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/MathExtras.h"
56 #include "llvm/Support/PatternMatch.h"
57 #include "llvm/Support/Compiler.h"
58 #include "llvm/Support/raw_ostream.h"
59 #include "llvm/ADT/DenseMap.h"
60 #include "llvm/ADT/SmallVector.h"
61 #include "llvm/ADT/SmallPtrSet.h"
62 #include "llvm/ADT/Statistic.h"
63 #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 Remove(Instruction *I) {
98 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
99 if (It == WorklistMap.end()) return; // Not in worklist.
101 // Don't bother moving everything down, just null out the slot.
102 Worklist[It->second] = 0;
104 WorklistMap.erase(It);
107 Instruction *RemoveOne() {
108 Instruction *I = Worklist.back();
110 WorklistMap.erase(I);
115 /// Zap - check that the worklist is empty and nuke the backing store for
116 /// the map if it is large.
118 assert(WorklistMap.empty() && "Worklist empty, but map not?");
120 // Do an explicit clear, this shrinks the map if needed.
124 } // end anonymous namespace.
128 class VISIBILITY_HIDDEN InstCombiner
129 : public FunctionPass,
130 public InstVisitor<InstCombiner, Instruction*> {
131 // Worklist of all of the instructions that need to be simplified.
132 InstCombineWorklist Worklist;
134 bool MustPreserveLCSSA;
136 static char ID; // Pass identification, replacement for typeid
137 InstCombiner() : FunctionPass(&ID) {}
139 LLVMContext *Context;
140 LLVMContext *getContext() const { return Context; }
142 /// AddToWorkList - Add the specified instruction to the worklist if it
143 /// isn't already in it.
144 void AddToWorkList(Instruction *I) {
148 // RemoveFromWorkList - remove I from the worklist if it exists.
149 void RemoveFromWorkList(Instruction *I) {
153 /// AddUsersToWorkList - When an instruction is simplified, add all users of
154 /// the instruction to the work lists because they might get more simplified
157 void AddUsersToWorkList(Value &I) {
158 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
160 AddToWorkList(cast<Instruction>(*UI));
163 /// AddUsesToWorkList - When an instruction is simplified, add operands to
164 /// the work lists because they might get more simplified now.
166 void AddUsesToWorkList(Instruction &I) {
167 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
168 if (Instruction *Op = dyn_cast<Instruction>(*i))
172 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
173 /// dead. Add all of its operands to the worklist, turning them into
174 /// undef's to reduce the number of uses of those instructions.
176 /// Return the specified operand before it is turned into an undef.
178 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
179 Value *R = I.getOperand(op);
181 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
182 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
184 // Set the operand to undef to drop the use.
185 *i = UndefValue::get(Op->getType());
192 virtual bool runOnFunction(Function &F);
194 bool DoOneIteration(Function &F, unsigned ItNum);
196 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
197 AU.addPreservedID(LCSSAID);
198 AU.setPreservesCFG();
201 TargetData *getTargetData() const { return TD; }
203 // Visitation implementation - Implement instruction combining for different
204 // instruction types. The semantics are as follows:
206 // null - No change was made
207 // I - Change was made, I is still valid, I may be dead though
208 // otherwise - Change was made, replace I with returned instruction
210 Instruction *visitAdd(BinaryOperator &I);
211 Instruction *visitFAdd(BinaryOperator &I);
212 Instruction *visitSub(BinaryOperator &I);
213 Instruction *visitFSub(BinaryOperator &I);
214 Instruction *visitMul(BinaryOperator &I);
215 Instruction *visitFMul(BinaryOperator &I);
216 Instruction *visitURem(BinaryOperator &I);
217 Instruction *visitSRem(BinaryOperator &I);
218 Instruction *visitFRem(BinaryOperator &I);
219 bool SimplifyDivRemOfSelect(BinaryOperator &I);
220 Instruction *commonRemTransforms(BinaryOperator &I);
221 Instruction *commonIRemTransforms(BinaryOperator &I);
222 Instruction *commonDivTransforms(BinaryOperator &I);
223 Instruction *commonIDivTransforms(BinaryOperator &I);
224 Instruction *visitUDiv(BinaryOperator &I);
225 Instruction *visitSDiv(BinaryOperator &I);
226 Instruction *visitFDiv(BinaryOperator &I);
227 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
228 Instruction *FoldAndOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
229 Instruction *visitAnd(BinaryOperator &I);
230 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
231 Instruction *FoldOrOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
232 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
233 Value *A, Value *B, Value *C);
234 Instruction *visitOr (BinaryOperator &I);
235 Instruction *visitXor(BinaryOperator &I);
236 Instruction *visitShl(BinaryOperator &I);
237 Instruction *visitAShr(BinaryOperator &I);
238 Instruction *visitLShr(BinaryOperator &I);
239 Instruction *commonShiftTransforms(BinaryOperator &I);
240 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
242 Instruction *visitFCmpInst(FCmpInst &I);
243 Instruction *visitICmpInst(ICmpInst &I);
244 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
245 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
248 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
249 ConstantInt *DivRHS);
251 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
252 ICmpInst::Predicate Cond, Instruction &I);
253 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
255 Instruction *commonCastTransforms(CastInst &CI);
256 Instruction *commonIntCastTransforms(CastInst &CI);
257 Instruction *commonPointerCastTransforms(CastInst &CI);
258 Instruction *visitTrunc(TruncInst &CI);
259 Instruction *visitZExt(ZExtInst &CI);
260 Instruction *visitSExt(SExtInst &CI);
261 Instruction *visitFPTrunc(FPTruncInst &CI);
262 Instruction *visitFPExt(CastInst &CI);
263 Instruction *visitFPToUI(FPToUIInst &FI);
264 Instruction *visitFPToSI(FPToSIInst &FI);
265 Instruction *visitUIToFP(CastInst &CI);
266 Instruction *visitSIToFP(CastInst &CI);
267 Instruction *visitPtrToInt(PtrToIntInst &CI);
268 Instruction *visitIntToPtr(IntToPtrInst &CI);
269 Instruction *visitBitCast(BitCastInst &CI);
270 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
272 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
273 Instruction *visitSelectInst(SelectInst &SI);
274 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
275 Instruction *visitCallInst(CallInst &CI);
276 Instruction *visitInvokeInst(InvokeInst &II);
277 Instruction *visitPHINode(PHINode &PN);
278 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
279 Instruction *visitAllocationInst(AllocationInst &AI);
280 Instruction *visitFreeInst(FreeInst &FI);
281 Instruction *visitLoadInst(LoadInst &LI);
282 Instruction *visitStoreInst(StoreInst &SI);
283 Instruction *visitBranchInst(BranchInst &BI);
284 Instruction *visitSwitchInst(SwitchInst &SI);
285 Instruction *visitInsertElementInst(InsertElementInst &IE);
286 Instruction *visitExtractElementInst(ExtractElementInst &EI);
287 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
288 Instruction *visitExtractValueInst(ExtractValueInst &EV);
290 // visitInstruction - Specify what to return for unhandled instructions...
291 Instruction *visitInstruction(Instruction &I) { return 0; }
294 Instruction *visitCallSite(CallSite CS);
295 bool transformConstExprCastCall(CallSite CS);
296 Instruction *transformCallThroughTrampoline(CallSite CS);
297 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
298 bool DoXform = true);
299 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
300 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
304 // InsertNewInstBefore - insert an instruction New before instruction Old
305 // in the program. Add the new instruction to the worklist.
307 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
308 assert(New && New->getParent() == 0 &&
309 "New instruction already inserted into a basic block!");
310 BasicBlock *BB = Old.getParent();
311 BB->getInstList().insert(&Old, New); // Insert inst
316 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
317 /// This also adds the cast to the worklist. Finally, this returns the
319 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
321 if (V->getType() == Ty) return V;
323 if (Constant *CV = dyn_cast<Constant>(V))
324 return ConstantExpr::getCast(opc, CV, Ty);
326 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
331 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
332 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
336 // ReplaceInstUsesWith - This method is to be used when an instruction is
337 // found to be dead, replacable with another preexisting expression. Here
338 // we add all uses of I to the worklist, replace all uses of I with the new
339 // value, then return I, so that the inst combiner will know that I was
342 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
343 AddUsersToWorkList(I); // Add all modified instrs to worklist
345 I.replaceAllUsesWith(V);
348 // If we are replacing the instruction with itself, this must be in a
349 // segment of unreachable code, so just clobber the instruction.
350 I.replaceAllUsesWith(UndefValue::get(I.getType()));
355 // EraseInstFromFunction - When dealing with an instruction that has side
356 // effects or produces a void value, we can't rely on DCE to delete the
357 // instruction. Instead, visit methods should return the value returned by
359 Instruction *EraseInstFromFunction(Instruction &I) {
360 assert(I.use_empty() && "Cannot erase instruction that is used!");
361 AddUsesToWorkList(I);
362 RemoveFromWorkList(&I);
364 return 0; // Don't do anything with FI
367 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
368 APInt &KnownOne, unsigned Depth = 0) const {
369 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
372 bool MaskedValueIsZero(Value *V, const APInt &Mask,
373 unsigned Depth = 0) const {
374 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
376 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
377 return llvm::ComputeNumSignBits(Op, TD, Depth);
382 /// SimplifyCommutative - This performs a few simplifications for
383 /// commutative operators.
384 bool SimplifyCommutative(BinaryOperator &I);
386 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
387 /// most-complex to least-complex order.
388 bool SimplifyCompare(CmpInst &I);
390 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
391 /// based on the demanded bits.
392 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
393 APInt& KnownZero, APInt& KnownOne,
395 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
396 APInt& KnownZero, APInt& KnownOne,
399 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
400 /// SimplifyDemandedBits knows about. See if the instruction has any
401 /// properties that allow us to simplify its operands.
402 bool SimplifyDemandedInstructionBits(Instruction &Inst);
404 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
405 APInt& UndefElts, unsigned Depth = 0);
407 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
408 // PHI node as operand #0, see if we can fold the instruction into the PHI
409 // (which is only possible if all operands to the PHI are constants).
410 Instruction *FoldOpIntoPhi(Instruction &I);
412 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
413 // operator and they all are only used by the PHI, PHI together their
414 // inputs, and do the operation once, to the result of the PHI.
415 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
416 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
417 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
420 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
421 ConstantInt *AndRHS, BinaryOperator &TheAnd);
423 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
424 bool isSub, Instruction &I);
425 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
426 bool isSigned, bool Inside, Instruction &IB);
427 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
428 Instruction *MatchBSwap(BinaryOperator &I);
429 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
430 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
431 Instruction *SimplifyMemSet(MemSetInst *MI);
434 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
436 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
437 unsigned CastOpc, int &NumCastsRemoved);
438 unsigned GetOrEnforceKnownAlignment(Value *V,
439 unsigned PrefAlign = 0);
442 } // end anonymous namespace
444 char InstCombiner::ID = 0;
445 static RegisterPass<InstCombiner>
446 X("instcombine", "Combine redundant instructions");
448 // getComplexity: Assign a complexity or rank value to LLVM Values...
449 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
450 static unsigned getComplexity(Value *V) {
451 if (isa<Instruction>(V)) {
452 if (BinaryOperator::isNeg(V) ||
453 BinaryOperator::isFNeg(V) ||
454 BinaryOperator::isNot(V))
458 if (isa<Argument>(V)) return 3;
459 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
462 // isOnlyUse - Return true if this instruction will be deleted if we stop using
464 static bool isOnlyUse(Value *V) {
465 return V->hasOneUse() || isa<Constant>(V);
468 // getPromotedType - Return the specified type promoted as it would be to pass
469 // though a va_arg area...
470 static const Type *getPromotedType(const Type *Ty) {
471 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
472 if (ITy->getBitWidth() < 32)
473 return Type::getInt32Ty(Ty->getContext());
478 /// getBitCastOperand - If the specified operand is a CastInst, a constant
479 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
480 /// operand value, otherwise return null.
481 static Value *getBitCastOperand(Value *V) {
482 if (Operator *O = dyn_cast<Operator>(V)) {
483 if (O->getOpcode() == Instruction::BitCast)
484 return O->getOperand(0);
485 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
486 if (GEP->hasAllZeroIndices())
487 return GEP->getPointerOperand();
492 /// This function is a wrapper around CastInst::isEliminableCastPair. It
493 /// simply extracts arguments and returns what that function returns.
494 static Instruction::CastOps
495 isEliminableCastPair(
496 const CastInst *CI, ///< The first cast instruction
497 unsigned opcode, ///< The opcode of the second cast instruction
498 const Type *DstTy, ///< The target type for the second cast instruction
499 TargetData *TD ///< The target data for pointer size
502 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
503 const Type *MidTy = CI->getType(); // B from above
505 // Get the opcodes of the two Cast instructions
506 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
507 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
509 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
511 TD ? TD->getIntPtrType(CI->getContext()) : 0);
513 // We don't want to form an inttoptr or ptrtoint that converts to an integer
514 // type that differs from the pointer size.
515 if ((Res == Instruction::IntToPtr &&
516 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
517 (Res == Instruction::PtrToInt &&
518 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
521 return Instruction::CastOps(Res);
524 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
525 /// in any code being generated. It does not require codegen if V is simple
526 /// enough or if the cast can be folded into other casts.
527 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
528 const Type *Ty, TargetData *TD) {
529 if (V->getType() == Ty || isa<Constant>(V)) return false;
531 // If this is another cast that can be eliminated, it isn't codegen either.
532 if (const CastInst *CI = dyn_cast<CastInst>(V))
533 if (isEliminableCastPair(CI, opcode, Ty, TD))
538 // SimplifyCommutative - This performs a few simplifications for commutative
541 // 1. Order operands such that they are listed from right (least complex) to
542 // left (most complex). This puts constants before unary operators before
545 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
546 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
548 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
549 bool Changed = false;
550 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
551 Changed = !I.swapOperands();
553 if (!I.isAssociative()) return Changed;
554 Instruction::BinaryOps Opcode = I.getOpcode();
555 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
556 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
557 if (isa<Constant>(I.getOperand(1))) {
558 Constant *Folded = ConstantExpr::get(I.getOpcode(),
559 cast<Constant>(I.getOperand(1)),
560 cast<Constant>(Op->getOperand(1)));
561 I.setOperand(0, Op->getOperand(0));
562 I.setOperand(1, Folded);
564 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
565 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
566 isOnlyUse(Op) && isOnlyUse(Op1)) {
567 Constant *C1 = cast<Constant>(Op->getOperand(1));
568 Constant *C2 = cast<Constant>(Op1->getOperand(1));
570 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
571 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
572 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
576 I.setOperand(0, New);
577 I.setOperand(1, Folded);
584 /// SimplifyCompare - For a CmpInst this function just orders the operands
585 /// so that theyare listed from right (least complex) to left (most complex).
586 /// This puts constants before unary operators before binary operators.
587 bool InstCombiner::SimplifyCompare(CmpInst &I) {
588 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
591 // Compare instructions are not associative so there's nothing else we can do.
595 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
596 // if the LHS is a constant zero (which is the 'negate' form).
598 static inline Value *dyn_castNegVal(Value *V) {
599 if (BinaryOperator::isNeg(V))
600 return BinaryOperator::getNegArgument(V);
602 // Constants can be considered to be negated values if they can be folded.
603 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
604 return ConstantExpr::getNeg(C);
606 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
607 if (C->getType()->getElementType()->isInteger())
608 return ConstantExpr::getNeg(C);
613 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
614 // instruction if the LHS is a constant negative zero (which is the 'negate'
617 static inline Value *dyn_castFNegVal(Value *V) {
618 if (BinaryOperator::isFNeg(V))
619 return BinaryOperator::getFNegArgument(V);
621 // Constants can be considered to be negated values if they can be folded.
622 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
623 return ConstantExpr::getFNeg(C);
625 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
626 if (C->getType()->getElementType()->isFloatingPoint())
627 return ConstantExpr::getFNeg(C);
632 static inline Value *dyn_castNotVal(Value *V) {
633 if (BinaryOperator::isNot(V))
634 return BinaryOperator::getNotArgument(V);
636 // Constants can be considered to be not'ed values...
637 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
638 return ConstantInt::get(C->getType(), ~C->getValue());
642 // dyn_castFoldableMul - If this value is a multiply that can be folded into
643 // other computations (because it has a constant operand), return the
644 // non-constant operand of the multiply, and set CST to point to the multiplier.
645 // Otherwise, return null.
647 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
648 if (V->hasOneUse() && V->getType()->isInteger())
649 if (Instruction *I = dyn_cast<Instruction>(V)) {
650 if (I->getOpcode() == Instruction::Mul)
651 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
652 return I->getOperand(0);
653 if (I->getOpcode() == Instruction::Shl)
654 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
655 // The multiplier is really 1 << CST.
656 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
657 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
658 CST = ConstantInt::get(V->getType()->getContext(),
659 APInt(BitWidth, 1).shl(CSTVal));
660 return I->getOperand(0);
666 /// AddOne - Add one to a ConstantInt
667 static Constant *AddOne(Constant *C) {
668 return ConstantExpr::getAdd(C,
669 ConstantInt::get(C->getType(), 1));
671 /// SubOne - Subtract one from a ConstantInt
672 static Constant *SubOne(ConstantInt *C) {
673 return ConstantExpr::getSub(C,
674 ConstantInt::get(C->getType(), 1));
676 /// MultiplyOverflows - True if the multiply can not be expressed in an int
678 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
679 uint32_t W = C1->getBitWidth();
680 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
689 APInt MulExt = LHSExt * RHSExt;
692 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
693 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
694 return MulExt.slt(Min) || MulExt.sgt(Max);
696 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
700 /// ShrinkDemandedConstant - Check to see if the specified operand of the
701 /// specified instruction is a constant integer. If so, check to see if there
702 /// are any bits set in the constant that are not demanded. If so, shrink the
703 /// constant and return true.
704 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
706 assert(I && "No instruction?");
707 assert(OpNo < I->getNumOperands() && "Operand index too large");
709 // If the operand is not a constant integer, nothing to do.
710 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
711 if (!OpC) return false;
713 // If there are no bits set that aren't demanded, nothing to do.
714 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
715 if ((~Demanded & OpC->getValue()) == 0)
718 // This instruction is producing bits that are not demanded. Shrink the RHS.
719 Demanded &= OpC->getValue();
720 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Demanded));
724 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
725 // set of known zero and one bits, compute the maximum and minimum values that
726 // could have the specified known zero and known one bits, returning them in
728 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
729 const APInt& KnownOne,
730 APInt& Min, APInt& Max) {
731 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
732 KnownZero.getBitWidth() == Min.getBitWidth() &&
733 KnownZero.getBitWidth() == Max.getBitWidth() &&
734 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
735 APInt UnknownBits = ~(KnownZero|KnownOne);
737 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
738 // bit if it is unknown.
740 Max = KnownOne|UnknownBits;
742 if (UnknownBits.isNegative()) { // Sign bit is unknown
743 Min.set(Min.getBitWidth()-1);
744 Max.clear(Max.getBitWidth()-1);
748 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
749 // a set of known zero and one bits, compute the maximum and minimum values that
750 // could have the specified known zero and known one bits, returning them in
752 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
753 const APInt &KnownOne,
754 APInt &Min, APInt &Max) {
755 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
756 KnownZero.getBitWidth() == Min.getBitWidth() &&
757 KnownZero.getBitWidth() == Max.getBitWidth() &&
758 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
759 APInt UnknownBits = ~(KnownZero|KnownOne);
761 // The minimum value is when the unknown bits are all zeros.
763 // The maximum value is when the unknown bits are all ones.
764 Max = KnownOne|UnknownBits;
767 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
768 /// SimplifyDemandedBits knows about. See if the instruction has any
769 /// properties that allow us to simplify its operands.
770 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
771 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
772 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
773 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
775 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
776 KnownZero, KnownOne, 0);
777 if (V == 0) return false;
778 if (V == &Inst) return true;
779 ReplaceInstUsesWith(Inst, V);
783 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
784 /// specified instruction operand if possible, updating it in place. It returns
785 /// true if it made any change and false otherwise.
786 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
787 APInt &KnownZero, APInt &KnownOne,
789 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
790 KnownZero, KnownOne, Depth);
791 if (NewVal == 0) return false;
797 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
798 /// value based on the demanded bits. When this function is called, it is known
799 /// that only the bits set in DemandedMask of the result of V are ever used
800 /// downstream. Consequently, depending on the mask and V, it may be possible
801 /// to replace V with a constant or one of its operands. In such cases, this
802 /// function does the replacement and returns true. In all other cases, it
803 /// returns false after analyzing the expression and setting KnownOne and known
804 /// to be one in the expression. KnownZero contains all the bits that are known
805 /// to be zero in the expression. These are provided to potentially allow the
806 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
807 /// the expression. KnownOne and KnownZero always follow the invariant that
808 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
809 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
810 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
811 /// and KnownOne must all be the same.
813 /// This returns null if it did not change anything and it permits no
814 /// simplification. This returns V itself if it did some simplification of V's
815 /// operands based on the information about what bits are demanded. This returns
816 /// some other non-null value if it found out that V is equal to another value
817 /// in the context where the specified bits are demanded, but not for all users.
818 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
819 APInt &KnownZero, APInt &KnownOne,
821 assert(V != 0 && "Null pointer of Value???");
822 assert(Depth <= 6 && "Limit Search Depth");
823 uint32_t BitWidth = DemandedMask.getBitWidth();
824 const Type *VTy = V->getType();
825 assert((TD || !isa<PointerType>(VTy)) &&
826 "SimplifyDemandedBits needs to know bit widths!");
827 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
828 (!VTy->isIntOrIntVector() ||
829 VTy->getScalarSizeInBits() == BitWidth) &&
830 KnownZero.getBitWidth() == BitWidth &&
831 KnownOne.getBitWidth() == BitWidth &&
832 "Value *V, DemandedMask, KnownZero and KnownOne "
833 "must have same BitWidth");
834 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
835 // We know all of the bits for a constant!
836 KnownOne = CI->getValue() & DemandedMask;
837 KnownZero = ~KnownOne & DemandedMask;
840 if (isa<ConstantPointerNull>(V)) {
841 // We know all of the bits for a constant!
843 KnownZero = DemandedMask;
849 if (DemandedMask == 0) { // Not demanding any bits from V.
850 if (isa<UndefValue>(V))
852 return UndefValue::get(VTy);
855 if (Depth == 6) // Limit search depth.
858 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
859 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
861 Instruction *I = dyn_cast<Instruction>(V);
863 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
864 return 0; // Only analyze instructions.
867 // If there are multiple uses of this value and we aren't at the root, then
868 // we can't do any simplifications of the operands, because DemandedMask
869 // only reflects the bits demanded by *one* of the users.
870 if (Depth != 0 && !I->hasOneUse()) {
871 // Despite the fact that we can't simplify this instruction in all User's
872 // context, we can at least compute the knownzero/knownone bits, and we can
873 // do simplifications that apply to *just* the one user if we know that
874 // this instruction has a simpler value in that context.
875 if (I->getOpcode() == Instruction::And) {
876 // If either the LHS or the RHS are Zero, the result is zero.
877 ComputeMaskedBits(I->getOperand(1), DemandedMask,
878 RHSKnownZero, RHSKnownOne, Depth+1);
879 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
880 LHSKnownZero, LHSKnownOne, Depth+1);
882 // If all of the demanded bits are known 1 on one side, return the other.
883 // These bits cannot contribute to the result of the 'and' in this
885 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
886 (DemandedMask & ~LHSKnownZero))
887 return I->getOperand(0);
888 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
889 (DemandedMask & ~RHSKnownZero))
890 return I->getOperand(1);
892 // If all of the demanded bits in the inputs are known zeros, return zero.
893 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
894 return Constant::getNullValue(VTy);
896 } else if (I->getOpcode() == Instruction::Or) {
897 // We can simplify (X|Y) -> X or Y in the user's context if we know that
898 // only bits from X or Y are demanded.
900 // If either the LHS or the RHS are One, the result is One.
901 ComputeMaskedBits(I->getOperand(1), DemandedMask,
902 RHSKnownZero, RHSKnownOne, Depth+1);
903 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
904 LHSKnownZero, LHSKnownOne, Depth+1);
906 // If all of the demanded bits are known zero on one side, return the
907 // other. These bits cannot contribute to the result of the 'or' in this
909 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
910 (DemandedMask & ~LHSKnownOne))
911 return I->getOperand(0);
912 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
913 (DemandedMask & ~RHSKnownOne))
914 return I->getOperand(1);
916 // If all of the potentially set bits on one side are known to be set on
917 // the other side, just use the 'other' side.
918 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
919 (DemandedMask & (~RHSKnownZero)))
920 return I->getOperand(0);
921 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
922 (DemandedMask & (~LHSKnownZero)))
923 return I->getOperand(1);
926 // Compute the KnownZero/KnownOne bits to simplify things downstream.
927 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
931 // If this is the root being simplified, allow it to have multiple uses,
932 // just set the DemandedMask to all bits so that we can try to simplify the
933 // operands. This allows visitTruncInst (for example) to simplify the
934 // operand of a trunc without duplicating all the logic below.
935 if (Depth == 0 && !V->hasOneUse())
936 DemandedMask = APInt::getAllOnesValue(BitWidth);
938 switch (I->getOpcode()) {
940 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
942 case Instruction::And:
943 // If either the LHS or the RHS are Zero, the result is zero.
944 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
945 RHSKnownZero, RHSKnownOne, Depth+1) ||
946 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
947 LHSKnownZero, LHSKnownOne, Depth+1))
949 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
950 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
952 // If all of the demanded bits are known 1 on one side, return the other.
953 // These bits cannot contribute to the result of the 'and'.
954 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
955 (DemandedMask & ~LHSKnownZero))
956 return I->getOperand(0);
957 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
958 (DemandedMask & ~RHSKnownZero))
959 return I->getOperand(1);
961 // If all of the demanded bits in the inputs are known zeros, return zero.
962 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
963 return Constant::getNullValue(VTy);
965 // If the RHS is a constant, see if we can simplify it.
966 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
969 // Output known-1 bits are only known if set in both the LHS & RHS.
970 RHSKnownOne &= LHSKnownOne;
971 // Output known-0 are known to be clear if zero in either the LHS | RHS.
972 RHSKnownZero |= LHSKnownZero;
974 case Instruction::Or:
975 // If either the LHS or the RHS are One, the result is One.
976 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
977 RHSKnownZero, RHSKnownOne, Depth+1) ||
978 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
979 LHSKnownZero, LHSKnownOne, Depth+1))
981 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
982 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
984 // If all of the demanded bits are known zero on one side, return the other.
985 // These bits cannot contribute to the result of the 'or'.
986 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
987 (DemandedMask & ~LHSKnownOne))
988 return I->getOperand(0);
989 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
990 (DemandedMask & ~RHSKnownOne))
991 return I->getOperand(1);
993 // If all of the potentially set bits on one side are known to be set on
994 // the other side, just use the 'other' side.
995 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
996 (DemandedMask & (~RHSKnownZero)))
997 return I->getOperand(0);
998 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
999 (DemandedMask & (~LHSKnownZero)))
1000 return I->getOperand(1);
1002 // If the RHS is a constant, see if we can simplify it.
1003 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1006 // Output known-0 bits are only known if clear in both the LHS & RHS.
1007 RHSKnownZero &= LHSKnownZero;
1008 // Output known-1 are known to be set if set in either the LHS | RHS.
1009 RHSKnownOne |= LHSKnownOne;
1011 case Instruction::Xor: {
1012 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1013 RHSKnownZero, RHSKnownOne, Depth+1) ||
1014 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1015 LHSKnownZero, LHSKnownOne, Depth+1))
1017 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1018 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1020 // If all of the demanded bits are known zero on one side, return the other.
1021 // These bits cannot contribute to the result of the 'xor'.
1022 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1023 return I->getOperand(0);
1024 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1025 return I->getOperand(1);
1027 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1028 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1029 (RHSKnownOne & LHSKnownOne);
1030 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1031 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1032 (RHSKnownOne & LHSKnownZero);
1034 // If all of the demanded bits are known to be zero on one side or the
1035 // other, turn this into an *inclusive* or.
1036 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1037 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1039 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1041 return InsertNewInstBefore(Or, *I);
1044 // If all of the demanded bits on one side are known, and all of the set
1045 // bits on that side are also known to be set on the other side, turn this
1046 // into an AND, as we know the bits will be cleared.
1047 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1048 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1050 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1051 Constant *AndC = Constant::getIntegerValue(VTy,
1052 ~RHSKnownOne & DemandedMask);
1054 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1055 return InsertNewInstBefore(And, *I);
1059 // If the RHS is a constant, see if we can simplify it.
1060 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1061 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1064 RHSKnownZero = KnownZeroOut;
1065 RHSKnownOne = KnownOneOut;
1068 case Instruction::Select:
1069 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1070 RHSKnownZero, RHSKnownOne, Depth+1) ||
1071 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1072 LHSKnownZero, LHSKnownOne, Depth+1))
1074 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1075 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1077 // If the operands are constants, see if we can simplify them.
1078 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1079 ShrinkDemandedConstant(I, 2, DemandedMask))
1082 // Only known if known in both the LHS and RHS.
1083 RHSKnownOne &= LHSKnownOne;
1084 RHSKnownZero &= LHSKnownZero;
1086 case Instruction::Trunc: {
1087 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1088 DemandedMask.zext(truncBf);
1089 RHSKnownZero.zext(truncBf);
1090 RHSKnownOne.zext(truncBf);
1091 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1092 RHSKnownZero, RHSKnownOne, Depth+1))
1094 DemandedMask.trunc(BitWidth);
1095 RHSKnownZero.trunc(BitWidth);
1096 RHSKnownOne.trunc(BitWidth);
1097 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1100 case Instruction::BitCast:
1101 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1102 return false; // vector->int or fp->int?
1104 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1105 if (const VectorType *SrcVTy =
1106 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1107 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1108 // Don't touch a bitcast between vectors of different element counts.
1111 // Don't touch a scalar-to-vector bitcast.
1113 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1114 // Don't touch a vector-to-scalar bitcast.
1117 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1118 RHSKnownZero, RHSKnownOne, Depth+1))
1120 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1122 case Instruction::ZExt: {
1123 // Compute the bits in the result that are not present in the input.
1124 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1126 DemandedMask.trunc(SrcBitWidth);
1127 RHSKnownZero.trunc(SrcBitWidth);
1128 RHSKnownOne.trunc(SrcBitWidth);
1129 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1130 RHSKnownZero, RHSKnownOne, Depth+1))
1132 DemandedMask.zext(BitWidth);
1133 RHSKnownZero.zext(BitWidth);
1134 RHSKnownOne.zext(BitWidth);
1135 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1136 // The top bits are known to be zero.
1137 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1140 case Instruction::SExt: {
1141 // Compute the bits in the result that are not present in the input.
1142 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1144 APInt InputDemandedBits = DemandedMask &
1145 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1147 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1148 // If any of the sign extended bits are demanded, we know that the sign
1150 if ((NewBits & DemandedMask) != 0)
1151 InputDemandedBits.set(SrcBitWidth-1);
1153 InputDemandedBits.trunc(SrcBitWidth);
1154 RHSKnownZero.trunc(SrcBitWidth);
1155 RHSKnownOne.trunc(SrcBitWidth);
1156 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1157 RHSKnownZero, RHSKnownOne, Depth+1))
1159 InputDemandedBits.zext(BitWidth);
1160 RHSKnownZero.zext(BitWidth);
1161 RHSKnownOne.zext(BitWidth);
1162 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1164 // If the sign bit of the input is known set or clear, then we know the
1165 // top bits of the result.
1167 // If the input sign bit is known zero, or if the NewBits are not demanded
1168 // convert this into a zero extension.
1169 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1170 // Convert to ZExt cast
1171 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1172 return InsertNewInstBefore(NewCast, *I);
1173 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1174 RHSKnownOne |= NewBits;
1178 case Instruction::Add: {
1179 // Figure out what the input bits are. If the top bits of the and result
1180 // are not demanded, then the add doesn't demand them from its input
1182 unsigned NLZ = DemandedMask.countLeadingZeros();
1184 // If there is a constant on the RHS, there are a variety of xformations
1186 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1187 // If null, this should be simplified elsewhere. Some of the xforms here
1188 // won't work if the RHS is zero.
1192 // If the top bit of the output is demanded, demand everything from the
1193 // input. Otherwise, we demand all the input bits except NLZ top bits.
1194 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1196 // Find information about known zero/one bits in the input.
1197 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1198 LHSKnownZero, LHSKnownOne, Depth+1))
1201 // If the RHS of the add has bits set that can't affect the input, reduce
1203 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1206 // Avoid excess work.
1207 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1210 // Turn it into OR if input bits are zero.
1211 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1213 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1215 return InsertNewInstBefore(Or, *I);
1218 // We can say something about the output known-zero and known-one bits,
1219 // depending on potential carries from the input constant and the
1220 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1221 // bits set and the RHS constant is 0x01001, then we know we have a known
1222 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1224 // To compute this, we first compute the potential carry bits. These are
1225 // the bits which may be modified. I'm not aware of a better way to do
1227 const APInt &RHSVal = RHS->getValue();
1228 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1230 // Now that we know which bits have carries, compute the known-1/0 sets.
1232 // Bits are known one if they are known zero in one operand and one in the
1233 // other, and there is no input carry.
1234 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1235 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1237 // Bits are known zero if they are known zero in both operands and there
1238 // is no input carry.
1239 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1241 // If the high-bits of this ADD are not demanded, then it does not demand
1242 // the high bits of its LHS or RHS.
1243 if (DemandedMask[BitWidth-1] == 0) {
1244 // Right fill the mask of bits for this ADD to demand the most
1245 // significant bit and all those below it.
1246 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1247 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1248 LHSKnownZero, LHSKnownOne, Depth+1) ||
1249 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1250 LHSKnownZero, LHSKnownOne, Depth+1))
1256 case Instruction::Sub:
1257 // If the high-bits of this SUB are not demanded, then it does not demand
1258 // the high bits of its LHS or RHS.
1259 if (DemandedMask[BitWidth-1] == 0) {
1260 // Right fill the mask of bits for this SUB to demand the most
1261 // significant bit and all those below it.
1262 uint32_t NLZ = DemandedMask.countLeadingZeros();
1263 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1264 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1265 LHSKnownZero, LHSKnownOne, Depth+1) ||
1266 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1267 LHSKnownZero, LHSKnownOne, Depth+1))
1270 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1271 // the known zeros and ones.
1272 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1274 case Instruction::Shl:
1275 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1276 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1277 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1278 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1279 RHSKnownZero, RHSKnownOne, Depth+1))
1281 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1282 RHSKnownZero <<= ShiftAmt;
1283 RHSKnownOne <<= ShiftAmt;
1284 // low bits known zero.
1286 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1289 case Instruction::LShr:
1290 // For a logical shift right
1291 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1292 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1294 // Unsigned shift right.
1295 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1296 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1297 RHSKnownZero, RHSKnownOne, Depth+1))
1299 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1300 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1301 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1303 // Compute the new bits that are at the top now.
1304 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1305 RHSKnownZero |= HighBits; // high bits known zero.
1309 case Instruction::AShr:
1310 // If this is an arithmetic shift right and only the low-bit is set, we can
1311 // always convert this into a logical shr, even if the shift amount is
1312 // variable. The low bit of the shift cannot be an input sign bit unless
1313 // the shift amount is >= the size of the datatype, which is undefined.
1314 if (DemandedMask == 1) {
1315 // Perform the logical shift right.
1316 Instruction *NewVal = BinaryOperator::CreateLShr(
1317 I->getOperand(0), I->getOperand(1), I->getName());
1318 return InsertNewInstBefore(NewVal, *I);
1321 // If the sign bit is the only bit demanded by this ashr, then there is no
1322 // need to do it, the shift doesn't change the high bit.
1323 if (DemandedMask.isSignBit())
1324 return I->getOperand(0);
1326 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1327 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1329 // Signed shift right.
1330 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1331 // If any of the "high bits" are demanded, we should set the sign bit as
1333 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1334 DemandedMaskIn.set(BitWidth-1);
1335 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1336 RHSKnownZero, RHSKnownOne, Depth+1))
1338 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1339 // Compute the new bits that are at the top now.
1340 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1341 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1342 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1344 // Handle the sign bits.
1345 APInt SignBit(APInt::getSignBit(BitWidth));
1346 // Adjust to where it is now in the mask.
1347 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1349 // If the input sign bit is known to be zero, or if none of the top bits
1350 // are demanded, turn this into an unsigned shift right.
1351 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1352 (HighBits & ~DemandedMask) == HighBits) {
1353 // Perform the logical shift right.
1354 Instruction *NewVal = BinaryOperator::CreateLShr(
1355 I->getOperand(0), SA, I->getName());
1356 return InsertNewInstBefore(NewVal, *I);
1357 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1358 RHSKnownOne |= HighBits;
1362 case Instruction::SRem:
1363 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1364 APInt RA = Rem->getValue().abs();
1365 if (RA.isPowerOf2()) {
1366 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1367 return I->getOperand(0);
1369 APInt LowBits = RA - 1;
1370 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1371 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1372 LHSKnownZero, LHSKnownOne, Depth+1))
1375 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1376 LHSKnownZero |= ~LowBits;
1378 KnownZero |= LHSKnownZero & DemandedMask;
1380 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1384 case Instruction::URem: {
1385 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1386 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1387 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1388 KnownZero2, KnownOne2, Depth+1) ||
1389 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1390 KnownZero2, KnownOne2, Depth+1))
1393 unsigned Leaders = KnownZero2.countLeadingOnes();
1394 Leaders = std::max(Leaders,
1395 KnownZero2.countLeadingOnes());
1396 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1399 case Instruction::Call:
1400 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1401 switch (II->getIntrinsicID()) {
1403 case Intrinsic::bswap: {
1404 // If the only bits demanded come from one byte of the bswap result,
1405 // just shift the input byte into position to eliminate the bswap.
1406 unsigned NLZ = DemandedMask.countLeadingZeros();
1407 unsigned NTZ = DemandedMask.countTrailingZeros();
1409 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1410 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1411 // have 14 leading zeros, round to 8.
1414 // If we need exactly one byte, we can do this transformation.
1415 if (BitWidth-NLZ-NTZ == 8) {
1416 unsigned ResultBit = NTZ;
1417 unsigned InputBit = BitWidth-NTZ-8;
1419 // Replace this with either a left or right shift to get the byte into
1421 Instruction *NewVal;
1422 if (InputBit > ResultBit)
1423 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1424 ConstantInt::get(I->getType(), InputBit-ResultBit));
1426 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1427 ConstantInt::get(I->getType(), ResultBit-InputBit));
1428 NewVal->takeName(I);
1429 return InsertNewInstBefore(NewVal, *I);
1432 // TODO: Could compute known zero/one bits based on the input.
1437 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1441 // If the client is only demanding bits that we know, return the known
1443 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1444 return Constant::getIntegerValue(VTy, RHSKnownOne);
1449 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1450 /// any number of elements. DemandedElts contains the set of elements that are
1451 /// actually used by the caller. This method analyzes which elements of the
1452 /// operand are undef and returns that information in UndefElts.
1454 /// If the information about demanded elements can be used to simplify the
1455 /// operation, the operation is simplified, then the resultant value is
1456 /// returned. This returns null if no change was made.
1457 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1460 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1461 APInt EltMask(APInt::getAllOnesValue(VWidth));
1462 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1464 if (isa<UndefValue>(V)) {
1465 // If the entire vector is undefined, just return this info.
1466 UndefElts = EltMask;
1468 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1469 UndefElts = EltMask;
1470 return UndefValue::get(V->getType());
1474 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1475 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1476 Constant *Undef = UndefValue::get(EltTy);
1478 std::vector<Constant*> Elts;
1479 for (unsigned i = 0; i != VWidth; ++i)
1480 if (!DemandedElts[i]) { // If not demanded, set to undef.
1481 Elts.push_back(Undef);
1483 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1484 Elts.push_back(Undef);
1486 } else { // Otherwise, defined.
1487 Elts.push_back(CP->getOperand(i));
1490 // If we changed the constant, return it.
1491 Constant *NewCP = ConstantVector::get(Elts);
1492 return NewCP != CP ? NewCP : 0;
1493 } else if (isa<ConstantAggregateZero>(V)) {
1494 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1497 // Check if this is identity. If so, return 0 since we are not simplifying
1499 if (DemandedElts == ((1ULL << VWidth) -1))
1502 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1503 Constant *Zero = Constant::getNullValue(EltTy);
1504 Constant *Undef = UndefValue::get(EltTy);
1505 std::vector<Constant*> Elts;
1506 for (unsigned i = 0; i != VWidth; ++i) {
1507 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1508 Elts.push_back(Elt);
1510 UndefElts = DemandedElts ^ EltMask;
1511 return ConstantVector::get(Elts);
1514 // Limit search depth.
1518 // If multiple users are using the root value, procede with
1519 // simplification conservatively assuming that all elements
1521 if (!V->hasOneUse()) {
1522 // Quit if we find multiple users of a non-root value though.
1523 // They'll be handled when it's their turn to be visited by
1524 // the main instcombine process.
1526 // TODO: Just compute the UndefElts information recursively.
1529 // Conservatively assume that all elements are needed.
1530 DemandedElts = EltMask;
1533 Instruction *I = dyn_cast<Instruction>(V);
1534 if (!I) return 0; // Only analyze instructions.
1536 bool MadeChange = false;
1537 APInt UndefElts2(VWidth, 0);
1539 switch (I->getOpcode()) {
1542 case Instruction::InsertElement: {
1543 // If this is a variable index, we don't know which element it overwrites.
1544 // demand exactly the same input as we produce.
1545 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1547 // Note that we can't propagate undef elt info, because we don't know
1548 // which elt is getting updated.
1549 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1550 UndefElts2, Depth+1);
1551 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1555 // If this is inserting an element that isn't demanded, remove this
1557 unsigned IdxNo = Idx->getZExtValue();
1558 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1559 return AddSoonDeadInstToWorklist(*I, 0);
1561 // Otherwise, the element inserted overwrites whatever was there, so the
1562 // input demanded set is simpler than the output set.
1563 APInt DemandedElts2 = DemandedElts;
1564 DemandedElts2.clear(IdxNo);
1565 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1566 UndefElts, Depth+1);
1567 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1569 // The inserted element is defined.
1570 UndefElts.clear(IdxNo);
1573 case Instruction::ShuffleVector: {
1574 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1575 uint64_t LHSVWidth =
1576 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1577 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1578 for (unsigned i = 0; i < VWidth; i++) {
1579 if (DemandedElts[i]) {
1580 unsigned MaskVal = Shuffle->getMaskValue(i);
1581 if (MaskVal != -1u) {
1582 assert(MaskVal < LHSVWidth * 2 &&
1583 "shufflevector mask index out of range!");
1584 if (MaskVal < LHSVWidth)
1585 LeftDemanded.set(MaskVal);
1587 RightDemanded.set(MaskVal - LHSVWidth);
1592 APInt UndefElts4(LHSVWidth, 0);
1593 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1594 UndefElts4, Depth+1);
1595 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1597 APInt UndefElts3(LHSVWidth, 0);
1598 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1599 UndefElts3, Depth+1);
1600 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1602 bool NewUndefElts = false;
1603 for (unsigned i = 0; i < VWidth; i++) {
1604 unsigned MaskVal = Shuffle->getMaskValue(i);
1605 if (MaskVal == -1u) {
1607 } else if (MaskVal < LHSVWidth) {
1608 if (UndefElts4[MaskVal]) {
1609 NewUndefElts = true;
1613 if (UndefElts3[MaskVal - LHSVWidth]) {
1614 NewUndefElts = true;
1621 // Add additional discovered undefs.
1622 std::vector<Constant*> Elts;
1623 for (unsigned i = 0; i < VWidth; ++i) {
1625 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
1627 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
1628 Shuffle->getMaskValue(i)));
1630 I->setOperand(2, ConstantVector::get(Elts));
1635 case Instruction::BitCast: {
1636 // Vector->vector casts only.
1637 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1639 unsigned InVWidth = VTy->getNumElements();
1640 APInt InputDemandedElts(InVWidth, 0);
1643 if (VWidth == InVWidth) {
1644 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1645 // elements as are demanded of us.
1647 InputDemandedElts = DemandedElts;
1648 } else if (VWidth > InVWidth) {
1652 // If there are more elements in the result than there are in the source,
1653 // then an input element is live if any of the corresponding output
1654 // elements are live.
1655 Ratio = VWidth/InVWidth;
1656 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1657 if (DemandedElts[OutIdx])
1658 InputDemandedElts.set(OutIdx/Ratio);
1664 // If there are more elements in the source than there are in the result,
1665 // then an input element is live if the corresponding output element is
1667 Ratio = InVWidth/VWidth;
1668 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1669 if (DemandedElts[InIdx/Ratio])
1670 InputDemandedElts.set(InIdx);
1673 // div/rem demand all inputs, because they don't want divide by zero.
1674 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1675 UndefElts2, Depth+1);
1677 I->setOperand(0, TmpV);
1681 UndefElts = UndefElts2;
1682 if (VWidth > InVWidth) {
1683 llvm_unreachable("Unimp");
1684 // If there are more elements in the result than there are in the source,
1685 // then an output element is undef if the corresponding input element is
1687 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1688 if (UndefElts2[OutIdx/Ratio])
1689 UndefElts.set(OutIdx);
1690 } else if (VWidth < InVWidth) {
1691 llvm_unreachable("Unimp");
1692 // If there are more elements in the source than there are in the result,
1693 // then a result element is undef if all of the corresponding input
1694 // elements are undef.
1695 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1696 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1697 if (!UndefElts2[InIdx]) // Not undef?
1698 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1702 case Instruction::And:
1703 case Instruction::Or:
1704 case Instruction::Xor:
1705 case Instruction::Add:
1706 case Instruction::Sub:
1707 case Instruction::Mul:
1708 // div/rem demand all inputs, because they don't want divide by zero.
1709 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1710 UndefElts, Depth+1);
1711 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1712 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1713 UndefElts2, Depth+1);
1714 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1716 // Output elements are undefined if both are undefined. Consider things
1717 // like undef&0. The result is known zero, not undef.
1718 UndefElts &= UndefElts2;
1721 case Instruction::Call: {
1722 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1724 switch (II->getIntrinsicID()) {
1727 // Binary vector operations that work column-wise. A dest element is a
1728 // function of the corresponding input elements from the two inputs.
1729 case Intrinsic::x86_sse_sub_ss:
1730 case Intrinsic::x86_sse_mul_ss:
1731 case Intrinsic::x86_sse_min_ss:
1732 case Intrinsic::x86_sse_max_ss:
1733 case Intrinsic::x86_sse2_sub_sd:
1734 case Intrinsic::x86_sse2_mul_sd:
1735 case Intrinsic::x86_sse2_min_sd:
1736 case Intrinsic::x86_sse2_max_sd:
1737 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1738 UndefElts, Depth+1);
1739 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1740 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1741 UndefElts2, Depth+1);
1742 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1744 // If only the low elt is demanded and this is a scalarizable intrinsic,
1745 // scalarize it now.
1746 if (DemandedElts == 1) {
1747 switch (II->getIntrinsicID()) {
1749 case Intrinsic::x86_sse_sub_ss:
1750 case Intrinsic::x86_sse_mul_ss:
1751 case Intrinsic::x86_sse2_sub_sd:
1752 case Intrinsic::x86_sse2_mul_sd:
1753 // TODO: Lower MIN/MAX/ABS/etc
1754 Value *LHS = II->getOperand(1);
1755 Value *RHS = II->getOperand(2);
1756 // Extract the element as scalars.
1757 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1758 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1759 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1760 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1762 switch (II->getIntrinsicID()) {
1763 default: llvm_unreachable("Case stmts out of sync!");
1764 case Intrinsic::x86_sse_sub_ss:
1765 case Intrinsic::x86_sse2_sub_sd:
1766 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1767 II->getName()), *II);
1769 case Intrinsic::x86_sse_mul_ss:
1770 case Intrinsic::x86_sse2_mul_sd:
1771 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1772 II->getName()), *II);
1777 InsertElementInst::Create(
1778 UndefValue::get(II->getType()), TmpV,
1779 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), II->getName());
1780 InsertNewInstBefore(New, *II);
1781 AddSoonDeadInstToWorklist(*II, 0);
1786 // Output elements are undefined if both are undefined. Consider things
1787 // like undef&0. The result is known zero, not undef.
1788 UndefElts &= UndefElts2;
1794 return MadeChange ? I : 0;
1798 /// AssociativeOpt - Perform an optimization on an associative operator. This
1799 /// function is designed to check a chain of associative operators for a
1800 /// potential to apply a certain optimization. Since the optimization may be
1801 /// applicable if the expression was reassociated, this checks the chain, then
1802 /// reassociates the expression as necessary to expose the optimization
1803 /// opportunity. This makes use of a special Functor, which must define
1804 /// 'shouldApply' and 'apply' methods.
1806 template<typename Functor>
1807 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1808 unsigned Opcode = Root.getOpcode();
1809 Value *LHS = Root.getOperand(0);
1811 // Quick check, see if the immediate LHS matches...
1812 if (F.shouldApply(LHS))
1813 return F.apply(Root);
1815 // Otherwise, if the LHS is not of the same opcode as the root, return.
1816 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1817 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1818 // Should we apply this transform to the RHS?
1819 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1821 // If not to the RHS, check to see if we should apply to the LHS...
1822 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1823 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1827 // If the functor wants to apply the optimization to the RHS of LHSI,
1828 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1830 // Now all of the instructions are in the current basic block, go ahead
1831 // and perform the reassociation.
1832 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1834 // First move the selected RHS to the LHS of the root...
1835 Root.setOperand(0, LHSI->getOperand(1));
1837 // Make what used to be the LHS of the root be the user of the root...
1838 Value *ExtraOperand = TmpLHSI->getOperand(1);
1839 if (&Root == TmpLHSI) {
1840 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1843 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1844 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1845 BasicBlock::iterator ARI = &Root; ++ARI;
1846 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1849 // Now propagate the ExtraOperand down the chain of instructions until we
1851 while (TmpLHSI != LHSI) {
1852 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1853 // Move the instruction to immediately before the chain we are
1854 // constructing to avoid breaking dominance properties.
1855 NextLHSI->moveBefore(ARI);
1858 Value *NextOp = NextLHSI->getOperand(1);
1859 NextLHSI->setOperand(1, ExtraOperand);
1861 ExtraOperand = NextOp;
1864 // Now that the instructions are reassociated, have the functor perform
1865 // the transformation...
1866 return F.apply(Root);
1869 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1876 // AddRHS - Implements: X + X --> X << 1
1879 explicit AddRHS(Value *rhs) : RHS(rhs) {}
1880 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1881 Instruction *apply(BinaryOperator &Add) const {
1882 return BinaryOperator::CreateShl(Add.getOperand(0),
1883 ConstantInt::get(Add.getType(), 1));
1887 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1889 struct AddMaskingAnd {
1891 explicit AddMaskingAnd(Constant *c) : C2(c) {}
1892 bool shouldApply(Value *LHS) const {
1894 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1895 ConstantExpr::getAnd(C1, C2)->isNullValue();
1897 Instruction *apply(BinaryOperator &Add) const {
1898 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1904 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1906 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1907 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1910 // Figure out if the constant is the left or the right argument.
1911 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1912 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1914 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1916 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1917 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1920 Value *Op0 = SO, *Op1 = ConstOperand;
1922 std::swap(Op0, Op1);
1924 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1925 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1926 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1927 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1928 Op0, Op1, SO->getName()+".cmp");
1930 llvm_unreachable("Unknown binary instruction type!");
1932 return IC->InsertNewInstBefore(New, I);
1935 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1936 // constant as the other operand, try to fold the binary operator into the
1937 // select arguments. This also works for Cast instructions, which obviously do
1938 // not have a second operand.
1939 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1941 // Don't modify shared select instructions
1942 if (!SI->hasOneUse()) return 0;
1943 Value *TV = SI->getOperand(1);
1944 Value *FV = SI->getOperand(2);
1946 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1947 // Bool selects with constant operands can be folded to logical ops.
1948 if (SI->getType() == Type::getInt1Ty(*IC->getContext())) return 0;
1950 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1951 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1953 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1960 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1961 /// node as operand #0, see if we can fold the instruction into the PHI (which
1962 /// is only possible if all operands to the PHI are constants).
1963 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1964 PHINode *PN = cast<PHINode>(I.getOperand(0));
1965 unsigned NumPHIValues = PN->getNumIncomingValues();
1966 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1968 // Check to see if all of the operands of the PHI are constants. If there is
1969 // one non-constant value, remember the BB it is. If there is more than one
1970 // or if *it* is a PHI, bail out.
1971 BasicBlock *NonConstBB = 0;
1972 for (unsigned i = 0; i != NumPHIValues; ++i)
1973 if (!isa<Constant>(PN->getIncomingValue(i))) {
1974 if (NonConstBB) return 0; // More than one non-const value.
1975 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1976 NonConstBB = PN->getIncomingBlock(i);
1978 // If the incoming non-constant value is in I's block, we have an infinite
1980 if (NonConstBB == I.getParent())
1984 // If there is exactly one non-constant value, we can insert a copy of the
1985 // operation in that block. However, if this is a critical edge, we would be
1986 // inserting the computation one some other paths (e.g. inside a loop). Only
1987 // do this if the pred block is unconditionally branching into the phi block.
1989 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1990 if (!BI || !BI->isUnconditional()) return 0;
1993 // Okay, we can do the transformation: create the new PHI node.
1994 PHINode *NewPN = PHINode::Create(I.getType(), "");
1995 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1996 InsertNewInstBefore(NewPN, *PN);
1997 NewPN->takeName(PN);
1999 // Next, add all of the operands to the PHI.
2000 if (I.getNumOperands() == 2) {
2001 Constant *C = cast<Constant>(I.getOperand(1));
2002 for (unsigned i = 0; i != NumPHIValues; ++i) {
2004 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2005 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2006 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2008 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2010 assert(PN->getIncomingBlock(i) == NonConstBB);
2011 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2012 InV = BinaryOperator::Create(BO->getOpcode(),
2013 PN->getIncomingValue(i), C, "phitmp",
2014 NonConstBB->getTerminator());
2015 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2016 InV = CmpInst::Create(CI->getOpcode(),
2018 PN->getIncomingValue(i), C, "phitmp",
2019 NonConstBB->getTerminator());
2021 llvm_unreachable("Unknown binop!");
2023 AddToWorkList(cast<Instruction>(InV));
2025 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2028 CastInst *CI = cast<CastInst>(&I);
2029 const Type *RetTy = CI->getType();
2030 for (unsigned i = 0; i != NumPHIValues; ++i) {
2032 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2033 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2035 assert(PN->getIncomingBlock(i) == NonConstBB);
2036 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2037 I.getType(), "phitmp",
2038 NonConstBB->getTerminator());
2039 AddToWorkList(cast<Instruction>(InV));
2041 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2044 return ReplaceInstUsesWith(I, NewPN);
2048 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2049 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2050 /// This basically requires proving that the add in the original type would not
2051 /// overflow to change the sign bit or have a carry out.
2052 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2053 // There are different heuristics we can use for this. Here are some simple
2056 // Add has the property that adding any two 2's complement numbers can only
2057 // have one carry bit which can change a sign. As such, if LHS and RHS each
2058 // have at least two sign bits, we know that the addition of the two values will
2059 // sign extend fine.
2060 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2064 // If one of the operands only has one non-zero bit, and if the other operand
2065 // has a known-zero bit in a more significant place than it (not including the
2066 // sign bit) the ripple may go up to and fill the zero, but won't change the
2067 // sign. For example, (X & ~4) + 1.
2075 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2076 bool Changed = SimplifyCommutative(I);
2077 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2079 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2080 // X + undef -> undef
2081 if (isa<UndefValue>(RHS))
2082 return ReplaceInstUsesWith(I, RHS);
2085 if (RHSC->isNullValue())
2086 return ReplaceInstUsesWith(I, LHS);
2088 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2089 // X + (signbit) --> X ^ signbit
2090 const APInt& Val = CI->getValue();
2091 uint32_t BitWidth = Val.getBitWidth();
2092 if (Val == APInt::getSignBit(BitWidth))
2093 return BinaryOperator::CreateXor(LHS, RHS);
2095 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2096 // (X & 254)+1 -> (X&254)|1
2097 if (SimplifyDemandedInstructionBits(I))
2100 // zext(bool) + C -> bool ? C + 1 : C
2101 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2102 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2103 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2106 if (isa<PHINode>(LHS))
2107 if (Instruction *NV = FoldOpIntoPhi(I))
2110 ConstantInt *XorRHS = 0;
2112 if (isa<ConstantInt>(RHSC) &&
2113 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2114 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2115 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2117 uint32_t Size = TySizeBits / 2;
2118 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2119 APInt CFF80Val(-C0080Val);
2121 if (TySizeBits > Size) {
2122 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2123 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2124 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2125 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2126 // This is a sign extend if the top bits are known zero.
2127 if (!MaskedValueIsZero(XorLHS,
2128 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2129 Size = 0; // Not a sign ext, but can't be any others either.
2134 C0080Val = APIntOps::lshr(C0080Val, Size);
2135 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2136 } while (Size >= 1);
2138 // FIXME: This shouldn't be necessary. When the backends can handle types
2139 // with funny bit widths then this switch statement should be removed. It
2140 // is just here to get the size of the "middle" type back up to something
2141 // that the back ends can handle.
2142 const Type *MiddleType = 0;
2145 case 32: MiddleType = Type::getInt32Ty(*Context); break;
2146 case 16: MiddleType = Type::getInt16Ty(*Context); break;
2147 case 8: MiddleType = Type::getInt8Ty(*Context); break;
2150 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2151 InsertNewInstBefore(NewTrunc, I);
2152 return new SExtInst(NewTrunc, I.getType(), I.getName());
2157 if (I.getType() == Type::getInt1Ty(*Context))
2158 return BinaryOperator::CreateXor(LHS, RHS);
2161 if (I.getType()->isInteger()) {
2162 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2165 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2166 if (RHSI->getOpcode() == Instruction::Sub)
2167 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2168 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2170 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2171 if (LHSI->getOpcode() == Instruction::Sub)
2172 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2173 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2178 // -A + -B --> -(A + B)
2179 if (Value *LHSV = dyn_castNegVal(LHS)) {
2180 if (LHS->getType()->isIntOrIntVector()) {
2181 if (Value *RHSV = dyn_castNegVal(RHS)) {
2182 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2183 InsertNewInstBefore(NewAdd, I);
2184 return BinaryOperator::CreateNeg(NewAdd);
2188 return BinaryOperator::CreateSub(RHS, LHSV);
2192 if (!isa<Constant>(RHS))
2193 if (Value *V = dyn_castNegVal(RHS))
2194 return BinaryOperator::CreateSub(LHS, V);
2198 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2199 if (X == RHS) // X*C + X --> X * (C+1)
2200 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2202 // X*C1 + X*C2 --> X * (C1+C2)
2204 if (X == dyn_castFoldableMul(RHS, C1))
2205 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2208 // X + X*C --> X * (C+1)
2209 if (dyn_castFoldableMul(RHS, C2) == LHS)
2210 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2212 // X + ~X --> -1 since ~X = -X-1
2213 if (dyn_castNotVal(LHS) == RHS ||
2214 dyn_castNotVal(RHS) == LHS)
2215 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2218 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2219 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2220 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2223 // A+B --> A|B iff A and B have no bits set in common.
2224 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2225 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2226 APInt LHSKnownOne(IT->getBitWidth(), 0);
2227 APInt LHSKnownZero(IT->getBitWidth(), 0);
2228 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2229 if (LHSKnownZero != 0) {
2230 APInt RHSKnownOne(IT->getBitWidth(), 0);
2231 APInt RHSKnownZero(IT->getBitWidth(), 0);
2232 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2234 // No bits in common -> bitwise or.
2235 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2236 return BinaryOperator::CreateOr(LHS, RHS);
2240 // W*X + Y*Z --> W * (X+Z) iff W == Y
2241 if (I.getType()->isIntOrIntVector()) {
2242 Value *W, *X, *Y, *Z;
2243 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2244 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2248 } else if (Y == X) {
2250 } else if (X == Z) {
2257 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2258 LHS->getName()), I);
2259 return BinaryOperator::CreateMul(W, NewAdd);
2264 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2266 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2267 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2269 // (X & FF00) + xx00 -> (X+xx00) & FF00
2270 if (LHS->hasOneUse() &&
2271 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2272 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2273 if (Anded == CRHS) {
2274 // See if all bits from the first bit set in the Add RHS up are included
2275 // in the mask. First, get the rightmost bit.
2276 const APInt& AddRHSV = CRHS->getValue();
2278 // Form a mask of all bits from the lowest bit added through the top.
2279 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2281 // See if the and mask includes all of these bits.
2282 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2284 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2285 // Okay, the xform is safe. Insert the new add pronto.
2286 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2287 LHS->getName()), I);
2288 return BinaryOperator::CreateAnd(NewAdd, C2);
2293 // Try to fold constant add into select arguments.
2294 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2295 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2299 // add (select X 0 (sub n A)) A --> select X A n
2301 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2304 SI = dyn_cast<SelectInst>(RHS);
2307 if (SI && SI->hasOneUse()) {
2308 Value *TV = SI->getTrueValue();
2309 Value *FV = SI->getFalseValue();
2312 // Can we fold the add into the argument of the select?
2313 // We check both true and false select arguments for a matching subtract.
2314 if (match(FV, m_Zero()) &&
2315 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2316 // Fold the add into the true select value.
2317 return SelectInst::Create(SI->getCondition(), N, A);
2318 if (match(TV, m_Zero()) &&
2319 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2320 // Fold the add into the false select value.
2321 return SelectInst::Create(SI->getCondition(), A, N);
2325 // Check for (add (sext x), y), see if we can merge this into an
2326 // integer add followed by a sext.
2327 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2328 // (add (sext x), cst) --> (sext (add x, cst'))
2329 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2331 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2332 if (LHSConv->hasOneUse() &&
2333 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2334 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2335 // Insert the new, smaller add.
2336 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2338 InsertNewInstBefore(NewAdd, I);
2339 return new SExtInst(NewAdd, I.getType());
2343 // (add (sext x), (sext y)) --> (sext (add int x, y))
2344 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2345 // Only do this if x/y have the same type, if at last one of them has a
2346 // single use (so we don't increase the number of sexts), and if the
2347 // integer add will not overflow.
2348 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2349 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2350 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2351 RHSConv->getOperand(0))) {
2352 // Insert the new integer add.
2353 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2354 RHSConv->getOperand(0),
2356 InsertNewInstBefore(NewAdd, I);
2357 return new SExtInst(NewAdd, I.getType());
2362 return Changed ? &I : 0;
2365 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2366 bool Changed = SimplifyCommutative(I);
2367 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2369 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2371 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2372 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2373 (I.getType())->getValueAPF()))
2374 return ReplaceInstUsesWith(I, LHS);
2377 if (isa<PHINode>(LHS))
2378 if (Instruction *NV = FoldOpIntoPhi(I))
2383 // -A + -B --> -(A + B)
2384 if (Value *LHSV = dyn_castFNegVal(LHS))
2385 return BinaryOperator::CreateFSub(RHS, LHSV);
2388 if (!isa<Constant>(RHS))
2389 if (Value *V = dyn_castFNegVal(RHS))
2390 return BinaryOperator::CreateFSub(LHS, V);
2392 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2393 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2394 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2395 return ReplaceInstUsesWith(I, LHS);
2397 // Check for (add double (sitofp x), y), see if we can merge this into an
2398 // integer add followed by a promotion.
2399 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2400 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2401 // ... if the constant fits in the integer value. This is useful for things
2402 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2403 // requires a constant pool load, and generally allows the add to be better
2405 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2407 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2408 if (LHSConv->hasOneUse() &&
2409 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2410 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2411 // Insert the new integer add.
2412 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2414 InsertNewInstBefore(NewAdd, I);
2415 return new SIToFPInst(NewAdd, I.getType());
2419 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2420 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2421 // Only do this if x/y have the same type, if at last one of them has a
2422 // single use (so we don't increase the number of int->fp conversions),
2423 // and if the integer add will not overflow.
2424 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2425 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2426 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2427 RHSConv->getOperand(0))) {
2428 // Insert the new integer add.
2429 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2430 RHSConv->getOperand(0),
2432 InsertNewInstBefore(NewAdd, I);
2433 return new SIToFPInst(NewAdd, I.getType());
2438 return Changed ? &I : 0;
2441 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2442 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2444 if (Op0 == Op1) // sub X, X -> 0
2445 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2447 // If this is a 'B = x-(-A)', change to B = x+A...
2448 if (Value *V = dyn_castNegVal(Op1))
2449 return BinaryOperator::CreateAdd(Op0, V);
2451 if (isa<UndefValue>(Op0))
2452 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2453 if (isa<UndefValue>(Op1))
2454 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2456 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2457 // Replace (-1 - A) with (~A)...
2458 if (C->isAllOnesValue())
2459 return BinaryOperator::CreateNot(Op1);
2461 // C - ~X == X + (1+C)
2463 if (match(Op1, m_Not(m_Value(X))))
2464 return BinaryOperator::CreateAdd(X, AddOne(C));
2466 // -(X >>u 31) -> (X >>s 31)
2467 // -(X >>s 31) -> (X >>u 31)
2469 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2470 if (SI->getOpcode() == Instruction::LShr) {
2471 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2472 // Check to see if we are shifting out everything but the sign bit.
2473 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2474 SI->getType()->getPrimitiveSizeInBits()-1) {
2475 // Ok, the transformation is safe. Insert AShr.
2476 return BinaryOperator::Create(Instruction::AShr,
2477 SI->getOperand(0), CU, SI->getName());
2481 else if (SI->getOpcode() == Instruction::AShr) {
2482 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2483 // Check to see if we are shifting out everything but the sign bit.
2484 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2485 SI->getType()->getPrimitiveSizeInBits()-1) {
2486 // Ok, the transformation is safe. Insert LShr.
2487 return BinaryOperator::CreateLShr(
2488 SI->getOperand(0), CU, SI->getName());
2495 // Try to fold constant sub into select arguments.
2496 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2497 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2500 // C - zext(bool) -> bool ? C - 1 : C
2501 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2502 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2503 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2506 if (I.getType() == Type::getInt1Ty(*Context))
2507 return BinaryOperator::CreateXor(Op0, Op1);
2509 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2510 if (Op1I->getOpcode() == Instruction::Add) {
2511 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2512 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2514 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2515 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2517 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2518 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2519 // C1-(X+C2) --> (C1-C2)-X
2520 return BinaryOperator::CreateSub(
2521 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2525 if (Op1I->hasOneUse()) {
2526 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2527 // is not used by anyone else...
2529 if (Op1I->getOpcode() == Instruction::Sub) {
2530 // Swap the two operands of the subexpr...
2531 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2532 Op1I->setOperand(0, IIOp1);
2533 Op1I->setOperand(1, IIOp0);
2535 // Create the new top level add instruction...
2536 return BinaryOperator::CreateAdd(Op0, Op1);
2539 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2541 if (Op1I->getOpcode() == Instruction::And &&
2542 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2543 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2546 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2547 return BinaryOperator::CreateAnd(Op0, NewNot);
2550 // 0 - (X sdiv C) -> (X sdiv -C)
2551 if (Op1I->getOpcode() == Instruction::SDiv)
2552 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2554 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2555 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2556 ConstantExpr::getNeg(DivRHS));
2558 // X - X*C --> X * (1-C)
2559 ConstantInt *C2 = 0;
2560 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2562 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2564 return BinaryOperator::CreateMul(Op0, CP1);
2569 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2570 if (Op0I->getOpcode() == Instruction::Add) {
2571 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2572 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2573 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2574 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2575 } else if (Op0I->getOpcode() == Instruction::Sub) {
2576 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2577 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2583 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2584 if (X == Op1) // X*C - X --> X * (C-1)
2585 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2587 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2588 if (X == dyn_castFoldableMul(Op1, C2))
2589 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2594 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2595 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2597 // If this is a 'B = x-(-A)', change to B = x+A...
2598 if (Value *V = dyn_castFNegVal(Op1))
2599 return BinaryOperator::CreateFAdd(Op0, V);
2601 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2602 if (Op1I->getOpcode() == Instruction::FAdd) {
2603 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2604 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
2606 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2607 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
2615 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2616 /// comparison only checks the sign bit. If it only checks the sign bit, set
2617 /// TrueIfSigned if the result of the comparison is true when the input value is
2619 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2620 bool &TrueIfSigned) {
2622 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2623 TrueIfSigned = true;
2624 return RHS->isZero();
2625 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2626 TrueIfSigned = true;
2627 return RHS->isAllOnesValue();
2628 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2629 TrueIfSigned = false;
2630 return RHS->isAllOnesValue();
2631 case ICmpInst::ICMP_UGT:
2632 // True if LHS u> RHS and RHS == high-bit-mask - 1
2633 TrueIfSigned = true;
2634 return RHS->getValue() ==
2635 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2636 case ICmpInst::ICMP_UGE:
2637 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2638 TrueIfSigned = true;
2639 return RHS->getValue().isSignBit();
2645 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2646 bool Changed = SimplifyCommutative(I);
2647 Value *Op0 = I.getOperand(0);
2649 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2650 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2652 // Simplify mul instructions with a constant RHS...
2653 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2654 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2656 // ((X << C1)*C2) == (X * (C2 << C1))
2657 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2658 if (SI->getOpcode() == Instruction::Shl)
2659 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2660 return BinaryOperator::CreateMul(SI->getOperand(0),
2661 ConstantExpr::getShl(CI, ShOp));
2664 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2665 if (CI->equalsInt(1)) // X * 1 == X
2666 return ReplaceInstUsesWith(I, Op0);
2667 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2668 return BinaryOperator::CreateNeg(Op0, I.getName());
2670 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2671 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2672 return BinaryOperator::CreateShl(Op0,
2673 ConstantInt::get(Op0->getType(), Val.logBase2()));
2675 } else if (isa<VectorType>(Op1->getType())) {
2676 if (Op1->isNullValue())
2677 return ReplaceInstUsesWith(I, Op1);
2679 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2680 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2681 return BinaryOperator::CreateNeg(Op0, I.getName());
2683 // As above, vector X*splat(1.0) -> X in all defined cases.
2684 if (Constant *Splat = Op1V->getSplatValue()) {
2685 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2686 if (CI->equalsInt(1))
2687 return ReplaceInstUsesWith(I, Op0);
2692 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2693 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2694 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2695 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2696 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2698 InsertNewInstBefore(Add, I);
2699 Value *C1C2 = ConstantExpr::getMul(Op1,
2700 cast<Constant>(Op0I->getOperand(1)));
2701 return BinaryOperator::CreateAdd(Add, C1C2);
2705 // Try to fold constant mul into select arguments.
2706 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2707 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2710 if (isa<PHINode>(Op0))
2711 if (Instruction *NV = FoldOpIntoPhi(I))
2715 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2716 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2717 return BinaryOperator::CreateMul(Op0v, Op1v);
2719 // (X / Y) * Y = X - (X % Y)
2720 // (X / Y) * -Y = (X % Y) - X
2722 Value *Op1 = I.getOperand(1);
2723 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2725 (BO->getOpcode() != Instruction::UDiv &&
2726 BO->getOpcode() != Instruction::SDiv)) {
2728 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2730 Value *Neg = dyn_castNegVal(Op1);
2731 if (BO && BO->hasOneUse() &&
2732 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2733 (BO->getOpcode() == Instruction::UDiv ||
2734 BO->getOpcode() == Instruction::SDiv)) {
2735 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2737 // If the division is exact, X % Y is zero.
2738 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
2739 if (SDiv->isExact()) {
2741 return ReplaceInstUsesWith(I, Op0BO);
2743 return BinaryOperator::CreateNeg(Op0BO);
2747 if (BO->getOpcode() == Instruction::UDiv)
2748 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2750 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2752 InsertNewInstBefore(Rem, I);
2756 return BinaryOperator::CreateSub(Op0BO, Rem);
2758 return BinaryOperator::CreateSub(Rem, Op0BO);
2762 if (I.getType() == Type::getInt1Ty(*Context))
2763 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2765 // If one of the operands of the multiply is a cast from a boolean value, then
2766 // we know the bool is either zero or one, so this is a 'masking' multiply.
2767 // See if we can simplify things based on how the boolean was originally
2769 CastInst *BoolCast = 0;
2770 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2771 if (CI->getOperand(0)->getType() == Type::getInt1Ty(*Context))
2774 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2775 if (CI->getOperand(0)->getType() == Type::getInt1Ty(*Context))
2778 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2779 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2780 const Type *SCOpTy = SCIOp0->getType();
2783 // If the icmp is true iff the sign bit of X is set, then convert this
2784 // multiply into a shift/and combination.
2785 if (isa<ConstantInt>(SCIOp1) &&
2786 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2788 // Shift the X value right to turn it into "all signbits".
2789 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2790 SCOpTy->getPrimitiveSizeInBits()-1);
2792 InsertNewInstBefore(
2793 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2794 BoolCast->getOperand(0)->getName()+
2797 // If the multiply type is not the same as the source type, sign extend
2798 // or truncate to the multiply type.
2799 if (I.getType() != V->getType()) {
2800 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2801 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2802 Instruction::CastOps opcode =
2803 (SrcBits == DstBits ? Instruction::BitCast :
2804 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2805 V = InsertCastBefore(opcode, V, I.getType(), I);
2808 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2809 return BinaryOperator::CreateAnd(V, OtherOp);
2814 return Changed ? &I : 0;
2817 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2818 bool Changed = SimplifyCommutative(I);
2819 Value *Op0 = I.getOperand(0);
2821 // Simplify mul instructions with a constant RHS...
2822 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2823 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2824 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2825 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2826 if (Op1F->isExactlyValue(1.0))
2827 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2828 } else if (isa<VectorType>(Op1->getType())) {
2829 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2830 // As above, vector X*splat(1.0) -> X in all defined cases.
2831 if (Constant *Splat = Op1V->getSplatValue()) {
2832 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2833 if (F->isExactlyValue(1.0))
2834 return ReplaceInstUsesWith(I, Op0);
2839 // Try to fold constant mul into select arguments.
2840 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2841 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2844 if (isa<PHINode>(Op0))
2845 if (Instruction *NV = FoldOpIntoPhi(I))
2849 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
2850 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1)))
2851 return BinaryOperator::CreateFMul(Op0v, Op1v);
2853 return Changed ? &I : 0;
2856 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2858 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2859 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2861 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2862 int NonNullOperand = -1;
2863 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2864 if (ST->isNullValue())
2866 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2867 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2868 if (ST->isNullValue())
2871 if (NonNullOperand == -1)
2874 Value *SelectCond = SI->getOperand(0);
2876 // Change the div/rem to use 'Y' instead of the select.
2877 I.setOperand(1, SI->getOperand(NonNullOperand));
2879 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2880 // problem. However, the select, or the condition of the select may have
2881 // multiple uses. Based on our knowledge that the operand must be non-zero,
2882 // propagate the known value for the select into other uses of it, and
2883 // propagate a known value of the condition into its other users.
2885 // If the select and condition only have a single use, don't bother with this,
2887 if (SI->use_empty() && SelectCond->hasOneUse())
2890 // Scan the current block backward, looking for other uses of SI.
2891 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2893 while (BBI != BBFront) {
2895 // If we found a call to a function, we can't assume it will return, so
2896 // information from below it cannot be propagated above it.
2897 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2900 // Replace uses of the select or its condition with the known values.
2901 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2904 *I = SI->getOperand(NonNullOperand);
2906 } else if (*I == SelectCond) {
2907 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
2908 ConstantInt::getFalse(*Context);
2913 // If we past the instruction, quit looking for it.
2916 if (&*BBI == SelectCond)
2919 // If we ran out of things to eliminate, break out of the loop.
2920 if (SelectCond == 0 && SI == 0)
2928 /// This function implements the transforms on div instructions that work
2929 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2930 /// used by the visitors to those instructions.
2931 /// @brief Transforms common to all three div instructions
2932 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2933 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2935 // undef / X -> 0 for integer.
2936 // undef / X -> undef for FP (the undef could be a snan).
2937 if (isa<UndefValue>(Op0)) {
2938 if (Op0->getType()->isFPOrFPVector())
2939 return ReplaceInstUsesWith(I, Op0);
2940 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2943 // X / undef -> undef
2944 if (isa<UndefValue>(Op1))
2945 return ReplaceInstUsesWith(I, Op1);
2950 /// This function implements the transforms common to both integer division
2951 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2952 /// division instructions.
2953 /// @brief Common integer divide transforms
2954 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2955 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2957 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2959 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2960 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
2961 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2962 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2965 Constant *CI = ConstantInt::get(I.getType(), 1);
2966 return ReplaceInstUsesWith(I, CI);
2969 if (Instruction *Common = commonDivTransforms(I))
2972 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2973 // This does not apply for fdiv.
2974 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2977 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2979 if (RHS->equalsInt(1))
2980 return ReplaceInstUsesWith(I, Op0);
2982 // (X / C1) / C2 -> X / (C1*C2)
2983 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2984 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2985 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2986 if (MultiplyOverflows(RHS, LHSRHS,
2987 I.getOpcode()==Instruction::SDiv))
2988 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2990 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2991 ConstantExpr::getMul(RHS, LHSRHS));
2994 if (!RHS->isZero()) { // avoid X udiv 0
2995 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2996 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2998 if (isa<PHINode>(Op0))
2999 if (Instruction *NV = FoldOpIntoPhi(I))
3004 // 0 / X == 0, we don't need to preserve faults!
3005 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3006 if (LHS->equalsInt(0))
3007 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3009 // It can't be division by zero, hence it must be division by one.
3010 if (I.getType() == Type::getInt1Ty(*Context))
3011 return ReplaceInstUsesWith(I, Op0);
3013 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3014 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3017 return ReplaceInstUsesWith(I, Op0);
3023 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3024 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3026 // Handle the integer div common cases
3027 if (Instruction *Common = commonIDivTransforms(I))
3030 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3031 // X udiv C^2 -> X >> C
3032 // Check to see if this is an unsigned division with an exact power of 2,
3033 // if so, convert to a right shift.
3034 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3035 return BinaryOperator::CreateLShr(Op0,
3036 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3038 // X udiv C, where C >= signbit
3039 if (C->getValue().isNegative()) {
3040 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
3042 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3043 ConstantInt::get(I.getType(), 1));
3047 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3048 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3049 if (RHSI->getOpcode() == Instruction::Shl &&
3050 isa<ConstantInt>(RHSI->getOperand(0))) {
3051 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3052 if (C1.isPowerOf2()) {
3053 Value *N = RHSI->getOperand(1);
3054 const Type *NTy = N->getType();
3055 if (uint32_t C2 = C1.logBase2()) {
3056 Constant *C2V = ConstantInt::get(NTy, C2);
3057 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3059 return BinaryOperator::CreateLShr(Op0, N);
3064 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3065 // where C1&C2 are powers of two.
3066 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3067 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3068 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3069 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3070 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3071 // Compute the shift amounts
3072 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3073 // Construct the "on true" case of the select
3074 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3075 Instruction *TSI = BinaryOperator::CreateLShr(
3076 Op0, TC, SI->getName()+".t");
3077 TSI = InsertNewInstBefore(TSI, I);
3079 // Construct the "on false" case of the select
3080 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3081 Instruction *FSI = BinaryOperator::CreateLShr(
3082 Op0, FC, SI->getName()+".f");
3083 FSI = InsertNewInstBefore(FSI, I);
3085 // construct the select instruction and return it.
3086 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3092 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3093 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3095 // Handle the integer div common cases
3096 if (Instruction *Common = commonIDivTransforms(I))
3099 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3101 if (RHS->isAllOnesValue())
3102 return BinaryOperator::CreateNeg(Op0);
3104 // sdiv X, C --> ashr X, log2(C)
3105 if (cast<SDivOperator>(&I)->isExact() &&
3106 RHS->getValue().isNonNegative() &&
3107 RHS->getValue().isPowerOf2()) {
3108 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3109 RHS->getValue().exactLogBase2());
3110 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3113 // -X/C --> X/-C provided the negation doesn't overflow.
3114 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
3115 if (isa<Constant>(Sub->getOperand(0)) &&
3116 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
3117 Sub->hasNoSignedWrap())
3118 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
3119 ConstantExpr::getNeg(RHS));
3122 // If the sign bits of both operands are zero (i.e. we can prove they are
3123 // unsigned inputs), turn this into a udiv.
3124 if (I.getType()->isInteger()) {
3125 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3126 if (MaskedValueIsZero(Op0, Mask)) {
3127 if (MaskedValueIsZero(Op1, Mask)) {
3128 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3129 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3131 ConstantInt *ShiftedInt;
3132 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3133 ShiftedInt->getValue().isPowerOf2()) {
3134 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3135 // Safe because the only negative value (1 << Y) can take on is
3136 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3137 // the sign bit set.
3138 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3146 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3147 return commonDivTransforms(I);
3150 /// This function implements the transforms on rem instructions that work
3151 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3152 /// is used by the visitors to those instructions.
3153 /// @brief Transforms common to all three rem instructions
3154 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3155 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3157 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3158 if (I.getType()->isFPOrFPVector())
3159 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3160 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3162 if (isa<UndefValue>(Op1))
3163 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3165 // Handle cases involving: rem X, (select Cond, Y, Z)
3166 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3172 /// This function implements the transforms common to both integer remainder
3173 /// instructions (urem and srem). It is called by the visitors to those integer
3174 /// remainder instructions.
3175 /// @brief Common integer remainder transforms
3176 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3177 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3179 if (Instruction *common = commonRemTransforms(I))
3182 // 0 % X == 0 for integer, we don't need to preserve faults!
3183 if (Constant *LHS = dyn_cast<Constant>(Op0))
3184 if (LHS->isNullValue())
3185 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3187 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3188 // X % 0 == undef, we don't need to preserve faults!
3189 if (RHS->equalsInt(0))
3190 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3192 if (RHS->equalsInt(1)) // X % 1 == 0
3193 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3195 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3196 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3197 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3199 } else if (isa<PHINode>(Op0I)) {
3200 if (Instruction *NV = FoldOpIntoPhi(I))
3204 // See if we can fold away this rem instruction.
3205 if (SimplifyDemandedInstructionBits(I))
3213 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3214 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3216 if (Instruction *common = commonIRemTransforms(I))
3219 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3220 // X urem C^2 -> X and C
3221 // Check to see if this is an unsigned remainder with an exact power of 2,
3222 // if so, convert to a bitwise and.
3223 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3224 if (C->getValue().isPowerOf2())
3225 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3228 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3229 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3230 if (RHSI->getOpcode() == Instruction::Shl &&
3231 isa<ConstantInt>(RHSI->getOperand(0))) {
3232 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3233 Constant *N1 = Constant::getAllOnesValue(I.getType());
3234 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3236 return BinaryOperator::CreateAnd(Op0, Add);
3241 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3242 // where C1&C2 are powers of two.
3243 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3244 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3245 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3246 // STO == 0 and SFO == 0 handled above.
3247 if ((STO->getValue().isPowerOf2()) &&
3248 (SFO->getValue().isPowerOf2())) {
3249 Value *TrueAnd = InsertNewInstBefore(
3250 BinaryOperator::CreateAnd(Op0, SubOne(STO),
3251 SI->getName()+".t"), I);
3252 Value *FalseAnd = InsertNewInstBefore(
3253 BinaryOperator::CreateAnd(Op0, SubOne(SFO),
3254 SI->getName()+".f"), I);
3255 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3263 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3264 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3266 // Handle the integer rem common cases
3267 if (Instruction *common = commonIRemTransforms(I))
3270 if (Value *RHSNeg = dyn_castNegVal(Op1))
3271 if (!isa<Constant>(RHSNeg) ||
3272 (isa<ConstantInt>(RHSNeg) &&
3273 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3275 AddUsesToWorkList(I);
3276 I.setOperand(1, RHSNeg);
3280 // If the sign bits of both operands are zero (i.e. we can prove they are
3281 // unsigned inputs), turn this into a urem.
3282 if (I.getType()->isInteger()) {
3283 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3284 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3285 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3286 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3290 // If it's a constant vector, flip any negative values positive.
3291 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3292 unsigned VWidth = RHSV->getNumOperands();
3294 bool hasNegative = false;
3295 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3296 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3297 if (RHS->getValue().isNegative())
3301 std::vector<Constant *> Elts(VWidth);
3302 for (unsigned i = 0; i != VWidth; ++i) {
3303 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3304 if (RHS->getValue().isNegative())
3305 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3311 Constant *NewRHSV = ConstantVector::get(Elts);
3312 if (NewRHSV != RHSV) {
3313 AddUsesToWorkList(I);
3314 I.setOperand(1, NewRHSV);
3323 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3324 return commonRemTransforms(I);
3327 // isOneBitSet - Return true if there is exactly one bit set in the specified
3329 static bool isOneBitSet(const ConstantInt *CI) {
3330 return CI->getValue().isPowerOf2();
3333 // isHighOnes - Return true if the constant is of the form 1+0+.
3334 // This is the same as lowones(~X).
3335 static bool isHighOnes(const ConstantInt *CI) {
3336 return (~CI->getValue() + 1).isPowerOf2();
3339 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3340 /// are carefully arranged to allow folding of expressions such as:
3342 /// (A < B) | (A > B) --> (A != B)
3344 /// Note that this is only valid if the first and second predicates have the
3345 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3347 /// Three bits are used to represent the condition, as follows:
3352 /// <=> Value Definition
3353 /// 000 0 Always false
3360 /// 111 7 Always true
3362 static unsigned getICmpCode(const ICmpInst *ICI) {
3363 switch (ICI->getPredicate()) {
3365 case ICmpInst::ICMP_UGT: return 1; // 001
3366 case ICmpInst::ICMP_SGT: return 1; // 001
3367 case ICmpInst::ICMP_EQ: return 2; // 010
3368 case ICmpInst::ICMP_UGE: return 3; // 011
3369 case ICmpInst::ICMP_SGE: return 3; // 011
3370 case ICmpInst::ICMP_ULT: return 4; // 100
3371 case ICmpInst::ICMP_SLT: return 4; // 100
3372 case ICmpInst::ICMP_NE: return 5; // 101
3373 case ICmpInst::ICMP_ULE: return 6; // 110
3374 case ICmpInst::ICMP_SLE: return 6; // 110
3377 llvm_unreachable("Invalid ICmp predicate!");
3382 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3383 /// predicate into a three bit mask. It also returns whether it is an ordered
3384 /// predicate by reference.
3385 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3388 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3389 case FCmpInst::FCMP_UNO: return 0; // 000
3390 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3391 case FCmpInst::FCMP_UGT: return 1; // 001
3392 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3393 case FCmpInst::FCMP_UEQ: return 2; // 010
3394 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3395 case FCmpInst::FCMP_UGE: return 3; // 011
3396 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3397 case FCmpInst::FCMP_ULT: return 4; // 100
3398 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3399 case FCmpInst::FCMP_UNE: return 5; // 101
3400 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3401 case FCmpInst::FCMP_ULE: return 6; // 110
3404 // Not expecting FCMP_FALSE and FCMP_TRUE;
3405 llvm_unreachable("Unexpected FCmp predicate!");
3410 /// getICmpValue - This is the complement of getICmpCode, which turns an
3411 /// opcode and two operands into either a constant true or false, or a brand
3412 /// new ICmp instruction. The sign is passed in to determine which kind
3413 /// of predicate to use in the new icmp instruction.
3414 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3415 LLVMContext *Context) {
3417 default: llvm_unreachable("Illegal ICmp code!");
3418 case 0: return ConstantInt::getFalse(*Context);
3421 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3423 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3424 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3427 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3429 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3432 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3434 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3435 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3438 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3440 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3441 case 7: return ConstantInt::getTrue(*Context);
3445 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3446 /// opcode and two operands into either a FCmp instruction. isordered is passed
3447 /// in to determine which kind of predicate to use in the new fcmp instruction.
3448 static Value *getFCmpValue(bool isordered, unsigned code,
3449 Value *LHS, Value *RHS, LLVMContext *Context) {
3451 default: llvm_unreachable("Illegal FCmp code!");
3454 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3456 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3459 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3461 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3464 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3466 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3469 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3471 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3474 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3476 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3479 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3481 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3484 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3486 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3487 case 7: return ConstantInt::getTrue(*Context);
3491 /// PredicatesFoldable - Return true if both predicates match sign or if at
3492 /// least one of them is an equality comparison (which is signless).
3493 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3494 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3495 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3496 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3500 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3501 struct FoldICmpLogical {
3504 ICmpInst::Predicate pred;
3505 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3506 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3507 pred(ICI->getPredicate()) {}
3508 bool shouldApply(Value *V) const {
3509 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3510 if (PredicatesFoldable(pred, ICI->getPredicate()))
3511 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3512 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3515 Instruction *apply(Instruction &Log) const {
3516 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3517 if (ICI->getOperand(0) != LHS) {
3518 assert(ICI->getOperand(1) == LHS);
3519 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3522 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3523 unsigned LHSCode = getICmpCode(ICI);
3524 unsigned RHSCode = getICmpCode(RHSICI);
3526 switch (Log.getOpcode()) {
3527 case Instruction::And: Code = LHSCode & RHSCode; break;
3528 case Instruction::Or: Code = LHSCode | RHSCode; break;
3529 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3530 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3533 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3534 ICmpInst::isSignedPredicate(ICI->getPredicate());
3536 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3537 if (Instruction *I = dyn_cast<Instruction>(RV))
3539 // Otherwise, it's a constant boolean value...
3540 return IC.ReplaceInstUsesWith(Log, RV);
3543 } // end anonymous namespace
3545 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3546 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3547 // guaranteed to be a binary operator.
3548 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3550 ConstantInt *AndRHS,
3551 BinaryOperator &TheAnd) {
3552 Value *X = Op->getOperand(0);
3553 Constant *Together = 0;
3555 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3557 switch (Op->getOpcode()) {
3558 case Instruction::Xor:
3559 if (Op->hasOneUse()) {
3560 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3561 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3562 InsertNewInstBefore(And, TheAnd);
3564 return BinaryOperator::CreateXor(And, Together);
3567 case Instruction::Or:
3568 if (Together == AndRHS) // (X | C) & C --> C
3569 return ReplaceInstUsesWith(TheAnd, AndRHS);
3571 if (Op->hasOneUse() && Together != OpRHS) {
3572 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3573 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3574 InsertNewInstBefore(Or, TheAnd);
3576 return BinaryOperator::CreateAnd(Or, AndRHS);
3579 case Instruction::Add:
3580 if (Op->hasOneUse()) {
3581 // Adding a one to a single bit bit-field should be turned into an XOR
3582 // of the bit. First thing to check is to see if this AND is with a
3583 // single bit constant.
3584 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3586 // If there is only one bit set...
3587 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3588 // Ok, at this point, we know that we are masking the result of the
3589 // ADD down to exactly one bit. If the constant we are adding has
3590 // no bits set below this bit, then we can eliminate the ADD.
3591 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3593 // Check to see if any bits below the one bit set in AndRHSV are set.
3594 if ((AddRHS & (AndRHSV-1)) == 0) {
3595 // If not, the only thing that can effect the output of the AND is
3596 // the bit specified by AndRHSV. If that bit is set, the effect of
3597 // the XOR is to toggle the bit. If it is clear, then the ADD has
3599 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3600 TheAnd.setOperand(0, X);
3603 // Pull the XOR out of the AND.
3604 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3605 InsertNewInstBefore(NewAnd, TheAnd);
3606 NewAnd->takeName(Op);
3607 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3614 case Instruction::Shl: {
3615 // We know that the AND will not produce any of the bits shifted in, so if
3616 // the anded constant includes them, clear them now!
3618 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3619 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3620 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3621 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3623 if (CI->getValue() == ShlMask) {
3624 // Masking out bits that the shift already masks
3625 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3626 } else if (CI != AndRHS) { // Reducing bits set in and.
3627 TheAnd.setOperand(1, CI);
3632 case Instruction::LShr:
3634 // We know that the AND will not produce any of the bits shifted in, so if
3635 // the anded constant includes them, clear them now! This only applies to
3636 // unsigned shifts, because a signed shr may bring in set bits!
3638 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3639 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3640 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3641 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3643 if (CI->getValue() == ShrMask) {
3644 // Masking out bits that the shift already masks.
3645 return ReplaceInstUsesWith(TheAnd, Op);
3646 } else if (CI != AndRHS) {
3647 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3652 case Instruction::AShr:
3654 // See if this is shifting in some sign extension, then masking it out
3656 if (Op->hasOneUse()) {
3657 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3658 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3659 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3660 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3661 if (C == AndRHS) { // Masking out bits shifted in.
3662 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3663 // Make the argument unsigned.
3664 Value *ShVal = Op->getOperand(0);
3665 ShVal = InsertNewInstBefore(
3666 BinaryOperator::CreateLShr(ShVal, OpRHS,
3667 Op->getName()), TheAnd);
3668 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3677 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3678 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3679 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3680 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3681 /// insert new instructions.
3682 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3683 bool isSigned, bool Inside,
3685 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3686 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3687 "Lo is not <= Hi in range emission code!");
3690 if (Lo == Hi) // Trivially false.
3691 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3693 // V >= Min && V < Hi --> V < Hi
3694 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3695 ICmpInst::Predicate pred = (isSigned ?
3696 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3697 return new ICmpInst(pred, V, Hi);
3700 // Emit V-Lo <u Hi-Lo
3701 Constant *NegLo = ConstantExpr::getNeg(Lo);
3702 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3703 InsertNewInstBefore(Add, IB);
3704 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3705 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3708 if (Lo == Hi) // Trivially true.
3709 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3711 // V < Min || V >= Hi -> V > Hi-1
3712 Hi = SubOne(cast<ConstantInt>(Hi));
3713 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3714 ICmpInst::Predicate pred = (isSigned ?
3715 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3716 return new ICmpInst(pred, V, Hi);
3719 // Emit V-Lo >u Hi-1-Lo
3720 // Note that Hi has already had one subtracted from it, above.
3721 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3722 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3723 InsertNewInstBefore(Add, IB);
3724 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3725 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3728 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3729 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3730 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3731 // not, since all 1s are not contiguous.
3732 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3733 const APInt& V = Val->getValue();
3734 uint32_t BitWidth = Val->getType()->getBitWidth();
3735 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3737 // look for the first zero bit after the run of ones
3738 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3739 // look for the first non-zero bit
3740 ME = V.getActiveBits();
3744 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3745 /// where isSub determines whether the operator is a sub. If we can fold one of
3746 /// the following xforms:
3748 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3749 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3750 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3752 /// return (A +/- B).
3754 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3755 ConstantInt *Mask, bool isSub,
3757 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3758 if (!LHSI || LHSI->getNumOperands() != 2 ||
3759 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3761 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3763 switch (LHSI->getOpcode()) {
3765 case Instruction::And:
3766 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3767 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3768 if ((Mask->getValue().countLeadingZeros() +
3769 Mask->getValue().countPopulation()) ==
3770 Mask->getValue().getBitWidth())
3773 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3774 // part, we don't need any explicit masks to take them out of A. If that
3775 // is all N is, ignore it.
3776 uint32_t MB = 0, ME = 0;
3777 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3778 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3779 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3780 if (MaskedValueIsZero(RHS, Mask))
3785 case Instruction::Or:
3786 case Instruction::Xor:
3787 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3788 if ((Mask->getValue().countLeadingZeros() +
3789 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3790 && ConstantExpr::getAnd(N, Mask)->isNullValue())
3797 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3799 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3800 return InsertNewInstBefore(New, I);
3803 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3804 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3805 ICmpInst *LHS, ICmpInst *RHS) {
3807 ConstantInt *LHSCst, *RHSCst;
3808 ICmpInst::Predicate LHSCC, RHSCC;
3810 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3811 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3812 m_ConstantInt(LHSCst))) ||
3813 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3814 m_ConstantInt(RHSCst))))
3817 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3818 // where C is a power of 2
3819 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3820 LHSCst->getValue().isPowerOf2()) {
3821 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3822 InsertNewInstBefore(NewOr, I);
3823 return new ICmpInst(LHSCC, NewOr, LHSCst);
3826 // From here on, we only handle:
3827 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3828 if (Val != Val2) return 0;
3830 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3831 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3832 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3833 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3834 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3837 // We can't fold (ugt x, C) & (sgt x, C2).
3838 if (!PredicatesFoldable(LHSCC, RHSCC))
3841 // Ensure that the larger constant is on the RHS.
3843 if (ICmpInst::isSignedPredicate(LHSCC) ||
3844 (ICmpInst::isEquality(LHSCC) &&
3845 ICmpInst::isSignedPredicate(RHSCC)))
3846 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3848 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3851 std::swap(LHS, RHS);
3852 std::swap(LHSCst, RHSCst);
3853 std::swap(LHSCC, RHSCC);
3856 // At this point, we know we have have two icmp instructions
3857 // comparing a value against two constants and and'ing the result
3858 // together. Because of the above check, we know that we only have
3859 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3860 // (from the FoldICmpLogical check above), that the two constants
3861 // are not equal and that the larger constant is on the RHS
3862 assert(LHSCst != RHSCst && "Compares not folded above?");
3865 default: llvm_unreachable("Unknown integer condition code!");
3866 case ICmpInst::ICMP_EQ:
3868 default: llvm_unreachable("Unknown integer condition code!");
3869 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3870 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3871 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3872 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3873 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3874 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3875 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3876 return ReplaceInstUsesWith(I, LHS);
3878 case ICmpInst::ICMP_NE:
3880 default: llvm_unreachable("Unknown integer condition code!");
3881 case ICmpInst::ICMP_ULT:
3882 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3883 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3884 break; // (X != 13 & X u< 15) -> no change
3885 case ICmpInst::ICMP_SLT:
3886 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3887 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3888 break; // (X != 13 & X s< 15) -> no change
3889 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3890 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3891 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3892 return ReplaceInstUsesWith(I, RHS);
3893 case ICmpInst::ICMP_NE:
3894 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3895 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3896 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3897 Val->getName()+".off");
3898 InsertNewInstBefore(Add, I);
3899 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3900 ConstantInt::get(Add->getType(), 1));
3902 break; // (X != 13 & X != 15) -> no change
3905 case ICmpInst::ICMP_ULT:
3907 default: llvm_unreachable("Unknown integer condition code!");
3908 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3909 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3910 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3911 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3913 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3914 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3915 return ReplaceInstUsesWith(I, LHS);
3916 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3920 case ICmpInst::ICMP_SLT:
3922 default: llvm_unreachable("Unknown integer condition code!");
3923 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3924 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3925 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3926 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3928 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3929 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3930 return ReplaceInstUsesWith(I, LHS);
3931 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3935 case ICmpInst::ICMP_UGT:
3937 default: llvm_unreachable("Unknown integer condition code!");
3938 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3939 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3940 return ReplaceInstUsesWith(I, RHS);
3941 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3943 case ICmpInst::ICMP_NE:
3944 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3945 return new ICmpInst(LHSCC, Val, RHSCst);
3946 break; // (X u> 13 & X != 15) -> no change
3947 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3948 return InsertRangeTest(Val, AddOne(LHSCst),
3949 RHSCst, false, true, I);
3950 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3954 case ICmpInst::ICMP_SGT:
3956 default: llvm_unreachable("Unknown integer condition code!");
3957 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3958 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3959 return ReplaceInstUsesWith(I, RHS);
3960 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3962 case ICmpInst::ICMP_NE:
3963 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3964 return new ICmpInst(LHSCC, Val, RHSCst);
3965 break; // (X s> 13 & X != 15) -> no change
3966 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3967 return InsertRangeTest(Val, AddOne(LHSCst),
3968 RHSCst, true, true, I);
3969 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3978 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
3981 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3982 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3983 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3984 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3985 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3986 // If either of the constants are nans, then the whole thing returns
3988 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3989 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3990 return new FCmpInst(FCmpInst::FCMP_ORD,
3991 LHS->getOperand(0), RHS->getOperand(0));
3994 // Handle vector zeros. This occurs because the canonical form of
3995 // "fcmp ord x,x" is "fcmp ord x, 0".
3996 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
3997 isa<ConstantAggregateZero>(RHS->getOperand(1)))
3998 return new FCmpInst(FCmpInst::FCMP_ORD,
3999 LHS->getOperand(0), RHS->getOperand(0));
4003 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4004 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4005 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4008 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4009 // Swap RHS operands to match LHS.
4010 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4011 std::swap(Op1LHS, Op1RHS);
4014 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4015 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4017 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
4019 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
4020 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4021 if (Op0CC == FCmpInst::FCMP_TRUE)
4022 return ReplaceInstUsesWith(I, RHS);
4023 if (Op1CC == FCmpInst::FCMP_TRUE)
4024 return ReplaceInstUsesWith(I, LHS);
4028 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4029 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4031 std::swap(LHS, RHS);
4032 std::swap(Op0Pred, Op1Pred);
4033 std::swap(Op0Ordered, Op1Ordered);
4036 // uno && ueq -> uno && (uno || eq) -> ueq
4037 // ord && olt -> ord && (ord && lt) -> olt
4038 if (Op0Ordered == Op1Ordered)
4039 return ReplaceInstUsesWith(I, RHS);
4041 // uno && oeq -> uno && (ord && eq) -> false
4042 // uno && ord -> false
4044 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4045 // ord && ueq -> ord && (uno || eq) -> oeq
4046 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4047 Op0LHS, Op0RHS, Context));
4055 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4056 bool Changed = SimplifyCommutative(I);
4057 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4059 if (isa<UndefValue>(Op1)) // X & undef -> 0
4060 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4064 return ReplaceInstUsesWith(I, Op1);
4066 // See if we can simplify any instructions used by the instruction whose sole
4067 // purpose is to compute bits we don't care about.
4068 if (SimplifyDemandedInstructionBits(I))
4070 if (isa<VectorType>(I.getType())) {
4071 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4072 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4073 return ReplaceInstUsesWith(I, I.getOperand(0));
4074 } else if (isa<ConstantAggregateZero>(Op1)) {
4075 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4079 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4080 const APInt& AndRHSMask = AndRHS->getValue();
4081 APInt NotAndRHS(~AndRHSMask);
4083 // Optimize a variety of ((val OP C1) & C2) combinations...
4084 if (isa<BinaryOperator>(Op0)) {
4085 Instruction *Op0I = cast<Instruction>(Op0);
4086 Value *Op0LHS = Op0I->getOperand(0);
4087 Value *Op0RHS = Op0I->getOperand(1);
4088 switch (Op0I->getOpcode()) {
4089 case Instruction::Xor:
4090 case Instruction::Or:
4091 // If the mask is only needed on one incoming arm, push it up.
4092 if (Op0I->hasOneUse()) {
4093 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4094 // Not masking anything out for the LHS, move to RHS.
4095 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4096 Op0RHS->getName()+".masked");
4097 InsertNewInstBefore(NewRHS, I);
4098 return BinaryOperator::Create(
4099 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4101 if (!isa<Constant>(Op0RHS) &&
4102 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4103 // Not masking anything out for the RHS, move to LHS.
4104 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4105 Op0LHS->getName()+".masked");
4106 InsertNewInstBefore(NewLHS, I);
4107 return BinaryOperator::Create(
4108 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4113 case Instruction::Add:
4114 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4115 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4116 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4117 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4118 return BinaryOperator::CreateAnd(V, AndRHS);
4119 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4120 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4123 case Instruction::Sub:
4124 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4125 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4126 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4127 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4128 return BinaryOperator::CreateAnd(V, AndRHS);
4130 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4131 // has 1's for all bits that the subtraction with A might affect.
4132 if (Op0I->hasOneUse()) {
4133 uint32_t BitWidth = AndRHSMask.getBitWidth();
4134 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4135 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4137 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4138 if (!(A && A->isZero()) && // avoid infinite recursion.
4139 MaskedValueIsZero(Op0LHS, Mask)) {
4140 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
4141 InsertNewInstBefore(NewNeg, I);
4142 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4147 case Instruction::Shl:
4148 case Instruction::LShr:
4149 // (1 << x) & 1 --> zext(x == 0)
4150 // (1 >> x) & 1 --> zext(x == 0)
4151 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4152 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ,
4153 Op0RHS, Constant::getNullValue(I.getType()));
4154 InsertNewInstBefore(NewICmp, I);
4155 return new ZExtInst(NewICmp, I.getType());
4160 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4161 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4163 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4164 // If this is an integer truncation or change from signed-to-unsigned, and
4165 // if the source is an and/or with immediate, transform it. This
4166 // frequently occurs for bitfield accesses.
4167 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4168 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4169 CastOp->getNumOperands() == 2)
4170 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4171 if (CastOp->getOpcode() == Instruction::And) {
4172 // Change: and (cast (and X, C1) to T), C2
4173 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4174 // This will fold the two constants together, which may allow
4175 // other simplifications.
4176 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4177 CastOp->getOperand(0), I.getType(),
4178 CastOp->getName()+".shrunk");
4179 NewCast = InsertNewInstBefore(NewCast, I);
4180 // trunc_or_bitcast(C1)&C2
4182 ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4183 C3 = ConstantExpr::getAnd(C3, AndRHS);
4184 return BinaryOperator::CreateAnd(NewCast, C3);
4185 } else if (CastOp->getOpcode() == Instruction::Or) {
4186 // Change: and (cast (or X, C1) to T), C2
4187 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4189 ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4190 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4192 return ReplaceInstUsesWith(I, AndRHS);
4198 // Try to fold constant and into select arguments.
4199 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4200 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4202 if (isa<PHINode>(Op0))
4203 if (Instruction *NV = FoldOpIntoPhi(I))
4207 Value *Op0NotVal = dyn_castNotVal(Op0);
4208 Value *Op1NotVal = dyn_castNotVal(Op1);
4210 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4211 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4213 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4214 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4215 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4216 I.getName()+".demorgan");
4217 InsertNewInstBefore(Or, I);
4218 return BinaryOperator::CreateNot(Or);
4222 Value *A = 0, *B = 0, *C = 0, *D = 0;
4223 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4224 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4225 return ReplaceInstUsesWith(I, Op1);
4227 // (A|B) & ~(A&B) -> A^B
4228 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4229 if ((A == C && B == D) || (A == D && B == C))
4230 return BinaryOperator::CreateXor(A, B);
4234 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4235 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4236 return ReplaceInstUsesWith(I, Op0);
4238 // ~(A&B) & (A|B) -> A^B
4239 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4240 if ((A == C && B == D) || (A == D && B == C))
4241 return BinaryOperator::CreateXor(A, B);
4245 if (Op0->hasOneUse() &&
4246 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4247 if (A == Op1) { // (A^B)&A -> A&(A^B)
4248 I.swapOperands(); // Simplify below
4249 std::swap(Op0, Op1);
4250 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4251 cast<BinaryOperator>(Op0)->swapOperands();
4252 I.swapOperands(); // Simplify below
4253 std::swap(Op0, Op1);
4257 if (Op1->hasOneUse() &&
4258 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4259 if (B == Op0) { // B&(A^B) -> B&(B^A)
4260 cast<BinaryOperator>(Op1)->swapOperands();
4263 if (A == Op0) { // A&(A^B) -> A & ~B
4264 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4265 InsertNewInstBefore(NotB, I);
4266 return BinaryOperator::CreateAnd(A, NotB);
4270 // (A&((~A)|B)) -> A&B
4271 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4272 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4273 return BinaryOperator::CreateAnd(A, Op1);
4274 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4275 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4276 return BinaryOperator::CreateAnd(A, Op0);
4279 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4280 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4281 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4284 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4285 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4289 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4290 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4291 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4292 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4293 const Type *SrcTy = Op0C->getOperand(0)->getType();
4294 if (SrcTy == Op1C->getOperand(0)->getType() &&
4295 SrcTy->isIntOrIntVector() &&
4296 // Only do this if the casts both really cause code to be generated.
4297 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4299 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4301 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4302 Op1C->getOperand(0),
4304 InsertNewInstBefore(NewOp, I);
4305 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4309 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4310 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4311 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4312 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4313 SI0->getOperand(1) == SI1->getOperand(1) &&
4314 (SI0->hasOneUse() || SI1->hasOneUse())) {
4315 Instruction *NewOp =
4316 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4318 SI0->getName()), I);
4319 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4320 SI1->getOperand(1));
4324 // If and'ing two fcmp, try combine them into one.
4325 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4326 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4327 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4331 return Changed ? &I : 0;
4334 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4335 /// capable of providing pieces of a bswap. The subexpression provides pieces
4336 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4337 /// the expression came from the corresponding "byte swapped" byte in some other
4338 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4339 /// we know that the expression deposits the low byte of %X into the high byte
4340 /// of the bswap result and that all other bytes are zero. This expression is
4341 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4344 /// This function returns true if the match was unsuccessful and false if so.
4345 /// On entry to the function the "OverallLeftShift" is a signed integer value
4346 /// indicating the number of bytes that the subexpression is later shifted. For
4347 /// example, if the expression is later right shifted by 16 bits, the
4348 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4349 /// byte of ByteValues is actually being set.
4351 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4352 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4353 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4354 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4355 /// always in the local (OverallLeftShift) coordinate space.
4357 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4358 SmallVector<Value*, 8> &ByteValues) {
4359 if (Instruction *I = dyn_cast<Instruction>(V)) {
4360 // If this is an or instruction, it may be an inner node of the bswap.
4361 if (I->getOpcode() == Instruction::Or) {
4362 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4364 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4368 // If this is a logical shift by a constant multiple of 8, recurse with
4369 // OverallLeftShift and ByteMask adjusted.
4370 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4372 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4373 // Ensure the shift amount is defined and of a byte value.
4374 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4377 unsigned ByteShift = ShAmt >> 3;
4378 if (I->getOpcode() == Instruction::Shl) {
4379 // X << 2 -> collect(X, +2)
4380 OverallLeftShift += ByteShift;
4381 ByteMask >>= ByteShift;
4383 // X >>u 2 -> collect(X, -2)
4384 OverallLeftShift -= ByteShift;
4385 ByteMask <<= ByteShift;
4386 ByteMask &= (~0U >> (32-ByteValues.size()));
4389 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4390 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4392 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4396 // If this is a logical 'and' with a mask that clears bytes, clear the
4397 // corresponding bytes in ByteMask.
4398 if (I->getOpcode() == Instruction::And &&
4399 isa<ConstantInt>(I->getOperand(1))) {
4400 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4401 unsigned NumBytes = ByteValues.size();
4402 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4403 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4405 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4406 // If this byte is masked out by a later operation, we don't care what
4408 if ((ByteMask & (1 << i)) == 0)
4411 // If the AndMask is all zeros for this byte, clear the bit.
4412 APInt MaskB = AndMask & Byte;
4414 ByteMask &= ~(1U << i);
4418 // If the AndMask is not all ones for this byte, it's not a bytezap.
4422 // Otherwise, this byte is kept.
4425 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4430 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4431 // the input value to the bswap. Some observations: 1) if more than one byte
4432 // is demanded from this input, then it could not be successfully assembled
4433 // into a byteswap. At least one of the two bytes would not be aligned with
4434 // their ultimate destination.
4435 if (!isPowerOf2_32(ByteMask)) return true;
4436 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4438 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4439 // is demanded, it needs to go into byte 0 of the result. This means that the
4440 // byte needs to be shifted until it lands in the right byte bucket. The
4441 // shift amount depends on the position: if the byte is coming from the high
4442 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4443 // low part, it must be shifted left.
4444 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4445 if (InputByteNo < ByteValues.size()/2) {
4446 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4449 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4453 // If the destination byte value is already defined, the values are or'd
4454 // together, which isn't a bswap (unless it's an or of the same bits).
4455 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4457 ByteValues[DestByteNo] = V;
4461 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4462 /// If so, insert the new bswap intrinsic and return it.
4463 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4464 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4465 if (!ITy || ITy->getBitWidth() % 16 ||
4466 // ByteMask only allows up to 32-byte values.
4467 ITy->getBitWidth() > 32*8)
4468 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4470 /// ByteValues - For each byte of the result, we keep track of which value
4471 /// defines each byte.
4472 SmallVector<Value*, 8> ByteValues;
4473 ByteValues.resize(ITy->getBitWidth()/8);
4475 // Try to find all the pieces corresponding to the bswap.
4476 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4477 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4480 // Check to see if all of the bytes come from the same value.
4481 Value *V = ByteValues[0];
4482 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4484 // Check to make sure that all of the bytes come from the same value.
4485 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4486 if (ByteValues[i] != V)
4488 const Type *Tys[] = { ITy };
4489 Module *M = I.getParent()->getParent()->getParent();
4490 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4491 return CallInst::Create(F, V);
4494 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4495 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4496 /// we can simplify this expression to "cond ? C : D or B".
4497 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4499 LLVMContext *Context) {
4500 // If A is not a select of -1/0, this cannot match.
4502 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4505 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4506 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4507 return SelectInst::Create(Cond, C, B);
4508 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4509 return SelectInst::Create(Cond, C, B);
4510 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4511 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4512 return SelectInst::Create(Cond, C, D);
4513 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4514 return SelectInst::Create(Cond, C, D);
4518 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4519 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4520 ICmpInst *LHS, ICmpInst *RHS) {
4522 ConstantInt *LHSCst, *RHSCst;
4523 ICmpInst::Predicate LHSCC, RHSCC;
4525 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4526 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4527 m_ConstantInt(LHSCst))) ||
4528 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4529 m_ConstantInt(RHSCst))))
4532 // From here on, we only handle:
4533 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4534 if (Val != Val2) return 0;
4536 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4537 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4538 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4539 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4540 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4543 // We can't fold (ugt x, C) | (sgt x, C2).
4544 if (!PredicatesFoldable(LHSCC, RHSCC))
4547 // Ensure that the larger constant is on the RHS.
4549 if (ICmpInst::isSignedPredicate(LHSCC) ||
4550 (ICmpInst::isEquality(LHSCC) &&
4551 ICmpInst::isSignedPredicate(RHSCC)))
4552 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4554 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4557 std::swap(LHS, RHS);
4558 std::swap(LHSCst, RHSCst);
4559 std::swap(LHSCC, RHSCC);
4562 // At this point, we know we have have two icmp instructions
4563 // comparing a value against two constants and or'ing the result
4564 // together. Because of the above check, we know that we only have
4565 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4566 // FoldICmpLogical check above), that the two constants are not
4568 assert(LHSCst != RHSCst && "Compares not folded above?");
4571 default: llvm_unreachable("Unknown integer condition code!");
4572 case ICmpInst::ICMP_EQ:
4574 default: llvm_unreachable("Unknown integer condition code!");
4575 case ICmpInst::ICMP_EQ:
4576 if (LHSCst == SubOne(RHSCst)) {
4577 // (X == 13 | X == 14) -> X-13 <u 2
4578 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4579 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4580 Val->getName()+".off");
4581 InsertNewInstBefore(Add, I);
4582 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4583 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4585 break; // (X == 13 | X == 15) -> no change
4586 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4587 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4589 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4590 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4591 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4592 return ReplaceInstUsesWith(I, RHS);
4595 case ICmpInst::ICMP_NE:
4597 default: llvm_unreachable("Unknown integer condition code!");
4598 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4599 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4600 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4601 return ReplaceInstUsesWith(I, LHS);
4602 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4603 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4604 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4605 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4608 case ICmpInst::ICMP_ULT:
4610 default: llvm_unreachable("Unknown integer condition code!");
4611 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4613 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4614 // If RHSCst is [us]MAXINT, it is always false. Not handling
4615 // this can cause overflow.
4616 if (RHSCst->isMaxValue(false))
4617 return ReplaceInstUsesWith(I, LHS);
4618 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4620 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4622 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4623 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4624 return ReplaceInstUsesWith(I, RHS);
4625 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4629 case ICmpInst::ICMP_SLT:
4631 default: llvm_unreachable("Unknown integer condition code!");
4632 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4634 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4635 // If RHSCst is [us]MAXINT, it is always false. Not handling
4636 // this can cause overflow.
4637 if (RHSCst->isMaxValue(true))
4638 return ReplaceInstUsesWith(I, LHS);
4639 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4641 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4643 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4644 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4645 return ReplaceInstUsesWith(I, RHS);
4646 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4650 case ICmpInst::ICMP_UGT:
4652 default: llvm_unreachable("Unknown integer condition code!");
4653 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4654 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4655 return ReplaceInstUsesWith(I, LHS);
4656 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4658 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4659 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4660 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4661 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4665 case ICmpInst::ICMP_SGT:
4667 default: llvm_unreachable("Unknown integer condition code!");
4668 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4669 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4670 return ReplaceInstUsesWith(I, LHS);
4671 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4673 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4674 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4675 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4676 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4684 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
4686 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4687 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4688 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4689 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4690 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4691 // If either of the constants are nans, then the whole thing returns
4693 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4694 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4696 // Otherwise, no need to compare the two constants, compare the
4698 return new FCmpInst(FCmpInst::FCMP_UNO,
4699 LHS->getOperand(0), RHS->getOperand(0));
4702 // Handle vector zeros. This occurs because the canonical form of
4703 // "fcmp uno x,x" is "fcmp uno x, 0".
4704 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4705 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4706 return new FCmpInst(FCmpInst::FCMP_UNO,
4707 LHS->getOperand(0), RHS->getOperand(0));
4712 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4713 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4714 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4716 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4717 // Swap RHS operands to match LHS.
4718 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4719 std::swap(Op1LHS, Op1RHS);
4721 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4722 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4724 return new FCmpInst((FCmpInst::Predicate)Op0CC,
4726 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
4727 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4728 if (Op0CC == FCmpInst::FCMP_FALSE)
4729 return ReplaceInstUsesWith(I, RHS);
4730 if (Op1CC == FCmpInst::FCMP_FALSE)
4731 return ReplaceInstUsesWith(I, LHS);
4734 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4735 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4736 if (Op0Ordered == Op1Ordered) {
4737 // If both are ordered or unordered, return a new fcmp with
4738 // or'ed predicates.
4739 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4740 Op0LHS, Op0RHS, Context);
4741 if (Instruction *I = dyn_cast<Instruction>(RV))
4743 // Otherwise, it's a constant boolean value...
4744 return ReplaceInstUsesWith(I, RV);
4750 /// FoldOrWithConstants - This helper function folds:
4752 /// ((A | B) & C1) | (B & C2)
4758 /// when the XOR of the two constants is "all ones" (-1).
4759 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4760 Value *A, Value *B, Value *C) {
4761 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4765 ConstantInt *CI2 = 0;
4766 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4768 APInt Xor = CI1->getValue() ^ CI2->getValue();
4769 if (!Xor.isAllOnesValue()) return 0;
4771 if (V1 == A || V1 == B) {
4772 Instruction *NewOp =
4773 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4774 return BinaryOperator::CreateOr(NewOp, V1);
4780 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4781 bool Changed = SimplifyCommutative(I);
4782 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4784 if (isa<UndefValue>(Op1)) // X | undef -> -1
4785 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4789 return ReplaceInstUsesWith(I, Op0);
4791 // See if we can simplify any instructions used by the instruction whose sole
4792 // purpose is to compute bits we don't care about.
4793 if (SimplifyDemandedInstructionBits(I))
4795 if (isa<VectorType>(I.getType())) {
4796 if (isa<ConstantAggregateZero>(Op1)) {
4797 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4798 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4799 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4800 return ReplaceInstUsesWith(I, I.getOperand(1));
4805 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4806 ConstantInt *C1 = 0; Value *X = 0;
4807 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4808 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
4810 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4811 InsertNewInstBefore(Or, I);
4813 return BinaryOperator::CreateAnd(Or,
4814 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
4817 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4818 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
4820 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4821 InsertNewInstBefore(Or, I);
4823 return BinaryOperator::CreateXor(Or,
4824 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
4827 // Try to fold constant and into select arguments.
4828 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4829 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4831 if (isa<PHINode>(Op0))
4832 if (Instruction *NV = FoldOpIntoPhi(I))
4836 Value *A = 0, *B = 0;
4837 ConstantInt *C1 = 0, *C2 = 0;
4839 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4840 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4841 return ReplaceInstUsesWith(I, Op1);
4842 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4843 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4844 return ReplaceInstUsesWith(I, Op0);
4846 // (A | B) | C and A | (B | C) -> bswap if possible.
4847 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4848 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4849 match(Op1, m_Or(m_Value(), m_Value())) ||
4850 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4851 match(Op1, m_Shift(m_Value(), m_Value())))) {
4852 if (Instruction *BSwap = MatchBSwap(I))
4856 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4857 if (Op0->hasOneUse() &&
4858 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4859 MaskedValueIsZero(Op1, C1->getValue())) {
4860 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4861 InsertNewInstBefore(NOr, I);
4863 return BinaryOperator::CreateXor(NOr, C1);
4866 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4867 if (Op1->hasOneUse() &&
4868 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4869 MaskedValueIsZero(Op0, C1->getValue())) {
4870 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4871 InsertNewInstBefore(NOr, I);
4873 return BinaryOperator::CreateXor(NOr, C1);
4877 Value *C = 0, *D = 0;
4878 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4879 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4880 Value *V1 = 0, *V2 = 0, *V3 = 0;
4881 C1 = dyn_cast<ConstantInt>(C);
4882 C2 = dyn_cast<ConstantInt>(D);
4883 if (C1 && C2) { // (A & C1)|(B & C2)
4884 // If we have: ((V + N) & C1) | (V & C2)
4885 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4886 // replace with V+N.
4887 if (C1->getValue() == ~C2->getValue()) {
4888 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4889 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4890 // Add commutes, try both ways.
4891 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4892 return ReplaceInstUsesWith(I, A);
4893 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4894 return ReplaceInstUsesWith(I, A);
4896 // Or commutes, try both ways.
4897 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4898 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4899 // Add commutes, try both ways.
4900 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4901 return ReplaceInstUsesWith(I, B);
4902 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4903 return ReplaceInstUsesWith(I, B);
4906 V1 = 0; V2 = 0; V3 = 0;
4909 // Check to see if we have any common things being and'ed. If so, find the
4910 // terms for V1 & (V2|V3).
4911 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4912 if (A == B) // (A & C)|(A & D) == A & (C|D)
4913 V1 = A, V2 = C, V3 = D;
4914 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4915 V1 = A, V2 = B, V3 = C;
4916 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4917 V1 = C, V2 = A, V3 = D;
4918 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4919 V1 = C, V2 = A, V3 = B;
4923 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4924 return BinaryOperator::CreateAnd(V1, Or);
4928 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4929 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4931 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4933 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4935 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4938 // ((A&~B)|(~A&B)) -> A^B
4939 if ((match(C, m_Not(m_Specific(D))) &&
4940 match(B, m_Not(m_Specific(A)))))
4941 return BinaryOperator::CreateXor(A, D);
4942 // ((~B&A)|(~A&B)) -> A^B
4943 if ((match(A, m_Not(m_Specific(D))) &&
4944 match(B, m_Not(m_Specific(C)))))
4945 return BinaryOperator::CreateXor(C, D);
4946 // ((A&~B)|(B&~A)) -> A^B
4947 if ((match(C, m_Not(m_Specific(B))) &&
4948 match(D, m_Not(m_Specific(A)))))
4949 return BinaryOperator::CreateXor(A, B);
4950 // ((~B&A)|(B&~A)) -> A^B
4951 if ((match(A, m_Not(m_Specific(B))) &&
4952 match(D, m_Not(m_Specific(C)))))
4953 return BinaryOperator::CreateXor(C, B);
4956 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4957 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4958 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4959 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4960 SI0->getOperand(1) == SI1->getOperand(1) &&
4961 (SI0->hasOneUse() || SI1->hasOneUse())) {
4962 Instruction *NewOp =
4963 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4965 SI0->getName()), I);
4966 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4967 SI1->getOperand(1));
4971 // ((A|B)&1)|(B&-2) -> (A&1) | B
4972 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4973 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4974 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4975 if (Ret) return Ret;
4977 // (B&-2)|((A|B)&1) -> (A&1) | B
4978 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4979 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4980 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4981 if (Ret) return Ret;
4984 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4985 if (A == Op1) // ~A | A == -1
4986 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4990 // Note, A is still live here!
4991 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4993 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4995 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4996 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4997 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4998 I.getName()+".demorgan"), I);
4999 return BinaryOperator::CreateNot(And);
5003 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
5004 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
5005 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5008 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
5009 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
5013 // fold (or (cast A), (cast B)) -> (cast (or A, B))
5014 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5015 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5016 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
5017 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
5018 !isa<ICmpInst>(Op1C->getOperand(0))) {
5019 const Type *SrcTy = Op0C->getOperand(0)->getType();
5020 if (SrcTy == Op1C->getOperand(0)->getType() &&
5021 SrcTy->isIntOrIntVector() &&
5022 // Only do this if the casts both really cause code to be
5024 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5026 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5028 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
5029 Op1C->getOperand(0),
5031 InsertNewInstBefore(NewOp, I);
5032 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5039 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
5040 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
5041 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
5042 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
5046 return Changed ? &I : 0;
5051 // XorSelf - Implements: X ^ X --> 0
5054 XorSelf(Value *rhs) : RHS(rhs) {}
5055 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5056 Instruction *apply(BinaryOperator &Xor) const {
5063 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5064 bool Changed = SimplifyCommutative(I);
5065 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5067 if (isa<UndefValue>(Op1)) {
5068 if (isa<UndefValue>(Op0))
5069 // Handle undef ^ undef -> 0 special case. This is a common
5071 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5072 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5075 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5076 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
5077 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5078 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5081 // See if we can simplify any instructions used by the instruction whose sole
5082 // purpose is to compute bits we don't care about.
5083 if (SimplifyDemandedInstructionBits(I))
5085 if (isa<VectorType>(I.getType()))
5086 if (isa<ConstantAggregateZero>(Op1))
5087 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5089 // Is this a ~ operation?
5090 if (Value *NotOp = dyn_castNotVal(&I)) {
5091 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5092 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5093 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5094 if (Op0I->getOpcode() == Instruction::And ||
5095 Op0I->getOpcode() == Instruction::Or) {
5096 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
5097 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5099 BinaryOperator::CreateNot(Op0I->getOperand(1),
5100 Op0I->getOperand(1)->getName()+".not");
5101 InsertNewInstBefore(NotY, I);
5102 if (Op0I->getOpcode() == Instruction::And)
5103 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5105 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5112 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5113 if (RHS == ConstantInt::getTrue(*Context) && Op0->hasOneUse()) {
5114 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5115 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5116 return new ICmpInst(ICI->getInversePredicate(),
5117 ICI->getOperand(0), ICI->getOperand(1));
5119 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5120 return new FCmpInst(FCI->getInversePredicate(),
5121 FCI->getOperand(0), FCI->getOperand(1));
5124 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5125 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5126 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5127 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5128 Instruction::CastOps Opcode = Op0C->getOpcode();
5129 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
5130 if (RHS == ConstantExpr::getCast(Opcode,
5131 ConstantInt::getTrue(*Context),
5132 Op0C->getDestTy())) {
5133 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
5134 CI->getOpcode(), CI->getInversePredicate(),
5135 CI->getOperand(0), CI->getOperand(1)), I);
5136 NewCI->takeName(CI);
5137 return CastInst::Create(Opcode, NewCI, Op0C->getType());
5144 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5145 // ~(c-X) == X-c-1 == X+(-c-1)
5146 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5147 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5148 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5149 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5150 ConstantInt::get(I.getType(), 1));
5151 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5154 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5155 if (Op0I->getOpcode() == Instruction::Add) {
5156 // ~(X-c) --> (-c-1)-X
5157 if (RHS->isAllOnesValue()) {
5158 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5159 return BinaryOperator::CreateSub(
5160 ConstantExpr::getSub(NegOp0CI,
5161 ConstantInt::get(I.getType(), 1)),
5162 Op0I->getOperand(0));
5163 } else if (RHS->getValue().isSignBit()) {
5164 // (X + C) ^ signbit -> (X + C + signbit)
5165 Constant *C = ConstantInt::get(*Context,
5166 RHS->getValue() + Op0CI->getValue());
5167 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5170 } else if (Op0I->getOpcode() == Instruction::Or) {
5171 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5172 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5173 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5174 // Anything in both C1 and C2 is known to be zero, remove it from
5176 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5177 NewRHS = ConstantExpr::getAnd(NewRHS,
5178 ConstantExpr::getNot(CommonBits));
5179 AddToWorkList(Op0I);
5180 I.setOperand(0, Op0I->getOperand(0));
5181 I.setOperand(1, NewRHS);
5188 // Try to fold constant and into select arguments.
5189 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5190 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5192 if (isa<PHINode>(Op0))
5193 if (Instruction *NV = FoldOpIntoPhi(I))
5197 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5199 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5201 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5203 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5206 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5209 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5210 if (A == Op0) { // B^(B|A) == (A|B)^B
5211 Op1I->swapOperands();
5213 std::swap(Op0, Op1);
5214 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5215 I.swapOperands(); // Simplified below.
5216 std::swap(Op0, Op1);
5218 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5219 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5220 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5221 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5222 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5224 if (A == Op0) { // A^(A&B) -> A^(B&A)
5225 Op1I->swapOperands();
5228 if (B == Op0) { // A^(B&A) -> (B&A)^A
5229 I.swapOperands(); // Simplified below.
5230 std::swap(Op0, Op1);
5235 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5238 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5239 Op0I->hasOneUse()) {
5240 if (A == Op1) // (B|A)^B == (A|B)^B
5242 if (B == Op1) { // (A|B)^B == A & ~B
5244 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5245 return BinaryOperator::CreateAnd(A, NotB);
5247 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5248 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5249 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5250 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5251 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5253 if (A == Op1) // (A&B)^A -> (B&A)^A
5255 if (B == Op1 && // (B&A)^A == ~B & A
5256 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5258 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5259 return BinaryOperator::CreateAnd(N, Op1);
5264 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5265 if (Op0I && Op1I && Op0I->isShift() &&
5266 Op0I->getOpcode() == Op1I->getOpcode() &&
5267 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5268 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5269 Instruction *NewOp =
5270 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5271 Op1I->getOperand(0),
5272 Op0I->getName()), I);
5273 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5274 Op1I->getOperand(1));
5278 Value *A, *B, *C, *D;
5279 // (A & B)^(A | B) -> A ^ B
5280 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5281 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5282 if ((A == C && B == D) || (A == D && B == C))
5283 return BinaryOperator::CreateXor(A, B);
5285 // (A | B)^(A & B) -> A ^ B
5286 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5287 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5288 if ((A == C && B == D) || (A == D && B == C))
5289 return BinaryOperator::CreateXor(A, B);
5293 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5294 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5295 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5296 // (X & Y)^(X & Y) -> (Y^Z) & X
5297 Value *X = 0, *Y = 0, *Z = 0;
5299 X = A, Y = B, Z = D;
5301 X = A, Y = B, Z = C;
5303 X = B, Y = A, Z = D;
5305 X = B, Y = A, Z = C;
5308 Instruction *NewOp =
5309 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5310 return BinaryOperator::CreateAnd(NewOp, X);
5315 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5316 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5317 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5320 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5321 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5322 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5323 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5324 const Type *SrcTy = Op0C->getOperand(0)->getType();
5325 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5326 // Only do this if the casts both really cause code to be generated.
5327 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5329 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5331 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5332 Op1C->getOperand(0),
5334 InsertNewInstBefore(NewOp, I);
5335 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5340 return Changed ? &I : 0;
5343 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5344 LLVMContext *Context) {
5345 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5348 static bool HasAddOverflow(ConstantInt *Result,
5349 ConstantInt *In1, ConstantInt *In2,
5352 if (In2->getValue().isNegative())
5353 return Result->getValue().sgt(In1->getValue());
5355 return Result->getValue().slt(In1->getValue());
5357 return Result->getValue().ult(In1->getValue());
5360 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5361 /// overflowed for this type.
5362 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5363 Constant *In2, LLVMContext *Context,
5364 bool IsSigned = false) {
5365 Result = ConstantExpr::getAdd(In1, In2);
5367 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5368 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5369 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5370 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5371 ExtractElement(In1, Idx, Context),
5372 ExtractElement(In2, Idx, Context),
5379 return HasAddOverflow(cast<ConstantInt>(Result),
5380 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5384 static bool HasSubOverflow(ConstantInt *Result,
5385 ConstantInt *In1, ConstantInt *In2,
5388 if (In2->getValue().isNegative())
5389 return Result->getValue().slt(In1->getValue());
5391 return Result->getValue().sgt(In1->getValue());
5393 return Result->getValue().ugt(In1->getValue());
5396 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5397 /// overflowed for this type.
5398 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5399 Constant *In2, LLVMContext *Context,
5400 bool IsSigned = false) {
5401 Result = ConstantExpr::getSub(In1, In2);
5403 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5404 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5405 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5406 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5407 ExtractElement(In1, Idx, Context),
5408 ExtractElement(In2, Idx, Context),
5415 return HasSubOverflow(cast<ConstantInt>(Result),
5416 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5420 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5421 /// code necessary to compute the offset from the base pointer (without adding
5422 /// in the base pointer). Return the result as a signed integer of intptr size.
5423 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5424 TargetData &TD = *IC.getTargetData();
5425 gep_type_iterator GTI = gep_type_begin(GEP);
5426 const Type *IntPtrTy = TD.getIntPtrType(I.getContext());
5427 LLVMContext *Context = IC.getContext();
5428 Value *Result = Constant::getNullValue(IntPtrTy);
5430 // Build a mask for high order bits.
5431 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5432 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5434 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5437 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5438 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5439 if (OpC->isZero()) continue;
5441 // Handle a struct index, which adds its field offset to the pointer.
5442 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5443 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5445 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5447 ConstantInt::get(*Context,
5448 RC->getValue() + APInt(IntPtrWidth, Size));
5450 Result = IC.InsertNewInstBefore(
5451 BinaryOperator::CreateAdd(Result,
5452 ConstantInt::get(IntPtrTy, Size),
5453 GEP->getName()+".offs"), I);
5457 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5459 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5460 Scale = ConstantExpr::getMul(OC, Scale);
5461 if (Constant *RC = dyn_cast<Constant>(Result))
5462 Result = ConstantExpr::getAdd(RC, Scale);
5464 // Emit an add instruction.
5465 Result = IC.InsertNewInstBefore(
5466 BinaryOperator::CreateAdd(Result, Scale,
5467 GEP->getName()+".offs"), I);
5471 // Convert to correct type.
5472 if (Op->getType() != IntPtrTy) {
5473 if (Constant *OpC = dyn_cast<Constant>(Op))
5474 Op = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true);
5476 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5478 Op->getName()+".c"), I);
5481 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5482 if (Constant *OpC = dyn_cast<Constant>(Op))
5483 Op = ConstantExpr::getMul(OpC, Scale);
5484 else // We'll let instcombine(mul) convert this to a shl if possible.
5485 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5486 GEP->getName()+".idx"), I);
5489 // Emit an add instruction.
5490 if (isa<Constant>(Op) && isa<Constant>(Result))
5491 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5492 cast<Constant>(Result));
5494 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5495 GEP->getName()+".offs"), I);
5501 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5502 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5503 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5504 /// be complex, and scales are involved. The above expression would also be
5505 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5506 /// This later form is less amenable to optimization though, and we are allowed
5507 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5509 /// If we can't emit an optimized form for this expression, this returns null.
5511 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5513 TargetData &TD = *IC.getTargetData();
5514 gep_type_iterator GTI = gep_type_begin(GEP);
5516 // Check to see if this gep only has a single variable index. If so, and if
5517 // any constant indices are a multiple of its scale, then we can compute this
5518 // in terms of the scale of the variable index. For example, if the GEP
5519 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5520 // because the expression will cross zero at the same point.
5521 unsigned i, e = GEP->getNumOperands();
5523 for (i = 1; i != e; ++i, ++GTI) {
5524 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5525 // Compute the aggregate offset of constant indices.
5526 if (CI->isZero()) continue;
5528 // Handle a struct index, which adds its field offset to the pointer.
5529 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5530 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5532 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5533 Offset += Size*CI->getSExtValue();
5536 // Found our variable index.
5541 // If there are no variable indices, we must have a constant offset, just
5542 // evaluate it the general way.
5543 if (i == e) return 0;
5545 Value *VariableIdx = GEP->getOperand(i);
5546 // Determine the scale factor of the variable element. For example, this is
5547 // 4 if the variable index is into an array of i32.
5548 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5550 // Verify that there are no other variable indices. If so, emit the hard way.
5551 for (++i, ++GTI; i != e; ++i, ++GTI) {
5552 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5555 // Compute the aggregate offset of constant indices.
5556 if (CI->isZero()) continue;
5558 // Handle a struct index, which adds its field offset to the pointer.
5559 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5560 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5562 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5563 Offset += Size*CI->getSExtValue();
5567 // Okay, we know we have a single variable index, which must be a
5568 // pointer/array/vector index. If there is no offset, life is simple, return
5570 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5572 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5573 // we don't need to bother extending: the extension won't affect where the
5574 // computation crosses zero.
5575 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5576 VariableIdx = new TruncInst(VariableIdx,
5577 TD.getIntPtrType(VariableIdx->getContext()),
5578 VariableIdx->getName(), &I);
5582 // Otherwise, there is an index. The computation we will do will be modulo
5583 // the pointer size, so get it.
5584 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5586 Offset &= PtrSizeMask;
5587 VariableScale &= PtrSizeMask;
5589 // To do this transformation, any constant index must be a multiple of the
5590 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5591 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5592 // multiple of the variable scale.
5593 int64_t NewOffs = Offset / (int64_t)VariableScale;
5594 if (Offset != NewOffs*(int64_t)VariableScale)
5597 // Okay, we can do this evaluation. Start by converting the index to intptr.
5598 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
5599 if (VariableIdx->getType() != IntPtrTy)
5600 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5602 VariableIdx->getName(), &I);
5603 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5604 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5608 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5609 /// else. At this point we know that the GEP is on the LHS of the comparison.
5610 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5611 ICmpInst::Predicate Cond,
5613 // Look through bitcasts.
5614 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5615 RHS = BCI->getOperand(0);
5617 Value *PtrBase = GEPLHS->getOperand(0);
5618 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5619 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5620 // This transformation (ignoring the base and scales) is valid because we
5621 // know pointers can't overflow since the gep is inbounds. See if we can
5622 // output an optimized form.
5623 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5625 // If not, synthesize the offset the hard way.
5627 Offset = EmitGEPOffset(GEPLHS, I, *this);
5628 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5629 Constant::getNullValue(Offset->getType()));
5630 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5631 // If the base pointers are different, but the indices are the same, just
5632 // compare the base pointer.
5633 if (PtrBase != GEPRHS->getOperand(0)) {
5634 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5635 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5636 GEPRHS->getOperand(0)->getType();
5638 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5639 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5640 IndicesTheSame = false;
5644 // If all indices are the same, just compare the base pointers.
5646 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5647 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5649 // Otherwise, the base pointers are different and the indices are
5650 // different, bail out.
5654 // If one of the GEPs has all zero indices, recurse.
5655 bool AllZeros = true;
5656 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5657 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5658 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5663 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5664 ICmpInst::getSwappedPredicate(Cond), I);
5666 // If the other GEP has all zero indices, recurse.
5668 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5669 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5670 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5675 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5677 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5678 // If the GEPs only differ by one index, compare it.
5679 unsigned NumDifferences = 0; // Keep track of # differences.
5680 unsigned DiffOperand = 0; // The operand that differs.
5681 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5682 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5683 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5684 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5685 // Irreconcilable differences.
5689 if (NumDifferences++) break;
5694 if (NumDifferences == 0) // SAME GEP?
5695 return ReplaceInstUsesWith(I, // No comparison is needed here.
5696 ConstantInt::get(Type::getInt1Ty(*Context),
5697 ICmpInst::isTrueWhenEqual(Cond)));
5699 else if (NumDifferences == 1) {
5700 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5701 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5702 // Make sure we do a signed comparison here.
5703 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5707 // Only lower this if the icmp is the only user of the GEP or if we expect
5708 // the result to fold to a constant!
5710 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5711 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5712 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5713 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5714 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5715 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5721 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5723 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5726 if (!isa<ConstantFP>(RHSC)) return 0;
5727 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5729 // Get the width of the mantissa. We don't want to hack on conversions that
5730 // might lose information from the integer, e.g. "i64 -> float"
5731 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5732 if (MantissaWidth == -1) return 0; // Unknown.
5734 // Check to see that the input is converted from an integer type that is small
5735 // enough that preserves all bits. TODO: check here for "known" sign bits.
5736 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5737 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5739 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5740 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5744 // If the conversion would lose info, don't hack on this.
5745 if ((int)InputSize > MantissaWidth)
5748 // Otherwise, we can potentially simplify the comparison. We know that it
5749 // will always come through as an integer value and we know the constant is
5750 // not a NAN (it would have been previously simplified).
5751 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5753 ICmpInst::Predicate Pred;
5754 switch (I.getPredicate()) {
5755 default: llvm_unreachable("Unexpected predicate!");
5756 case FCmpInst::FCMP_UEQ:
5757 case FCmpInst::FCMP_OEQ:
5758 Pred = ICmpInst::ICMP_EQ;
5760 case FCmpInst::FCMP_UGT:
5761 case FCmpInst::FCMP_OGT:
5762 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5764 case FCmpInst::FCMP_UGE:
5765 case FCmpInst::FCMP_OGE:
5766 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5768 case FCmpInst::FCMP_ULT:
5769 case FCmpInst::FCMP_OLT:
5770 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5772 case FCmpInst::FCMP_ULE:
5773 case FCmpInst::FCMP_OLE:
5774 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5776 case FCmpInst::FCMP_UNE:
5777 case FCmpInst::FCMP_ONE:
5778 Pred = ICmpInst::ICMP_NE;
5780 case FCmpInst::FCMP_ORD:
5781 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5782 case FCmpInst::FCMP_UNO:
5783 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5786 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5788 // Now we know that the APFloat is a normal number, zero or inf.
5790 // See if the FP constant is too large for the integer. For example,
5791 // comparing an i8 to 300.0.
5792 unsigned IntWidth = IntTy->getScalarSizeInBits();
5795 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5796 // and large values.
5797 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5798 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5799 APFloat::rmNearestTiesToEven);
5800 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5801 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5802 Pred == ICmpInst::ICMP_SLE)
5803 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5804 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5807 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5808 // +INF and large values.
5809 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5810 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5811 APFloat::rmNearestTiesToEven);
5812 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5813 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5814 Pred == ICmpInst::ICMP_ULE)
5815 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5816 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5821 // See if the RHS value is < SignedMin.
5822 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5823 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5824 APFloat::rmNearestTiesToEven);
5825 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5826 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5827 Pred == ICmpInst::ICMP_SGE)
5828 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5829 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5833 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5834 // [0, UMAX], but it may still be fractional. See if it is fractional by
5835 // casting the FP value to the integer value and back, checking for equality.
5836 // Don't do this for zero, because -0.0 is not fractional.
5837 Constant *RHSInt = LHSUnsigned
5838 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5839 : ConstantExpr::getFPToSI(RHSC, IntTy);
5840 if (!RHS.isZero()) {
5841 bool Equal = LHSUnsigned
5842 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5843 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5845 // If we had a comparison against a fractional value, we have to adjust
5846 // the compare predicate and sometimes the value. RHSC is rounded towards
5847 // zero at this point.
5849 default: llvm_unreachable("Unexpected integer comparison!");
5850 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5851 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5852 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5853 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5854 case ICmpInst::ICMP_ULE:
5855 // (float)int <= 4.4 --> int <= 4
5856 // (float)int <= -4.4 --> false
5857 if (RHS.isNegative())
5858 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5860 case ICmpInst::ICMP_SLE:
5861 // (float)int <= 4.4 --> int <= 4
5862 // (float)int <= -4.4 --> int < -4
5863 if (RHS.isNegative())
5864 Pred = ICmpInst::ICMP_SLT;
5866 case ICmpInst::ICMP_ULT:
5867 // (float)int < -4.4 --> false
5868 // (float)int < 4.4 --> int <= 4
5869 if (RHS.isNegative())
5870 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5871 Pred = ICmpInst::ICMP_ULE;
5873 case ICmpInst::ICMP_SLT:
5874 // (float)int < -4.4 --> int < -4
5875 // (float)int < 4.4 --> int <= 4
5876 if (!RHS.isNegative())
5877 Pred = ICmpInst::ICMP_SLE;
5879 case ICmpInst::ICMP_UGT:
5880 // (float)int > 4.4 --> int > 4
5881 // (float)int > -4.4 --> true
5882 if (RHS.isNegative())
5883 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5885 case ICmpInst::ICMP_SGT:
5886 // (float)int > 4.4 --> int > 4
5887 // (float)int > -4.4 --> int >= -4
5888 if (RHS.isNegative())
5889 Pred = ICmpInst::ICMP_SGE;
5891 case ICmpInst::ICMP_UGE:
5892 // (float)int >= -4.4 --> true
5893 // (float)int >= 4.4 --> int > 4
5894 if (!RHS.isNegative())
5895 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5896 Pred = ICmpInst::ICMP_UGT;
5898 case ICmpInst::ICMP_SGE:
5899 // (float)int >= -4.4 --> int >= -4
5900 // (float)int >= 4.4 --> int > 4
5901 if (!RHS.isNegative())
5902 Pred = ICmpInst::ICMP_SGT;
5908 // Lower this FP comparison into an appropriate integer version of the
5910 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5913 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5914 bool Changed = SimplifyCompare(I);
5915 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5917 // Fold trivial predicates.
5918 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5919 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5920 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5921 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5923 // Simplify 'fcmp pred X, X'
5925 switch (I.getPredicate()) {
5926 default: llvm_unreachable("Unknown predicate!");
5927 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5928 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5929 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5930 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5931 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5932 case FCmpInst::FCMP_OLT: // True if ordered and less than
5933 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5934 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5936 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5937 case FCmpInst::FCMP_ULT: // True if unordered or less than
5938 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5939 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5940 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5941 I.setPredicate(FCmpInst::FCMP_UNO);
5942 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5945 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5946 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5947 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5948 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5949 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5950 I.setPredicate(FCmpInst::FCMP_ORD);
5951 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5956 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5957 return ReplaceInstUsesWith(I, UndefValue::get(Type::getInt1Ty(*Context)));
5959 // Handle fcmp with constant RHS
5960 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5961 // If the constant is a nan, see if we can fold the comparison based on it.
5962 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5963 if (CFP->getValueAPF().isNaN()) {
5964 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5965 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5966 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5967 "Comparison must be either ordered or unordered!");
5968 // True if unordered.
5969 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5973 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5974 switch (LHSI->getOpcode()) {
5975 case Instruction::PHI:
5976 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5977 // block. If in the same block, we're encouraging jump threading. If
5978 // not, we are just pessimizing the code by making an i1 phi.
5979 if (LHSI->getParent() == I.getParent())
5980 if (Instruction *NV = FoldOpIntoPhi(I))
5983 case Instruction::SIToFP:
5984 case Instruction::UIToFP:
5985 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5988 case Instruction::Select:
5989 // If either operand of the select is a constant, we can fold the
5990 // comparison into the select arms, which will cause one to be
5991 // constant folded and the select turned into a bitwise or.
5992 Value *Op1 = 0, *Op2 = 0;
5993 if (LHSI->hasOneUse()) {
5994 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5995 // Fold the known value into the constant operand.
5996 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5997 // Insert a new FCmp of the other select operand.
5998 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5999 LHSI->getOperand(2), RHSC,
6001 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6002 // Fold the known value into the constant operand.
6003 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
6004 // Insert a new FCmp of the other select operand.
6005 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
6006 LHSI->getOperand(1), RHSC,
6012 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6017 return Changed ? &I : 0;
6020 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
6021 bool Changed = SimplifyCompare(I);
6022 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6023 const Type *Ty = Op0->getType();
6027 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
6028 I.isTrueWhenEqual()));
6030 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
6031 return ReplaceInstUsesWith(I, UndefValue::get(Type::getInt1Ty(*Context)));
6033 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
6034 // addresses never equal each other! We already know that Op0 != Op1.
6035 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
6036 isa<ConstantPointerNull>(Op0)) &&
6037 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
6038 isa<ConstantPointerNull>(Op1)))
6039 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
6040 !I.isTrueWhenEqual()));
6042 // icmp's with boolean values can always be turned into bitwise operations
6043 if (Ty == Type::getInt1Ty(*Context)) {
6044 switch (I.getPredicate()) {
6045 default: llvm_unreachable("Invalid icmp instruction!");
6046 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
6047 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
6048 InsertNewInstBefore(Xor, I);
6049 return BinaryOperator::CreateNot(Xor);
6051 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6052 return BinaryOperator::CreateXor(Op0, Op1);
6054 case ICmpInst::ICMP_UGT:
6055 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6057 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6058 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
6059 InsertNewInstBefore(Not, I);
6060 return BinaryOperator::CreateAnd(Not, Op1);
6062 case ICmpInst::ICMP_SGT:
6063 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6065 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6066 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
6067 InsertNewInstBefore(Not, I);
6068 return BinaryOperator::CreateAnd(Not, Op0);
6070 case ICmpInst::ICMP_UGE:
6071 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6073 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6074 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
6075 InsertNewInstBefore(Not, I);
6076 return BinaryOperator::CreateOr(Not, Op1);
6078 case ICmpInst::ICMP_SGE:
6079 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6081 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6082 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
6083 InsertNewInstBefore(Not, I);
6084 return BinaryOperator::CreateOr(Not, Op0);
6089 unsigned BitWidth = 0;
6091 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6092 else if (Ty->isIntOrIntVector())
6093 BitWidth = Ty->getScalarSizeInBits();
6095 bool isSignBit = false;
6097 // See if we are doing a comparison with a constant.
6098 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6099 Value *A = 0, *B = 0;
6101 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6102 if (I.isEquality() && CI->isNullValue() &&
6103 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
6104 // (icmp cond A B) if cond is equality
6105 return new ICmpInst(I.getPredicate(), A, B);
6108 // If we have an icmp le or icmp ge instruction, turn it into the
6109 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6110 // them being folded in the code below.
6111 switch (I.getPredicate()) {
6113 case ICmpInst::ICMP_ULE:
6114 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6115 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6116 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
6118 case ICmpInst::ICMP_SLE:
6119 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6120 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6121 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6123 case ICmpInst::ICMP_UGE:
6124 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6125 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6126 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
6128 case ICmpInst::ICMP_SGE:
6129 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6130 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6131 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6135 // If this comparison is a normal comparison, it demands all
6136 // bits, if it is a sign bit comparison, it only demands the sign bit.
6138 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6141 // See if we can fold the comparison based on range information we can get
6142 // by checking whether bits are known to be zero or one in the input.
6143 if (BitWidth != 0) {
6144 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6145 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6147 if (SimplifyDemandedBits(I.getOperandUse(0),
6148 isSignBit ? APInt::getSignBit(BitWidth)
6149 : APInt::getAllOnesValue(BitWidth),
6150 Op0KnownZero, Op0KnownOne, 0))
6152 if (SimplifyDemandedBits(I.getOperandUse(1),
6153 APInt::getAllOnesValue(BitWidth),
6154 Op1KnownZero, Op1KnownOne, 0))
6157 // Given the known and unknown bits, compute a range that the LHS could be
6158 // in. Compute the Min, Max and RHS values based on the known bits. For the
6159 // EQ and NE we use unsigned values.
6160 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6161 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6162 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6163 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6165 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6168 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6170 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6174 // If Min and Max are known to be the same, then SimplifyDemandedBits
6175 // figured out that the LHS is a constant. Just constant fold this now so
6176 // that code below can assume that Min != Max.
6177 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6178 return new ICmpInst(I.getPredicate(),
6179 ConstantInt::get(*Context, Op0Min), Op1);
6180 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6181 return new ICmpInst(I.getPredicate(), Op0,
6182 ConstantInt::get(*Context, Op1Min));
6184 // Based on the range information we know about the LHS, see if we can
6185 // simplify this comparison. For example, (x&4) < 8 is always true.
6186 switch (I.getPredicate()) {
6187 default: llvm_unreachable("Unknown icmp opcode!");
6188 case ICmpInst::ICMP_EQ:
6189 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6190 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6192 case ICmpInst::ICMP_NE:
6193 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6194 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6196 case ICmpInst::ICMP_ULT:
6197 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6198 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6199 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6200 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6201 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6202 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6203 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6204 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6205 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6208 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6209 if (CI->isMinValue(true))
6210 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6211 Constant::getAllOnesValue(Op0->getType()));
6214 case ICmpInst::ICMP_UGT:
6215 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6216 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6217 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6218 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6220 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6221 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6222 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6223 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6224 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6227 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6228 if (CI->isMaxValue(true))
6229 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6230 Constant::getNullValue(Op0->getType()));
6233 case ICmpInst::ICMP_SLT:
6234 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6235 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6236 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6237 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6238 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6239 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6240 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6241 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6242 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6246 case ICmpInst::ICMP_SGT:
6247 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6248 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6249 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6250 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6252 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6253 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6254 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6255 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6256 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6260 case ICmpInst::ICMP_SGE:
6261 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6262 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6263 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6264 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6265 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6267 case ICmpInst::ICMP_SLE:
6268 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6269 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6270 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6271 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6272 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6274 case ICmpInst::ICMP_UGE:
6275 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6276 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6277 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6278 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6279 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6281 case ICmpInst::ICMP_ULE:
6282 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6283 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6284 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6285 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6286 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6290 // Turn a signed comparison into an unsigned one if both operands
6291 // are known to have the same sign.
6292 if (I.isSignedPredicate() &&
6293 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6294 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6295 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
6298 // Test if the ICmpInst instruction is used exclusively by a select as
6299 // part of a minimum or maximum operation. If so, refrain from doing
6300 // any other folding. This helps out other analyses which understand
6301 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6302 // and CodeGen. And in this case, at least one of the comparison
6303 // operands has at least one user besides the compare (the select),
6304 // which would often largely negate the benefit of folding anyway.
6306 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6307 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6308 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6311 // See if we are doing a comparison between a constant and an instruction that
6312 // can be folded into the comparison.
6313 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6314 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6315 // instruction, see if that instruction also has constants so that the
6316 // instruction can be folded into the icmp
6317 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6318 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6322 // Handle icmp with constant (but not simple integer constant) RHS
6323 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6324 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6325 switch (LHSI->getOpcode()) {
6326 case Instruction::GetElementPtr:
6327 if (RHSC->isNullValue()) {
6328 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6329 bool isAllZeros = true;
6330 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6331 if (!isa<Constant>(LHSI->getOperand(i)) ||
6332 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6337 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6338 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6342 case Instruction::PHI:
6343 // Only fold icmp into the PHI if the phi and fcmp are in the same
6344 // block. If in the same block, we're encouraging jump threading. If
6345 // not, we are just pessimizing the code by making an i1 phi.
6346 if (LHSI->getParent() == I.getParent())
6347 if (Instruction *NV = FoldOpIntoPhi(I))
6350 case Instruction::Select: {
6351 // If either operand of the select is a constant, we can fold the
6352 // comparison into the select arms, which will cause one to be
6353 // constant folded and the select turned into a bitwise or.
6354 Value *Op1 = 0, *Op2 = 0;
6355 if (LHSI->hasOneUse()) {
6356 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6357 // Fold the known value into the constant operand.
6358 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6359 // Insert a new ICmp of the other select operand.
6360 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6361 LHSI->getOperand(2), RHSC,
6363 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6364 // Fold the known value into the constant operand.
6365 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6366 // Insert a new ICmp of the other select operand.
6367 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6368 LHSI->getOperand(1), RHSC,
6374 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6377 case Instruction::Malloc:
6378 // If we have (malloc != null), and if the malloc has a single use, we
6379 // can assume it is successful and remove the malloc.
6380 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6381 AddToWorkList(LHSI);
6382 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
6383 !I.isTrueWhenEqual()));
6389 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6390 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6391 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6393 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6394 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6395 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6398 // Test to see if the operands of the icmp are casted versions of other
6399 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6401 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6402 if (isa<PointerType>(Op0->getType()) &&
6403 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6404 // We keep moving the cast from the left operand over to the right
6405 // operand, where it can often be eliminated completely.
6406 Op0 = CI->getOperand(0);
6408 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6409 // so eliminate it as well.
6410 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6411 Op1 = CI2->getOperand(0);
6413 // If Op1 is a constant, we can fold the cast into the constant.
6414 if (Op0->getType() != Op1->getType()) {
6415 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6416 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6418 // Otherwise, cast the RHS right before the icmp
6419 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6422 return new ICmpInst(I.getPredicate(), Op0, Op1);
6426 if (isa<CastInst>(Op0)) {
6427 // Handle the special case of: icmp (cast bool to X), <cst>
6428 // This comes up when you have code like
6431 // For generality, we handle any zero-extension of any operand comparison
6432 // with a constant or another cast from the same type.
6433 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6434 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6438 // See if it's the same type of instruction on the left and right.
6439 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6440 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6441 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6442 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6443 switch (Op0I->getOpcode()) {
6445 case Instruction::Add:
6446 case Instruction::Sub:
6447 case Instruction::Xor:
6448 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6449 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6450 Op1I->getOperand(0));
6451 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6452 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6453 if (CI->getValue().isSignBit()) {
6454 ICmpInst::Predicate Pred = I.isSignedPredicate()
6455 ? I.getUnsignedPredicate()
6456 : I.getSignedPredicate();
6457 return new ICmpInst(Pred, Op0I->getOperand(0),
6458 Op1I->getOperand(0));
6461 if (CI->getValue().isMaxSignedValue()) {
6462 ICmpInst::Predicate Pred = I.isSignedPredicate()
6463 ? I.getUnsignedPredicate()
6464 : I.getSignedPredicate();
6465 Pred = I.getSwappedPredicate(Pred);
6466 return new ICmpInst(Pred, Op0I->getOperand(0),
6467 Op1I->getOperand(0));
6471 case Instruction::Mul:
6472 if (!I.isEquality())
6475 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6476 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6477 // Mask = -1 >> count-trailing-zeros(Cst).
6478 if (!CI->isZero() && !CI->isOne()) {
6479 const APInt &AP = CI->getValue();
6480 ConstantInt *Mask = ConstantInt::get(*Context,
6481 APInt::getLowBitsSet(AP.getBitWidth(),
6483 AP.countTrailingZeros()));
6484 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6486 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6488 InsertNewInstBefore(And1, I);
6489 InsertNewInstBefore(And2, I);
6490 return new ICmpInst(I.getPredicate(), And1, And2);
6499 // ~x < ~y --> y < x
6501 if (match(Op0, m_Not(m_Value(A))) &&
6502 match(Op1, m_Not(m_Value(B))))
6503 return new ICmpInst(I.getPredicate(), B, A);
6506 if (I.isEquality()) {
6507 Value *A, *B, *C, *D;
6509 // -x == -y --> x == y
6510 if (match(Op0, m_Neg(m_Value(A))) &&
6511 match(Op1, m_Neg(m_Value(B))))
6512 return new ICmpInst(I.getPredicate(), A, B);
6514 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6515 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6516 Value *OtherVal = A == Op1 ? B : A;
6517 return new ICmpInst(I.getPredicate(), OtherVal,
6518 Constant::getNullValue(A->getType()));
6521 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6522 // A^c1 == C^c2 --> A == C^(c1^c2)
6523 ConstantInt *C1, *C2;
6524 if (match(B, m_ConstantInt(C1)) &&
6525 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6527 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6528 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6529 return new ICmpInst(I.getPredicate(), A,
6530 InsertNewInstBefore(Xor, I));
6533 // A^B == A^D -> B == D
6534 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6535 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6536 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6537 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6541 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6542 (A == Op0 || B == Op0)) {
6543 // A == (A^B) -> B == 0
6544 Value *OtherVal = A == Op0 ? B : A;
6545 return new ICmpInst(I.getPredicate(), OtherVal,
6546 Constant::getNullValue(A->getType()));
6549 // (A-B) == A -> B == 0
6550 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6551 return new ICmpInst(I.getPredicate(), B,
6552 Constant::getNullValue(B->getType()));
6554 // A == (A-B) -> B == 0
6555 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6556 return new ICmpInst(I.getPredicate(), B,
6557 Constant::getNullValue(B->getType()));
6559 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6560 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6561 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6562 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6563 Value *X = 0, *Y = 0, *Z = 0;
6566 X = B; Y = D; Z = A;
6567 } else if (A == D) {
6568 X = B; Y = C; Z = A;
6569 } else if (B == C) {
6570 X = A; Y = D; Z = B;
6571 } else if (B == D) {
6572 X = A; Y = C; Z = B;
6575 if (X) { // Build (X^Y) & Z
6576 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6577 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6578 I.setOperand(0, Op1);
6579 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6584 return Changed ? &I : 0;
6588 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6589 /// and CmpRHS are both known to be integer constants.
6590 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6591 ConstantInt *DivRHS) {
6592 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6593 const APInt &CmpRHSV = CmpRHS->getValue();
6595 // FIXME: If the operand types don't match the type of the divide
6596 // then don't attempt this transform. The code below doesn't have the
6597 // logic to deal with a signed divide and an unsigned compare (and
6598 // vice versa). This is because (x /s C1) <s C2 produces different
6599 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6600 // (x /u C1) <u C2. Simply casting the operands and result won't
6601 // work. :( The if statement below tests that condition and bails
6603 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6604 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6606 if (DivRHS->isZero())
6607 return 0; // The ProdOV computation fails on divide by zero.
6608 if (DivIsSigned && DivRHS->isAllOnesValue())
6609 return 0; // The overflow computation also screws up here
6610 if (DivRHS->isOne())
6611 return 0; // Not worth bothering, and eliminates some funny cases
6614 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6615 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6616 // C2 (CI). By solving for X we can turn this into a range check
6617 // instead of computing a divide.
6618 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6620 // Determine if the product overflows by seeing if the product is
6621 // not equal to the divide. Make sure we do the same kind of divide
6622 // as in the LHS instruction that we're folding.
6623 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6624 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6626 // Get the ICmp opcode
6627 ICmpInst::Predicate Pred = ICI.getPredicate();
6629 // Figure out the interval that is being checked. For example, a comparison
6630 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6631 // Compute this interval based on the constants involved and the signedness of
6632 // the compare/divide. This computes a half-open interval, keeping track of
6633 // whether either value in the interval overflows. After analysis each
6634 // overflow variable is set to 0 if it's corresponding bound variable is valid
6635 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6636 int LoOverflow = 0, HiOverflow = 0;
6637 Constant *LoBound = 0, *HiBound = 0;
6639 if (!DivIsSigned) { // udiv
6640 // e.g. X/5 op 3 --> [15, 20)
6642 HiOverflow = LoOverflow = ProdOV;
6644 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6645 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6646 if (CmpRHSV == 0) { // (X / pos) op 0
6647 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6648 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6650 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6651 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6652 HiOverflow = LoOverflow = ProdOV;
6654 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6655 } else { // (X / pos) op neg
6656 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6657 HiBound = AddOne(Prod);
6658 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6660 ConstantInt* DivNeg =
6661 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6662 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6666 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6667 if (CmpRHSV == 0) { // (X / neg) op 0
6668 // e.g. X/-5 op 0 --> [-4, 5)
6669 LoBound = AddOne(DivRHS);
6670 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6671 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6672 HiOverflow = 1; // [INTMIN+1, overflow)
6673 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6675 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6676 // e.g. X/-5 op 3 --> [-19, -14)
6677 HiBound = AddOne(Prod);
6678 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6680 LoOverflow = AddWithOverflow(LoBound, HiBound,
6681 DivRHS, Context, true) ? -1 : 0;
6682 } else { // (X / neg) op neg
6683 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6684 LoOverflow = HiOverflow = ProdOV;
6686 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6689 // Dividing by a negative swaps the condition. LT <-> GT
6690 Pred = ICmpInst::getSwappedPredicate(Pred);
6693 Value *X = DivI->getOperand(0);
6695 default: llvm_unreachable("Unhandled icmp opcode!");
6696 case ICmpInst::ICMP_EQ:
6697 if (LoOverflow && HiOverflow)
6698 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6699 else if (HiOverflow)
6700 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6701 ICmpInst::ICMP_UGE, X, LoBound);
6702 else if (LoOverflow)
6703 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6704 ICmpInst::ICMP_ULT, X, HiBound);
6706 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6707 case ICmpInst::ICMP_NE:
6708 if (LoOverflow && HiOverflow)
6709 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6710 else if (HiOverflow)
6711 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6712 ICmpInst::ICMP_ULT, X, LoBound);
6713 else if (LoOverflow)
6714 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6715 ICmpInst::ICMP_UGE, X, HiBound);
6717 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6718 case ICmpInst::ICMP_ULT:
6719 case ICmpInst::ICMP_SLT:
6720 if (LoOverflow == +1) // Low bound is greater than input range.
6721 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6722 if (LoOverflow == -1) // Low bound is less than input range.
6723 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6724 return new ICmpInst(Pred, X, LoBound);
6725 case ICmpInst::ICMP_UGT:
6726 case ICmpInst::ICMP_SGT:
6727 if (HiOverflow == +1) // High bound greater than input range.
6728 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6729 else if (HiOverflow == -1) // High bound less than input range.
6730 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6731 if (Pred == ICmpInst::ICMP_UGT)
6732 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6734 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6739 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6741 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6744 const APInt &RHSV = RHS->getValue();
6746 switch (LHSI->getOpcode()) {
6747 case Instruction::Trunc:
6748 if (ICI.isEquality() && LHSI->hasOneUse()) {
6749 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6750 // of the high bits truncated out of x are known.
6751 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6752 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6753 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6754 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6755 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6757 // If all the high bits are known, we can do this xform.
6758 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6759 // Pull in the high bits from known-ones set.
6760 APInt NewRHS(RHS->getValue());
6761 NewRHS.zext(SrcBits);
6763 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6764 ConstantInt::get(*Context, NewRHS));
6769 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6770 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6771 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6773 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6774 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6775 Value *CompareVal = LHSI->getOperand(0);
6777 // If the sign bit of the XorCST is not set, there is no change to
6778 // the operation, just stop using the Xor.
6779 if (!XorCST->getValue().isNegative()) {
6780 ICI.setOperand(0, CompareVal);
6781 AddToWorkList(LHSI);
6785 // Was the old condition true if the operand is positive?
6786 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6788 // If so, the new one isn't.
6789 isTrueIfPositive ^= true;
6791 if (isTrueIfPositive)
6792 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
6795 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
6799 if (LHSI->hasOneUse()) {
6800 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6801 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6802 const APInt &SignBit = XorCST->getValue();
6803 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6804 ? ICI.getUnsignedPredicate()
6805 : ICI.getSignedPredicate();
6806 return new ICmpInst(Pred, LHSI->getOperand(0),
6807 ConstantInt::get(*Context, RHSV ^ SignBit));
6810 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6811 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6812 const APInt &NotSignBit = XorCST->getValue();
6813 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6814 ? ICI.getUnsignedPredicate()
6815 : ICI.getSignedPredicate();
6816 Pred = ICI.getSwappedPredicate(Pred);
6817 return new ICmpInst(Pred, LHSI->getOperand(0),
6818 ConstantInt::get(*Context, RHSV ^ NotSignBit));
6823 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6824 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6825 LHSI->getOperand(0)->hasOneUse()) {
6826 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6828 // If the LHS is an AND of a truncating cast, we can widen the
6829 // and/compare to be the input width without changing the value
6830 // produced, eliminating a cast.
6831 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6832 // We can do this transformation if either the AND constant does not
6833 // have its sign bit set or if it is an equality comparison.
6834 // Extending a relational comparison when we're checking the sign
6835 // bit would not work.
6836 if (Cast->hasOneUse() &&
6837 (ICI.isEquality() ||
6838 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6840 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6841 APInt NewCST = AndCST->getValue();
6842 NewCST.zext(BitWidth);
6844 NewCI.zext(BitWidth);
6845 Instruction *NewAnd =
6846 BinaryOperator::CreateAnd(Cast->getOperand(0),
6847 ConstantInt::get(*Context, NewCST), LHSI->getName());
6848 InsertNewInstBefore(NewAnd, ICI);
6849 return new ICmpInst(ICI.getPredicate(), NewAnd,
6850 ConstantInt::get(*Context, NewCI));
6854 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6855 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6856 // happens a LOT in code produced by the C front-end, for bitfield
6858 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6859 if (Shift && !Shift->isShift())
6863 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6864 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6865 const Type *AndTy = AndCST->getType(); // Type of the and.
6867 // We can fold this as long as we can't shift unknown bits
6868 // into the mask. This can only happen with signed shift
6869 // rights, as they sign-extend.
6871 bool CanFold = Shift->isLogicalShift();
6873 // To test for the bad case of the signed shr, see if any
6874 // of the bits shifted in could be tested after the mask.
6875 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6876 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6878 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6879 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6880 AndCST->getValue()) == 0)
6886 if (Shift->getOpcode() == Instruction::Shl)
6887 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6889 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6891 // Check to see if we are shifting out any of the bits being
6893 if (ConstantExpr::get(Shift->getOpcode(),
6894 NewCst, ShAmt) != RHS) {
6895 // If we shifted bits out, the fold is not going to work out.
6896 // As a special case, check to see if this means that the
6897 // result is always true or false now.
6898 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6899 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6900 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6901 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6903 ICI.setOperand(1, NewCst);
6904 Constant *NewAndCST;
6905 if (Shift->getOpcode() == Instruction::Shl)
6906 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6908 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6909 LHSI->setOperand(1, NewAndCST);
6910 LHSI->setOperand(0, Shift->getOperand(0));
6911 AddToWorkList(Shift); // Shift is dead.
6912 AddUsesToWorkList(ICI);
6918 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6919 // preferable because it allows the C<<Y expression to be hoisted out
6920 // of a loop if Y is invariant and X is not.
6921 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6922 ICI.isEquality() && !Shift->isArithmeticShift() &&
6923 !isa<Constant>(Shift->getOperand(0))) {
6926 if (Shift->getOpcode() == Instruction::LShr) {
6927 NS = BinaryOperator::CreateShl(AndCST,
6928 Shift->getOperand(1), "tmp");
6930 // Insert a logical shift.
6931 NS = BinaryOperator::CreateLShr(AndCST,
6932 Shift->getOperand(1), "tmp");
6934 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6936 // Compute X & (C << Y).
6937 Instruction *NewAnd =
6938 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6939 InsertNewInstBefore(NewAnd, ICI);
6941 ICI.setOperand(0, NewAnd);
6947 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6948 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6951 uint32_t TypeBits = RHSV.getBitWidth();
6953 // Check that the shift amount is in range. If not, don't perform
6954 // undefined shifts. When the shift is visited it will be
6956 if (ShAmt->uge(TypeBits))
6959 if (ICI.isEquality()) {
6960 // If we are comparing against bits always shifted out, the
6961 // comparison cannot succeed.
6963 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
6965 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6966 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6967 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6968 return ReplaceInstUsesWith(ICI, Cst);
6971 if (LHSI->hasOneUse()) {
6972 // Otherwise strength reduce the shift into an and.
6973 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6975 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
6976 TypeBits-ShAmtVal));
6979 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6980 Mask, LHSI->getName()+".mask");
6981 Value *And = InsertNewInstBefore(AndI, ICI);
6982 return new ICmpInst(ICI.getPredicate(), And,
6983 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
6987 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6988 bool TrueIfSigned = false;
6989 if (LHSI->hasOneUse() &&
6990 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6991 // (X << 31) <s 0 --> (X&1) != 0
6992 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
6993 (TypeBits-ShAmt->getZExtValue()-1));
6995 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6996 Mask, LHSI->getName()+".mask");
6997 Value *And = InsertNewInstBefore(AndI, ICI);
6999 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
7000 And, Constant::getNullValue(And->getType()));
7005 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
7006 case Instruction::AShr: {
7007 // Only handle equality comparisons of shift-by-constant.
7008 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7009 if (!ShAmt || !ICI.isEquality()) break;
7011 // Check that the shift amount is in range. If not, don't perform
7012 // undefined shifts. When the shift is visited it will be
7014 uint32_t TypeBits = RHSV.getBitWidth();
7015 if (ShAmt->uge(TypeBits))
7018 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
7020 // If we are comparing against bits always shifted out, the
7021 // comparison cannot succeed.
7022 APInt Comp = RHSV << ShAmtVal;
7023 if (LHSI->getOpcode() == Instruction::LShr)
7024 Comp = Comp.lshr(ShAmtVal);
7026 Comp = Comp.ashr(ShAmtVal);
7028 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
7029 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7030 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
7031 return ReplaceInstUsesWith(ICI, Cst);
7034 // Otherwise, check to see if the bits shifted out are known to be zero.
7035 // If so, we can compare against the unshifted value:
7036 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
7037 if (LHSI->hasOneUse() &&
7038 MaskedValueIsZero(LHSI->getOperand(0),
7039 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
7040 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
7041 ConstantExpr::getShl(RHS, ShAmt));
7044 if (LHSI->hasOneUse()) {
7045 // Otherwise strength reduce the shift into an and.
7046 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7047 Constant *Mask = ConstantInt::get(*Context, Val);
7050 BinaryOperator::CreateAnd(LHSI->getOperand(0),
7051 Mask, LHSI->getName()+".mask");
7052 Value *And = InsertNewInstBefore(AndI, ICI);
7053 return new ICmpInst(ICI.getPredicate(), And,
7054 ConstantExpr::getShl(RHS, ShAmt));
7059 case Instruction::SDiv:
7060 case Instruction::UDiv:
7061 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7062 // Fold this div into the comparison, producing a range check.
7063 // Determine, based on the divide type, what the range is being
7064 // checked. If there is an overflow on the low or high side, remember
7065 // it, otherwise compute the range [low, hi) bounding the new value.
7066 // See: InsertRangeTest above for the kinds of replacements possible.
7067 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7068 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7073 case Instruction::Add:
7074 // Fold: icmp pred (add, X, C1), C2
7076 if (!ICI.isEquality()) {
7077 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7079 const APInt &LHSV = LHSC->getValue();
7081 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7084 if (ICI.isSignedPredicate()) {
7085 if (CR.getLower().isSignBit()) {
7086 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7087 ConstantInt::get(*Context, CR.getUpper()));
7088 } else if (CR.getUpper().isSignBit()) {
7089 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7090 ConstantInt::get(*Context, CR.getLower()));
7093 if (CR.getLower().isMinValue()) {
7094 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7095 ConstantInt::get(*Context, CR.getUpper()));
7096 } else if (CR.getUpper().isMinValue()) {
7097 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7098 ConstantInt::get(*Context, CR.getLower()));
7105 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7106 if (ICI.isEquality()) {
7107 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7109 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7110 // the second operand is a constant, simplify a bit.
7111 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7112 switch (BO->getOpcode()) {
7113 case Instruction::SRem:
7114 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7115 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7116 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7117 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7118 Instruction *NewRem =
7119 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
7121 InsertNewInstBefore(NewRem, ICI);
7122 return new ICmpInst(ICI.getPredicate(), NewRem,
7123 Constant::getNullValue(BO->getType()));
7127 case Instruction::Add:
7128 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7129 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7130 if (BO->hasOneUse())
7131 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7132 ConstantExpr::getSub(RHS, BOp1C));
7133 } else if (RHSV == 0) {
7134 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7135 // efficiently invertible, or if the add has just this one use.
7136 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7138 if (Value *NegVal = dyn_castNegVal(BOp1))
7139 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
7140 else if (Value *NegVal = dyn_castNegVal(BOp0))
7141 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
7142 else if (BO->hasOneUse()) {
7143 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
7144 InsertNewInstBefore(Neg, ICI);
7146 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
7150 case Instruction::Xor:
7151 // For the xor case, we can xor two constants together, eliminating
7152 // the explicit xor.
7153 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7154 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7155 ConstantExpr::getXor(RHS, BOC));
7158 case Instruction::Sub:
7159 // Replace (([sub|xor] A, B) != 0) with (A != B)
7161 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7165 case Instruction::Or:
7166 // If bits are being or'd in that are not present in the constant we
7167 // are comparing against, then the comparison could never succeed!
7168 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7169 Constant *NotCI = ConstantExpr::getNot(RHS);
7170 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7171 return ReplaceInstUsesWith(ICI,
7172 ConstantInt::get(Type::getInt1Ty(*Context),
7177 case Instruction::And:
7178 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7179 // If bits are being compared against that are and'd out, then the
7180 // comparison can never succeed!
7181 if ((RHSV & ~BOC->getValue()) != 0)
7182 return ReplaceInstUsesWith(ICI,
7183 ConstantInt::get(Type::getInt1Ty(*Context),
7186 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7187 if (RHS == BOC && RHSV.isPowerOf2())
7188 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
7189 ICmpInst::ICMP_NE, LHSI,
7190 Constant::getNullValue(RHS->getType()));
7192 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7193 if (BOC->getValue().isSignBit()) {
7194 Value *X = BO->getOperand(0);
7195 Constant *Zero = Constant::getNullValue(X->getType());
7196 ICmpInst::Predicate pred = isICMP_NE ?
7197 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7198 return new ICmpInst(pred, X, Zero);
7201 // ((X & ~7) == 0) --> X < 8
7202 if (RHSV == 0 && isHighOnes(BOC)) {
7203 Value *X = BO->getOperand(0);
7204 Constant *NegX = ConstantExpr::getNeg(BOC);
7205 ICmpInst::Predicate pred = isICMP_NE ?
7206 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7207 return new ICmpInst(pred, X, NegX);
7212 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7213 // Handle icmp {eq|ne} <intrinsic>, intcst.
7214 if (II->getIntrinsicID() == Intrinsic::bswap) {
7216 ICI.setOperand(0, II->getOperand(1));
7217 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7225 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7226 /// We only handle extending casts so far.
7228 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7229 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7230 Value *LHSCIOp = LHSCI->getOperand(0);
7231 const Type *SrcTy = LHSCIOp->getType();
7232 const Type *DestTy = LHSCI->getType();
7235 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7236 // integer type is the same size as the pointer type.
7237 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7238 TD->getPointerSizeInBits() ==
7239 cast<IntegerType>(DestTy)->getBitWidth()) {
7241 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7242 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7243 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7244 RHSOp = RHSC->getOperand(0);
7245 // If the pointer types don't match, insert a bitcast.
7246 if (LHSCIOp->getType() != RHSOp->getType())
7247 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
7251 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
7254 // The code below only handles extension cast instructions, so far.
7256 if (LHSCI->getOpcode() != Instruction::ZExt &&
7257 LHSCI->getOpcode() != Instruction::SExt)
7260 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7261 bool isSignedCmp = ICI.isSignedPredicate();
7263 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7264 // Not an extension from the same type?
7265 RHSCIOp = CI->getOperand(0);
7266 if (RHSCIOp->getType() != LHSCIOp->getType())
7269 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7270 // and the other is a zext), then we can't handle this.
7271 if (CI->getOpcode() != LHSCI->getOpcode())
7274 // Deal with equality cases early.
7275 if (ICI.isEquality())
7276 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7278 // A signed comparison of sign extended values simplifies into a
7279 // signed comparison.
7280 if (isSignedCmp && isSignedExt)
7281 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7283 // The other three cases all fold into an unsigned comparison.
7284 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7287 // If we aren't dealing with a constant on the RHS, exit early
7288 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7292 // Compute the constant that would happen if we truncated to SrcTy then
7293 // reextended to DestTy.
7294 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7295 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7298 // If the re-extended constant didn't change...
7300 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7301 // For example, we might have:
7302 // %A = sext i16 %X to i32
7303 // %B = icmp ugt i32 %A, 1330
7304 // It is incorrect to transform this into
7305 // %B = icmp ugt i16 %X, 1330
7306 // because %A may have negative value.
7308 // However, we allow this when the compare is EQ/NE, because they are
7310 if (isSignedExt == isSignedCmp || ICI.isEquality())
7311 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7315 // The re-extended constant changed so the constant cannot be represented
7316 // in the shorter type. Consequently, we cannot emit a simple comparison.
7318 // First, handle some easy cases. We know the result cannot be equal at this
7319 // point so handle the ICI.isEquality() cases
7320 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7321 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7322 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7323 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7325 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7326 // should have been folded away previously and not enter in here.
7329 // We're performing a signed comparison.
7330 if (cast<ConstantInt>(CI)->getValue().isNegative())
7331 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7333 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7335 // We're performing an unsigned comparison.
7337 // We're performing an unsigned comp with a sign extended value.
7338 // This is true if the input is >= 0. [aka >s -1]
7339 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7340 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT,
7341 LHSCIOp, NegOne, ICI.getName()), ICI);
7343 // Unsigned extend & unsigned compare -> always true.
7344 Result = ConstantInt::getTrue(*Context);
7348 // Finally, return the value computed.
7349 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7350 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7351 return ReplaceInstUsesWith(ICI, Result);
7353 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7354 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7355 "ICmp should be folded!");
7356 if (Constant *CI = dyn_cast<Constant>(Result))
7357 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7358 return BinaryOperator::CreateNot(Result);
7361 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7362 return commonShiftTransforms(I);
7365 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7366 return commonShiftTransforms(I);
7369 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7370 if (Instruction *R = commonShiftTransforms(I))
7373 Value *Op0 = I.getOperand(0);
7375 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7376 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7377 if (CSI->isAllOnesValue())
7378 return ReplaceInstUsesWith(I, CSI);
7380 // See if we can turn a signed shr into an unsigned shr.
7381 if (MaskedValueIsZero(Op0,
7382 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7383 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7385 // Arithmetic shifting an all-sign-bit value is a no-op.
7386 unsigned NumSignBits = ComputeNumSignBits(Op0);
7387 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7388 return ReplaceInstUsesWith(I, Op0);
7393 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7394 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7395 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7397 // shl X, 0 == X and shr X, 0 == X
7398 // shl 0, X == 0 and shr 0, X == 0
7399 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7400 Op0 == Constant::getNullValue(Op0->getType()))
7401 return ReplaceInstUsesWith(I, Op0);
7403 if (isa<UndefValue>(Op0)) {
7404 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7405 return ReplaceInstUsesWith(I, Op0);
7406 else // undef << X -> 0, undef >>u X -> 0
7407 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7409 if (isa<UndefValue>(Op1)) {
7410 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7411 return ReplaceInstUsesWith(I, Op0);
7412 else // X << undef, X >>u undef -> 0
7413 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7416 // See if we can fold away this shift.
7417 if (SimplifyDemandedInstructionBits(I))
7420 // Try to fold constant and into select arguments.
7421 if (isa<Constant>(Op0))
7422 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7423 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7426 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7427 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7432 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7433 BinaryOperator &I) {
7434 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7436 // See if we can simplify any instructions used by the instruction whose sole
7437 // purpose is to compute bits we don't care about.
7438 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7440 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7443 if (Op1->uge(TypeBits)) {
7444 if (I.getOpcode() != Instruction::AShr)
7445 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7447 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7452 // ((X*C1) << C2) == (X * (C1 << C2))
7453 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7454 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7455 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7456 return BinaryOperator::CreateMul(BO->getOperand(0),
7457 ConstantExpr::getShl(BOOp, Op1));
7459 // Try to fold constant and into select arguments.
7460 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7461 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7463 if (isa<PHINode>(Op0))
7464 if (Instruction *NV = FoldOpIntoPhi(I))
7467 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7468 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7469 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7470 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7471 // place. Don't try to do this transformation in this case. Also, we
7472 // require that the input operand is a shift-by-constant so that we have
7473 // confidence that the shifts will get folded together. We could do this
7474 // xform in more cases, but it is unlikely to be profitable.
7475 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7476 isa<ConstantInt>(TrOp->getOperand(1))) {
7477 // Okay, we'll do this xform. Make the shift of shift.
7478 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7479 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7481 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7483 // For logical shifts, the truncation has the effect of making the high
7484 // part of the register be zeros. Emulate this by inserting an AND to
7485 // clear the top bits as needed. This 'and' will usually be zapped by
7486 // other xforms later if dead.
7487 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7488 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7489 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7491 // The mask we constructed says what the trunc would do if occurring
7492 // between the shifts. We want to know the effect *after* the second
7493 // shift. We know that it is a logical shift by a constant, so adjust the
7494 // mask as appropriate.
7495 if (I.getOpcode() == Instruction::Shl)
7496 MaskV <<= Op1->getZExtValue();
7498 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7499 MaskV = MaskV.lshr(Op1->getZExtValue());
7503 BinaryOperator::CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7505 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7507 // Return the value truncated to the interesting size.
7508 return new TruncInst(And, I.getType());
7512 if (Op0->hasOneUse()) {
7513 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7514 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7517 switch (Op0BO->getOpcode()) {
7519 case Instruction::Add:
7520 case Instruction::And:
7521 case Instruction::Or:
7522 case Instruction::Xor: {
7523 // These operators commute.
7524 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7525 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7526 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7528 Instruction *YS = BinaryOperator::CreateShl(
7529 Op0BO->getOperand(0), Op1,
7531 InsertNewInstBefore(YS, I); // (Y << C)
7533 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7534 Op0BO->getOperand(1)->getName());
7535 InsertNewInstBefore(X, I); // (X + (Y << C))
7536 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7537 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7538 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7541 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7542 Value *Op0BOOp1 = Op0BO->getOperand(1);
7543 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7545 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7546 m_ConstantInt(CC))) &&
7547 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7548 Instruction *YS = BinaryOperator::CreateShl(
7549 Op0BO->getOperand(0), Op1,
7551 InsertNewInstBefore(YS, I); // (Y << C)
7553 BinaryOperator::CreateAnd(V1,
7554 ConstantExpr::getShl(CC, Op1),
7555 V1->getName()+".mask");
7556 InsertNewInstBefore(XM, I); // X & (CC << C)
7558 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7563 case Instruction::Sub: {
7564 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7565 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7566 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7567 m_Specific(Op1)))) {
7568 Instruction *YS = BinaryOperator::CreateShl(
7569 Op0BO->getOperand(1), Op1,
7571 InsertNewInstBefore(YS, I); // (Y << C)
7573 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7574 Op0BO->getOperand(0)->getName());
7575 InsertNewInstBefore(X, I); // (X + (Y << C))
7576 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7577 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7578 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7581 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7582 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7583 match(Op0BO->getOperand(0),
7584 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7585 m_ConstantInt(CC))) && V2 == Op1 &&
7586 cast<BinaryOperator>(Op0BO->getOperand(0))
7587 ->getOperand(0)->hasOneUse()) {
7588 Instruction *YS = BinaryOperator::CreateShl(
7589 Op0BO->getOperand(1), Op1,
7591 InsertNewInstBefore(YS, I); // (Y << C)
7593 BinaryOperator::CreateAnd(V1,
7594 ConstantExpr::getShl(CC, Op1),
7595 V1->getName()+".mask");
7596 InsertNewInstBefore(XM, I); // X & (CC << C)
7598 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7606 // If the operand is an bitwise operator with a constant RHS, and the
7607 // shift is the only use, we can pull it out of the shift.
7608 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7609 bool isValid = true; // Valid only for And, Or, Xor
7610 bool highBitSet = false; // Transform if high bit of constant set?
7612 switch (Op0BO->getOpcode()) {
7613 default: isValid = false; break; // Do not perform transform!
7614 case Instruction::Add:
7615 isValid = isLeftShift;
7617 case Instruction::Or:
7618 case Instruction::Xor:
7621 case Instruction::And:
7626 // If this is a signed shift right, and the high bit is modified
7627 // by the logical operation, do not perform the transformation.
7628 // The highBitSet boolean indicates the value of the high bit of
7629 // the constant which would cause it to be modified for this
7632 if (isValid && I.getOpcode() == Instruction::AShr)
7633 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7636 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7638 Instruction *NewShift =
7639 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7640 InsertNewInstBefore(NewShift, I);
7641 NewShift->takeName(Op0BO);
7643 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7650 // Find out if this is a shift of a shift by a constant.
7651 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7652 if (ShiftOp && !ShiftOp->isShift())
7655 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7656 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7657 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7658 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7659 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7660 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7661 Value *X = ShiftOp->getOperand(0);
7663 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7665 const IntegerType *Ty = cast<IntegerType>(I.getType());
7667 // Check for (X << c1) << c2 and (X >> c1) >> c2
7668 if (I.getOpcode() == ShiftOp->getOpcode()) {
7669 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7671 if (AmtSum >= TypeBits) {
7672 if (I.getOpcode() != Instruction::AShr)
7673 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7674 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7677 return BinaryOperator::Create(I.getOpcode(), X,
7678 ConstantInt::get(Ty, AmtSum));
7679 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7680 I.getOpcode() == Instruction::AShr) {
7681 if (AmtSum >= TypeBits)
7682 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7684 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7685 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7686 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7687 I.getOpcode() == Instruction::LShr) {
7688 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7689 if (AmtSum >= TypeBits)
7690 AmtSum = TypeBits-1;
7692 Instruction *Shift =
7693 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7694 InsertNewInstBefore(Shift, I);
7696 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7697 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7700 // Okay, if we get here, one shift must be left, and the other shift must be
7701 // right. See if the amounts are equal.
7702 if (ShiftAmt1 == ShiftAmt2) {
7703 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7704 if (I.getOpcode() == Instruction::Shl) {
7705 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7706 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7708 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7709 if (I.getOpcode() == Instruction::LShr) {
7710 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7711 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7713 // We can simplify ((X << C) >>s C) into a trunc + sext.
7714 // NOTE: we could do this for any C, but that would make 'unusual' integer
7715 // types. For now, just stick to ones well-supported by the code
7717 const Type *SExtType = 0;
7718 switch (Ty->getBitWidth() - ShiftAmt1) {
7725 SExtType = IntegerType::get(*Context, Ty->getBitWidth() - ShiftAmt1);
7730 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7731 InsertNewInstBefore(NewTrunc, I);
7732 return new SExtInst(NewTrunc, Ty);
7734 // Otherwise, we can't handle it yet.
7735 } else if (ShiftAmt1 < ShiftAmt2) {
7736 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7738 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7739 if (I.getOpcode() == Instruction::Shl) {
7740 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7741 ShiftOp->getOpcode() == Instruction::AShr);
7742 Instruction *Shift =
7743 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7744 InsertNewInstBefore(Shift, I);
7746 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7747 return BinaryOperator::CreateAnd(Shift,
7748 ConstantInt::get(*Context, Mask));
7751 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7752 if (I.getOpcode() == Instruction::LShr) {
7753 assert(ShiftOp->getOpcode() == Instruction::Shl);
7754 Instruction *Shift =
7755 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7756 InsertNewInstBefore(Shift, I);
7758 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7759 return BinaryOperator::CreateAnd(Shift,
7760 ConstantInt::get(*Context, Mask));
7763 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7765 assert(ShiftAmt2 < ShiftAmt1);
7766 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7768 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7769 if (I.getOpcode() == Instruction::Shl) {
7770 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7771 ShiftOp->getOpcode() == Instruction::AShr);
7772 Instruction *Shift =
7773 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7774 ConstantInt::get(Ty, ShiftDiff));
7775 InsertNewInstBefore(Shift, I);
7777 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7778 return BinaryOperator::CreateAnd(Shift,
7779 ConstantInt::get(*Context, Mask));
7782 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7783 if (I.getOpcode() == Instruction::LShr) {
7784 assert(ShiftOp->getOpcode() == Instruction::Shl);
7785 Instruction *Shift =
7786 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7787 InsertNewInstBefore(Shift, I);
7789 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7790 return BinaryOperator::CreateAnd(Shift,
7791 ConstantInt::get(*Context, Mask));
7794 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7801 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7802 /// expression. If so, decompose it, returning some value X, such that Val is
7805 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7806 int &Offset, LLVMContext *Context) {
7807 assert(Val->getType() == Type::getInt32Ty(*Context) && "Unexpected allocation size type!");
7808 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7809 Offset = CI->getZExtValue();
7811 return ConstantInt::get(Type::getInt32Ty(*Context), 0);
7812 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7813 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7814 if (I->getOpcode() == Instruction::Shl) {
7815 // This is a value scaled by '1 << the shift amt'.
7816 Scale = 1U << RHS->getZExtValue();
7818 return I->getOperand(0);
7819 } else if (I->getOpcode() == Instruction::Mul) {
7820 // This value is scaled by 'RHS'.
7821 Scale = RHS->getZExtValue();
7823 return I->getOperand(0);
7824 } else if (I->getOpcode() == Instruction::Add) {
7825 // We have X+C. Check to see if we really have (X*C2)+C1,
7826 // where C1 is divisible by C2.
7829 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7831 Offset += RHS->getZExtValue();
7838 // Otherwise, we can't look past this.
7845 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7846 /// try to eliminate the cast by moving the type information into the alloc.
7847 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7848 AllocationInst &AI) {
7849 const PointerType *PTy = cast<PointerType>(CI.getType());
7851 // Remove any uses of AI that are dead.
7852 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7854 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7855 Instruction *User = cast<Instruction>(*UI++);
7856 if (isInstructionTriviallyDead(User)) {
7857 while (UI != E && *UI == User)
7858 ++UI; // If this instruction uses AI more than once, don't break UI.
7861 DEBUG(errs() << "IC: DCE: " << *User << '\n');
7862 EraseInstFromFunction(*User);
7866 // This requires TargetData to get the alloca alignment and size information.
7869 // Get the type really allocated and the type casted to.
7870 const Type *AllocElTy = AI.getAllocatedType();
7871 const Type *CastElTy = PTy->getElementType();
7872 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7874 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7875 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7876 if (CastElTyAlign < AllocElTyAlign) return 0;
7878 // If the allocation has multiple uses, only promote it if we are strictly
7879 // increasing the alignment of the resultant allocation. If we keep it the
7880 // same, we open the door to infinite loops of various kinds. (A reference
7881 // from a dbg.declare doesn't count as a use for this purpose.)
7882 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7883 CastElTyAlign == AllocElTyAlign) return 0;
7885 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7886 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7887 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7889 // See if we can satisfy the modulus by pulling a scale out of the array
7891 unsigned ArraySizeScale;
7893 Value *NumElements = // See if the array size is a decomposable linear expr.
7894 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7895 ArrayOffset, Context);
7897 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7899 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7900 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7902 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7907 // If the allocation size is constant, form a constant mul expression
7908 Amt = ConstantInt::get(Type::getInt32Ty(*Context), Scale);
7909 if (isa<ConstantInt>(NumElements))
7910 Amt = ConstantExpr::getMul(cast<ConstantInt>(NumElements),
7911 cast<ConstantInt>(Amt));
7912 // otherwise multiply the amount and the number of elements
7914 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7915 Amt = InsertNewInstBefore(Tmp, AI);
7919 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7920 Value *Off = ConstantInt::get(Type::getInt32Ty(*Context), Offset, true);
7921 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7922 Amt = InsertNewInstBefore(Tmp, AI);
7925 AllocationInst *New;
7926 if (isa<MallocInst>(AI))
7927 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7929 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7930 InsertNewInstBefore(New, AI);
7933 // If the allocation has one real use plus a dbg.declare, just remove the
7935 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7936 EraseInstFromFunction(*DI);
7938 // If the allocation has multiple real uses, insert a cast and change all
7939 // things that used it to use the new cast. This will also hack on CI, but it
7941 else if (!AI.hasOneUse()) {
7942 AddUsesToWorkList(AI);
7943 // New is the allocation instruction, pointer typed. AI is the original
7944 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7945 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7946 InsertNewInstBefore(NewCast, AI);
7947 AI.replaceAllUsesWith(NewCast);
7949 return ReplaceInstUsesWith(CI, New);
7952 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7953 /// and return it as type Ty without inserting any new casts and without
7954 /// changing the computed value. This is used by code that tries to decide
7955 /// whether promoting or shrinking integer operations to wider or smaller types
7956 /// will allow us to eliminate a truncate or extend.
7958 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7959 /// extension operation if Ty is larger.
7961 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7962 /// should return true if trunc(V) can be computed by computing V in the smaller
7963 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7964 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7965 /// efficiently truncated.
7967 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7968 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7969 /// the final result.
7970 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7972 int &NumCastsRemoved){
7973 // We can always evaluate constants in another type.
7974 if (isa<Constant>(V))
7977 Instruction *I = dyn_cast<Instruction>(V);
7978 if (!I) return false;
7980 const Type *OrigTy = V->getType();
7982 // If this is an extension or truncate, we can often eliminate it.
7983 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7984 // If this is a cast from the destination type, we can trivially eliminate
7985 // it, and this will remove a cast overall.
7986 if (I->getOperand(0)->getType() == Ty) {
7987 // If the first operand is itself a cast, and is eliminable, do not count
7988 // this as an eliminable cast. We would prefer to eliminate those two
7990 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7996 // We can't extend or shrink something that has multiple uses: doing so would
7997 // require duplicating the instruction in general, which isn't profitable.
7998 if (!I->hasOneUse()) return false;
8000 unsigned Opc = I->getOpcode();
8002 case Instruction::Add:
8003 case Instruction::Sub:
8004 case Instruction::Mul:
8005 case Instruction::And:
8006 case Instruction::Or:
8007 case Instruction::Xor:
8008 // These operators can all arbitrarily be extended or truncated.
8009 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8011 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
8014 case Instruction::UDiv:
8015 case Instruction::URem: {
8016 // UDiv and URem can be truncated if all the truncated bits are zero.
8017 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8018 uint32_t BitWidth = Ty->getScalarSizeInBits();
8019 if (BitWidth < OrigBitWidth) {
8020 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
8021 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
8022 MaskedValueIsZero(I->getOperand(1), Mask)) {
8023 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8025 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
8031 case Instruction::Shl:
8032 // If we are truncating the result of this SHL, and if it's a shift of a
8033 // constant amount, we can always perform a SHL in a smaller type.
8034 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8035 uint32_t BitWidth = Ty->getScalarSizeInBits();
8036 if (BitWidth < OrigTy->getScalarSizeInBits() &&
8037 CI->getLimitedValue(BitWidth) < BitWidth)
8038 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8042 case Instruction::LShr:
8043 // If this is a truncate of a logical shr, we can truncate it to a smaller
8044 // lshr iff we know that the bits we would otherwise be shifting in are
8046 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8047 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8048 uint32_t BitWidth = Ty->getScalarSizeInBits();
8049 if (BitWidth < OrigBitWidth &&
8050 MaskedValueIsZero(I->getOperand(0),
8051 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
8052 CI->getLimitedValue(BitWidth) < BitWidth) {
8053 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8058 case Instruction::ZExt:
8059 case Instruction::SExt:
8060 case Instruction::Trunc:
8061 // If this is the same kind of case as our original (e.g. zext+zext), we
8062 // can safely replace it. Note that replacing it does not reduce the number
8063 // of casts in the input.
8067 // sext (zext ty1), ty2 -> zext ty2
8068 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8071 case Instruction::Select: {
8072 SelectInst *SI = cast<SelectInst>(I);
8073 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8075 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8078 case Instruction::PHI: {
8079 // We can change a phi if we can change all operands.
8080 PHINode *PN = cast<PHINode>(I);
8081 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8082 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8088 // TODO: Can handle more cases here.
8095 /// EvaluateInDifferentType - Given an expression that
8096 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8097 /// evaluate the expression.
8098 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8100 if (Constant *C = dyn_cast<Constant>(V))
8101 return ConstantExpr::getIntegerCast(C, Ty,
8102 isSigned /*Sext or ZExt*/);
8104 // Otherwise, it must be an instruction.
8105 Instruction *I = cast<Instruction>(V);
8106 Instruction *Res = 0;
8107 unsigned Opc = I->getOpcode();
8109 case Instruction::Add:
8110 case Instruction::Sub:
8111 case Instruction::Mul:
8112 case Instruction::And:
8113 case Instruction::Or:
8114 case Instruction::Xor:
8115 case Instruction::AShr:
8116 case Instruction::LShr:
8117 case Instruction::Shl:
8118 case Instruction::UDiv:
8119 case Instruction::URem: {
8120 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8121 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8122 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8125 case Instruction::Trunc:
8126 case Instruction::ZExt:
8127 case Instruction::SExt:
8128 // If the source type of the cast is the type we're trying for then we can
8129 // just return the source. There's no need to insert it because it is not
8131 if (I->getOperand(0)->getType() == Ty)
8132 return I->getOperand(0);
8134 // Otherwise, must be the same type of cast, so just reinsert a new one.
8135 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8138 case Instruction::Select: {
8139 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8140 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8141 Res = SelectInst::Create(I->getOperand(0), True, False);
8144 case Instruction::PHI: {
8145 PHINode *OPN = cast<PHINode>(I);
8146 PHINode *NPN = PHINode::Create(Ty);
8147 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8148 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8149 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8155 // TODO: Can handle more cases here.
8156 llvm_unreachable("Unreachable!");
8161 return InsertNewInstBefore(Res, *I);
8164 /// @brief Implement the transforms common to all CastInst visitors.
8165 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8166 Value *Src = CI.getOperand(0);
8168 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8169 // eliminate it now.
8170 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8171 if (Instruction::CastOps opc =
8172 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8173 // The first cast (CSrc) is eliminable so we need to fix up or replace
8174 // the second cast (CI). CSrc will then have a good chance of being dead.
8175 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8179 // If we are casting a select then fold the cast into the select
8180 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8181 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8184 // If we are casting a PHI then fold the cast into the PHI
8185 if (isa<PHINode>(Src))
8186 if (Instruction *NV = FoldOpIntoPhi(CI))
8192 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8193 /// or not there is a sequence of GEP indices into the type that will land us at
8194 /// the specified offset. If so, fill them into NewIndices and return the
8195 /// resultant element type, otherwise return null.
8196 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8197 SmallVectorImpl<Value*> &NewIndices,
8198 const TargetData *TD,
8199 LLVMContext *Context) {
8201 if (!Ty->isSized()) return 0;
8203 // Start with the index over the outer type. Note that the type size
8204 // might be zero (even if the offset isn't zero) if the indexed type
8205 // is something like [0 x {int, int}]
8206 const Type *IntPtrTy = TD->getIntPtrType(*Context);
8207 int64_t FirstIdx = 0;
8208 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8209 FirstIdx = Offset/TySize;
8210 Offset -= FirstIdx*TySize;
8212 // Handle hosts where % returns negative instead of values [0..TySize).
8216 assert(Offset >= 0);
8218 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8221 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8223 // Index into the types. If we fail, set OrigBase to null.
8225 // Indexing into tail padding between struct/array elements.
8226 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8229 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8230 const StructLayout *SL = TD->getStructLayout(STy);
8231 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8232 "Offset must stay within the indexed type");
8234 unsigned Elt = SL->getElementContainingOffset(Offset);
8235 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Elt));
8237 Offset -= SL->getElementOffset(Elt);
8238 Ty = STy->getElementType(Elt);
8239 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8240 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8241 assert(EltSize && "Cannot index into a zero-sized array");
8242 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8244 Ty = AT->getElementType();
8246 // Otherwise, we can't index into the middle of this atomic type, bail.
8254 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8255 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8256 Value *Src = CI.getOperand(0);
8258 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8259 // If casting the result of a getelementptr instruction with no offset, turn
8260 // this into a cast of the original pointer!
8261 if (GEP->hasAllZeroIndices()) {
8262 // Changing the cast operand is usually not a good idea but it is safe
8263 // here because the pointer operand is being replaced with another
8264 // pointer operand so the opcode doesn't need to change.
8266 CI.setOperand(0, GEP->getOperand(0));
8270 // If the GEP has a single use, and the base pointer is a bitcast, and the
8271 // GEP computes a constant offset, see if we can convert these three
8272 // instructions into fewer. This typically happens with unions and other
8273 // non-type-safe code.
8274 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8275 if (GEP->hasAllConstantIndices()) {
8276 // We are guaranteed to get a constant from EmitGEPOffset.
8277 ConstantInt *OffsetV =
8278 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8279 int64_t Offset = OffsetV->getSExtValue();
8281 // Get the base pointer input of the bitcast, and the type it points to.
8282 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8283 const Type *GEPIdxTy =
8284 cast<PointerType>(OrigBase->getType())->getElementType();
8285 SmallVector<Value*, 8> NewIndices;
8286 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8287 // If we were able to index down into an element, create the GEP
8288 // and bitcast the result. This eliminates one bitcast, potentially
8290 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
8292 NewIndices.end(), "");
8293 InsertNewInstBefore(NGEP, CI);
8294 NGEP->takeName(GEP);
8295 if (cast<GEPOperator>(GEP)->isInBounds())
8296 cast<GEPOperator>(NGEP)->setIsInBounds(true);
8298 if (isa<BitCastInst>(CI))
8299 return new BitCastInst(NGEP, CI.getType());
8300 assert(isa<PtrToIntInst>(CI));
8301 return new PtrToIntInst(NGEP, CI.getType());
8307 return commonCastTransforms(CI);
8310 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8311 /// type like i42. We don't want to introduce operations on random non-legal
8312 /// integer types where they don't already exist in the code. In the future,
8313 /// we should consider making this based off target-data, so that 32-bit targets
8314 /// won't get i64 operations etc.
8315 static bool isSafeIntegerType(const Type *Ty) {
8316 switch (Ty->getPrimitiveSizeInBits()) {
8327 /// commonIntCastTransforms - This function implements the common transforms
8328 /// for trunc, zext, and sext.
8329 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8330 if (Instruction *Result = commonCastTransforms(CI))
8333 Value *Src = CI.getOperand(0);
8334 const Type *SrcTy = Src->getType();
8335 const Type *DestTy = CI.getType();
8336 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8337 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8339 // See if we can simplify any instructions used by the LHS whose sole
8340 // purpose is to compute bits we don't care about.
8341 if (SimplifyDemandedInstructionBits(CI))
8344 // If the source isn't an instruction or has more than one use then we
8345 // can't do anything more.
8346 Instruction *SrcI = dyn_cast<Instruction>(Src);
8347 if (!SrcI || !Src->hasOneUse())
8350 // Attempt to propagate the cast into the instruction for int->int casts.
8351 int NumCastsRemoved = 0;
8352 // Only do this if the dest type is a simple type, don't convert the
8353 // expression tree to something weird like i93 unless the source is also
8355 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8356 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8357 CanEvaluateInDifferentType(SrcI, DestTy,
8358 CI.getOpcode(), NumCastsRemoved)) {
8359 // If this cast is a truncate, evaluting in a different type always
8360 // eliminates the cast, so it is always a win. If this is a zero-extension,
8361 // we need to do an AND to maintain the clear top-part of the computation,
8362 // so we require that the input have eliminated at least one cast. If this
8363 // is a sign extension, we insert two new casts (to do the extension) so we
8364 // require that two casts have been eliminated.
8365 bool DoXForm = false;
8366 bool JustReplace = false;
8367 switch (CI.getOpcode()) {
8369 // All the others use floating point so we shouldn't actually
8370 // get here because of the check above.
8371 llvm_unreachable("Unknown cast type");
8372 case Instruction::Trunc:
8375 case Instruction::ZExt: {
8376 DoXForm = NumCastsRemoved >= 1;
8377 if (!DoXForm && 0) {
8378 // If it's unnecessary to issue an AND to clear the high bits, it's
8379 // always profitable to do this xform.
8380 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8381 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8382 if (MaskedValueIsZero(TryRes, Mask))
8383 return ReplaceInstUsesWith(CI, TryRes);
8385 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8386 if (TryI->use_empty())
8387 EraseInstFromFunction(*TryI);
8391 case Instruction::SExt: {
8392 DoXForm = NumCastsRemoved >= 2;
8393 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8394 // If we do not have to emit the truncate + sext pair, then it's always
8395 // profitable to do this xform.
8397 // It's not safe to eliminate the trunc + sext pair if one of the
8398 // eliminated cast is a truncate. e.g.
8399 // t2 = trunc i32 t1 to i16
8400 // t3 = sext i16 t2 to i32
8403 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8404 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8405 if (NumSignBits > (DestBitSize - SrcBitSize))
8406 return ReplaceInstUsesWith(CI, TryRes);
8408 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8409 if (TryI->use_empty())
8410 EraseInstFromFunction(*TryI);
8417 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
8418 " to avoid cast: " << CI);
8419 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8420 CI.getOpcode() == Instruction::SExt);
8422 // Just replace this cast with the result.
8423 return ReplaceInstUsesWith(CI, Res);
8425 assert(Res->getType() == DestTy);
8426 switch (CI.getOpcode()) {
8427 default: llvm_unreachable("Unknown cast type!");
8428 case Instruction::Trunc:
8429 // Just replace this cast with the result.
8430 return ReplaceInstUsesWith(CI, Res);
8431 case Instruction::ZExt: {
8432 assert(SrcBitSize < DestBitSize && "Not a zext?");
8434 // If the high bits are already zero, just replace this cast with the
8436 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8437 if (MaskedValueIsZero(Res, Mask))
8438 return ReplaceInstUsesWith(CI, Res);
8440 // We need to emit an AND to clear the high bits.
8441 Constant *C = ConstantInt::get(*Context,
8442 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8443 return BinaryOperator::CreateAnd(Res, C);
8445 case Instruction::SExt: {
8446 // If the high bits are already filled with sign bit, just replace this
8447 // cast with the result.
8448 unsigned NumSignBits = ComputeNumSignBits(Res);
8449 if (NumSignBits > (DestBitSize - SrcBitSize))
8450 return ReplaceInstUsesWith(CI, Res);
8452 // We need to emit a cast to truncate, then a cast to sext.
8453 return CastInst::Create(Instruction::SExt,
8454 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8461 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8462 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8464 switch (SrcI->getOpcode()) {
8465 case Instruction::Add:
8466 case Instruction::Mul:
8467 case Instruction::And:
8468 case Instruction::Or:
8469 case Instruction::Xor:
8470 // If we are discarding information, rewrite.
8471 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8472 // Don't insert two casts unless at least one can be eliminated.
8473 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8474 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8475 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8476 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8477 return BinaryOperator::Create(
8478 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8482 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8483 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8484 SrcI->getOpcode() == Instruction::Xor &&
8485 Op1 == ConstantInt::getTrue(*Context) &&
8486 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8487 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8488 return BinaryOperator::CreateXor(New,
8489 ConstantInt::get(CI.getType(), 1));
8493 case Instruction::Shl: {
8494 // Canonicalize trunc inside shl, if we can.
8495 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8496 if (CI && DestBitSize < SrcBitSize &&
8497 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8498 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8499 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8500 return BinaryOperator::CreateShl(Op0c, Op1c);
8508 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8509 if (Instruction *Result = commonIntCastTransforms(CI))
8512 Value *Src = CI.getOperand(0);
8513 const Type *Ty = CI.getType();
8514 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8515 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8517 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8518 if (DestBitWidth == 1) {
8519 Constant *One = ConstantInt::get(Src->getType(), 1);
8520 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8521 Value *Zero = Constant::getNullValue(Src->getType());
8522 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8525 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8526 ConstantInt *ShAmtV = 0;
8528 if (Src->hasOneUse() &&
8529 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8530 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8532 // Get a mask for the bits shifting in.
8533 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8534 if (MaskedValueIsZero(ShiftOp, Mask)) {
8535 if (ShAmt >= DestBitWidth) // All zeros.
8536 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8538 // Okay, we can shrink this. Truncate the input, then return a new
8540 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8541 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8542 return BinaryOperator::CreateLShr(V1, V2);
8549 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8550 /// in order to eliminate the icmp.
8551 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8553 // If we are just checking for a icmp eq of a single bit and zext'ing it
8554 // to an integer, then shift the bit to the appropriate place and then
8555 // cast to integer to avoid the comparison.
8556 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8557 const APInt &Op1CV = Op1C->getValue();
8559 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8560 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8561 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8562 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8563 if (!DoXform) return ICI;
8565 Value *In = ICI->getOperand(0);
8566 Value *Sh = ConstantInt::get(In->getType(),
8567 In->getType()->getScalarSizeInBits()-1);
8568 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8569 In->getName()+".lobit"),
8571 if (In->getType() != CI.getType())
8572 In = CastInst::CreateIntegerCast(In, CI.getType(),
8573 false/*ZExt*/, "tmp", &CI);
8575 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8576 Constant *One = ConstantInt::get(In->getType(), 1);
8577 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8578 In->getName()+".not"),
8582 return ReplaceInstUsesWith(CI, In);
8587 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8588 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8589 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8590 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8591 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8592 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8593 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8594 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8595 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8596 // This only works for EQ and NE
8597 ICI->isEquality()) {
8598 // If Op1C some other power of two, convert:
8599 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8600 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8601 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8602 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8604 APInt KnownZeroMask(~KnownZero);
8605 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8606 if (!DoXform) return ICI;
8608 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8609 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8610 // (X&4) == 2 --> false
8611 // (X&4) != 2 --> true
8612 Constant *Res = ConstantInt::get(Type::getInt1Ty(*Context), isNE);
8613 Res = ConstantExpr::getZExt(Res, CI.getType());
8614 return ReplaceInstUsesWith(CI, Res);
8617 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8618 Value *In = ICI->getOperand(0);
8620 // Perform a logical shr by shiftamt.
8621 // Insert the shift to put the result in the low bit.
8622 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8623 ConstantInt::get(In->getType(), ShiftAmt),
8624 In->getName()+".lobit"), CI);
8627 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8628 Constant *One = ConstantInt::get(In->getType(), 1);
8629 In = BinaryOperator::CreateXor(In, One, "tmp");
8630 InsertNewInstBefore(cast<Instruction>(In), CI);
8633 if (CI.getType() == In->getType())
8634 return ReplaceInstUsesWith(CI, In);
8636 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8644 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8645 // If one of the common conversion will work ..
8646 if (Instruction *Result = commonIntCastTransforms(CI))
8649 Value *Src = CI.getOperand(0);
8651 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8652 // types and if the sizes are just right we can convert this into a logical
8653 // 'and' which will be much cheaper than the pair of casts.
8654 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8655 // Get the sizes of the types involved. We know that the intermediate type
8656 // will be smaller than A or C, but don't know the relation between A and C.
8657 Value *A = CSrc->getOperand(0);
8658 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8659 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8660 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8661 // If we're actually extending zero bits, then if
8662 // SrcSize < DstSize: zext(a & mask)
8663 // SrcSize == DstSize: a & mask
8664 // SrcSize > DstSize: trunc(a) & mask
8665 if (SrcSize < DstSize) {
8666 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8667 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8669 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8670 InsertNewInstBefore(And, CI);
8671 return new ZExtInst(And, CI.getType());
8672 } else if (SrcSize == DstSize) {
8673 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8674 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8676 } else if (SrcSize > DstSize) {
8677 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8678 InsertNewInstBefore(Trunc, CI);
8679 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8680 return BinaryOperator::CreateAnd(Trunc,
8681 ConstantInt::get(Trunc->getType(),
8686 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8687 return transformZExtICmp(ICI, CI);
8689 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8690 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8691 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8692 // of the (zext icmp) will be transformed.
8693 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8694 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8695 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8696 (transformZExtICmp(LHS, CI, false) ||
8697 transformZExtICmp(RHS, CI, false))) {
8698 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8699 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8700 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8704 // zext(trunc(t) & C) -> (t & zext(C)).
8705 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8706 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8707 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8708 Value *TI0 = TI->getOperand(0);
8709 if (TI0->getType() == CI.getType())
8711 BinaryOperator::CreateAnd(TI0,
8712 ConstantExpr::getZExt(C, CI.getType()));
8715 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8716 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8717 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8718 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8719 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8720 And->getOperand(1) == C)
8721 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8722 Value *TI0 = TI->getOperand(0);
8723 if (TI0->getType() == CI.getType()) {
8724 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8725 Instruction *NewAnd = BinaryOperator::CreateAnd(TI0, ZC, "tmp");
8726 InsertNewInstBefore(NewAnd, *And);
8727 return BinaryOperator::CreateXor(NewAnd, ZC);
8734 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8735 if (Instruction *I = commonIntCastTransforms(CI))
8738 Value *Src = CI.getOperand(0);
8740 // Canonicalize sign-extend from i1 to a select.
8741 if (Src->getType() == Type::getInt1Ty(*Context))
8742 return SelectInst::Create(Src,
8743 Constant::getAllOnesValue(CI.getType()),
8744 Constant::getNullValue(CI.getType()));
8746 // See if the value being truncated is already sign extended. If so, just
8747 // eliminate the trunc/sext pair.
8748 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8749 Value *Op = cast<User>(Src)->getOperand(0);
8750 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8751 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8752 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8753 unsigned NumSignBits = ComputeNumSignBits(Op);
8755 if (OpBits == DestBits) {
8756 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8757 // bits, it is already ready.
8758 if (NumSignBits > DestBits-MidBits)
8759 return ReplaceInstUsesWith(CI, Op);
8760 } else if (OpBits < DestBits) {
8761 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8762 // bits, just sext from i32.
8763 if (NumSignBits > OpBits-MidBits)
8764 return new SExtInst(Op, CI.getType(), "tmp");
8766 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8767 // bits, just truncate to i32.
8768 if (NumSignBits > OpBits-MidBits)
8769 return new TruncInst(Op, CI.getType(), "tmp");
8773 // If the input is a shl/ashr pair of a same constant, then this is a sign
8774 // extension from a smaller value. If we could trust arbitrary bitwidth
8775 // integers, we could turn this into a truncate to the smaller bit and then
8776 // use a sext for the whole extension. Since we don't, look deeper and check
8777 // for a truncate. If the source and dest are the same type, eliminate the
8778 // trunc and extend and just do shifts. For example, turn:
8779 // %a = trunc i32 %i to i8
8780 // %b = shl i8 %a, 6
8781 // %c = ashr i8 %b, 6
8782 // %d = sext i8 %c to i32
8784 // %a = shl i32 %i, 30
8785 // %d = ashr i32 %a, 30
8787 ConstantInt *BA = 0, *CA = 0;
8788 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8789 m_ConstantInt(CA))) &&
8790 BA == CA && isa<TruncInst>(A)) {
8791 Value *I = cast<TruncInst>(A)->getOperand(0);
8792 if (I->getType() == CI.getType()) {
8793 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8794 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8795 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8796 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8797 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8799 return BinaryOperator::CreateAShr(I, ShAmtV);
8806 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8807 /// in the specified FP type without changing its value.
8808 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8809 LLVMContext *Context) {
8811 APFloat F = CFP->getValueAPF();
8812 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8814 return ConstantFP::get(*Context, F);
8818 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8819 /// through it until we get the source value.
8820 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8821 if (Instruction *I = dyn_cast<Instruction>(V))
8822 if (I->getOpcode() == Instruction::FPExt)
8823 return LookThroughFPExtensions(I->getOperand(0), Context);
8825 // If this value is a constant, return the constant in the smallest FP type
8826 // that can accurately represent it. This allows us to turn
8827 // (float)((double)X+2.0) into x+2.0f.
8828 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8829 if (CFP->getType() == Type::getPPC_FP128Ty(*Context))
8830 return V; // No constant folding of this.
8831 // See if the value can be truncated to float and then reextended.
8832 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8834 if (CFP->getType() == Type::getDoubleTy(*Context))
8835 return V; // Won't shrink.
8836 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8838 // Don't try to shrink to various long double types.
8844 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8845 if (Instruction *I = commonCastTransforms(CI))
8848 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8849 // smaller than the destination type, we can eliminate the truncate by doing
8850 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8851 // many builtins (sqrt, etc).
8852 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8853 if (OpI && OpI->hasOneUse()) {
8854 switch (OpI->getOpcode()) {
8856 case Instruction::FAdd:
8857 case Instruction::FSub:
8858 case Instruction::FMul:
8859 case Instruction::FDiv:
8860 case Instruction::FRem:
8861 const Type *SrcTy = OpI->getType();
8862 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8863 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8864 if (LHSTrunc->getType() != SrcTy &&
8865 RHSTrunc->getType() != SrcTy) {
8866 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8867 // If the source types were both smaller than the destination type of
8868 // the cast, do this xform.
8869 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8870 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8871 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8873 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8875 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8884 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8885 return commonCastTransforms(CI);
8888 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8889 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8891 return commonCastTransforms(FI);
8893 // fptoui(uitofp(X)) --> X
8894 // fptoui(sitofp(X)) --> X
8895 // This is safe if the intermediate type has enough bits in its mantissa to
8896 // accurately represent all values of X. For example, do not do this with
8897 // i64->float->i64. This is also safe for sitofp case, because any negative
8898 // 'X' value would cause an undefined result for the fptoui.
8899 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8900 OpI->getOperand(0)->getType() == FI.getType() &&
8901 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8902 OpI->getType()->getFPMantissaWidth())
8903 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8905 return commonCastTransforms(FI);
8908 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8909 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8911 return commonCastTransforms(FI);
8913 // fptosi(sitofp(X)) --> X
8914 // fptosi(uitofp(X)) --> X
8915 // This is safe if the intermediate type has enough bits in its mantissa to
8916 // accurately represent all values of X. For example, do not do this with
8917 // i64->float->i64. This is also safe for sitofp case, because any negative
8918 // 'X' value would cause an undefined result for the fptoui.
8919 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8920 OpI->getOperand(0)->getType() == FI.getType() &&
8921 (int)FI.getType()->getScalarSizeInBits() <=
8922 OpI->getType()->getFPMantissaWidth())
8923 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8925 return commonCastTransforms(FI);
8928 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8929 return commonCastTransforms(CI);
8932 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8933 return commonCastTransforms(CI);
8936 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8937 // If the destination integer type is smaller than the intptr_t type for
8938 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8939 // trunc to be exposed to other transforms. Don't do this for extending
8940 // ptrtoint's, because we don't know if the target sign or zero extends its
8943 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8944 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8945 TD->getIntPtrType(CI.getContext()),
8947 return new TruncInst(P, CI.getType());
8950 return commonPointerCastTransforms(CI);
8953 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8954 // If the source integer type is larger than the intptr_t type for
8955 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8956 // allows the trunc to be exposed to other transforms. Don't do this for
8957 // extending inttoptr's, because we don't know if the target sign or zero
8958 // extends to pointers.
8960 CI.getOperand(0)->getType()->getScalarSizeInBits() >
8961 TD->getPointerSizeInBits()) {
8962 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8963 TD->getIntPtrType(CI.getContext()),
8965 return new IntToPtrInst(P, CI.getType());
8968 if (Instruction *I = commonCastTransforms(CI))
8974 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8975 // If the operands are integer typed then apply the integer transforms,
8976 // otherwise just apply the common ones.
8977 Value *Src = CI.getOperand(0);
8978 const Type *SrcTy = Src->getType();
8979 const Type *DestTy = CI.getType();
8981 if (isa<PointerType>(SrcTy)) {
8982 if (Instruction *I = commonPointerCastTransforms(CI))
8985 if (Instruction *Result = commonCastTransforms(CI))
8990 // Get rid of casts from one type to the same type. These are useless and can
8991 // be replaced by the operand.
8992 if (DestTy == Src->getType())
8993 return ReplaceInstUsesWith(CI, Src);
8995 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8996 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8997 const Type *DstElTy = DstPTy->getElementType();
8998 const Type *SrcElTy = SrcPTy->getElementType();
9000 // If the address spaces don't match, don't eliminate the bitcast, which is
9001 // required for changing types.
9002 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
9005 // If we are casting a malloc or alloca to a pointer to a type of the same
9006 // size, rewrite the allocation instruction to allocate the "right" type.
9007 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
9008 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
9011 // If the source and destination are pointers, and this cast is equivalent
9012 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
9013 // This can enhance SROA and other transforms that want type-safe pointers.
9014 Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(*Context));
9015 unsigned NumZeros = 0;
9016 while (SrcElTy != DstElTy &&
9017 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
9018 SrcElTy->getNumContainedTypes() /* not "{}" */) {
9019 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
9023 // If we found a path from the src to dest, create the getelementptr now.
9024 if (SrcElTy == DstElTy) {
9025 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
9026 Instruction *GEP = GetElementPtrInst::Create(Src,
9027 Idxs.begin(), Idxs.end(), "",
9028 ((Instruction*) NULL));
9029 cast<GEPOperator>(GEP)->setIsInBounds(true);
9034 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
9035 if (DestVTy->getNumElements() == 1) {
9036 if (!isa<VectorType>(SrcTy)) {
9037 Value *Elem = InsertCastBefore(Instruction::BitCast, Src,
9038 DestVTy->getElementType(), CI);
9039 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
9040 Constant::getNullValue(Type::getInt32Ty(*Context)));
9042 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
9046 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
9047 if (SrcVTy->getNumElements() == 1) {
9048 if (!isa<VectorType>(DestTy)) {
9050 ExtractElementInst::Create(Src, Constant::getNullValue(Type::getInt32Ty(*Context)));
9051 InsertNewInstBefore(Elem, CI);
9052 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
9057 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9058 if (SVI->hasOneUse()) {
9059 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9060 // a bitconvert to a vector with the same # elts.
9061 if (isa<VectorType>(DestTy) &&
9062 cast<VectorType>(DestTy)->getNumElements() ==
9063 SVI->getType()->getNumElements() &&
9064 SVI->getType()->getNumElements() ==
9065 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9067 // If either of the operands is a cast from CI.getType(), then
9068 // evaluating the shuffle in the casted destination's type will allow
9069 // us to eliminate at least one cast.
9070 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9071 Tmp->getOperand(0)->getType() == DestTy) ||
9072 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9073 Tmp->getOperand(0)->getType() == DestTy)) {
9074 Value *LHS = InsertCastBefore(Instruction::BitCast,
9075 SVI->getOperand(0), DestTy, CI);
9076 Value *RHS = InsertCastBefore(Instruction::BitCast,
9077 SVI->getOperand(1), DestTy, CI);
9078 // Return a new shuffle vector. Use the same element ID's, as we
9079 // know the vector types match #elts.
9080 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9088 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9090 /// %D = select %cond, %C, %A
9092 /// %C = select %cond, %B, 0
9095 /// Assuming that the specified instruction is an operand to the select, return
9096 /// a bitmask indicating which operands of this instruction are foldable if they
9097 /// equal the other incoming value of the select.
9099 static unsigned GetSelectFoldableOperands(Instruction *I) {
9100 switch (I->getOpcode()) {
9101 case Instruction::Add:
9102 case Instruction::Mul:
9103 case Instruction::And:
9104 case Instruction::Or:
9105 case Instruction::Xor:
9106 return 3; // Can fold through either operand.
9107 case Instruction::Sub: // Can only fold on the amount subtracted.
9108 case Instruction::Shl: // Can only fold on the shift amount.
9109 case Instruction::LShr:
9110 case Instruction::AShr:
9113 return 0; // Cannot fold
9117 /// GetSelectFoldableConstant - For the same transformation as the previous
9118 /// function, return the identity constant that goes into the select.
9119 static Constant *GetSelectFoldableConstant(Instruction *I,
9120 LLVMContext *Context) {
9121 switch (I->getOpcode()) {
9122 default: llvm_unreachable("This cannot happen!");
9123 case Instruction::Add:
9124 case Instruction::Sub:
9125 case Instruction::Or:
9126 case Instruction::Xor:
9127 case Instruction::Shl:
9128 case Instruction::LShr:
9129 case Instruction::AShr:
9130 return Constant::getNullValue(I->getType());
9131 case Instruction::And:
9132 return Constant::getAllOnesValue(I->getType());
9133 case Instruction::Mul:
9134 return ConstantInt::get(I->getType(), 1);
9138 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9139 /// have the same opcode and only one use each. Try to simplify this.
9140 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9142 if (TI->getNumOperands() == 1) {
9143 // If this is a non-volatile load or a cast from the same type,
9146 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9149 return 0; // unknown unary op.
9152 // Fold this by inserting a select from the input values.
9153 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9154 FI->getOperand(0), SI.getName()+".v");
9155 InsertNewInstBefore(NewSI, SI);
9156 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9160 // Only handle binary operators here.
9161 if (!isa<BinaryOperator>(TI))
9164 // Figure out if the operations have any operands in common.
9165 Value *MatchOp, *OtherOpT, *OtherOpF;
9167 if (TI->getOperand(0) == FI->getOperand(0)) {
9168 MatchOp = TI->getOperand(0);
9169 OtherOpT = TI->getOperand(1);
9170 OtherOpF = FI->getOperand(1);
9171 MatchIsOpZero = true;
9172 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9173 MatchOp = TI->getOperand(1);
9174 OtherOpT = TI->getOperand(0);
9175 OtherOpF = FI->getOperand(0);
9176 MatchIsOpZero = false;
9177 } else if (!TI->isCommutative()) {
9179 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9180 MatchOp = TI->getOperand(0);
9181 OtherOpT = TI->getOperand(1);
9182 OtherOpF = FI->getOperand(0);
9183 MatchIsOpZero = true;
9184 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9185 MatchOp = TI->getOperand(1);
9186 OtherOpT = TI->getOperand(0);
9187 OtherOpF = FI->getOperand(1);
9188 MatchIsOpZero = true;
9193 // If we reach here, they do have operations in common.
9194 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9195 OtherOpF, SI.getName()+".v");
9196 InsertNewInstBefore(NewSI, SI);
9198 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9200 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9202 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9204 llvm_unreachable("Shouldn't get here");
9208 static bool isSelect01(Constant *C1, Constant *C2) {
9209 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9212 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9215 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9218 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9219 /// facilitate further optimization.
9220 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9222 // See the comment above GetSelectFoldableOperands for a description of the
9223 // transformation we are doing here.
9224 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9225 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9226 !isa<Constant>(FalseVal)) {
9227 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9228 unsigned OpToFold = 0;
9229 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9231 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9236 Constant *C = GetSelectFoldableConstant(TVI, Context);
9237 Value *OOp = TVI->getOperand(2-OpToFold);
9238 // Avoid creating select between 2 constants unless it's selecting
9240 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9241 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9242 InsertNewInstBefore(NewSel, SI);
9243 NewSel->takeName(TVI);
9244 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9245 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9246 llvm_unreachable("Unknown instruction!!");
9253 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9254 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9255 !isa<Constant>(TrueVal)) {
9256 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9257 unsigned OpToFold = 0;
9258 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9260 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9265 Constant *C = GetSelectFoldableConstant(FVI, Context);
9266 Value *OOp = FVI->getOperand(2-OpToFold);
9267 // Avoid creating select between 2 constants unless it's selecting
9269 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9270 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9271 InsertNewInstBefore(NewSel, SI);
9272 NewSel->takeName(FVI);
9273 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9274 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9275 llvm_unreachable("Unknown instruction!!");
9285 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9286 /// ICmpInst as its first operand.
9288 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9290 bool Changed = false;
9291 ICmpInst::Predicate Pred = ICI->getPredicate();
9292 Value *CmpLHS = ICI->getOperand(0);
9293 Value *CmpRHS = ICI->getOperand(1);
9294 Value *TrueVal = SI.getTrueValue();
9295 Value *FalseVal = SI.getFalseValue();
9297 // Check cases where the comparison is with a constant that
9298 // can be adjusted to fit the min/max idiom. We may edit ICI in
9299 // place here, so make sure the select is the only user.
9300 if (ICI->hasOneUse())
9301 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9304 case ICmpInst::ICMP_ULT:
9305 case ICmpInst::ICMP_SLT: {
9306 // X < MIN ? T : F --> F
9307 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9308 return ReplaceInstUsesWith(SI, FalseVal);
9309 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9310 Constant *AdjustedRHS = SubOne(CI);
9311 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9312 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9313 Pred = ICmpInst::getSwappedPredicate(Pred);
9314 CmpRHS = AdjustedRHS;
9315 std::swap(FalseVal, TrueVal);
9316 ICI->setPredicate(Pred);
9317 ICI->setOperand(1, CmpRHS);
9318 SI.setOperand(1, TrueVal);
9319 SI.setOperand(2, FalseVal);
9324 case ICmpInst::ICMP_UGT:
9325 case ICmpInst::ICMP_SGT: {
9326 // X > MAX ? T : F --> F
9327 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9328 return ReplaceInstUsesWith(SI, FalseVal);
9329 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9330 Constant *AdjustedRHS = AddOne(CI);
9331 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9332 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9333 Pred = ICmpInst::getSwappedPredicate(Pred);
9334 CmpRHS = AdjustedRHS;
9335 std::swap(FalseVal, TrueVal);
9336 ICI->setPredicate(Pred);
9337 ICI->setOperand(1, CmpRHS);
9338 SI.setOperand(1, TrueVal);
9339 SI.setOperand(2, FalseVal);
9346 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9347 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9348 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9349 if (match(TrueVal, m_ConstantInt<-1>()) &&
9350 match(FalseVal, m_ConstantInt<0>()))
9351 Pred = ICI->getPredicate();
9352 else if (match(TrueVal, m_ConstantInt<0>()) &&
9353 match(FalseVal, m_ConstantInt<-1>()))
9354 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9356 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9357 // If we are just checking for a icmp eq of a single bit and zext'ing it
9358 // to an integer, then shift the bit to the appropriate place and then
9359 // cast to integer to avoid the comparison.
9360 const APInt &Op1CV = CI->getValue();
9362 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9363 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9364 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9365 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9366 Value *In = ICI->getOperand(0);
9367 Value *Sh = ConstantInt::get(In->getType(),
9368 In->getType()->getScalarSizeInBits()-1);
9369 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9370 In->getName()+".lobit"),
9372 if (In->getType() != SI.getType())
9373 In = CastInst::CreateIntegerCast(In, SI.getType(),
9374 true/*SExt*/, "tmp", ICI);
9376 if (Pred == ICmpInst::ICMP_SGT)
9377 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9378 In->getName()+".not"), *ICI);
9380 return ReplaceInstUsesWith(SI, In);
9385 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9386 // Transform (X == Y) ? X : Y -> Y
9387 if (Pred == ICmpInst::ICMP_EQ)
9388 return ReplaceInstUsesWith(SI, FalseVal);
9389 // Transform (X != Y) ? X : Y -> X
9390 if (Pred == ICmpInst::ICMP_NE)
9391 return ReplaceInstUsesWith(SI, TrueVal);
9392 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9394 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9395 // Transform (X == Y) ? Y : X -> X
9396 if (Pred == ICmpInst::ICMP_EQ)
9397 return ReplaceInstUsesWith(SI, FalseVal);
9398 // Transform (X != Y) ? Y : X -> Y
9399 if (Pred == ICmpInst::ICMP_NE)
9400 return ReplaceInstUsesWith(SI, TrueVal);
9401 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9404 /// NOTE: if we wanted to, this is where to detect integer ABS
9406 return Changed ? &SI : 0;
9409 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9410 Value *CondVal = SI.getCondition();
9411 Value *TrueVal = SI.getTrueValue();
9412 Value *FalseVal = SI.getFalseValue();
9414 // select true, X, Y -> X
9415 // select false, X, Y -> Y
9416 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9417 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9419 // select C, X, X -> X
9420 if (TrueVal == FalseVal)
9421 return ReplaceInstUsesWith(SI, TrueVal);
9423 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9424 return ReplaceInstUsesWith(SI, FalseVal);
9425 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9426 return ReplaceInstUsesWith(SI, TrueVal);
9427 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9428 if (isa<Constant>(TrueVal))
9429 return ReplaceInstUsesWith(SI, TrueVal);
9431 return ReplaceInstUsesWith(SI, FalseVal);
9434 if (SI.getType() == Type::getInt1Ty(*Context)) {
9435 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9436 if (C->getZExtValue()) {
9437 // Change: A = select B, true, C --> A = or B, C
9438 return BinaryOperator::CreateOr(CondVal, FalseVal);
9440 // Change: A = select B, false, C --> A = and !B, C
9442 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9443 "not."+CondVal->getName()), SI);
9444 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9446 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9447 if (C->getZExtValue() == false) {
9448 // Change: A = select B, C, false --> A = and B, C
9449 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9451 // Change: A = select B, C, true --> A = or !B, C
9453 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9454 "not."+CondVal->getName()), SI);
9455 return BinaryOperator::CreateOr(NotCond, TrueVal);
9459 // select a, b, a -> a&b
9460 // select a, a, b -> a|b
9461 if (CondVal == TrueVal)
9462 return BinaryOperator::CreateOr(CondVal, FalseVal);
9463 else if (CondVal == FalseVal)
9464 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9467 // Selecting between two integer constants?
9468 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9469 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9470 // select C, 1, 0 -> zext C to int
9471 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9472 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9473 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9474 // select C, 0, 1 -> zext !C to int
9476 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9477 "not."+CondVal->getName()), SI);
9478 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9481 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9482 // If one of the constants is zero (we know they can't both be) and we
9483 // have an icmp instruction with zero, and we have an 'and' with the
9484 // non-constant value, eliminate this whole mess. This corresponds to
9485 // cases like this: ((X & 27) ? 27 : 0)
9486 if (TrueValC->isZero() || FalseValC->isZero())
9487 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9488 cast<Constant>(IC->getOperand(1))->isNullValue())
9489 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9490 if (ICA->getOpcode() == Instruction::And &&
9491 isa<ConstantInt>(ICA->getOperand(1)) &&
9492 (ICA->getOperand(1) == TrueValC ||
9493 ICA->getOperand(1) == FalseValC) &&
9494 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9495 // Okay, now we know that everything is set up, we just don't
9496 // know whether we have a icmp_ne or icmp_eq and whether the
9497 // true or false val is the zero.
9498 bool ShouldNotVal = !TrueValC->isZero();
9499 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9502 V = InsertNewInstBefore(BinaryOperator::Create(
9503 Instruction::Xor, V, ICA->getOperand(1)), SI);
9504 return ReplaceInstUsesWith(SI, V);
9509 // See if we are selecting two values based on a comparison of the two values.
9510 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9511 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9512 // Transform (X == Y) ? X : Y -> Y
9513 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9514 // This is not safe in general for floating point:
9515 // consider X== -0, Y== +0.
9516 // It becomes safe if either operand is a nonzero constant.
9517 ConstantFP *CFPt, *CFPf;
9518 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9519 !CFPt->getValueAPF().isZero()) ||
9520 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9521 !CFPf->getValueAPF().isZero()))
9522 return ReplaceInstUsesWith(SI, FalseVal);
9524 // Transform (X != Y) ? X : Y -> X
9525 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9526 return ReplaceInstUsesWith(SI, TrueVal);
9527 // NOTE: if we wanted to, this is where to detect MIN/MAX
9529 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9530 // Transform (X == Y) ? Y : X -> X
9531 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9532 // This is not safe in general for floating point:
9533 // consider X== -0, Y== +0.
9534 // It becomes safe if either operand is a nonzero constant.
9535 ConstantFP *CFPt, *CFPf;
9536 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9537 !CFPt->getValueAPF().isZero()) ||
9538 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9539 !CFPf->getValueAPF().isZero()))
9540 return ReplaceInstUsesWith(SI, FalseVal);
9542 // Transform (X != Y) ? Y : X -> Y
9543 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9544 return ReplaceInstUsesWith(SI, TrueVal);
9545 // NOTE: if we wanted to, this is where to detect MIN/MAX
9547 // NOTE: if we wanted to, this is where to detect ABS
9550 // See if we are selecting two values based on a comparison of the two values.
9551 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9552 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9555 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9556 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9557 if (TI->hasOneUse() && FI->hasOneUse()) {
9558 Instruction *AddOp = 0, *SubOp = 0;
9560 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9561 if (TI->getOpcode() == FI->getOpcode())
9562 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9565 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9566 // even legal for FP.
9567 if ((TI->getOpcode() == Instruction::Sub &&
9568 FI->getOpcode() == Instruction::Add) ||
9569 (TI->getOpcode() == Instruction::FSub &&
9570 FI->getOpcode() == Instruction::FAdd)) {
9571 AddOp = FI; SubOp = TI;
9572 } else if ((FI->getOpcode() == Instruction::Sub &&
9573 TI->getOpcode() == Instruction::Add) ||
9574 (FI->getOpcode() == Instruction::FSub &&
9575 TI->getOpcode() == Instruction::FAdd)) {
9576 AddOp = TI; SubOp = FI;
9580 Value *OtherAddOp = 0;
9581 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9582 OtherAddOp = AddOp->getOperand(1);
9583 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9584 OtherAddOp = AddOp->getOperand(0);
9588 // So at this point we know we have (Y -> OtherAddOp):
9589 // select C, (add X, Y), (sub X, Z)
9590 Value *NegVal; // Compute -Z
9591 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9592 NegVal = ConstantExpr::getNeg(C);
9594 NegVal = InsertNewInstBefore(
9595 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9599 Value *NewTrueOp = OtherAddOp;
9600 Value *NewFalseOp = NegVal;
9602 std::swap(NewTrueOp, NewFalseOp);
9603 Instruction *NewSel =
9604 SelectInst::Create(CondVal, NewTrueOp,
9605 NewFalseOp, SI.getName() + ".p");
9607 NewSel = InsertNewInstBefore(NewSel, SI);
9608 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9613 // See if we can fold the select into one of our operands.
9614 if (SI.getType()->isInteger()) {
9615 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9620 if (BinaryOperator::isNot(CondVal)) {
9621 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9622 SI.setOperand(1, FalseVal);
9623 SI.setOperand(2, TrueVal);
9630 /// EnforceKnownAlignment - If the specified pointer points to an object that
9631 /// we control, modify the object's alignment to PrefAlign. This isn't
9632 /// often possible though. If alignment is important, a more reliable approach
9633 /// is to simply align all global variables and allocation instructions to
9634 /// their preferred alignment from the beginning.
9636 static unsigned EnforceKnownAlignment(Value *V,
9637 unsigned Align, unsigned PrefAlign) {
9639 User *U = dyn_cast<User>(V);
9640 if (!U) return Align;
9642 switch (Operator::getOpcode(U)) {
9644 case Instruction::BitCast:
9645 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9646 case Instruction::GetElementPtr: {
9647 // If all indexes are zero, it is just the alignment of the base pointer.
9648 bool AllZeroOperands = true;
9649 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9650 if (!isa<Constant>(*i) ||
9651 !cast<Constant>(*i)->isNullValue()) {
9652 AllZeroOperands = false;
9656 if (AllZeroOperands) {
9657 // Treat this like a bitcast.
9658 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9664 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9665 // If there is a large requested alignment and we can, bump up the alignment
9667 if (!GV->isDeclaration()) {
9668 if (GV->getAlignment() >= PrefAlign)
9669 Align = GV->getAlignment();
9671 GV->setAlignment(PrefAlign);
9675 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9676 // If there is a requested alignment and if this is an alloca, round up. We
9677 // don't do this for malloc, because some systems can't respect the request.
9678 if (isa<AllocaInst>(AI)) {
9679 if (AI->getAlignment() >= PrefAlign)
9680 Align = AI->getAlignment();
9682 AI->setAlignment(PrefAlign);
9691 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9692 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9693 /// and it is more than the alignment of the ultimate object, see if we can
9694 /// increase the alignment of the ultimate object, making this check succeed.
9695 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9696 unsigned PrefAlign) {
9697 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9698 sizeof(PrefAlign) * CHAR_BIT;
9699 APInt Mask = APInt::getAllOnesValue(BitWidth);
9700 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9701 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9702 unsigned TrailZ = KnownZero.countTrailingOnes();
9703 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9705 if (PrefAlign > Align)
9706 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9708 // We don't need to make any adjustment.
9712 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9713 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9714 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9715 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9716 unsigned CopyAlign = MI->getAlignment();
9718 if (CopyAlign < MinAlign) {
9719 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9724 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9726 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9727 if (MemOpLength == 0) return 0;
9729 // Source and destination pointer types are always "i8*" for intrinsic. See
9730 // if the size is something we can handle with a single primitive load/store.
9731 // A single load+store correctly handles overlapping memory in the memmove
9733 unsigned Size = MemOpLength->getZExtValue();
9734 if (Size == 0) return MI; // Delete this mem transfer.
9736 if (Size > 8 || (Size&(Size-1)))
9737 return 0; // If not 1/2/4/8 bytes, exit.
9739 // Use an integer load+store unless we can find something better.
9741 PointerType::getUnqual(IntegerType::get(*Context, Size<<3));
9743 // Memcpy forces the use of i8* for the source and destination. That means
9744 // that if you're using memcpy to move one double around, you'll get a cast
9745 // from double* to i8*. We'd much rather use a double load+store rather than
9746 // an i64 load+store, here because this improves the odds that the source or
9747 // dest address will be promotable. See if we can find a better type than the
9748 // integer datatype.
9749 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9750 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9751 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9752 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9753 // down through these levels if so.
9754 while (!SrcETy->isSingleValueType()) {
9755 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9756 if (STy->getNumElements() == 1)
9757 SrcETy = STy->getElementType(0);
9760 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9761 if (ATy->getNumElements() == 1)
9762 SrcETy = ATy->getElementType();
9769 if (SrcETy->isSingleValueType())
9770 NewPtrTy = PointerType::getUnqual(SrcETy);
9775 // If the memcpy/memmove provides better alignment info than we can
9777 SrcAlign = std::max(SrcAlign, CopyAlign);
9778 DstAlign = std::max(DstAlign, CopyAlign);
9780 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9781 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9782 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9783 InsertNewInstBefore(L, *MI);
9784 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9786 // Set the size of the copy to 0, it will be deleted on the next iteration.
9787 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9791 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9792 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9793 if (MI->getAlignment() < Alignment) {
9794 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9799 // Extract the length and alignment and fill if they are constant.
9800 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9801 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9802 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(*Context))
9804 uint64_t Len = LenC->getZExtValue();
9805 Alignment = MI->getAlignment();
9807 // If the length is zero, this is a no-op
9808 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9810 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9811 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9812 const Type *ITy = IntegerType::get(*Context, Len*8); // n=1 -> i8.
9814 Value *Dest = MI->getDest();
9815 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9817 // Alignment 0 is identity for alignment 1 for memset, but not store.
9818 if (Alignment == 0) Alignment = 1;
9820 // Extract the fill value and store.
9821 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9822 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
9823 Dest, false, Alignment), *MI);
9825 // Set the size of the copy to 0, it will be deleted on the next iteration.
9826 MI->setLength(Constant::getNullValue(LenC->getType()));
9834 /// visitCallInst - CallInst simplification. This mostly only handles folding
9835 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9836 /// the heavy lifting.
9838 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9839 // If the caller function is nounwind, mark the call as nounwind, even if the
9841 if (CI.getParent()->getParent()->doesNotThrow() &&
9842 !CI.doesNotThrow()) {
9843 CI.setDoesNotThrow();
9849 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9850 if (!II) return visitCallSite(&CI);
9852 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9854 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9855 bool Changed = false;
9857 // memmove/cpy/set of zero bytes is a noop.
9858 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9859 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9861 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9862 if (CI->getZExtValue() == 1) {
9863 // Replace the instruction with just byte operations. We would
9864 // transform other cases to loads/stores, but we don't know if
9865 // alignment is sufficient.
9869 // If we have a memmove and the source operation is a constant global,
9870 // then the source and dest pointers can't alias, so we can change this
9871 // into a call to memcpy.
9872 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9873 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9874 if (GVSrc->isConstant()) {
9875 Module *M = CI.getParent()->getParent()->getParent();
9876 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9878 Tys[0] = CI.getOperand(3)->getType();
9880 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9884 // memmove(x,x,size) -> noop.
9885 if (MMI->getSource() == MMI->getDest())
9886 return EraseInstFromFunction(CI);
9889 // If we can determine a pointer alignment that is bigger than currently
9890 // set, update the alignment.
9891 if (isa<MemTransferInst>(MI)) {
9892 if (Instruction *I = SimplifyMemTransfer(MI))
9894 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9895 if (Instruction *I = SimplifyMemSet(MSI))
9899 if (Changed) return II;
9902 switch (II->getIntrinsicID()) {
9904 case Intrinsic::bswap:
9905 // bswap(bswap(x)) -> x
9906 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9907 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9908 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9910 case Intrinsic::ppc_altivec_lvx:
9911 case Intrinsic::ppc_altivec_lvxl:
9912 case Intrinsic::x86_sse_loadu_ps:
9913 case Intrinsic::x86_sse2_loadu_pd:
9914 case Intrinsic::x86_sse2_loadu_dq:
9915 // Turn PPC lvx -> load if the pointer is known aligned.
9916 // Turn X86 loadups -> load if the pointer is known aligned.
9917 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9918 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9919 PointerType::getUnqual(II->getType()),
9921 return new LoadInst(Ptr);
9924 case Intrinsic::ppc_altivec_stvx:
9925 case Intrinsic::ppc_altivec_stvxl:
9926 // Turn stvx -> store if the pointer is known aligned.
9927 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9928 const Type *OpPtrTy =
9929 PointerType::getUnqual(II->getOperand(1)->getType());
9930 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9931 return new StoreInst(II->getOperand(1), Ptr);
9934 case Intrinsic::x86_sse_storeu_ps:
9935 case Intrinsic::x86_sse2_storeu_pd:
9936 case Intrinsic::x86_sse2_storeu_dq:
9937 // Turn X86 storeu -> store if the pointer is known aligned.
9938 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9939 const Type *OpPtrTy =
9940 PointerType::getUnqual(II->getOperand(2)->getType());
9941 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9942 return new StoreInst(II->getOperand(2), Ptr);
9946 case Intrinsic::x86_sse_cvttss2si: {
9947 // These intrinsics only demands the 0th element of its input vector. If
9948 // we can simplify the input based on that, do so now.
9950 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9951 APInt DemandedElts(VWidth, 1);
9952 APInt UndefElts(VWidth, 0);
9953 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9955 II->setOperand(1, V);
9961 case Intrinsic::ppc_altivec_vperm:
9962 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9963 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9964 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9966 // Check that all of the elements are integer constants or undefs.
9967 bool AllEltsOk = true;
9968 for (unsigned i = 0; i != 16; ++i) {
9969 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9970 !isa<UndefValue>(Mask->getOperand(i))) {
9977 // Cast the input vectors to byte vectors.
9978 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9979 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9980 Value *Result = UndefValue::get(Op0->getType());
9982 // Only extract each element once.
9983 Value *ExtractedElts[32];
9984 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9986 for (unsigned i = 0; i != 16; ++i) {
9987 if (isa<UndefValue>(Mask->getOperand(i)))
9989 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9990 Idx &= 31; // Match the hardware behavior.
9992 if (ExtractedElts[Idx] == 0) {
9994 ExtractElementInst::Create(Idx < 16 ? Op0 : Op1,
9995 ConstantInt::get(Type::getInt32Ty(*Context), Idx&15, false), "tmp");
9996 InsertNewInstBefore(Elt, CI);
9997 ExtractedElts[Idx] = Elt;
10000 // Insert this value into the result vector.
10001 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
10002 ConstantInt::get(Type::getInt32Ty(*Context), i, false),
10004 InsertNewInstBefore(cast<Instruction>(Result), CI);
10006 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
10011 case Intrinsic::stackrestore: {
10012 // If the save is right next to the restore, remove the restore. This can
10013 // happen when variable allocas are DCE'd.
10014 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
10015 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
10016 BasicBlock::iterator BI = SS;
10018 return EraseInstFromFunction(CI);
10022 // Scan down this block to see if there is another stack restore in the
10023 // same block without an intervening call/alloca.
10024 BasicBlock::iterator BI = II;
10025 TerminatorInst *TI = II->getParent()->getTerminator();
10026 bool CannotRemove = false;
10027 for (++BI; &*BI != TI; ++BI) {
10028 if (isa<AllocaInst>(BI)) {
10029 CannotRemove = true;
10032 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
10033 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
10034 // If there is a stackrestore below this one, remove this one.
10035 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10036 return EraseInstFromFunction(CI);
10037 // Otherwise, ignore the intrinsic.
10039 // If we found a non-intrinsic call, we can't remove the stack
10041 CannotRemove = true;
10047 // If the stack restore is in a return/unwind block and if there are no
10048 // allocas or calls between the restore and the return, nuke the restore.
10049 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10050 return EraseInstFromFunction(CI);
10055 return visitCallSite(II);
10058 // InvokeInst simplification
10060 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10061 return visitCallSite(&II);
10064 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10065 /// passed through the varargs area, we can eliminate the use of the cast.
10066 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10067 const CastInst * const CI,
10068 const TargetData * const TD,
10070 if (!CI->isLosslessCast())
10073 // The size of ByVal arguments is derived from the type, so we
10074 // can't change to a type with a different size. If the size were
10075 // passed explicitly we could avoid this check.
10076 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10079 const Type* SrcTy =
10080 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10081 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10082 if (!SrcTy->isSized() || !DstTy->isSized())
10084 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10089 // visitCallSite - Improvements for call and invoke instructions.
10091 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10092 bool Changed = false;
10094 // If the callee is a constexpr cast of a function, attempt to move the cast
10095 // to the arguments of the call/invoke.
10096 if (transformConstExprCastCall(CS)) return 0;
10098 Value *Callee = CS.getCalledValue();
10100 if (Function *CalleeF = dyn_cast<Function>(Callee))
10101 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10102 Instruction *OldCall = CS.getInstruction();
10103 // If the call and callee calling conventions don't match, this call must
10104 // be unreachable, as the call is undefined.
10105 new StoreInst(ConstantInt::getTrue(*Context),
10106 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))),
10108 if (!OldCall->use_empty())
10109 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
10110 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10111 return EraseInstFromFunction(*OldCall);
10115 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10116 // This instruction is not reachable, just remove it. We insert a store to
10117 // undef so that we know that this code is not reachable, despite the fact
10118 // that we can't modify the CFG here.
10119 new StoreInst(ConstantInt::getTrue(*Context),
10120 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))),
10121 CS.getInstruction());
10123 if (!CS.getInstruction()->use_empty())
10124 CS.getInstruction()->
10125 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
10127 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10128 // Don't break the CFG, insert a dummy cond branch.
10129 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10130 ConstantInt::getTrue(*Context), II);
10132 return EraseInstFromFunction(*CS.getInstruction());
10135 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10136 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10137 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10138 return transformCallThroughTrampoline(CS);
10140 const PointerType *PTy = cast<PointerType>(Callee->getType());
10141 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10142 if (FTy->isVarArg()) {
10143 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10144 // See if we can optimize any arguments passed through the varargs area of
10146 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10147 E = CS.arg_end(); I != E; ++I, ++ix) {
10148 CastInst *CI = dyn_cast<CastInst>(*I);
10149 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10150 *I = CI->getOperand(0);
10156 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10157 // Inline asm calls cannot throw - mark them 'nounwind'.
10158 CS.setDoesNotThrow();
10162 return Changed ? CS.getInstruction() : 0;
10165 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10166 // attempt to move the cast to the arguments of the call/invoke.
10168 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10169 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10170 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10171 if (CE->getOpcode() != Instruction::BitCast ||
10172 !isa<Function>(CE->getOperand(0)))
10174 Function *Callee = cast<Function>(CE->getOperand(0));
10175 Instruction *Caller = CS.getInstruction();
10176 const AttrListPtr &CallerPAL = CS.getAttributes();
10178 // Okay, this is a cast from a function to a different type. Unless doing so
10179 // would cause a type conversion of one of our arguments, change this call to
10180 // be a direct call with arguments casted to the appropriate types.
10182 const FunctionType *FT = Callee->getFunctionType();
10183 const Type *OldRetTy = Caller->getType();
10184 const Type *NewRetTy = FT->getReturnType();
10186 if (isa<StructType>(NewRetTy))
10187 return false; // TODO: Handle multiple return values.
10189 // Check to see if we are changing the return type...
10190 if (OldRetTy != NewRetTy) {
10191 if (Callee->isDeclaration() &&
10192 // Conversion is ok if changing from one pointer type to another or from
10193 // a pointer to an integer of the same size.
10194 !((isa<PointerType>(OldRetTy) || !TD ||
10195 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
10196 (isa<PointerType>(NewRetTy) || !TD ||
10197 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
10198 return false; // Cannot transform this return value.
10200 if (!Caller->use_empty() &&
10201 // void -> non-void is handled specially
10202 NewRetTy != Type::getVoidTy(*Context) && !CastInst::isCastable(NewRetTy, OldRetTy))
10203 return false; // Cannot transform this return value.
10205 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10206 Attributes RAttrs = CallerPAL.getRetAttributes();
10207 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10208 return false; // Attribute not compatible with transformed value.
10211 // If the callsite is an invoke instruction, and the return value is used by
10212 // a PHI node in a successor, we cannot change the return type of the call
10213 // because there is no place to put the cast instruction (without breaking
10214 // the critical edge). Bail out in this case.
10215 if (!Caller->use_empty())
10216 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10217 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10219 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10220 if (PN->getParent() == II->getNormalDest() ||
10221 PN->getParent() == II->getUnwindDest())
10225 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10226 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10228 CallSite::arg_iterator AI = CS.arg_begin();
10229 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10230 const Type *ParamTy = FT->getParamType(i);
10231 const Type *ActTy = (*AI)->getType();
10233 if (!CastInst::isCastable(ActTy, ParamTy))
10234 return false; // Cannot transform this parameter value.
10236 if (CallerPAL.getParamAttributes(i + 1)
10237 & Attribute::typeIncompatible(ParamTy))
10238 return false; // Attribute not compatible with transformed value.
10240 // Converting from one pointer type to another or between a pointer and an
10241 // integer of the same size is safe even if we do not have a body.
10242 bool isConvertible = ActTy == ParamTy ||
10243 (TD && ((isa<PointerType>(ParamTy) ||
10244 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
10245 (isa<PointerType>(ActTy) ||
10246 ActTy == TD->getIntPtrType(Caller->getContext()))));
10247 if (Callee->isDeclaration() && !isConvertible) return false;
10250 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10251 Callee->isDeclaration())
10252 return false; // Do not delete arguments unless we have a function body.
10254 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10255 !CallerPAL.isEmpty())
10256 // In this case we have more arguments than the new function type, but we
10257 // won't be dropping them. Check that these extra arguments have attributes
10258 // that are compatible with being a vararg call argument.
10259 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10260 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10262 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10263 if (PAttrs & Attribute::VarArgsIncompatible)
10267 // Okay, we decided that this is a safe thing to do: go ahead and start
10268 // inserting cast instructions as necessary...
10269 std::vector<Value*> Args;
10270 Args.reserve(NumActualArgs);
10271 SmallVector<AttributeWithIndex, 8> attrVec;
10272 attrVec.reserve(NumCommonArgs);
10274 // Get any return attributes.
10275 Attributes RAttrs = CallerPAL.getRetAttributes();
10277 // If the return value is not being used, the type may not be compatible
10278 // with the existing attributes. Wipe out any problematic attributes.
10279 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10281 // Add the new return attributes.
10283 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10285 AI = CS.arg_begin();
10286 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10287 const Type *ParamTy = FT->getParamType(i);
10288 if ((*AI)->getType() == ParamTy) {
10289 Args.push_back(*AI);
10291 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10292 false, ParamTy, false);
10293 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
10294 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
10297 // Add any parameter attributes.
10298 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10299 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10302 // If the function takes more arguments than the call was taking, add them
10304 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10305 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10307 // If we are removing arguments to the function, emit an obnoxious warning...
10308 if (FT->getNumParams() < NumActualArgs) {
10309 if (!FT->isVarArg()) {
10310 errs() << "WARNING: While resolving call to function '"
10311 << Callee->getName() << "' arguments were dropped!\n";
10313 // Add all of the arguments in their promoted form to the arg list...
10314 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10315 const Type *PTy = getPromotedType((*AI)->getType());
10316 if (PTy != (*AI)->getType()) {
10317 // Must promote to pass through va_arg area!
10318 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
10320 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
10321 InsertNewInstBefore(Cast, *Caller);
10322 Args.push_back(Cast);
10324 Args.push_back(*AI);
10327 // Add any parameter attributes.
10328 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10329 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10334 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10335 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10337 if (NewRetTy == Type::getVoidTy(*Context))
10338 Caller->setName(""); // Void type should not have a name.
10340 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10344 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10345 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10346 Args.begin(), Args.end(),
10347 Caller->getName(), Caller);
10348 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10349 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10351 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10352 Caller->getName(), Caller);
10353 CallInst *CI = cast<CallInst>(Caller);
10354 if (CI->isTailCall())
10355 cast<CallInst>(NC)->setTailCall();
10356 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10357 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10360 // Insert a cast of the return type as necessary.
10362 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10363 if (NV->getType() != Type::getVoidTy(*Context)) {
10364 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10366 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10368 // If this is an invoke instruction, we should insert it after the first
10369 // non-phi, instruction in the normal successor block.
10370 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10371 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10372 InsertNewInstBefore(NC, *I);
10374 // Otherwise, it's a call, just insert cast right after the call instr
10375 InsertNewInstBefore(NC, *Caller);
10377 AddUsersToWorkList(*Caller);
10379 NV = UndefValue::get(Caller->getType());
10383 if (Caller->getType() != Type::getVoidTy(*Context) && !Caller->use_empty())
10384 Caller->replaceAllUsesWith(NV);
10385 Caller->eraseFromParent();
10386 RemoveFromWorkList(Caller);
10390 // transformCallThroughTrampoline - Turn a call to a function created by the
10391 // init_trampoline intrinsic into a direct call to the underlying function.
10393 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10394 Value *Callee = CS.getCalledValue();
10395 const PointerType *PTy = cast<PointerType>(Callee->getType());
10396 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10397 const AttrListPtr &Attrs = CS.getAttributes();
10399 // If the call already has the 'nest' attribute somewhere then give up -
10400 // otherwise 'nest' would occur twice after splicing in the chain.
10401 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10404 IntrinsicInst *Tramp =
10405 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10407 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10408 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10409 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10411 const AttrListPtr &NestAttrs = NestF->getAttributes();
10412 if (!NestAttrs.isEmpty()) {
10413 unsigned NestIdx = 1;
10414 const Type *NestTy = 0;
10415 Attributes NestAttr = Attribute::None;
10417 // Look for a parameter marked with the 'nest' attribute.
10418 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10419 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10420 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10421 // Record the parameter type and any other attributes.
10423 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10428 Instruction *Caller = CS.getInstruction();
10429 std::vector<Value*> NewArgs;
10430 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10432 SmallVector<AttributeWithIndex, 8> NewAttrs;
10433 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10435 // Insert the nest argument into the call argument list, which may
10436 // mean appending it. Likewise for attributes.
10438 // Add any result attributes.
10439 if (Attributes Attr = Attrs.getRetAttributes())
10440 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10444 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10446 if (Idx == NestIdx) {
10447 // Add the chain argument and attributes.
10448 Value *NestVal = Tramp->getOperand(3);
10449 if (NestVal->getType() != NestTy)
10450 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10451 NewArgs.push_back(NestVal);
10452 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10458 // Add the original argument and attributes.
10459 NewArgs.push_back(*I);
10460 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10462 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10468 // Add any function attributes.
10469 if (Attributes Attr = Attrs.getFnAttributes())
10470 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10472 // The trampoline may have been bitcast to a bogus type (FTy).
10473 // Handle this by synthesizing a new function type, equal to FTy
10474 // with the chain parameter inserted.
10476 std::vector<const Type*> NewTypes;
10477 NewTypes.reserve(FTy->getNumParams()+1);
10479 // Insert the chain's type into the list of parameter types, which may
10480 // mean appending it.
10483 FunctionType::param_iterator I = FTy->param_begin(),
10484 E = FTy->param_end();
10487 if (Idx == NestIdx)
10488 // Add the chain's type.
10489 NewTypes.push_back(NestTy);
10494 // Add the original type.
10495 NewTypes.push_back(*I);
10501 // Replace the trampoline call with a direct call. Let the generic
10502 // code sort out any function type mismatches.
10503 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10505 Constant *NewCallee =
10506 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10507 NestF : ConstantExpr::getBitCast(NestF,
10508 PointerType::getUnqual(NewFTy));
10509 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10512 Instruction *NewCaller;
10513 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10514 NewCaller = InvokeInst::Create(NewCallee,
10515 II->getNormalDest(), II->getUnwindDest(),
10516 NewArgs.begin(), NewArgs.end(),
10517 Caller->getName(), Caller);
10518 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10519 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10521 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10522 Caller->getName(), Caller);
10523 if (cast<CallInst>(Caller)->isTailCall())
10524 cast<CallInst>(NewCaller)->setTailCall();
10525 cast<CallInst>(NewCaller)->
10526 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10527 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10529 if (Caller->getType() != Type::getVoidTy(*Context) && !Caller->use_empty())
10530 Caller->replaceAllUsesWith(NewCaller);
10531 Caller->eraseFromParent();
10532 RemoveFromWorkList(Caller);
10537 // Replace the trampoline call with a direct call. Since there is no 'nest'
10538 // parameter, there is no need to adjust the argument list. Let the generic
10539 // code sort out any function type mismatches.
10540 Constant *NewCallee =
10541 NestF->getType() == PTy ? NestF :
10542 ConstantExpr::getBitCast(NestF, PTy);
10543 CS.setCalledFunction(NewCallee);
10544 return CS.getInstruction();
10547 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10548 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10549 /// and a single binop.
10550 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10551 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10552 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10553 unsigned Opc = FirstInst->getOpcode();
10554 Value *LHSVal = FirstInst->getOperand(0);
10555 Value *RHSVal = FirstInst->getOperand(1);
10557 const Type *LHSType = LHSVal->getType();
10558 const Type *RHSType = RHSVal->getType();
10560 // Scan to see if all operands are the same opcode, all have one use, and all
10561 // kill their operands (i.e. the operands have one use).
10562 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10563 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10564 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10565 // Verify type of the LHS matches so we don't fold cmp's of different
10566 // types or GEP's with different index types.
10567 I->getOperand(0)->getType() != LHSType ||
10568 I->getOperand(1)->getType() != RHSType)
10571 // If they are CmpInst instructions, check their predicates
10572 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10573 if (cast<CmpInst>(I)->getPredicate() !=
10574 cast<CmpInst>(FirstInst)->getPredicate())
10577 // Keep track of which operand needs a phi node.
10578 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10579 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10582 // Otherwise, this is safe to transform!
10584 Value *InLHS = FirstInst->getOperand(0);
10585 Value *InRHS = FirstInst->getOperand(1);
10586 PHINode *NewLHS = 0, *NewRHS = 0;
10588 NewLHS = PHINode::Create(LHSType,
10589 FirstInst->getOperand(0)->getName() + ".pn");
10590 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10591 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10592 InsertNewInstBefore(NewLHS, PN);
10597 NewRHS = PHINode::Create(RHSType,
10598 FirstInst->getOperand(1)->getName() + ".pn");
10599 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10600 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10601 InsertNewInstBefore(NewRHS, PN);
10605 // Add all operands to the new PHIs.
10606 if (NewLHS || NewRHS) {
10607 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10608 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10610 Value *NewInLHS = InInst->getOperand(0);
10611 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10614 Value *NewInRHS = InInst->getOperand(1);
10615 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10620 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10621 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10622 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10623 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10627 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10628 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10630 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10631 FirstInst->op_end());
10632 // This is true if all GEP bases are allocas and if all indices into them are
10634 bool AllBasePointersAreAllocas = true;
10636 // Scan to see if all operands are the same opcode, all have one use, and all
10637 // kill their operands (i.e. the operands have one use).
10638 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10639 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10640 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10641 GEP->getNumOperands() != FirstInst->getNumOperands())
10644 // Keep track of whether or not all GEPs are of alloca pointers.
10645 if (AllBasePointersAreAllocas &&
10646 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10647 !GEP->hasAllConstantIndices()))
10648 AllBasePointersAreAllocas = false;
10650 // Compare the operand lists.
10651 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10652 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10655 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10656 // if one of the PHIs has a constant for the index. The index may be
10657 // substantially cheaper to compute for the constants, so making it a
10658 // variable index could pessimize the path. This also handles the case
10659 // for struct indices, which must always be constant.
10660 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10661 isa<ConstantInt>(GEP->getOperand(op)))
10664 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10666 FixedOperands[op] = 0; // Needs a PHI.
10670 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10671 // bother doing this transformation. At best, this will just save a bit of
10672 // offset calculation, but all the predecessors will have to materialize the
10673 // stack address into a register anyway. We'd actually rather *clone* the
10674 // load up into the predecessors so that we have a load of a gep of an alloca,
10675 // which can usually all be folded into the load.
10676 if (AllBasePointersAreAllocas)
10679 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10680 // that is variable.
10681 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10683 bool HasAnyPHIs = false;
10684 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10685 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10686 Value *FirstOp = FirstInst->getOperand(i);
10687 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10688 FirstOp->getName()+".pn");
10689 InsertNewInstBefore(NewPN, PN);
10691 NewPN->reserveOperandSpace(e);
10692 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10693 OperandPhis[i] = NewPN;
10694 FixedOperands[i] = NewPN;
10699 // Add all operands to the new PHIs.
10701 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10702 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10703 BasicBlock *InBB = PN.getIncomingBlock(i);
10705 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10706 if (PHINode *OpPhi = OperandPhis[op])
10707 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10711 Value *Base = FixedOperands[0];
10712 GetElementPtrInst *GEP =
10713 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10714 FixedOperands.end());
10715 if (cast<GEPOperator>(FirstInst)->isInBounds())
10716 cast<GEPOperator>(GEP)->setIsInBounds(true);
10721 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10722 /// sink the load out of the block that defines it. This means that it must be
10723 /// obvious the value of the load is not changed from the point of the load to
10724 /// the end of the block it is in.
10726 /// Finally, it is safe, but not profitable, to sink a load targetting a
10727 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10729 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10730 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10732 for (++BBI; BBI != E; ++BBI)
10733 if (BBI->mayWriteToMemory())
10736 // Check for non-address taken alloca. If not address-taken already, it isn't
10737 // profitable to do this xform.
10738 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10739 bool isAddressTaken = false;
10740 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10742 if (isa<LoadInst>(UI)) continue;
10743 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10744 // If storing TO the alloca, then the address isn't taken.
10745 if (SI->getOperand(1) == AI) continue;
10747 isAddressTaken = true;
10751 if (!isAddressTaken && AI->isStaticAlloca())
10755 // If this load is a load from a GEP with a constant offset from an alloca,
10756 // then we don't want to sink it. In its present form, it will be
10757 // load [constant stack offset]. Sinking it will cause us to have to
10758 // materialize the stack addresses in each predecessor in a register only to
10759 // do a shared load from register in the successor.
10760 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10761 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10762 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10769 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10770 // operator and they all are only used by the PHI, PHI together their
10771 // inputs, and do the operation once, to the result of the PHI.
10772 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10773 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10775 // Scan the instruction, looking for input operations that can be folded away.
10776 // If all input operands to the phi are the same instruction (e.g. a cast from
10777 // the same type or "+42") we can pull the operation through the PHI, reducing
10778 // code size and simplifying code.
10779 Constant *ConstantOp = 0;
10780 const Type *CastSrcTy = 0;
10781 bool isVolatile = false;
10782 if (isa<CastInst>(FirstInst)) {
10783 CastSrcTy = FirstInst->getOperand(0)->getType();
10784 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10785 // Can fold binop, compare or shift here if the RHS is a constant,
10786 // otherwise call FoldPHIArgBinOpIntoPHI.
10787 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10788 if (ConstantOp == 0)
10789 return FoldPHIArgBinOpIntoPHI(PN);
10790 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10791 isVolatile = LI->isVolatile();
10792 // We can't sink the load if the loaded value could be modified between the
10793 // load and the PHI.
10794 if (LI->getParent() != PN.getIncomingBlock(0) ||
10795 !isSafeAndProfitableToSinkLoad(LI))
10798 // If the PHI is of volatile loads and the load block has multiple
10799 // successors, sinking it would remove a load of the volatile value from
10800 // the path through the other successor.
10802 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10805 } else if (isa<GetElementPtrInst>(FirstInst)) {
10806 return FoldPHIArgGEPIntoPHI(PN);
10808 return 0; // Cannot fold this operation.
10811 // Check to see if all arguments are the same operation.
10812 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10813 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10814 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10815 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10818 if (I->getOperand(0)->getType() != CastSrcTy)
10819 return 0; // Cast operation must match.
10820 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10821 // We can't sink the load if the loaded value could be modified between
10822 // the load and the PHI.
10823 if (LI->isVolatile() != isVolatile ||
10824 LI->getParent() != PN.getIncomingBlock(i) ||
10825 !isSafeAndProfitableToSinkLoad(LI))
10828 // If the PHI is of volatile loads and the load block has multiple
10829 // successors, sinking it would remove a load of the volatile value from
10830 // the path through the other successor.
10832 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10835 } else if (I->getOperand(1) != ConstantOp) {
10840 // Okay, they are all the same operation. Create a new PHI node of the
10841 // correct type, and PHI together all of the LHS's of the instructions.
10842 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10843 PN.getName()+".in");
10844 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10846 Value *InVal = FirstInst->getOperand(0);
10847 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10849 // Add all operands to the new PHI.
10850 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10851 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10852 if (NewInVal != InVal)
10854 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10859 // The new PHI unions all of the same values together. This is really
10860 // common, so we handle it intelligently here for compile-time speed.
10864 InsertNewInstBefore(NewPN, PN);
10868 // Insert and return the new operation.
10869 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10870 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10871 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10872 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10873 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10874 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10875 PhiVal, ConstantOp);
10876 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10878 // If this was a volatile load that we are merging, make sure to loop through
10879 // and mark all the input loads as non-volatile. If we don't do this, we will
10880 // insert a new volatile load and the old ones will not be deletable.
10882 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10883 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10885 return new LoadInst(PhiVal, "", isVolatile);
10888 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10890 static bool DeadPHICycle(PHINode *PN,
10891 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10892 if (PN->use_empty()) return true;
10893 if (!PN->hasOneUse()) return false;
10895 // Remember this node, and if we find the cycle, return.
10896 if (!PotentiallyDeadPHIs.insert(PN))
10899 // Don't scan crazily complex things.
10900 if (PotentiallyDeadPHIs.size() == 16)
10903 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10904 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10909 /// PHIsEqualValue - Return true if this phi node is always equal to
10910 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10911 /// z = some value; x = phi (y, z); y = phi (x, z)
10912 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10913 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10914 // See if we already saw this PHI node.
10915 if (!ValueEqualPHIs.insert(PN))
10918 // Don't scan crazily complex things.
10919 if (ValueEqualPHIs.size() == 16)
10922 // Scan the operands to see if they are either phi nodes or are equal to
10924 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10925 Value *Op = PN->getIncomingValue(i);
10926 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10927 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10929 } else if (Op != NonPhiInVal)
10937 // PHINode simplification
10939 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10940 // If LCSSA is around, don't mess with Phi nodes
10941 if (MustPreserveLCSSA) return 0;
10943 if (Value *V = PN.hasConstantValue())
10944 return ReplaceInstUsesWith(PN, V);
10946 // If all PHI operands are the same operation, pull them through the PHI,
10947 // reducing code size.
10948 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10949 isa<Instruction>(PN.getIncomingValue(1)) &&
10950 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10951 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10952 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10953 // than themselves more than once.
10954 PN.getIncomingValue(0)->hasOneUse())
10955 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10958 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10959 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10960 // PHI)... break the cycle.
10961 if (PN.hasOneUse()) {
10962 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10963 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10964 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10965 PotentiallyDeadPHIs.insert(&PN);
10966 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10967 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10970 // If this phi has a single use, and if that use just computes a value for
10971 // the next iteration of a loop, delete the phi. This occurs with unused
10972 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10973 // common case here is good because the only other things that catch this
10974 // are induction variable analysis (sometimes) and ADCE, which is only run
10976 if (PHIUser->hasOneUse() &&
10977 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10978 PHIUser->use_back() == &PN) {
10979 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10983 // We sometimes end up with phi cycles that non-obviously end up being the
10984 // same value, for example:
10985 // z = some value; x = phi (y, z); y = phi (x, z)
10986 // where the phi nodes don't necessarily need to be in the same block. Do a
10987 // quick check to see if the PHI node only contains a single non-phi value, if
10988 // so, scan to see if the phi cycle is actually equal to that value.
10990 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10991 // Scan for the first non-phi operand.
10992 while (InValNo != NumOperandVals &&
10993 isa<PHINode>(PN.getIncomingValue(InValNo)))
10996 if (InValNo != NumOperandVals) {
10997 Value *NonPhiInVal = PN.getOperand(InValNo);
10999 // Scan the rest of the operands to see if there are any conflicts, if so
11000 // there is no need to recursively scan other phis.
11001 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
11002 Value *OpVal = PN.getIncomingValue(InValNo);
11003 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
11007 // If we scanned over all operands, then we have one unique value plus
11008 // phi values. Scan PHI nodes to see if they all merge in each other or
11010 if (InValNo == NumOperandVals) {
11011 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
11012 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
11013 return ReplaceInstUsesWith(PN, NonPhiInVal);
11020 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
11021 Value *PtrOp = GEP.getOperand(0);
11022 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
11023 // If so, eliminate the noop.
11024 if (GEP.getNumOperands() == 1)
11025 return ReplaceInstUsesWith(GEP, PtrOp);
11027 if (isa<UndefValue>(GEP.getOperand(0)))
11028 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
11030 bool HasZeroPointerIndex = false;
11031 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
11032 HasZeroPointerIndex = C->isNullValue();
11034 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11035 return ReplaceInstUsesWith(GEP, PtrOp);
11037 // Eliminate unneeded casts for indices.
11039 bool MadeChange = false;
11040 unsigned PtrSize = TD->getPointerSizeInBits();
11042 gep_type_iterator GTI = gep_type_begin(GEP);
11043 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
11044 I != E; ++I, ++GTI) {
11045 if (!isa<SequentialType>(*GTI)) continue;
11047 // If we are using a wider index than needed for this platform, shrink it
11048 // to what we need. If narrower, sign-extend it to what we need. This
11049 // explicit cast can make subsequent optimizations more obvious.
11050 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
11052 if (OpBits == PtrSize)
11055 Instruction::CastOps Opc =
11056 OpBits > PtrSize ? Instruction::Trunc : Instruction::SExt;
11057 *I = InsertCastBefore(Opc, *I, TD->getIntPtrType(GEP.getContext()), GEP);
11060 if (MadeChange) return &GEP;
11063 // Combine Indices - If the source pointer to this getelementptr instruction
11064 // is a getelementptr instruction, combine the indices of the two
11065 // getelementptr instructions into a single instruction.
11067 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
11068 // Note that if our source is a gep chain itself that we wait for that
11069 // chain to be resolved before we perform this transformation. This
11070 // avoids us creating a TON of code in some cases.
11072 if (GetElementPtrInst *SrcGEP =
11073 dyn_cast<GetElementPtrInst>(Src->getOperand(0)))
11074 if (SrcGEP->getNumOperands() == 2)
11075 return 0; // Wait until our source is folded to completion.
11077 SmallVector<Value*, 8> Indices;
11079 // Find out whether the last index in the source GEP is a sequential idx.
11080 bool EndsWithSequential = false;
11081 for (gep_type_iterator I = gep_type_begin(*Src), E = gep_type_end(*Src);
11083 EndsWithSequential = !isa<StructType>(*I);
11085 // Can we combine the two pointer arithmetics offsets?
11086 if (EndsWithSequential) {
11087 // Replace: gep (gep %P, long B), long A, ...
11088 // With: T = long A+B; gep %P, T, ...
11091 Value *SO1 = Src->getOperand(Src->getNumOperands()-1);
11092 Value *GO1 = GEP.getOperand(1);
11093 if (SO1 == Constant::getNullValue(SO1->getType())) {
11095 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
11098 // If they aren't the same type, then the input hasn't been processed
11099 // by the loop above yet (which canonicalizes sequential index types to
11100 // intptr_t). Just avoid transforming this until the input has been
11102 if (SO1->getType() != GO1->getType())
11104 if (isa<Constant>(SO1) && isa<Constant>(GO1))
11105 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
11107 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11108 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
11112 // Update the GEP in place if possible.
11113 if (Src->getNumOperands() == 2) {
11114 GEP.setOperand(0, Src->getOperand(0));
11115 GEP.setOperand(1, Sum);
11118 Indices.append(Src->op_begin()+1, Src->op_end()-1);
11119 Indices.push_back(Sum);
11120 Indices.append(GEP.op_begin()+2, GEP.op_end());
11121 } else if (isa<Constant>(*GEP.idx_begin()) &&
11122 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11123 Src->getNumOperands() != 1) {
11124 // Otherwise we can do the fold if the first index of the GEP is a zero
11125 Indices.append(Src->op_begin()+1, Src->op_end());
11126 Indices.append(GEP.idx_begin()+1, GEP.idx_end());
11129 if (!Indices.empty()) {
11130 GetElementPtrInst *NewGEP =
11131 GetElementPtrInst::Create(Src->getOperand(0), Indices.begin(),
11132 Indices.end(), GEP.getName());
11133 if (cast<GEPOperator>(&GEP)->isInBounds() && Src->isInBounds())
11134 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11139 // Handle gep(bitcast x) and gep(gep x, 0, 0, 0).
11140 if (Value *X = getBitCastOperand(PtrOp)) {
11141 assert(isa<PointerType>(X->getType()) && "Must be cast from pointer");
11143 if (HasZeroPointerIndex) {
11144 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11145 // into : GEP [10 x i8]* X, i32 0, ...
11147 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11148 // into : GEP i8* X, ...
11150 // This occurs when the program declares an array extern like "int X[];"
11151 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11152 const PointerType *XTy = cast<PointerType>(X->getType());
11153 if (const ArrayType *CATy =
11154 dyn_cast<ArrayType>(CPTy->getElementType())) {
11155 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11156 if (CATy->getElementType() == XTy->getElementType()) {
11157 // -> GEP i8* X, ...
11158 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11159 GetElementPtrInst *NewGEP =
11160 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11162 if (cast<GEPOperator>(&GEP)->isInBounds())
11163 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11165 } else if (const ArrayType *XATy =
11166 dyn_cast<ArrayType>(XTy->getElementType())) {
11167 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11168 if (CATy->getElementType() == XATy->getElementType()) {
11169 // -> GEP [10 x i8]* X, i32 0, ...
11170 // At this point, we know that the cast source type is a pointer
11171 // to an array of the same type as the destination pointer
11172 // array. Because the array type is never stepped over (there
11173 // is a leading zero) we can fold the cast into this GEP.
11174 GEP.setOperand(0, X);
11179 } else if (GEP.getNumOperands() == 2) {
11180 // Transform things like:
11181 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11182 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11183 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11184 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11185 if (TD && isa<ArrayType>(SrcElTy) &&
11186 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11187 TD->getTypeAllocSize(ResElTy)) {
11189 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11190 Idx[1] = GEP.getOperand(1);
11191 GetElementPtrInst *NewGEP =
11192 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11193 if (cast<GEPOperator>(&GEP)->isInBounds())
11194 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11195 Value *V = InsertNewInstBefore(NewGEP, GEP);
11196 // V and GEP are both pointer types --> BitCast
11197 return new BitCastInst(V, GEP.getType());
11200 // Transform things like:
11201 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11202 // (where tmp = 8*tmp2) into:
11203 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11205 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::getInt8Ty(*Context)) {
11206 uint64_t ArrayEltSize =
11207 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11209 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11210 // allow either a mul, shift, or constant here.
11212 ConstantInt *Scale = 0;
11213 if (ArrayEltSize == 1) {
11214 NewIdx = GEP.getOperand(1);
11215 Scale = ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11216 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11217 NewIdx = ConstantInt::get(CI->getType(), 1);
11219 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11220 if (Inst->getOpcode() == Instruction::Shl &&
11221 isa<ConstantInt>(Inst->getOperand(1))) {
11222 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11223 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11224 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11226 NewIdx = Inst->getOperand(0);
11227 } else if (Inst->getOpcode() == Instruction::Mul &&
11228 isa<ConstantInt>(Inst->getOperand(1))) {
11229 Scale = cast<ConstantInt>(Inst->getOperand(1));
11230 NewIdx = Inst->getOperand(0);
11234 // If the index will be to exactly the right offset with the scale taken
11235 // out, perform the transformation. Note, we don't know whether Scale is
11236 // signed or not. We'll use unsigned version of division/modulo
11237 // operation after making sure Scale doesn't have the sign bit set.
11238 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11239 Scale->getZExtValue() % ArrayEltSize == 0) {
11240 Scale = ConstantInt::get(Scale->getType(),
11241 Scale->getZExtValue() / ArrayEltSize);
11242 if (Scale->getZExtValue() != 1) {
11243 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11245 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
11246 NewIdx = InsertNewInstBefore(Sc, GEP);
11249 // Insert the new GEP instruction.
11251 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11253 Instruction *NewGEP =
11254 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11255 if (cast<GEPOperator>(&GEP)->isInBounds())
11256 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11257 NewGEP = InsertNewInstBefore(NewGEP, GEP);
11258 // The NewGEP must be pointer typed, so must the old one -> BitCast
11259 return new BitCastInst(NewGEP, GEP.getType());
11265 /// See if we can simplify:
11266 /// X = bitcast A* to B*
11267 /// Y = gep X, <...constant indices...>
11268 /// into a gep of the original struct. This is important for SROA and alias
11269 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11270 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11272 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11273 // Determine how much the GEP moves the pointer. We are guaranteed to get
11274 // a constant back from EmitGEPOffset.
11275 ConstantInt *OffsetV =
11276 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11277 int64_t Offset = OffsetV->getSExtValue();
11279 // If this GEP instruction doesn't move the pointer, just replace the GEP
11280 // with a bitcast of the real input to the dest type.
11282 // If the bitcast is of an allocation, and the allocation will be
11283 // converted to match the type of the cast, don't touch this.
11284 if (isa<AllocationInst>(BCI->getOperand(0))) {
11285 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11286 if (Instruction *I = visitBitCast(*BCI)) {
11289 BCI->getParent()->getInstList().insert(BCI, I);
11290 ReplaceInstUsesWith(*BCI, I);
11295 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11298 // Otherwise, if the offset is non-zero, we need to find out if there is a
11299 // field at Offset in 'A's type. If so, we can pull the cast through the
11301 SmallVector<Value*, 8> NewIndices;
11303 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11304 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11305 Instruction *NGEP =
11306 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11308 if (NGEP->getType() == GEP.getType()) return NGEP;
11309 if (cast<GEPOperator>(&GEP)->isInBounds())
11310 cast<GEPOperator>(NGEP)->setIsInBounds(true);
11311 InsertNewInstBefore(NGEP, GEP);
11312 NGEP->takeName(&GEP);
11313 return new BitCastInst(NGEP, GEP.getType());
11321 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11322 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11323 if (AI.isArrayAllocation()) { // Check C != 1
11324 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11325 const Type *NewTy =
11326 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11327 AllocationInst *New = 0;
11329 // Create and insert the replacement instruction...
11330 if (isa<MallocInst>(AI))
11331 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11333 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11334 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11337 InsertNewInstBefore(New, AI);
11339 // Scan to the end of the allocation instructions, to skip over a block of
11340 // allocas if possible...also skip interleaved debug info
11342 BasicBlock::iterator It = New;
11343 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11345 // Now that I is pointing to the first non-allocation-inst in the block,
11346 // insert our getelementptr instruction...
11348 Value *NullIdx = Constant::getNullValue(Type::getInt32Ty(*Context));
11352 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11353 New->getName()+".sub", It);
11354 cast<GEPOperator>(V)->setIsInBounds(true);
11356 // Now make everything use the getelementptr instead of the original
11358 return ReplaceInstUsesWith(AI, V);
11359 } else if (isa<UndefValue>(AI.getArraySize())) {
11360 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11364 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11365 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11366 // Note that we only do this for alloca's, because malloc should allocate
11367 // and return a unique pointer, even for a zero byte allocation.
11368 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11369 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11371 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11372 if (AI.getAlignment() == 0)
11373 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11379 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11380 Value *Op = FI.getOperand(0);
11382 // free undef -> unreachable.
11383 if (isa<UndefValue>(Op)) {
11384 // Insert a new store to null because we cannot modify the CFG here.
11385 new StoreInst(ConstantInt::getTrue(*Context),
11386 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))), &FI);
11387 return EraseInstFromFunction(FI);
11390 // If we have 'free null' delete the instruction. This can happen in stl code
11391 // when lots of inlining happens.
11392 if (isa<ConstantPointerNull>(Op))
11393 return EraseInstFromFunction(FI);
11395 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11396 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11397 FI.setOperand(0, CI->getOperand(0));
11401 // Change free (gep X, 0,0,0,0) into free(X)
11402 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11403 if (GEPI->hasAllZeroIndices()) {
11404 AddToWorkList(GEPI);
11405 FI.setOperand(0, GEPI->getOperand(0));
11410 // Change free(malloc) into nothing, if the malloc has a single use.
11411 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11412 if (MI->hasOneUse()) {
11413 EraseInstFromFunction(FI);
11414 return EraseInstFromFunction(*MI);
11421 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11422 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11423 const TargetData *TD) {
11424 User *CI = cast<User>(LI.getOperand(0));
11425 Value *CastOp = CI->getOperand(0);
11426 LLVMContext *Context = IC.getContext();
11429 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11430 // Instead of loading constant c string, use corresponding integer value
11431 // directly if string length is small enough.
11433 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11434 unsigned len = Str.length();
11435 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11436 unsigned numBits = Ty->getPrimitiveSizeInBits();
11437 // Replace LI with immediate integer store.
11438 if ((numBits >> 3) == len + 1) {
11439 APInt StrVal(numBits, 0);
11440 APInt SingleChar(numBits, 0);
11441 if (TD->isLittleEndian()) {
11442 for (signed i = len-1; i >= 0; i--) {
11443 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11444 StrVal = (StrVal << 8) | SingleChar;
11447 for (unsigned i = 0; i < len; i++) {
11448 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11449 StrVal = (StrVal << 8) | SingleChar;
11451 // Append NULL at the end.
11453 StrVal = (StrVal << 8) | SingleChar;
11455 Value *NL = ConstantInt::get(*Context, StrVal);
11456 return IC.ReplaceInstUsesWith(LI, NL);
11462 const PointerType *DestTy = cast<PointerType>(CI->getType());
11463 const Type *DestPTy = DestTy->getElementType();
11464 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11466 // If the address spaces don't match, don't eliminate the cast.
11467 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11470 const Type *SrcPTy = SrcTy->getElementType();
11472 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11473 isa<VectorType>(DestPTy)) {
11474 // If the source is an array, the code below will not succeed. Check to
11475 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11477 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11478 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11479 if (ASrcTy->getNumElements() != 0) {
11481 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::getInt32Ty(*Context));
11482 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11483 SrcTy = cast<PointerType>(CastOp->getType());
11484 SrcPTy = SrcTy->getElementType();
11487 if (IC.getTargetData() &&
11488 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11489 isa<VectorType>(SrcPTy)) &&
11490 // Do not allow turning this into a load of an integer, which is then
11491 // casted to a pointer, this pessimizes pointer analysis a lot.
11492 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11493 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
11494 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
11496 // Okay, we are casting from one integer or pointer type to another of
11497 // the same size. Instead of casting the pointer before the load, cast
11498 // the result of the loaded value.
11499 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11501 LI.isVolatile()),LI);
11502 // Now cast the result of the load.
11503 return new BitCastInst(NewLoad, LI.getType());
11510 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11511 Value *Op = LI.getOperand(0);
11513 // Attempt to improve the alignment.
11515 unsigned KnownAlign =
11516 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11518 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11519 LI.getAlignment()))
11520 LI.setAlignment(KnownAlign);
11523 // load (cast X) --> cast (load X) iff safe
11524 if (isa<CastInst>(Op))
11525 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11528 // None of the following transforms are legal for volatile loads.
11529 if (LI.isVolatile()) return 0;
11531 // Do really simple store-to-load forwarding and load CSE, to catch cases
11532 // where there are several consequtive memory accesses to the same location,
11533 // separated by a few arithmetic operations.
11534 BasicBlock::iterator BBI = &LI;
11535 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11536 return ReplaceInstUsesWith(LI, AvailableVal);
11538 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11539 const Value *GEPI0 = GEPI->getOperand(0);
11540 // TODO: Consider a target hook for valid address spaces for this xform.
11541 if (isa<ConstantPointerNull>(GEPI0) &&
11542 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11543 // Insert a new store to null instruction before the load to indicate
11544 // that this code is not reachable. We do this instead of inserting
11545 // an unreachable instruction directly because we cannot modify the
11547 new StoreInst(UndefValue::get(LI.getType()),
11548 Constant::getNullValue(Op->getType()), &LI);
11549 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11553 if (Constant *C = dyn_cast<Constant>(Op)) {
11554 // load null/undef -> undef
11555 // TODO: Consider a target hook for valid address spaces for this xform.
11556 if (isa<UndefValue>(C) || (C->isNullValue() &&
11557 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11558 // Insert a new store to null instruction before the load to indicate that
11559 // this code is not reachable. We do this instead of inserting an
11560 // unreachable instruction directly because we cannot modify the CFG.
11561 new StoreInst(UndefValue::get(LI.getType()),
11562 Constant::getNullValue(Op->getType()), &LI);
11563 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11566 // Instcombine load (constant global) into the value loaded.
11567 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11568 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11569 return ReplaceInstUsesWith(LI, GV->getInitializer());
11571 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11572 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11573 if (CE->getOpcode() == Instruction::GetElementPtr) {
11574 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11575 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11577 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11579 return ReplaceInstUsesWith(LI, V);
11580 if (CE->getOperand(0)->isNullValue()) {
11581 // Insert a new store to null instruction before the load to indicate
11582 // that this code is not reachable. We do this instead of inserting
11583 // an unreachable instruction directly because we cannot modify the
11585 new StoreInst(UndefValue::get(LI.getType()),
11586 Constant::getNullValue(Op->getType()), &LI);
11587 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11590 } else if (CE->isCast()) {
11591 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11597 // If this load comes from anywhere in a constant global, and if the global
11598 // is all undef or zero, we know what it loads.
11599 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11600 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11601 if (GV->getInitializer()->isNullValue())
11602 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11603 else if (isa<UndefValue>(GV->getInitializer()))
11604 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11608 if (Op->hasOneUse()) {
11609 // Change select and PHI nodes to select values instead of addresses: this
11610 // helps alias analysis out a lot, allows many others simplifications, and
11611 // exposes redundancy in the code.
11613 // Note that we cannot do the transformation unless we know that the
11614 // introduced loads cannot trap! Something like this is valid as long as
11615 // the condition is always false: load (select bool %C, int* null, int* %G),
11616 // but it would not be valid if we transformed it to load from null
11617 // unconditionally.
11619 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11620 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11621 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11622 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11623 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11624 SI->getOperand(1)->getName()+".val"), LI);
11625 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11626 SI->getOperand(2)->getName()+".val"), LI);
11627 return SelectInst::Create(SI->getCondition(), V1, V2);
11630 // load (select (cond, null, P)) -> load P
11631 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11632 if (C->isNullValue()) {
11633 LI.setOperand(0, SI->getOperand(2));
11637 // load (select (cond, P, null)) -> load P
11638 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11639 if (C->isNullValue()) {
11640 LI.setOperand(0, SI->getOperand(1));
11648 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11649 /// when possible. This makes it generally easy to do alias analysis and/or
11650 /// SROA/mem2reg of the memory object.
11651 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11652 User *CI = cast<User>(SI.getOperand(1));
11653 Value *CastOp = CI->getOperand(0);
11655 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11656 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11657 if (SrcTy == 0) return 0;
11659 const Type *SrcPTy = SrcTy->getElementType();
11661 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11664 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11665 /// to its first element. This allows us to handle things like:
11666 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11667 /// on 32-bit hosts.
11668 SmallVector<Value*, 4> NewGEPIndices;
11670 // If the source is an array, the code below will not succeed. Check to
11671 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11673 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11674 // Index through pointer.
11675 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(*IC.getContext()));
11676 NewGEPIndices.push_back(Zero);
11679 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11680 if (!STy->getNumElements()) /* Struct can be empty {} */
11682 NewGEPIndices.push_back(Zero);
11683 SrcPTy = STy->getElementType(0);
11684 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11685 NewGEPIndices.push_back(Zero);
11686 SrcPTy = ATy->getElementType();
11692 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11695 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11698 // If the pointers point into different address spaces or if they point to
11699 // values with different sizes, we can't do the transformation.
11700 if (!IC.getTargetData() ||
11701 SrcTy->getAddressSpace() !=
11702 cast<PointerType>(CI->getType())->getAddressSpace() ||
11703 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
11704 IC.getTargetData()->getTypeSizeInBits(DestPTy))
11707 // Okay, we are casting from one integer or pointer type to another of
11708 // the same size. Instead of casting the pointer before
11709 // the store, cast the value to be stored.
11711 Value *SIOp0 = SI.getOperand(0);
11712 Instruction::CastOps opcode = Instruction::BitCast;
11713 const Type* CastSrcTy = SIOp0->getType();
11714 const Type* CastDstTy = SrcPTy;
11715 if (isa<PointerType>(CastDstTy)) {
11716 if (CastSrcTy->isInteger())
11717 opcode = Instruction::IntToPtr;
11718 } else if (isa<IntegerType>(CastDstTy)) {
11719 if (isa<PointerType>(SIOp0->getType()))
11720 opcode = Instruction::PtrToInt;
11723 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11724 // emit a GEP to index into its first field.
11725 if (!NewGEPIndices.empty()) {
11726 if (Constant *C = dyn_cast<Constant>(CastOp))
11727 CastOp = ConstantExpr::getGetElementPtr(C, &NewGEPIndices[0],
11728 NewGEPIndices.size());
11730 CastOp = IC.InsertNewInstBefore(
11731 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11732 NewGEPIndices.end()), SI);
11733 cast<GEPOperator>(CastOp)->setIsInBounds(true);
11736 if (Constant *C = dyn_cast<Constant>(SIOp0))
11737 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11739 NewCast = IC.InsertNewInstBefore(
11740 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11742 return new StoreInst(NewCast, CastOp);
11745 /// equivalentAddressValues - Test if A and B will obviously have the same
11746 /// value. This includes recognizing that %t0 and %t1 will have the same
11747 /// value in code like this:
11748 /// %t0 = getelementptr \@a, 0, 3
11749 /// store i32 0, i32* %t0
11750 /// %t1 = getelementptr \@a, 0, 3
11751 /// %t2 = load i32* %t1
11753 static bool equivalentAddressValues(Value *A, Value *B) {
11754 // Test if the values are trivially equivalent.
11755 if (A == B) return true;
11757 // Test if the values come form identical arithmetic instructions.
11758 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
11759 // its only used to compare two uses within the same basic block, which
11760 // means that they'll always either have the same value or one of them
11761 // will have an undefined value.
11762 if (isa<BinaryOperator>(A) ||
11763 isa<CastInst>(A) ||
11765 isa<GetElementPtrInst>(A))
11766 if (Instruction *BI = dyn_cast<Instruction>(B))
11767 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
11770 // Otherwise they may not be equivalent.
11774 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11775 // return the llvm.dbg.declare.
11776 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11777 if (!V->hasNUses(2))
11779 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11781 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11783 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11784 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11791 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11792 Value *Val = SI.getOperand(0);
11793 Value *Ptr = SI.getOperand(1);
11795 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11796 EraseInstFromFunction(SI);
11801 // If the RHS is an alloca with a single use, zapify the store, making the
11803 // If the RHS is an alloca with a two uses, the other one being a
11804 // llvm.dbg.declare, zapify the store and the declare, making the
11805 // alloca dead. We must do this to prevent declare's from affecting
11807 if (!SI.isVolatile()) {
11808 if (Ptr->hasOneUse()) {
11809 if (isa<AllocaInst>(Ptr)) {
11810 EraseInstFromFunction(SI);
11814 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11815 if (isa<AllocaInst>(GEP->getOperand(0))) {
11816 if (GEP->getOperand(0)->hasOneUse()) {
11817 EraseInstFromFunction(SI);
11821 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11822 EraseInstFromFunction(*DI);
11823 EraseInstFromFunction(SI);
11830 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11831 EraseInstFromFunction(*DI);
11832 EraseInstFromFunction(SI);
11838 // Attempt to improve the alignment.
11840 unsigned KnownAlign =
11841 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11843 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11844 SI.getAlignment()))
11845 SI.setAlignment(KnownAlign);
11848 // Do really simple DSE, to catch cases where there are several consecutive
11849 // stores to the same location, separated by a few arithmetic operations. This
11850 // situation often occurs with bitfield accesses.
11851 BasicBlock::iterator BBI = &SI;
11852 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11855 // Don't count debug info directives, lest they affect codegen,
11856 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11857 // It is necessary for correctness to skip those that feed into a
11858 // llvm.dbg.declare, as these are not present when debugging is off.
11859 if (isa<DbgInfoIntrinsic>(BBI) ||
11860 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11865 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11866 // Prev store isn't volatile, and stores to the same location?
11867 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11868 SI.getOperand(1))) {
11871 EraseInstFromFunction(*PrevSI);
11877 // If this is a load, we have to stop. However, if the loaded value is from
11878 // the pointer we're loading and is producing the pointer we're storing,
11879 // then *this* store is dead (X = load P; store X -> P).
11880 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11881 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11882 !SI.isVolatile()) {
11883 EraseInstFromFunction(SI);
11887 // Otherwise, this is a load from some other location. Stores before it
11888 // may not be dead.
11892 // Don't skip over loads or things that can modify memory.
11893 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11898 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11900 // store X, null -> turns into 'unreachable' in SimplifyCFG
11901 if (isa<ConstantPointerNull>(Ptr) &&
11902 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11903 if (!isa<UndefValue>(Val)) {
11904 SI.setOperand(0, UndefValue::get(Val->getType()));
11905 if (Instruction *U = dyn_cast<Instruction>(Val))
11906 AddToWorkList(U); // Dropped a use.
11909 return 0; // Do not modify these!
11912 // store undef, Ptr -> noop
11913 if (isa<UndefValue>(Val)) {
11914 EraseInstFromFunction(SI);
11919 // If the pointer destination is a cast, see if we can fold the cast into the
11921 if (isa<CastInst>(Ptr))
11922 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11924 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11926 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11930 // If this store is the last instruction in the basic block (possibly
11931 // excepting debug info instructions and the pointer bitcasts that feed
11932 // into them), and if the block ends with an unconditional branch, try
11933 // to move it to the successor block.
11937 } while (isa<DbgInfoIntrinsic>(BBI) ||
11938 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11939 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11940 if (BI->isUnconditional())
11941 if (SimplifyStoreAtEndOfBlock(SI))
11942 return 0; // xform done!
11947 /// SimplifyStoreAtEndOfBlock - Turn things like:
11948 /// if () { *P = v1; } else { *P = v2 }
11949 /// into a phi node with a store in the successor.
11951 /// Simplify things like:
11952 /// *P = v1; if () { *P = v2; }
11953 /// into a phi node with a store in the successor.
11955 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11956 BasicBlock *StoreBB = SI.getParent();
11958 // Check to see if the successor block has exactly two incoming edges. If
11959 // so, see if the other predecessor contains a store to the same location.
11960 // if so, insert a PHI node (if needed) and move the stores down.
11961 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11963 // Determine whether Dest has exactly two predecessors and, if so, compute
11964 // the other predecessor.
11965 pred_iterator PI = pred_begin(DestBB);
11966 BasicBlock *OtherBB = 0;
11967 if (*PI != StoreBB)
11970 if (PI == pred_end(DestBB))
11973 if (*PI != StoreBB) {
11978 if (++PI != pred_end(DestBB))
11981 // Bail out if all the relevant blocks aren't distinct (this can happen,
11982 // for example, if SI is in an infinite loop)
11983 if (StoreBB == DestBB || OtherBB == DestBB)
11986 // Verify that the other block ends in a branch and is not otherwise empty.
11987 BasicBlock::iterator BBI = OtherBB->getTerminator();
11988 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11989 if (!OtherBr || BBI == OtherBB->begin())
11992 // If the other block ends in an unconditional branch, check for the 'if then
11993 // else' case. there is an instruction before the branch.
11994 StoreInst *OtherStore = 0;
11995 if (OtherBr->isUnconditional()) {
11997 // Skip over debugging info.
11998 while (isa<DbgInfoIntrinsic>(BBI) ||
11999 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12000 if (BBI==OtherBB->begin())
12004 // If this isn't a store, or isn't a store to the same location, bail out.
12005 OtherStore = dyn_cast<StoreInst>(BBI);
12006 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
12009 // Otherwise, the other block ended with a conditional branch. If one of the
12010 // destinations is StoreBB, then we have the if/then case.
12011 if (OtherBr->getSuccessor(0) != StoreBB &&
12012 OtherBr->getSuccessor(1) != StoreBB)
12015 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12016 // if/then triangle. See if there is a store to the same ptr as SI that
12017 // lives in OtherBB.
12019 // Check to see if we find the matching store.
12020 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12021 if (OtherStore->getOperand(1) != SI.getOperand(1))
12025 // If we find something that may be using or overwriting the stored
12026 // value, or if we run out of instructions, we can't do the xform.
12027 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12028 BBI == OtherBB->begin())
12032 // In order to eliminate the store in OtherBr, we have to
12033 // make sure nothing reads or overwrites the stored value in
12035 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12036 // FIXME: This should really be AA driven.
12037 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12042 // Insert a PHI node now if we need it.
12043 Value *MergedVal = OtherStore->getOperand(0);
12044 if (MergedVal != SI.getOperand(0)) {
12045 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12046 PN->reserveOperandSpace(2);
12047 PN->addIncoming(SI.getOperand(0), SI.getParent());
12048 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12049 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12052 // Advance to a place where it is safe to insert the new store and
12054 BBI = DestBB->getFirstNonPHI();
12055 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12056 OtherStore->isVolatile()), *BBI);
12058 // Nuke the old stores.
12059 EraseInstFromFunction(SI);
12060 EraseInstFromFunction(*OtherStore);
12066 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12067 // Change br (not X), label True, label False to: br X, label False, True
12069 BasicBlock *TrueDest;
12070 BasicBlock *FalseDest;
12071 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
12072 !isa<Constant>(X)) {
12073 // Swap Destinations and condition...
12074 BI.setCondition(X);
12075 BI.setSuccessor(0, FalseDest);
12076 BI.setSuccessor(1, TrueDest);
12080 // Cannonicalize fcmp_one -> fcmp_oeq
12081 FCmpInst::Predicate FPred; Value *Y;
12082 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12083 TrueDest, FalseDest)))
12084 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12085 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
12086 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
12087 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
12088 Instruction *NewSCC = new FCmpInst(I, NewPred, X, Y, "");
12089 NewSCC->takeName(I);
12090 // Swap Destinations and condition...
12091 BI.setCondition(NewSCC);
12092 BI.setSuccessor(0, FalseDest);
12093 BI.setSuccessor(1, TrueDest);
12094 RemoveFromWorkList(I);
12095 I->eraseFromParent();
12096 AddToWorkList(NewSCC);
12100 // Cannonicalize icmp_ne -> icmp_eq
12101 ICmpInst::Predicate IPred;
12102 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12103 TrueDest, FalseDest)))
12104 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12105 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12106 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
12107 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
12108 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
12109 Instruction *NewSCC = new ICmpInst(I, NewPred, X, Y, "");
12110 NewSCC->takeName(I);
12111 // Swap Destinations and condition...
12112 BI.setCondition(NewSCC);
12113 BI.setSuccessor(0, FalseDest);
12114 BI.setSuccessor(1, TrueDest);
12115 RemoveFromWorkList(I);
12116 I->eraseFromParent();;
12117 AddToWorkList(NewSCC);
12124 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12125 Value *Cond = SI.getCondition();
12126 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12127 if (I->getOpcode() == Instruction::Add)
12128 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12129 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12130 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12132 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
12134 SI.setOperand(0, I->getOperand(0));
12142 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12143 Value *Agg = EV.getAggregateOperand();
12145 if (!EV.hasIndices())
12146 return ReplaceInstUsesWith(EV, Agg);
12148 if (Constant *C = dyn_cast<Constant>(Agg)) {
12149 if (isa<UndefValue>(C))
12150 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12152 if (isa<ConstantAggregateZero>(C))
12153 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12155 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12156 // Extract the element indexed by the first index out of the constant
12157 Value *V = C->getOperand(*EV.idx_begin());
12158 if (EV.getNumIndices() > 1)
12159 // Extract the remaining indices out of the constant indexed by the
12161 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12163 return ReplaceInstUsesWith(EV, V);
12165 return 0; // Can't handle other constants
12167 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12168 // We're extracting from an insertvalue instruction, compare the indices
12169 const unsigned *exti, *exte, *insi, *inse;
12170 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12171 exte = EV.idx_end(), inse = IV->idx_end();
12172 exti != exte && insi != inse;
12174 if (*insi != *exti)
12175 // The insert and extract both reference distinctly different elements.
12176 // This means the extract is not influenced by the insert, and we can
12177 // replace the aggregate operand of the extract with the aggregate
12178 // operand of the insert. i.e., replace
12179 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12180 // %E = extractvalue { i32, { i32 } } %I, 0
12182 // %E = extractvalue { i32, { i32 } } %A, 0
12183 return ExtractValueInst::Create(IV->getAggregateOperand(),
12184 EV.idx_begin(), EV.idx_end());
12186 if (exti == exte && insi == inse)
12187 // Both iterators are at the end: Index lists are identical. Replace
12188 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12189 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12191 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12192 if (exti == exte) {
12193 // The extract list is a prefix of the insert list. i.e. replace
12194 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12195 // %E = extractvalue { i32, { i32 } } %I, 1
12197 // %X = extractvalue { i32, { i32 } } %A, 1
12198 // %E = insertvalue { i32 } %X, i32 42, 0
12199 // by switching the order of the insert and extract (though the
12200 // insertvalue should be left in, since it may have other uses).
12201 Value *NewEV = InsertNewInstBefore(
12202 ExtractValueInst::Create(IV->getAggregateOperand(),
12203 EV.idx_begin(), EV.idx_end()),
12205 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12209 // The insert list is a prefix of the extract list
12210 // We can simply remove the common indices from the extract and make it
12211 // operate on the inserted value instead of the insertvalue result.
12213 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12214 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12216 // %E extractvalue { i32 } { i32 42 }, 0
12217 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12220 // Can't simplify extracts from other values. Note that nested extracts are
12221 // already simplified implicitely by the above (extract ( extract (insert) )
12222 // will be translated into extract ( insert ( extract ) ) first and then just
12223 // the value inserted, if appropriate).
12227 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12228 /// is to leave as a vector operation.
12229 static bool CheapToScalarize(Value *V, bool isConstant) {
12230 if (isa<ConstantAggregateZero>(V))
12232 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12233 if (isConstant) return true;
12234 // If all elts are the same, we can extract.
12235 Constant *Op0 = C->getOperand(0);
12236 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12237 if (C->getOperand(i) != Op0)
12241 Instruction *I = dyn_cast<Instruction>(V);
12242 if (!I) return false;
12244 // Insert element gets simplified to the inserted element or is deleted if
12245 // this is constant idx extract element and its a constant idx insertelt.
12246 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12247 isa<ConstantInt>(I->getOperand(2)))
12249 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12251 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12252 if (BO->hasOneUse() &&
12253 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12254 CheapToScalarize(BO->getOperand(1), isConstant)))
12256 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12257 if (CI->hasOneUse() &&
12258 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12259 CheapToScalarize(CI->getOperand(1), isConstant)))
12265 /// Read and decode a shufflevector mask.
12267 /// It turns undef elements into values that are larger than the number of
12268 /// elements in the input.
12269 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12270 unsigned NElts = SVI->getType()->getNumElements();
12271 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12272 return std::vector<unsigned>(NElts, 0);
12273 if (isa<UndefValue>(SVI->getOperand(2)))
12274 return std::vector<unsigned>(NElts, 2*NElts);
12276 std::vector<unsigned> Result;
12277 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12278 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12279 if (isa<UndefValue>(*i))
12280 Result.push_back(NElts*2); // undef -> 8
12282 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12286 /// FindScalarElement - Given a vector and an element number, see if the scalar
12287 /// value is already around as a register, for example if it were inserted then
12288 /// extracted from the vector.
12289 static Value *FindScalarElement(Value *V, unsigned EltNo,
12290 LLVMContext *Context) {
12291 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12292 const VectorType *PTy = cast<VectorType>(V->getType());
12293 unsigned Width = PTy->getNumElements();
12294 if (EltNo >= Width) // Out of range access.
12295 return UndefValue::get(PTy->getElementType());
12297 if (isa<UndefValue>(V))
12298 return UndefValue::get(PTy->getElementType());
12299 else if (isa<ConstantAggregateZero>(V))
12300 return Constant::getNullValue(PTy->getElementType());
12301 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12302 return CP->getOperand(EltNo);
12303 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12304 // If this is an insert to a variable element, we don't know what it is.
12305 if (!isa<ConstantInt>(III->getOperand(2)))
12307 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12309 // If this is an insert to the element we are looking for, return the
12311 if (EltNo == IIElt)
12312 return III->getOperand(1);
12314 // Otherwise, the insertelement doesn't modify the value, recurse on its
12316 return FindScalarElement(III->getOperand(0), EltNo, Context);
12317 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12318 unsigned LHSWidth =
12319 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12320 unsigned InEl = getShuffleMask(SVI)[EltNo];
12321 if (InEl < LHSWidth)
12322 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12323 else if (InEl < LHSWidth*2)
12324 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12326 return UndefValue::get(PTy->getElementType());
12329 // Otherwise, we don't know.
12333 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12334 // If vector val is undef, replace extract with scalar undef.
12335 if (isa<UndefValue>(EI.getOperand(0)))
12336 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12338 // If vector val is constant 0, replace extract with scalar 0.
12339 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12340 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12342 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12343 // If vector val is constant with all elements the same, replace EI with
12344 // that element. When the elements are not identical, we cannot replace yet
12345 // (we do that below, but only when the index is constant).
12346 Constant *op0 = C->getOperand(0);
12347 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12348 if (C->getOperand(i) != op0) {
12353 return ReplaceInstUsesWith(EI, op0);
12356 // If extracting a specified index from the vector, see if we can recursively
12357 // find a previously computed scalar that was inserted into the vector.
12358 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12359 unsigned IndexVal = IdxC->getZExtValue();
12360 unsigned VectorWidth =
12361 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12363 // If this is extracting an invalid index, turn this into undef, to avoid
12364 // crashing the code below.
12365 if (IndexVal >= VectorWidth)
12366 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12368 // This instruction only demands the single element from the input vector.
12369 // If the input vector has a single use, simplify it based on this use
12371 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12372 APInt UndefElts(VectorWidth, 0);
12373 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12374 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12375 DemandedMask, UndefElts)) {
12376 EI.setOperand(0, V);
12381 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12382 return ReplaceInstUsesWith(EI, Elt);
12384 // If the this extractelement is directly using a bitcast from a vector of
12385 // the same number of elements, see if we can find the source element from
12386 // it. In this case, we will end up needing to bitcast the scalars.
12387 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12388 if (const VectorType *VT =
12389 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12390 if (VT->getNumElements() == VectorWidth)
12391 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12392 IndexVal, Context))
12393 return new BitCastInst(Elt, EI.getType());
12397 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12398 if (I->hasOneUse()) {
12399 // Push extractelement into predecessor operation if legal and
12400 // profitable to do so
12401 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12402 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12403 if (CheapToScalarize(BO, isConstantElt)) {
12404 ExtractElementInst *newEI0 =
12405 ExtractElementInst::Create(BO->getOperand(0), EI.getOperand(1),
12406 EI.getName()+".lhs");
12407 ExtractElementInst *newEI1 =
12408 ExtractElementInst::Create(BO->getOperand(1), EI.getOperand(1),
12409 EI.getName()+".rhs");
12410 InsertNewInstBefore(newEI0, EI);
12411 InsertNewInstBefore(newEI1, EI);
12412 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12414 } else if (isa<LoadInst>(I)) {
12416 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12417 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12418 PointerType::get(EI.getType(), AS),*I);
12419 GetElementPtrInst *GEP =
12420 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12421 cast<GEPOperator>(GEP)->setIsInBounds(true);
12422 InsertNewInstBefore(GEP, *I);
12423 LoadInst* Load = new LoadInst(GEP, "tmp");
12424 InsertNewInstBefore(Load, *I);
12425 return ReplaceInstUsesWith(EI, Load);
12428 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12429 // Extracting the inserted element?
12430 if (IE->getOperand(2) == EI.getOperand(1))
12431 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12432 // If the inserted and extracted elements are constants, they must not
12433 // be the same value, extract from the pre-inserted value instead.
12434 if (isa<Constant>(IE->getOperand(2)) &&
12435 isa<Constant>(EI.getOperand(1))) {
12436 AddUsesToWorkList(EI);
12437 EI.setOperand(0, IE->getOperand(0));
12440 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12441 // If this is extracting an element from a shufflevector, figure out where
12442 // it came from and extract from the appropriate input element instead.
12443 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12444 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12446 unsigned LHSWidth =
12447 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12449 if (SrcIdx < LHSWidth)
12450 Src = SVI->getOperand(0);
12451 else if (SrcIdx < LHSWidth*2) {
12452 SrcIdx -= LHSWidth;
12453 Src = SVI->getOperand(1);
12455 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12457 return ExtractElementInst::Create(Src,
12458 ConstantInt::get(Type::getInt32Ty(*Context), SrcIdx, false));
12461 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12466 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12467 /// elements from either LHS or RHS, return the shuffle mask and true.
12468 /// Otherwise, return false.
12469 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12470 std::vector<Constant*> &Mask,
12471 LLVMContext *Context) {
12472 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12473 "Invalid CollectSingleShuffleElements");
12474 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12476 if (isa<UndefValue>(V)) {
12477 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12479 } else if (V == LHS) {
12480 for (unsigned i = 0; i != NumElts; ++i)
12481 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12483 } else if (V == RHS) {
12484 for (unsigned i = 0; i != NumElts; ++i)
12485 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i+NumElts));
12487 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12488 // If this is an insert of an extract from some other vector, include it.
12489 Value *VecOp = IEI->getOperand(0);
12490 Value *ScalarOp = IEI->getOperand(1);
12491 Value *IdxOp = IEI->getOperand(2);
12493 if (!isa<ConstantInt>(IdxOp))
12495 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12497 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12498 // Okay, we can handle this if the vector we are insertinting into is
12499 // transitively ok.
12500 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12501 // If so, update the mask to reflect the inserted undef.
12502 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(*Context));
12505 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12506 if (isa<ConstantInt>(EI->getOperand(1)) &&
12507 EI->getOperand(0)->getType() == V->getType()) {
12508 unsigned ExtractedIdx =
12509 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12511 // This must be extracting from either LHS or RHS.
12512 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12513 // Okay, we can handle this if the vector we are insertinting into is
12514 // transitively ok.
12515 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12516 // If so, update the mask to reflect the inserted value.
12517 if (EI->getOperand(0) == LHS) {
12518 Mask[InsertedIdx % NumElts] =
12519 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12521 assert(EI->getOperand(0) == RHS);
12522 Mask[InsertedIdx % NumElts] =
12523 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx+NumElts);
12532 // TODO: Handle shufflevector here!
12537 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12538 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12539 /// that computes V and the LHS value of the shuffle.
12540 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12541 Value *&RHS, LLVMContext *Context) {
12542 assert(isa<VectorType>(V->getType()) &&
12543 (RHS == 0 || V->getType() == RHS->getType()) &&
12544 "Invalid shuffle!");
12545 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12547 if (isa<UndefValue>(V)) {
12548 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12550 } else if (isa<ConstantAggregateZero>(V)) {
12551 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(*Context), 0));
12553 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12554 // If this is an insert of an extract from some other vector, include it.
12555 Value *VecOp = IEI->getOperand(0);
12556 Value *ScalarOp = IEI->getOperand(1);
12557 Value *IdxOp = IEI->getOperand(2);
12559 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12560 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12561 EI->getOperand(0)->getType() == V->getType()) {
12562 unsigned ExtractedIdx =
12563 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12564 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12566 // Either the extracted from or inserted into vector must be RHSVec,
12567 // otherwise we'd end up with a shuffle of three inputs.
12568 if (EI->getOperand(0) == RHS || RHS == 0) {
12569 RHS = EI->getOperand(0);
12570 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12571 Mask[InsertedIdx % NumElts] =
12572 ConstantInt::get(Type::getInt32Ty(*Context), NumElts+ExtractedIdx);
12576 if (VecOp == RHS) {
12577 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12579 // Everything but the extracted element is replaced with the RHS.
12580 for (unsigned i = 0; i != NumElts; ++i) {
12581 if (i != InsertedIdx)
12582 Mask[i] = ConstantInt::get(Type::getInt32Ty(*Context), NumElts+i);
12587 // If this insertelement is a chain that comes from exactly these two
12588 // vectors, return the vector and the effective shuffle.
12589 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12591 return EI->getOperand(0);
12596 // TODO: Handle shufflevector here!
12598 // Otherwise, can't do anything fancy. Return an identity vector.
12599 for (unsigned i = 0; i != NumElts; ++i)
12600 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12604 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12605 Value *VecOp = IE.getOperand(0);
12606 Value *ScalarOp = IE.getOperand(1);
12607 Value *IdxOp = IE.getOperand(2);
12609 // Inserting an undef or into an undefined place, remove this.
12610 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12611 ReplaceInstUsesWith(IE, VecOp);
12613 // If the inserted element was extracted from some other vector, and if the
12614 // indexes are constant, try to turn this into a shufflevector operation.
12615 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12616 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12617 EI->getOperand(0)->getType() == IE.getType()) {
12618 unsigned NumVectorElts = IE.getType()->getNumElements();
12619 unsigned ExtractedIdx =
12620 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12621 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12623 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12624 return ReplaceInstUsesWith(IE, VecOp);
12626 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12627 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12629 // If we are extracting a value from a vector, then inserting it right
12630 // back into the same place, just use the input vector.
12631 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12632 return ReplaceInstUsesWith(IE, VecOp);
12634 // We could theoretically do this for ANY input. However, doing so could
12635 // turn chains of insertelement instructions into a chain of shufflevector
12636 // instructions, and right now we do not merge shufflevectors. As such,
12637 // only do this in a situation where it is clear that there is benefit.
12638 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12639 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12640 // the values of VecOp, except then one read from EIOp0.
12641 // Build a new shuffle mask.
12642 std::vector<Constant*> Mask;
12643 if (isa<UndefValue>(VecOp))
12644 Mask.assign(NumVectorElts, UndefValue::get(Type::getInt32Ty(*Context)));
12646 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12647 Mask.assign(NumVectorElts, ConstantInt::get(Type::getInt32Ty(*Context),
12650 Mask[InsertedIdx] =
12651 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12652 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12653 ConstantVector::get(Mask));
12656 // If this insertelement isn't used by some other insertelement, turn it
12657 // (and any insertelements it points to), into one big shuffle.
12658 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12659 std::vector<Constant*> Mask;
12661 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12662 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12663 // We now have a shuffle of LHS, RHS, Mask.
12664 return new ShuffleVectorInst(LHS, RHS,
12665 ConstantVector::get(Mask));
12670 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12671 APInt UndefElts(VWidth, 0);
12672 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12673 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12680 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12681 Value *LHS = SVI.getOperand(0);
12682 Value *RHS = SVI.getOperand(1);
12683 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12685 bool MadeChange = false;
12687 // Undefined shuffle mask -> undefined value.
12688 if (isa<UndefValue>(SVI.getOperand(2)))
12689 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12691 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12693 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12696 APInt UndefElts(VWidth, 0);
12697 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12698 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12699 LHS = SVI.getOperand(0);
12700 RHS = SVI.getOperand(1);
12704 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12705 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12706 if (LHS == RHS || isa<UndefValue>(LHS)) {
12707 if (isa<UndefValue>(LHS) && LHS == RHS) {
12708 // shuffle(undef,undef,mask) -> undef.
12709 return ReplaceInstUsesWith(SVI, LHS);
12712 // Remap any references to RHS to use LHS.
12713 std::vector<Constant*> Elts;
12714 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12715 if (Mask[i] >= 2*e)
12716 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12718 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12719 (Mask[i] < e && isa<UndefValue>(LHS))) {
12720 Mask[i] = 2*e; // Turn into undef.
12721 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12723 Mask[i] = Mask[i] % e; // Force to LHS.
12724 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Mask[i]));
12728 SVI.setOperand(0, SVI.getOperand(1));
12729 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12730 SVI.setOperand(2, ConstantVector::get(Elts));
12731 LHS = SVI.getOperand(0);
12732 RHS = SVI.getOperand(1);
12736 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12737 bool isLHSID = true, isRHSID = true;
12739 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12740 if (Mask[i] >= e*2) continue; // Ignore undef values.
12741 // Is this an identity shuffle of the LHS value?
12742 isLHSID &= (Mask[i] == i);
12744 // Is this an identity shuffle of the RHS value?
12745 isRHSID &= (Mask[i]-e == i);
12748 // Eliminate identity shuffles.
12749 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12750 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12752 // If the LHS is a shufflevector itself, see if we can combine it with this
12753 // one without producing an unusual shuffle. Here we are really conservative:
12754 // we are absolutely afraid of producing a shuffle mask not in the input
12755 // program, because the code gen may not be smart enough to turn a merged
12756 // shuffle into two specific shuffles: it may produce worse code. As such,
12757 // we only merge two shuffles if the result is one of the two input shuffle
12758 // masks. In this case, merging the shuffles just removes one instruction,
12759 // which we know is safe. This is good for things like turning:
12760 // (splat(splat)) -> splat.
12761 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12762 if (isa<UndefValue>(RHS)) {
12763 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12765 std::vector<unsigned> NewMask;
12766 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12767 if (Mask[i] >= 2*e)
12768 NewMask.push_back(2*e);
12770 NewMask.push_back(LHSMask[Mask[i]]);
12772 // If the result mask is equal to the src shuffle or this shuffle mask, do
12773 // the replacement.
12774 if (NewMask == LHSMask || NewMask == Mask) {
12775 unsigned LHSInNElts =
12776 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12777 std::vector<Constant*> Elts;
12778 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12779 if (NewMask[i] >= LHSInNElts*2) {
12780 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12782 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), NewMask[i]));
12785 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12786 LHSSVI->getOperand(1),
12787 ConstantVector::get(Elts));
12792 return MadeChange ? &SVI : 0;
12798 /// TryToSinkInstruction - Try to move the specified instruction from its
12799 /// current block into the beginning of DestBlock, which can only happen if it's
12800 /// safe to move the instruction past all of the instructions between it and the
12801 /// end of its block.
12802 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12803 assert(I->hasOneUse() && "Invariants didn't hold!");
12805 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12806 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12809 // Do not sink alloca instructions out of the entry block.
12810 if (isa<AllocaInst>(I) && I->getParent() ==
12811 &DestBlock->getParent()->getEntryBlock())
12814 // We can only sink load instructions if there is nothing between the load and
12815 // the end of block that could change the value.
12816 if (I->mayReadFromMemory()) {
12817 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12819 if (Scan->mayWriteToMemory())
12823 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12825 CopyPrecedingStopPoint(I, InsertPos);
12826 I->moveBefore(InsertPos);
12832 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12833 /// all reachable code to the worklist.
12835 /// This has a couple of tricks to make the code faster and more powerful. In
12836 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12837 /// them to the worklist (this significantly speeds up instcombine on code where
12838 /// many instructions are dead or constant). Additionally, if we find a branch
12839 /// whose condition is a known constant, we only visit the reachable successors.
12841 static void AddReachableCodeToWorklist(BasicBlock *BB,
12842 SmallPtrSet<BasicBlock*, 64> &Visited,
12844 const TargetData *TD) {
12845 SmallVector<BasicBlock*, 256> Worklist;
12846 Worklist.push_back(BB);
12848 while (!Worklist.empty()) {
12849 BB = Worklist.back();
12850 Worklist.pop_back();
12852 // We have now visited this block! If we've already been here, ignore it.
12853 if (!Visited.insert(BB)) continue;
12855 DbgInfoIntrinsic *DBI_Prev = NULL;
12856 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12857 Instruction *Inst = BBI++;
12859 // DCE instruction if trivially dead.
12860 if (isInstructionTriviallyDead(Inst)) {
12862 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
12863 Inst->eraseFromParent();
12867 // ConstantProp instruction if trivially constant.
12868 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12869 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
12871 Inst->replaceAllUsesWith(C);
12873 Inst->eraseFromParent();
12877 // If there are two consecutive llvm.dbg.stoppoint calls then
12878 // it is likely that the optimizer deleted code in between these
12880 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12883 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12884 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12885 IC.RemoveFromWorkList(DBI_Prev);
12886 DBI_Prev->eraseFromParent();
12888 DBI_Prev = DBI_Next;
12893 IC.AddToWorkList(Inst);
12896 // Recursively visit successors. If this is a branch or switch on a
12897 // constant, only visit the reachable successor.
12898 TerminatorInst *TI = BB->getTerminator();
12899 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12900 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12901 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12902 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12903 Worklist.push_back(ReachableBB);
12906 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12907 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12908 // See if this is an explicit destination.
12909 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12910 if (SI->getCaseValue(i) == Cond) {
12911 BasicBlock *ReachableBB = SI->getSuccessor(i);
12912 Worklist.push_back(ReachableBB);
12916 // Otherwise it is the default destination.
12917 Worklist.push_back(SI->getSuccessor(0));
12922 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12923 Worklist.push_back(TI->getSuccessor(i));
12927 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12928 bool Changed = false;
12929 TD = getAnalysisIfAvailable<TargetData>();
12931 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12932 << F.getNameStr() << "\n");
12935 // Do a depth-first traversal of the function, populate the worklist with
12936 // the reachable instructions. Ignore blocks that are not reachable. Keep
12937 // track of which blocks we visit.
12938 SmallPtrSet<BasicBlock*, 64> Visited;
12939 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12941 // Do a quick scan over the function. If we find any blocks that are
12942 // unreachable, remove any instructions inside of them. This prevents
12943 // the instcombine code from having to deal with some bad special cases.
12944 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12945 if (!Visited.count(BB)) {
12946 Instruction *Term = BB->getTerminator();
12947 while (Term != BB->begin()) { // Remove instrs bottom-up
12948 BasicBlock::iterator I = Term; --I;
12950 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12951 // A debug intrinsic shouldn't force another iteration if we weren't
12952 // going to do one without it.
12953 if (!isa<DbgInfoIntrinsic>(I)) {
12957 if (!I->use_empty())
12958 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12959 I->eraseFromParent();
12964 while (!Worklist.isEmpty()) {
12965 Instruction *I = Worklist.RemoveOne();
12966 if (I == 0) continue; // skip null values.
12968 // Check to see if we can DCE the instruction.
12969 if (isInstructionTriviallyDead(I)) {
12970 // Add operands to the worklist.
12971 if (I->getNumOperands() < 4)
12972 AddUsesToWorkList(*I);
12975 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12977 I->eraseFromParent();
12978 RemoveFromWorkList(I);
12983 // Instruction isn't dead, see if we can constant propagate it.
12984 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
12985 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
12987 // Add operands to the worklist.
12988 AddUsesToWorkList(*I);
12989 ReplaceInstUsesWith(*I, C);
12992 I->eraseFromParent();
12993 RemoveFromWorkList(I);
12999 // See if we can constant fold its operands.
13000 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
13001 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
13002 if (Constant *NewC = ConstantFoldConstantExpression(CE,
13003 F.getContext(), TD))
13010 // See if we can trivially sink this instruction to a successor basic block.
13011 if (I->hasOneUse()) {
13012 BasicBlock *BB = I->getParent();
13013 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
13014 if (UserParent != BB) {
13015 bool UserIsSuccessor = false;
13016 // See if the user is one of our successors.
13017 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13018 if (*SI == UserParent) {
13019 UserIsSuccessor = true;
13023 // If the user is one of our immediate successors, and if that successor
13024 // only has us as a predecessors (we'd have to split the critical edge
13025 // otherwise), we can keep going.
13026 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
13027 next(pred_begin(UserParent)) == pred_end(UserParent))
13028 // Okay, the CFG is simple enough, try to sink this instruction.
13029 Changed |= TryToSinkInstruction(I, UserParent);
13033 // Now that we have an instruction, try combining it to simplify it...
13037 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
13038 if (Instruction *Result = visit(*I)) {
13040 // Should we replace the old instruction with a new one?
13042 DEBUG(errs() << "IC: Old = " << *I << '\n'
13043 << " New = " << *Result << '\n');
13045 // Everything uses the new instruction now.
13046 I->replaceAllUsesWith(Result);
13048 // Push the new instruction and any users onto the worklist.
13049 AddToWorkList(Result);
13050 AddUsersToWorkList(*Result);
13052 // Move the name to the new instruction first.
13053 Result->takeName(I);
13055 // Insert the new instruction into the basic block...
13056 BasicBlock *InstParent = I->getParent();
13057 BasicBlock::iterator InsertPos = I;
13059 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13060 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13063 InstParent->getInstList().insert(InsertPos, Result);
13065 // Make sure that we reprocess all operands now that we reduced their
13067 AddUsesToWorkList(*I);
13069 // Instructions can end up on the worklist more than once. Make sure
13070 // we do not process an instruction that has been deleted.
13071 RemoveFromWorkList(I);
13073 // Erase the old instruction.
13074 InstParent->getInstList().erase(I);
13077 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
13078 << " New = " << *I << '\n');
13081 // If the instruction was modified, it's possible that it is now dead.
13082 // if so, remove it.
13083 if (isInstructionTriviallyDead(I)) {
13084 // Make sure we process all operands now that we are reducing their
13086 AddUsesToWorkList(*I);
13088 // Instructions may end up in the worklist more than once. Erase all
13089 // occurrences of this instruction.
13090 RemoveFromWorkList(I);
13091 I->eraseFromParent();
13094 AddUsersToWorkList(*I);
13106 bool InstCombiner::runOnFunction(Function &F) {
13107 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13108 Context = &F.getContext();
13110 bool EverMadeChange = false;
13112 // Iterate while there is work to do.
13113 unsigned Iteration = 0;
13114 while (DoOneIteration(F, Iteration++))
13115 EverMadeChange = true;
13116 return EverMadeChange;
13119 FunctionPass *llvm::createInstructionCombiningPass() {
13120 return new InstCombiner();