getOperand(0);
11209 LLVMContext *Context = IC.getContext();
11212 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11213 // Instead of loading constant c string, use corresponding integer value
11214 // directly if string length is small enough.
11216 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11217 unsigned len = Str.length();
11218 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11219 unsigned numBits = Ty->getPrimitiveSizeInBits();
11220 // Replace LI with immediate integer store.
11221 if ((numBits >> 3) == len + 1) {
11222 APInt StrVal(numBits, 0);
11223 APInt SingleChar(numBits, 0);
11224 if (TD->isLittleEndian()) {
11225 for (signed i = len-1; i >= 0; i--) {
11226 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11227 StrVal = (StrVal << 8) | SingleChar;
11230 for (unsigned i = 0; i < len; i++) {
11231 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11232 StrVal = (StrVal << 8) | SingleChar;
11234 // Append NULL at the end.
11236 StrVal = (StrVal << 8) | SingleChar;
11238 Value *NL = ConstantInt::get(*Context, StrVal);
11239 return IC.ReplaceInstUsesWith(LI, NL);
11245 const PointerType *DestTy = cast<PointerType>(CI->getType());
11246 const Type *DestPTy = DestTy->getElementType();
11247 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11249 // If the address spaces don't match, don't eliminate the cast.
11250 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11253 const Type *SrcPTy = SrcTy->getElementType();
11255 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11256 isa<VectorType>(DestPTy)) {
11257 // If the source is an array, the code below will not succeed. Check to
11258 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11260 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11261 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11262 if (ASrcTy->getNumElements() != 0) {
11264 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::getInt32Ty(*Context));
11265 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11266 SrcTy = cast<PointerType>(CastOp->getType());
11267 SrcPTy = SrcTy->getElementType();
11270 if (IC.getTargetData() &&
11271 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11272 isa<VectorType>(SrcPTy)) &&
11273 // Do not allow turning this into a load of an integer, which is then
11274 // casted to a pointer, this pessimizes pointer analysis a lot.
11275 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11276 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
11277 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
11279 // Okay, we are casting from one integer or pointer type to another of
11280 // the same size. Instead of casting the pointer before the load, cast
11281 // the result of the loaded value.
11283 IC.Builder->CreateLoad(CastOp, LI.isVolatile(), CI->getName());
11284 // Now cast the result of the load.
11285 return new BitCastInst(NewLoad, LI.getType());
11292 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11293 Value *Op = LI.getOperand(0);
11295 // Attempt to improve the alignment.
11297 unsigned KnownAlign =
11298 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11300 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11301 LI.getAlignment()))
11302 LI.setAlignment(KnownAlign);
11305 // load (cast X) --> cast (load X) iff safe
11306 if (isa<CastInst>(Op))
11307 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11310 // None of the following transforms are legal for volatile loads.
11311 if (LI.isVolatile()) return 0;
11313 // Do really simple store-to-load forwarding and load CSE, to catch cases
11314 // where there are several consequtive memory accesses to the same location,
11315 // separated by a few arithmetic operations.
11316 BasicBlock::iterator BBI = &LI;
11317 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11318 return ReplaceInstUsesWith(LI, AvailableVal);
11320 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11321 const Value *GEPI0 = GEPI->getOperand(0);
11322 // TODO: Consider a target hook for valid address spaces for this xform.
11323 if (isa<ConstantPointerNull>(GEPI0) &&
11324 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11325 // Insert a new store to null instruction before the load to indicate
11326 // that this code is not reachable. We do this instead of inserting
11327 // an unreachable instruction directly because we cannot modify the
11329 new StoreInst(UndefValue::get(LI.getType()),
11330 Constant::getNullValue(Op->getType()), &LI);
11331 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11335 if (Constant *C = dyn_cast<Constant>(Op)) {
11336 // load null/undef -> undef
11337 // TODO: Consider a target hook for valid address spaces for this xform.
11338 if (isa<UndefValue>(C) || (C->isNullValue() &&
11339 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11340 // Insert a new store to null instruction before the load to indicate that
11341 // this code is not reachable. We do this instead of inserting an
11342 // unreachable instruction directly because we cannot modify the CFG.
11343 new StoreInst(UndefValue::get(LI.getType()),
11344 Constant::getNullValue(Op->getType()), &LI);
11345 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11348 // Instcombine load (constant global) into the value loaded.
11349 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11350 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11351 return ReplaceInstUsesWith(LI, GV->getInitializer());
11353 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11354 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11355 if (CE->getOpcode() == Instruction::GetElementPtr) {
11356 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11357 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11359 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11361 return ReplaceInstUsesWith(LI, V);
11362 if (CE->getOperand(0)->isNullValue()) {
11363 // Insert a new store to null instruction before the load to indicate
11364 // that this code is not reachable. We do this instead of inserting
11365 // an unreachable instruction directly because we cannot modify the
11367 new StoreInst(UndefValue::get(LI.getType()),
11368 Constant::getNullValue(Op->getType()), &LI);
11369 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11372 } else if (CE->isCast()) {
11373 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11379 // If this load comes from anywhere in a constant global, and if the global
11380 // is all undef or zero, we know what it loads.
11381 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11382 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11383 if (GV->getInitializer()->isNullValue())
11384 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11385 else if (isa<UndefValue>(GV->getInitializer()))
11386 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11390 if (Op->hasOneUse()) {
11391 // Change select and PHI nodes to select values instead of addresses: this
11392 // helps alias analysis out a lot, allows many others simplifications, and
11393 // exposes redundancy in the code.
11395 // Note that we cannot do the transformation unless we know that the
11396 // introduced loads cannot trap! Something like this is valid as long as
11397 // the condition is always false: load (select bool %C, int* null, int* %G),
11398 // but it would not be valid if we transformed it to load from null
11399 // unconditionally.
11401 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11402 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11403 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11404 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11405 Value *V1 = Builder->CreateLoad(SI->getOperand(1),
11406 SI->getOperand(1)->getName()+".val");
11407 Value *V2 = Builder->CreateLoad(SI->getOperand(2),
11408 SI->getOperand(2)->getName()+".val");
11409 return SelectInst::Create(SI->getCondition(), V1, V2);
11412 // load (select (cond, null, P)) -> load P
11413 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11414 if (C->isNullValue()) {
11415 LI.setOperand(0, SI->getOperand(2));
11419 // load (select (cond, P, null)) -> load P
11420 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11421 if (C->isNullValue()) {
11422 LI.setOperand(0, SI->getOperand(1));
11430 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11431 /// when possible. This makes it generally easy to do alias analysis and/or
11432 /// SROA/mem2reg of the memory object.
11433 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11434 User *CI = cast<User>(SI.getOperand(1));
11435 Value *CastOp = CI->getOperand(0);
11437 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11438 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11439 if (SrcTy == 0) return 0;
11441 const Type *SrcPTy = SrcTy->getElementType();
11443 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11446 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11447 /// to its first element. This allows us to handle things like:
11448 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11449 /// on 32-bit hosts.
11450 SmallVector<Value*, 4> NewGEPIndices;
11452 // If the source is an array, the code below will not succeed. Check to
11453 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11455 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11456 // Index through pointer.
11457 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(*IC.getContext()));
11458 NewGEPIndices.push_back(Zero);
11461 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11462 if (!STy->getNumElements()) /* Struct can be empty {} */
11464 NewGEPIndices.push_back(Zero);
11465 SrcPTy = STy->getElementType(0);
11466 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11467 NewGEPIndices.push_back(Zero);
11468 SrcPTy = ATy->getElementType();
11474 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11477 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11480 // If the pointers point into different address spaces or if they point to
11481 // values with different sizes, we can't do the transformation.
11482 if (!IC.getTargetData() ||
11483 SrcTy->getAddressSpace() !=
11484 cast<PointerType>(CI->getType())->getAddressSpace() ||
11485 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
11486 IC.getTargetData()->getTypeSizeInBits(DestPTy))
11489 // Okay, we are casting from one integer or pointer type to another of
11490 // the same size. Instead of casting the pointer before
11491 // the store, cast the value to be stored.
11493 Value *SIOp0 = SI.getOperand(0);
11494 Instruction::CastOps opcode = Instruction::BitCast;
11495 const Type* CastSrcTy = SIOp0->getType();
11496 const Type* CastDstTy = SrcPTy;
11497 if (isa<PointerType>(CastDstTy)) {
11498 if (CastSrcTy->isInteger())
11499 opcode = Instruction::IntToPtr;
11500 } else if (isa<IntegerType>(CastDstTy)) {
11501 if (isa<PointerType>(SIOp0->getType()))
11502 opcode = Instruction::PtrToInt;
11505 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11506 // emit a GEP to index into its first field.
11507 if (!NewGEPIndices.empty()) {
11508 CastOp = IC.Builder->CreateGEP(CastOp, NewGEPIndices.begin(),
11509 NewGEPIndices.end());
11510 cast<GEPOperator>(CastOp)->setIsInBounds(true);
11513 NewCast = IC.Builder->CreateCast(opcode, SIOp0, CastDstTy,
11514 SIOp0->getName()+".c");
11515 return new StoreInst(NewCast, CastOp);
11518 /// equivalentAddressValues - Test if A and B will obviously have the same
11519 /// value. This includes recognizing that %t0 and %t1 will have the same
11520 /// value in code like this:
11521 /// %t0 = getelementptr \@a, 0, 3
11522 /// store i32 0, i32* %t0
11523 /// %t1 = getelementptr \@a, 0, 3
11524 /// %t2 = load i32* %t1
11526 static bool equivalentAddressValues(Value *A, Value *B) {
11527 // Test if the values are trivially equivalent.
11528 if (A == B) return true;
11530 // Test if the values come form identical arithmetic instructions.
11531 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
11532 // its only used to compare two uses within the same basic block, which
11533 // means that they'll always either have the same value or one of them
11534 // will have an undefined value.
11535 if (isa<BinaryOperator>(A) ||
11536 isa<CastInst>(A) ||
11538 isa<GetElementPtrInst>(A))
11539 if (Instruction *BI = dyn_cast<Instruction>(B))
11540 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
11543 // Otherwise they may not be equivalent.
11547 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11548 // return the llvm.dbg.declare.
11549 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11550 if (!V->hasNUses(2))
11552 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11554 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11556 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11557 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11564 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11565 Value *Val = SI.getOperand(0);
11566 Value *Ptr = SI.getOperand(1);
11568 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11569 EraseInstFromFunction(SI);
11574 // If the RHS is an alloca with a single use, zapify the store, making the
11576 // If the RHS is an alloca with a two uses, the other one being a
11577 // llvm.dbg.declare, zapify the store and the declare, making the
11578 // alloca dead. We must do this to prevent declare's from affecting
11580 if (!SI.isVolatile()) {
11581 if (Ptr->hasOneUse()) {
11582 if (isa<AllocaInst>(Ptr)) {
11583 EraseInstFromFunction(SI);
11587 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11588 if (isa<AllocaInst>(GEP->getOperand(0))) {
11589 if (GEP->getOperand(0)->hasOneUse()) {
11590 EraseInstFromFunction(SI);
11594 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11595 EraseInstFromFunction(*DI);
11596 EraseInstFromFunction(SI);
11603 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11604 EraseInstFromFunction(*DI);
11605 EraseInstFromFunction(SI);
11611 // Attempt to improve the alignment.
11613 unsigned KnownAlign =
11614 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11616 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11617 SI.getAlignment()))
11618 SI.setAlignment(KnownAlign);
11621 // Do really simple DSE, to catch cases where there are several consecutive
11622 // stores to the same location, separated by a few arithmetic operations. This
11623 // situation often occurs with bitfield accesses.
11624 BasicBlock::iterator BBI = &SI;
11625 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11628 // Don't count debug info directives, lest they affect codegen,
11629 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11630 // It is necessary for correctness to skip those that feed into a
11631 // llvm.dbg.declare, as these are not present when debugging is off.
11632 if (isa<DbgInfoIntrinsic>(BBI) ||
11633 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11638 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11639 // Prev store isn't volatile, and stores to the same location?
11640 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11641 SI.getOperand(1))) {
11644 EraseInstFromFunction(*PrevSI);
11650 // If this is a load, we have to stop. However, if the loaded value is from
11651 // the pointer we're loading and is producing the pointer we're storing,
11652 // then *this* store is dead (X = load P; store X -> P).
11653 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11654 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11655 !SI.isVolatile()) {
11656 EraseInstFromFunction(SI);
11660 // Otherwise, this is a load from some other location. Stores before it
11661 // may not be dead.
11665 // Don't skip over loads or things that can modify memory.
11666 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11671 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11673 // store X, null -> turns into 'unreachable' in SimplifyCFG
11674 if (isa<ConstantPointerNull>(Ptr) &&
11675 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11676 if (!isa<UndefValue>(Val)) {
11677 SI.setOperand(0, UndefValue::get(Val->getType()));
11678 if (Instruction *U = dyn_cast<Instruction>(Val))
11679 Worklist.Add(U); // Dropped a use.
11682 return 0; // Do not modify these!
11685 // store undef, Ptr -> noop
11686 if (isa<UndefValue>(Val)) {
11687 EraseInstFromFunction(SI);
11692 // If the pointer destination is a cast, see if we can fold the cast into the
11694 if (isa<CastInst>(Ptr))
11695 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11697 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11699 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11703 // If this store is the last instruction in the basic block (possibly
11704 // excepting debug info instructions and the pointer bitcasts that feed
11705 // into them), and if the block ends with an unconditional branch, try
11706 // to move it to the successor block.
11710 } while (isa<DbgInfoIntrinsic>(BBI) ||
11711 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11712 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11713 if (BI->isUnconditional())
11714 if (SimplifyStoreAtEndOfBlock(SI))
11715 return 0; // xform done!
11720 /// SimplifyStoreAtEndOfBlock - Turn things like:
11721 /// if () { *P = v1; } else { *P = v2 }
11722 /// into a phi node with a store in the successor.
11724 /// Simplify things like:
11725 /// *P = v1; if () { *P = v2; }
11726 /// into a phi node with a store in the successor.
11728 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11729 BasicBlock *StoreBB = SI.getParent();
11731 // Check to see if the successor block has exactly two incoming edges. If
11732 // so, see if the other predecessor contains a store to the same location.
11733 // if so, insert a PHI node (if needed) and move the stores down.
11734 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11736 // Determine whether Dest has exactly two predecessors and, if so, compute
11737 // the other predecessor.
11738 pred_iterator PI = pred_begin(DestBB);
11739 BasicBlock *OtherBB = 0;
11740 if (*PI != StoreBB)
11743 if (PI == pred_end(DestBB))
11746 if (*PI != StoreBB) {
11751 if (++PI != pred_end(DestBB))
11754 // Bail out if all the relevant blocks aren't distinct (this can happen,
11755 // for example, if SI is in an infinite loop)
11756 if (StoreBB == DestBB || OtherBB == DestBB)
11759 // Verify that the other block ends in a branch and is not otherwise empty.
11760 BasicBlock::iterator BBI = OtherBB->getTerminator();
11761 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11762 if (!OtherBr || BBI == OtherBB->begin())
11765 // If the other block ends in an unconditional branch, check for the 'if then
11766 // else' case. there is an instruction before the branch.
11767 StoreInst *OtherStore = 0;
11768 if (OtherBr->isUnconditional()) {
11770 // Skip over debugging info.
11771 while (isa<DbgInfoIntrinsic>(BBI) ||
11772 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11773 if (BBI==OtherBB->begin())
11777 // If this isn't a store, or isn't a store to the same location, bail out.
11778 OtherStore = dyn_cast<StoreInst>(BBI);
11779 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
11782 // Otherwise, the other block ended with a conditional branch. If one of the
11783 // destinations is StoreBB, then we have the if/then case.
11784 if (OtherBr->getSuccessor(0) != StoreBB &&
11785 OtherBr->getSuccessor(1) != StoreBB)
11788 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
11789 // if/then triangle. See if there is a store to the same ptr as SI that
11790 // lives in OtherBB.
11792 // Check to see if we find the matching store.
11793 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
11794 if (OtherStore->getOperand(1) != SI.getOperand(1))
11798 // If we find something that may be using or overwriting the stored
11799 // value, or if we run out of instructions, we can't do the xform.
11800 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
11801 BBI == OtherBB->begin())
11805 // In order to eliminate the store in OtherBr, we have to
11806 // make sure nothing reads or overwrites the stored value in
11808 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
11809 // FIXME: This should really be AA driven.
11810 if (I->mayReadFromMemory() || I->mayWriteToMemory())
11815 // Insert a PHI node now if we need it.
11816 Value *MergedVal = OtherStore->getOperand(0);
11817 if (MergedVal != SI.getOperand(0)) {
11818 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
11819 PN->reserveOperandSpace(2);
11820 PN->addIncoming(SI.getOperand(0), SI.getParent());
11821 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
11822 MergedVal = InsertNewInstBefore(PN, DestBB->front());
11825 // Advance to a place where it is safe to insert the new store and
11827 BBI = DestBB->getFirstNonPHI();
11828 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
11829 OtherStore->isVolatile()), *BBI);
11831 // Nuke the old stores.
11832 EraseInstFromFunction(SI);
11833 EraseInstFromFunction(*OtherStore);
11839 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
11840 // Change br (not X), label True, label False to: br X, label False, True
11842 BasicBlock *TrueDest;
11843 BasicBlock *FalseDest;
11844 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
11845 !isa<Constant>(X)) {
11846 // Swap Destinations and condition...
11847 BI.setCondition(X);
11848 BI.setSuccessor(0, FalseDest);
11849 BI.setSuccessor(1, TrueDest);
11853 // Cannonicalize fcmp_one -> fcmp_oeq
11854 FCmpInst::Predicate FPred; Value *Y;
11855 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
11856 TrueDest, FalseDest)) &&
11857 BI.getCondition()->hasOneUse())
11858 if (FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
11859 FPred == FCmpInst::FCMP_OGE) {
11860 FCmpInst *Cond = cast<FCmpInst>(BI.getCondition());
11861 Cond->setPredicate(FCmpInst::getInversePredicate(FPred));
11863 // Swap Destinations and condition.
11864 BI.setSuccessor(0, FalseDest);
11865 BI.setSuccessor(1, TrueDest);
11866 Worklist.Add(Cond);
11870 // Cannonicalize icmp_ne -> icmp_eq
11871 ICmpInst::Predicate IPred;
11872 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
11873 TrueDest, FalseDest)) &&
11874 BI.getCondition()->hasOneUse())
11875 if (IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
11876 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
11877 IPred == ICmpInst::ICMP_SGE) {
11878 ICmpInst *Cond = cast<ICmpInst>(BI.getCondition());
11879 Cond->setPredicate(ICmpInst::getInversePredicate(IPred));
11880 // Swap Destinations and condition.
11881 BI.setSuccessor(0, FalseDest);
11882 BI.setSuccessor(1, TrueDest);
11883 Worklist.Add(Cond);
11890 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
11891 Value *Cond = SI.getCondition();
11892 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
11893 if (I->getOpcode() == Instruction::Add)
11894 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
11895 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
11896 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
11898 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
11900 SI.setOperand(0, I->getOperand(0));
11908 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
11909 Value *Agg = EV.getAggregateOperand();
11911 if (!EV.hasIndices())
11912 return ReplaceInstUsesWith(EV, Agg);
11914 if (Constant *C = dyn_cast<Constant>(Agg)) {
11915 if (isa<UndefValue>(C))
11916 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
11918 if (isa<ConstantAggregateZero>(C))
11919 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
11921 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
11922 // Extract the element indexed by the first index out of the constant
11923 Value *V = C->getOperand(*EV.idx_begin());
11924 if (EV.getNumIndices() > 1)
11925 // Extract the remaining indices out of the constant indexed by the
11927 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
11929 return ReplaceInstUsesWith(EV, V);
11931 return 0; // Can't handle other constants
11933 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
11934 // We're extracting from an insertvalue instruction, compare the indices
11935 const unsigned *exti, *exte, *insi, *inse;
11936 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
11937 exte = EV.idx_end(), inse = IV->idx_end();
11938 exti != exte && insi != inse;
11940 if (*insi != *exti)
11941 // The insert and extract both reference distinctly different elements.
11942 // This means the extract is not influenced by the insert, and we can
11943 // replace the aggregate operand of the extract with the aggregate
11944 // operand of the insert. i.e., replace
11945 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11946 // %E = extractvalue { i32, { i32 } } %I, 0
11948 // %E = extractvalue { i32, { i32 } } %A, 0
11949 return ExtractValueInst::Create(IV->getAggregateOperand(),
11950 EV.idx_begin(), EV.idx_end());
11952 if (exti == exte && insi == inse)
11953 // Both iterators are at the end: Index lists are identical. Replace
11954 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11955 // %C = extractvalue { i32, { i32 } } %B, 1, 0
11957 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
11958 if (exti == exte) {
11959 // The extract list is a prefix of the insert list. i.e. replace
11960 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
11961 // %E = extractvalue { i32, { i32 } } %I, 1
11963 // %X = extractvalue { i32, { i32 } } %A, 1
11964 // %E = insertvalue { i32 } %X, i32 42, 0
11965 // by switching the order of the insert and extract (though the
11966 // insertvalue should be left in, since it may have other uses).
11967 Value *NewEV = Builder->CreateExtractValue(IV->getAggregateOperand(),
11968 EV.idx_begin(), EV.idx_end());
11969 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
11973 // The insert list is a prefix of the extract list
11974 // We can simply remove the common indices from the extract and make it
11975 // operate on the inserted value instead of the insertvalue result.
11977 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
11978 // %E = extractvalue { i32, { i32 } } %I, 1, 0
11980 // %E extractvalue { i32 } { i32 42 }, 0
11981 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
11984 // Can't simplify extracts from other values. Note that nested extracts are
11985 // already simplified implicitely by the above (extract ( extract (insert) )
11986 // will be translated into extract ( insert ( extract ) ) first and then just
11987 // the value inserted, if appropriate).
11991 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
11992 /// is to leave as a vector operation.
11993 static bool CheapToScalarize(Value *V, bool isConstant) {
11994 if (isa<ConstantAggregateZero>(V))
11996 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
11997 if (isConstant) return true;
11998 // If all elts are the same, we can extract.
11999 Constant *Op0 = C->getOperand(0);
12000 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12001 if (C->getOperand(i) != Op0)
12005 Instruction *I = dyn_cast<Instruction>(V);
12006 if (!I) return false;
12008 // Insert element gets simplified to the inserted element or is deleted if
12009 // this is constant idx extract element and its a constant idx insertelt.
12010 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12011 isa<ConstantInt>(I->getOperand(2)))
12013 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12015 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12016 if (BO->hasOneUse() &&
12017 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12018 CheapToScalarize(BO->getOperand(1), isConstant)))
12020 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12021 if (CI->hasOneUse() &&
12022 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12023 CheapToScalarize(CI->getOperand(1), isConstant)))
12029 /// Read and decode a shufflevector mask.
12031 /// It turns undef elements into values that are larger than the number of
12032 /// elements in the input.
12033 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12034 unsigned NElts = SVI->getType()->getNumElements();
12035 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12036 return std::vector<unsigned>(NElts, 0);
12037 if (isa<UndefValue>(SVI->getOperand(2)))
12038 return std::vector<unsigned>(NElts, 2*NElts);
12040 std::vector<unsigned> Result;
12041 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12042 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12043 if (isa<UndefValue>(*i))
12044 Result.push_back(NElts*2); // undef -> 8
12046 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12050 /// FindScalarElement - Given a vector and an element number, see if the scalar
12051 /// value is already around as a register, for example if it were inserted then
12052 /// extracted from the vector.
12053 static Value *FindScalarElement(Value *V, unsigned EltNo,
12054 LLVMContext *Context) {
12055 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12056 const VectorType *PTy = cast<VectorType>(V->getType());
12057 unsigned Width = PTy->getNumElements();
12058 if (EltNo >= Width) // Out of range access.
12059 return UndefValue::get(PTy->getElementType());
12061 if (isa<UndefValue>(V))
12062 return UndefValue::get(PTy->getElementType());
12063 else if (isa<ConstantAggregateZero>(V))
12064 return Constant::getNullValue(PTy->getElementType());
12065 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12066 return CP->getOperand(EltNo);
12067 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12068 // If this is an insert to a variable element, we don't know what it is.
12069 if (!isa<ConstantInt>(III->getOperand(2)))
12071 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12073 // If this is an insert to the element we are looking for, return the
12075 if (EltNo == IIElt)
12076 return III->getOperand(1);
12078 // Otherwise, the insertelement doesn't modify the value, recurse on its
12080 return FindScalarElement(III->getOperand(0), EltNo, Context);
12081 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12082 unsigned LHSWidth =
12083 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12084 unsigned InEl = getShuffleMask(SVI)[EltNo];
12085 if (InEl < LHSWidth)
12086 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12087 else if (InEl < LHSWidth*2)
12088 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12090 return UndefValue::get(PTy->getElementType());
12093 // Otherwise, we don't know.
12097 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12098 // If vector val is undef, replace extract with scalar undef.
12099 if (isa<UndefValue>(EI.getOperand(0)))
12100 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12102 // If vector val is constant 0, replace extract with scalar 0.
12103 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12104 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12106 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12107 // If vector val is constant with all elements the same, replace EI with
12108 // that element. When the elements are not identical, we cannot replace yet
12109 // (we do that below, but only when the index is constant).
12110 Constant *op0 = C->getOperand(0);
12111 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12112 if (C->getOperand(i) != op0) {
12117 return ReplaceInstUsesWith(EI, op0);
12120 // If extracting a specified index from the vector, see if we can recursively
12121 // find a previously computed scalar that was inserted into the vector.
12122 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12123 unsigned IndexVal = IdxC->getZExtValue();
12124 unsigned VectorWidth =
12125 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12127 // If this is extracting an invalid index, turn this into undef, to avoid
12128 // crashing the code below.
12129 if (IndexVal >= VectorWidth)
12130 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12132 // This instruction only demands the single element from the input vector.
12133 // If the input vector has a single use, simplify it based on this use
12135 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12136 APInt UndefElts(VectorWidth, 0);
12137 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12138 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12139 DemandedMask, UndefElts)) {
12140 EI.setOperand(0, V);
12145 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12146 return ReplaceInstUsesWith(EI, Elt);
12148 // If the this extractelement is directly using a bitcast from a vector of
12149 // the same number of elements, see if we can find the source element from
12150 // it. In this case, we will end up needing to bitcast the scalars.
12151 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12152 if (const VectorType *VT =
12153 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12154 if (VT->getNumElements() == VectorWidth)
12155 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12156 IndexVal, Context))
12157 return new BitCastInst(Elt, EI.getType());
12161 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12162 if (I->hasOneUse()) {
12163 // Push extractelement into predecessor operation if legal and
12164 // profitable to do so
12165 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12166 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12167 if (CheapToScalarize(BO, isConstantElt)) {
12169 Builder->CreateExtractElement(BO->getOperand(0), EI.getOperand(1),
12170 EI.getName()+".lhs");
12172 Builder->CreateExtractElement(BO->getOperand(1), EI.getOperand(1),
12173 EI.getName()+".rhs");
12174 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12176 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
12177 unsigned AS = LI->getPointerAddressSpace();
12178 Value *Ptr = Builder->CreateBitCast(I->getOperand(0),
12179 PointerType::get(EI.getType(), AS),
12180 I->getOperand(0)->getName());
12182 Builder->CreateGEP(Ptr, EI.getOperand(1), I->getName()+".gep");
12183 cast<GEPOperator>(GEP)->setIsInBounds(true);
12185 LoadInst *Load = Builder->CreateLoad(GEP, "tmp");
12187 // Make sure the Load goes before the load instruction in the source,
12188 // not wherever the extract happens to be.
12189 if (Instruction *P = dyn_cast<Instruction>(Ptr))
12191 if (Instruction *G = dyn_cast<Instruction>(GEP))
12193 Load->moveBefore(I);
12195 return ReplaceInstUsesWith(EI, Load);
12198 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12199 // Extracting the inserted element?
12200 if (IE->getOperand(2) == EI.getOperand(1))
12201 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12202 // If the inserted and extracted elements are constants, they must not
12203 // be the same value, extract from the pre-inserted value instead.
12204 if (isa<Constant>(IE->getOperand(2)) && isa<Constant>(EI.getOperand(1))) {
12205 Worklist.AddValue(EI.getOperand(0));
12206 EI.setOperand(0, IE->getOperand(0));
12209 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12210 // If this is extracting an element from a shufflevector, figure out where
12211 // it came from and extract from the appropriate input element instead.
12212 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12213 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12215 unsigned LHSWidth =
12216 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12218 if (SrcIdx < LHSWidth)
12219 Src = SVI->getOperand(0);
12220 else if (SrcIdx < LHSWidth*2) {
12221 SrcIdx -= LHSWidth;
12222 Src = SVI->getOperand(1);
12224 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12226 return ExtractElementInst::Create(Src,
12227 ConstantInt::get(Type::getInt32Ty(*Context), SrcIdx,
12231 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12236 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12237 /// elements from either LHS or RHS, return the shuffle mask and true.
12238 /// Otherwise, return false.
12239 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12240 std::vector<Constant*> &Mask,
12241 LLVMContext *Context) {
12242 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12243 "Invalid CollectSingleShuffleElements");
12244 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12246 if (isa<UndefValue>(V)) {
12247 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12249 } else if (V == LHS) {
12250 for (unsigned i = 0; i != NumElts; ++i)
12251 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12253 } else if (V == RHS) {
12254 for (unsigned i = 0; i != NumElts; ++i)
12255 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i+NumElts));
12257 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12258 // If this is an insert of an extract from some other vector, include it.
12259 Value *VecOp = IEI->getOperand(0);
12260 Value *ScalarOp = IEI->getOperand(1);
12261 Value *IdxOp = IEI->getOperand(2);
12263 if (!isa<ConstantInt>(IdxOp))
12265 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12267 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12268 // Okay, we can handle this if the vector we are insertinting into is
12269 // transitively ok.
12270 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12271 // If so, update the mask to reflect the inserted undef.
12272 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(*Context));
12275 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12276 if (isa<ConstantInt>(EI->getOperand(1)) &&
12277 EI->getOperand(0)->getType() == V->getType()) {
12278 unsigned ExtractedIdx =
12279 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12281 // This must be extracting from either LHS or RHS.
12282 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12283 // Okay, we can handle this if the vector we are insertinting into is
12284 // transitively ok.
12285 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12286 // If so, update the mask to reflect the inserted value.
12287 if (EI->getOperand(0) == LHS) {
12288 Mask[InsertedIdx % NumElts] =
12289 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12291 assert(EI->getOperand(0) == RHS);
12292 Mask[InsertedIdx % NumElts] =
12293 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx+NumElts);
12302 // TODO: Handle shufflevector here!
12307 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12308 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12309 /// that computes V and the LHS value of the shuffle.
12310 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12311 Value *&RHS, LLVMContext *Context) {
12312 assert(isa<VectorType>(V->getType()) &&
12313 (RHS == 0 || V->getType() == RHS->getType()) &&
12314 "Invalid shuffle!");
12315 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12317 if (isa<UndefValue>(V)) {
12318 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12320 } else if (isa<ConstantAggregateZero>(V)) {
12321 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(*Context), 0));
12323 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12324 // If this is an insert of an extract from some other vector, include it.
12325 Value *VecOp = IEI->getOperand(0);
12326 Value *ScalarOp = IEI->getOperand(1);
12327 Value *IdxOp = IEI->getOperand(2);
12329 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12330 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12331 EI->getOperand(0)->getType() == V->getType()) {
12332 unsigned ExtractedIdx =
12333 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12334 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12336 // Either the extracted from or inserted into vector must be RHSVec,
12337 // otherwise we'd end up with a shuffle of three inputs.
12338 if (EI->getOperand(0) == RHS || RHS == 0) {
12339 RHS = EI->getOperand(0);
12340 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12341 Mask[InsertedIdx % NumElts] =
12342 ConstantInt::get(Type::getInt32Ty(*Context), NumElts+ExtractedIdx);
12346 if (VecOp == RHS) {
12347 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12349 // Everything but the extracted element is replaced with the RHS.
12350 for (unsigned i = 0; i != NumElts; ++i) {
12351 if (i != InsertedIdx)
12352 Mask[i] = ConstantInt::get(Type::getInt32Ty(*Context), NumElts+i);
12357 // If this insertelement is a chain that comes from exactly these two
12358 // vectors, return the vector and the effective shuffle.
12359 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12361 return EI->getOperand(0);
12366 // TODO: Handle shufflevector here!
12368 // Otherwise, can't do anything fancy. Return an identity vector.
12369 for (unsigned i = 0; i != NumElts; ++i)
12370 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12374 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12375 Value *VecOp = IE.getOperand(0);
12376 Value *ScalarOp = IE.getOperand(1);
12377 Value *IdxOp = IE.getOperand(2);
12379 // Inserting an undef or into an undefined place, remove this.
12380 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12381 ReplaceInstUsesWith(IE, VecOp);
12383 // If the inserted element was extracted from some other vector, and if the
12384 // indexes are constant, try to turn this into a shufflevector operation.
12385 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12386 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12387 EI->getOperand(0)->getType() == IE.getType()) {
12388 unsigned NumVectorElts = IE.getType()->getNumElements();
12389 unsigned ExtractedIdx =
12390 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12391 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12393 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12394 return ReplaceInstUsesWith(IE, VecOp);
12396 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12397 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12399 // If we are extracting a value from a vector, then inserting it right
12400 // back into the same place, just use the input vector.
12401 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12402 return ReplaceInstUsesWith(IE, VecOp);
12404 // We could theoretically do this for ANY input. However, doing so could
12405 // turn chains of insertelement instructions into a chain of shufflevector
12406 // instructions, and right now we do not merge shufflevectors. As such,
12407 // only do this in a situation where it is clear that there is benefit.
12408 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12409 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12410 // the values of VecOp, except then one read from EIOp0.
12411 // Build a new shuffle mask.
12412 std::vector<Constant*> Mask;
12413 if (isa<UndefValue>(VecOp))
12414 Mask.assign(NumVectorElts, UndefValue::get(Type::getInt32Ty(*Context)));
12416 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12417 Mask.assign(NumVectorElts, ConstantInt::get(Type::getInt32Ty(*Context),
12420 Mask[InsertedIdx] =
12421 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12422 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12423 ConstantVector::get(Mask));
12426 // If this insertelement isn't used by some other insertelement, turn it
12427 // (and any insertelements it points to), into one big shuffle.
12428 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12429 std::vector<Constant*> Mask;
12431 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12432 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12433 // We now have a shuffle of LHS, RHS, Mask.
12434 return new ShuffleVectorInst(LHS, RHS,
12435 ConstantVector::get(Mask));
12440 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12441 APInt UndefElts(VWidth, 0);
12442 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12443 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12450 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12451 Value *LHS = SVI.getOperand(0);
12452 Value *RHS = SVI.getOperand(1);
12453 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12455 bool MadeChange = false;
12457 // Undefined shuffle mask -> undefined value.
12458 if (isa<UndefValue>(SVI.getOperand(2)))
12459 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12461 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12463 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12466 APInt UndefElts(VWidth, 0);
12467 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12468 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12469 LHS = SVI.getOperand(0);
12470 RHS = SVI.getOperand(1);
12474 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12475 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12476 if (LHS == RHS || isa<UndefValue>(LHS)) {
12477 if (isa<UndefValue>(LHS) && LHS == RHS) {
12478 // shuffle(undef,undef,mask) -> undef.
12479 return ReplaceInstUsesWith(SVI, LHS);
12482 // Remap any references to RHS to use LHS.
12483 std::vector<Constant*> Elts;
12484 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12485 if (Mask[i] >= 2*e)
12486 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12488 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12489 (Mask[i] < e && isa<UndefValue>(LHS))) {
12490 Mask[i] = 2*e; // Turn into undef.
12491 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12493 Mask[i] = Mask[i] % e; // Force to LHS.
12494 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Mask[i]));
12498 SVI.setOperand(0, SVI.getOperand(1));
12499 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12500 SVI.setOperand(2, ConstantVector::get(Elts));
12501 LHS = SVI.getOperand(0);
12502 RHS = SVI.getOperand(1);
12506 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12507 bool isLHSID = true, isRHSID = true;
12509 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12510 if (Mask[i] >= e*2) continue; // Ignore undef values.
12511 // Is this an identity shuffle of the LHS value?
12512 isLHSID &= (Mask[i] == i);
12514 // Is this an identity shuffle of the RHS value?
12515 isRHSID &= (Mask[i]-e == i);
12518 // Eliminate identity shuffles.
12519 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12520 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12522 // If the LHS is a shufflevector itself, see if we can combine it with this
12523 // one without producing an unusual shuffle. Here we are really conservative:
12524 // we are absolutely afraid of producing a shuffle mask not in the input
12525 // program, because the code gen may not be smart enough to turn a merged
12526 // shuffle into two specific shuffles: it may produce worse code. As such,
12527 // we only merge two shuffles if the result is one of the two input shuffle
12528 // masks. In this case, merging the shuffles just removes one instruction,
12529 // which we know is safe. This is good for things like turning:
12530 // (splat(splat)) -> splat.
12531 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12532 if (isa<UndefValue>(RHS)) {
12533 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12535 std::vector<unsigned> NewMask;
12536 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12537 if (Mask[i] >= 2*e)
12538 NewMask.push_back(2*e);
12540 NewMask.push_back(LHSMask[Mask[i]]);
12542 // If the result mask is equal to the src shuffle or this shuffle mask, do
12543 // the replacement.
12544 if (NewMask == LHSMask || NewMask == Mask) {
12545 unsigned LHSInNElts =
12546 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12547 std::vector<Constant*> Elts;
12548 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12549 if (NewMask[i] >= LHSInNElts*2) {
12550 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12552 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), NewMask[i]));
12555 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12556 LHSSVI->getOperand(1),
12557 ConstantVector::get(Elts));
12562 return MadeChange ? &SVI : 0;
12568 /// TryToSinkInstruction - Try to move the specified instruction from its
12569 /// current block into the beginning of DestBlock, which can only happen if it's
12570 /// safe to move the instruction past all of the instructions between it and the
12571 /// end of its block.
12572 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12573 assert(I->hasOneUse() && "Invariants didn't hold!");
12575 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12576 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12579 // Do not sink alloca instructions out of the entry block.
12580 if (isa<AllocaInst>(I) && I->getParent() ==
12581 &DestBlock->getParent()->getEntryBlock())
12584 // We can only sink load instructions if there is nothing between the load and
12585 // the end of block that could change the value.
12586 if (I->mayReadFromMemory()) {
12587 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12589 if (Scan->mayWriteToMemory())
12593 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12595 CopyPrecedingStopPoint(I, InsertPos);
12596 I->moveBefore(InsertPos);
12602 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12603 /// all reachable code to the worklist.
12605 /// This has a couple of tricks to make the code faster and more powerful. In
12606 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12607 /// them to the worklist (this significantly speeds up instcombine on code where
12608 /// many instructions are dead or constant). Additionally, if we find a branch
12609 /// whose condition is a known constant, we only visit the reachable successors.
12611 static void AddReachableCodeToWorklist(BasicBlock *BB,
12612 SmallPtrSet<BasicBlock*, 64> &Visited,
12614 const TargetData *TD) {
12615 SmallVector<BasicBlock*, 256> Worklist;
12616 Worklist.push_back(BB);
12618 while (!Worklist.empty()) {
12619 BB = Worklist.back();
12620 Worklist.pop_back();
12622 // We have now visited this block! If we've already been here, ignore it.
12623 if (!Visited.insert(BB)) continue;
12625 DbgInfoIntrinsic *DBI_Prev = NULL;
12626 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12627 Instruction *Inst = BBI++;
12629 // DCE instruction if trivially dead.
12630 if (isInstructionTriviallyDead(Inst)) {
12632 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
12633 Inst->eraseFromParent();
12637 // ConstantProp instruction if trivially constant.
12638 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12639 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
12641 Inst->replaceAllUsesWith(C);
12643 Inst->eraseFromParent();
12647 // If there are two consecutive llvm.dbg.stoppoint calls then
12648 // it is likely that the optimizer deleted code in between these
12650 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12653 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12654 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12655 IC.Worklist.Remove(DBI_Prev);
12656 DBI_Prev->eraseFromParent();
12658 DBI_Prev = DBI_Next;
12663 IC.Worklist.Add(Inst);
12666 // Recursively visit successors. If this is a branch or switch on a
12667 // constant, only visit the reachable successor.
12668 TerminatorInst *TI = BB->getTerminator();
12669 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12670 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12671 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12672 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12673 Worklist.push_back(ReachableBB);
12676 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12677 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12678 // See if this is an explicit destination.
12679 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12680 if (SI->getCaseValue(i) == Cond) {
12681 BasicBlock *ReachableBB = SI->getSuccessor(i);
12682 Worklist.push_back(ReachableBB);
12686 // Otherwise it is the default destination.
12687 Worklist.push_back(SI->getSuccessor(0));
12692 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12693 Worklist.push_back(TI->getSuccessor(i));
12697 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12698 bool Changed = false;
12699 TD = getAnalysisIfAvailable<TargetData>();
12701 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12702 << F.getNameStr() << "\n");
12705 // Do a depth-first traversal of the function, populate the worklist with
12706 // the reachable instructions. Ignore blocks that are not reachable. Keep
12707 // track of which blocks we visit.
12708 SmallPtrSet<BasicBlock*, 64> Visited;
12709 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12711 // Do a quick scan over the function. If we find any blocks that are
12712 // unreachable, remove any instructions inside of them. This prevents
12713 // the instcombine code from having to deal with some bad special cases.
12714 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12715 if (!Visited.count(BB)) {
12716 Instruction *Term = BB->getTerminator();
12717 while (Term != BB->begin()) { // Remove instrs bottom-up
12718 BasicBlock::iterator I = Term; --I;
12720 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12721 // A debug intrinsic shouldn't force another iteration if we weren't
12722 // going to do one without it.
12723 if (!isa<DbgInfoIntrinsic>(I)) {
12727 if (!I->use_empty())
12728 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12729 I->eraseFromParent();
12734 while (!Worklist.isEmpty()) {
12735 Instruction *I = Worklist.RemoveOne();
12736 if (I == 0) continue; // skip null values.
12738 // Check to see if we can DCE the instruction.
12739 if (isInstructionTriviallyDead(I)) {
12740 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12741 EraseInstFromFunction(*I);
12747 // Instruction isn't dead, see if we can constant propagate it.
12748 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
12749 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
12751 // Add operands to the worklist.
12752 ReplaceInstUsesWith(*I, C);
12754 EraseInstFromFunction(*I);
12760 // See if we can constant fold its operands.
12761 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
12762 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
12763 if (Constant *NewC = ConstantFoldConstantExpression(CE,
12764 F.getContext(), TD))
12771 // See if we can trivially sink this instruction to a successor basic block.
12772 if (I->hasOneUse()) {
12773 BasicBlock *BB = I->getParent();
12774 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
12775 if (UserParent != BB) {
12776 bool UserIsSuccessor = false;
12777 // See if the user is one of our successors.
12778 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
12779 if (*SI == UserParent) {
12780 UserIsSuccessor = true;
12784 // If the user is one of our immediate successors, and if that successor
12785 // only has us as a predecessors (we'd have to split the critical edge
12786 // otherwise), we can keep going.
12787 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
12788 next(pred_begin(UserParent)) == pred_end(UserParent))
12789 // Okay, the CFG is simple enough, try to sink this instruction.
12790 Changed |= TryToSinkInstruction(I, UserParent);
12794 // Now that we have an instruction, try combining it to simplify it.
12795 Builder->SetInsertPoint(I->getParent(), I);
12800 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
12802 if (Instruction *Result = visit(*I)) {
12804 // Should we replace the old instruction with a new one?
12806 DEBUG(errs() << "IC: Old = " << *I << '\n'
12807 << " New = " << *Result << '\n');
12809 // Everything uses the new instruction now.
12810 I->replaceAllUsesWith(Result);
12812 // Push the new instruction and any users onto the worklist.
12813 Worklist.Add(Result);
12814 Worklist.AddUsersToWorkList(*Result);
12816 // Move the name to the new instruction first.
12817 Result->takeName(I);
12819 // Insert the new instruction into the basic block...
12820 BasicBlock *InstParent = I->getParent();
12821 BasicBlock::iterator InsertPos = I;
12823 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
12824 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
12827 InstParent->getInstList().insert(InsertPos, Result);
12829 EraseInstFromFunction(*I);
12832 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
12833 << " New = " << *I << '\n');
12836 // If the instruction was modified, it's possible that it is now dead.
12837 // if so, remove it.
12838 if (isInstructionTriviallyDead(I)) {
12839 EraseInstFromFunction(*I);
12842 Worklist.AddUsersToWorkList(*I);
12854 bool InstCombiner::runOnFunction(Function &F) {
12855 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
12856 Context = &F.getContext();
12859 /// Builder - This is an IRBuilder that automatically inserts new
12860 /// instructions into the worklist when they are created.
12861 IRBuilder<true, ConstantFolder, InstCombineIRInserter>
12862 TheBuilder(F.getContext(), ConstantFolder(F.getContext()),
12863 InstCombineIRInserter(Worklist));
12864 Builder = &TheBuilder;
12866 bool EverMadeChange = false;
12868 // Iterate while there is work to do.
12869 unsigned Iteration = 0;
12870 while (DoOneIteration(F, Iteration++))
12871 EverMadeChange = true;
12874 return EverMadeChange;
12877 FunctionPass *llvm::createInstructionCombiningPass() {
12878 return new InstCombiner();