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 class VISIBILITY_HIDDEN InstCombiner
78 : public FunctionPass,
79 public InstVisitor<InstCombiner, Instruction*> {
80 // Worklist of all of the instructions that need to be simplified.
81 SmallVector<Instruction*, 256> Worklist;
82 DenseMap<Instruction*, unsigned> WorklistMap;
84 bool MustPreserveLCSSA;
86 static char ID; // Pass identification, replacement for typeid
87 InstCombiner() : FunctionPass(&ID) {}
90 LLVMContext *getContext() const { return Context; }
92 /// AddToWorkList - Add the specified instruction to the worklist if it
93 /// isn't already in it.
94 void AddToWorkList(Instruction *I) {
95 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
96 Worklist.push_back(I);
99 // RemoveFromWorkList - remove I from the worklist if it exists.
100 void RemoveFromWorkList(Instruction *I) {
101 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
102 if (It == WorklistMap.end()) return; // Not in worklist.
104 // Don't bother moving everything down, just null out the slot.
105 Worklist[It->second] = 0;
107 WorklistMap.erase(It);
110 Instruction *RemoveOneFromWorkList() {
111 Instruction *I = Worklist.back();
113 WorklistMap.erase(I);
118 /// AddUsersToWorkList - When an instruction is simplified, add all users of
119 /// the instruction to the work lists because they might get more simplified
122 void AddUsersToWorkList(Value &I) {
123 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
125 AddToWorkList(cast<Instruction>(*UI));
128 /// AddUsesToWorkList - When an instruction is simplified, add operands to
129 /// the work lists because they might get more simplified now.
131 void AddUsesToWorkList(Instruction &I) {
132 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
133 if (Instruction *Op = dyn_cast<Instruction>(*i))
137 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
138 /// dead. Add all of its operands to the worklist, turning them into
139 /// undef's to reduce the number of uses of those instructions.
141 /// Return the specified operand before it is turned into an undef.
143 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
144 Value *R = I.getOperand(op);
146 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
147 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
149 // Set the operand to undef to drop the use.
150 *i = UndefValue::get(Op->getType());
157 virtual bool runOnFunction(Function &F);
159 bool DoOneIteration(Function &F, unsigned ItNum);
161 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
162 AU.addPreservedID(LCSSAID);
163 AU.setPreservesCFG();
166 TargetData *getTargetData() const { return TD; }
168 // Visitation implementation - Implement instruction combining for different
169 // instruction types. The semantics are as follows:
171 // null - No change was made
172 // I - Change was made, I is still valid, I may be dead though
173 // otherwise - Change was made, replace I with returned instruction
175 Instruction *visitAdd(BinaryOperator &I);
176 Instruction *visitFAdd(BinaryOperator &I);
177 Instruction *visitSub(BinaryOperator &I);
178 Instruction *visitFSub(BinaryOperator &I);
179 Instruction *visitMul(BinaryOperator &I);
180 Instruction *visitFMul(BinaryOperator &I);
181 Instruction *visitURem(BinaryOperator &I);
182 Instruction *visitSRem(BinaryOperator &I);
183 Instruction *visitFRem(BinaryOperator &I);
184 bool SimplifyDivRemOfSelect(BinaryOperator &I);
185 Instruction *commonRemTransforms(BinaryOperator &I);
186 Instruction *commonIRemTransforms(BinaryOperator &I);
187 Instruction *commonDivTransforms(BinaryOperator &I);
188 Instruction *commonIDivTransforms(BinaryOperator &I);
189 Instruction *visitUDiv(BinaryOperator &I);
190 Instruction *visitSDiv(BinaryOperator &I);
191 Instruction *visitFDiv(BinaryOperator &I);
192 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
193 Instruction *FoldAndOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
194 Instruction *visitAnd(BinaryOperator &I);
195 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
196 Instruction *FoldOrOfFCmps(Instruction &I, FCmpInst *LHS, FCmpInst *RHS);
197 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
198 Value *A, Value *B, Value *C);
199 Instruction *visitOr (BinaryOperator &I);
200 Instruction *visitXor(BinaryOperator &I);
201 Instruction *visitShl(BinaryOperator &I);
202 Instruction *visitAShr(BinaryOperator &I);
203 Instruction *visitLShr(BinaryOperator &I);
204 Instruction *commonShiftTransforms(BinaryOperator &I);
205 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
207 Instruction *visitFCmpInst(FCmpInst &I);
208 Instruction *visitICmpInst(ICmpInst &I);
209 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
210 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
213 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
214 ConstantInt *DivRHS);
216 Instruction *FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
217 ICmpInst::Predicate Cond, Instruction &I);
218 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
220 Instruction *commonCastTransforms(CastInst &CI);
221 Instruction *commonIntCastTransforms(CastInst &CI);
222 Instruction *commonPointerCastTransforms(CastInst &CI);
223 Instruction *visitTrunc(TruncInst &CI);
224 Instruction *visitZExt(ZExtInst &CI);
225 Instruction *visitSExt(SExtInst &CI);
226 Instruction *visitFPTrunc(FPTruncInst &CI);
227 Instruction *visitFPExt(CastInst &CI);
228 Instruction *visitFPToUI(FPToUIInst &FI);
229 Instruction *visitFPToSI(FPToSIInst &FI);
230 Instruction *visitUIToFP(CastInst &CI);
231 Instruction *visitSIToFP(CastInst &CI);
232 Instruction *visitPtrToInt(PtrToIntInst &CI);
233 Instruction *visitIntToPtr(IntToPtrInst &CI);
234 Instruction *visitBitCast(BitCastInst &CI);
235 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
237 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
238 Instruction *visitSelectInst(SelectInst &SI);
239 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
240 Instruction *visitCallInst(CallInst &CI);
241 Instruction *visitInvokeInst(InvokeInst &II);
242 Instruction *visitPHINode(PHINode &PN);
243 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
244 Instruction *visitAllocationInst(AllocationInst &AI);
245 Instruction *visitFreeInst(FreeInst &FI);
246 Instruction *visitLoadInst(LoadInst &LI);
247 Instruction *visitStoreInst(StoreInst &SI);
248 Instruction *visitBranchInst(BranchInst &BI);
249 Instruction *visitSwitchInst(SwitchInst &SI);
250 Instruction *visitInsertElementInst(InsertElementInst &IE);
251 Instruction *visitExtractElementInst(ExtractElementInst &EI);
252 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
253 Instruction *visitExtractValueInst(ExtractValueInst &EV);
255 // visitInstruction - Specify what to return for unhandled instructions...
256 Instruction *visitInstruction(Instruction &I) { return 0; }
259 Instruction *visitCallSite(CallSite CS);
260 bool transformConstExprCastCall(CallSite CS);
261 Instruction *transformCallThroughTrampoline(CallSite CS);
262 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
263 bool DoXform = true);
264 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
265 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
269 // InsertNewInstBefore - insert an instruction New before instruction Old
270 // in the program. Add the new instruction to the worklist.
272 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
273 assert(New && New->getParent() == 0 &&
274 "New instruction already inserted into a basic block!");
275 BasicBlock *BB = Old.getParent();
276 BB->getInstList().insert(&Old, New); // Insert inst
281 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
282 /// This also adds the cast to the worklist. Finally, this returns the
284 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
286 if (V->getType() == Ty) return V;
288 if (Constant *CV = dyn_cast<Constant>(V))
289 return ConstantExpr::getCast(opc, CV, Ty);
291 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
296 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
297 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
301 // ReplaceInstUsesWith - This method is to be used when an instruction is
302 // found to be dead, replacable with another preexisting expression. Here
303 // we add all uses of I to the worklist, replace all uses of I with the new
304 // value, then return I, so that the inst combiner will know that I was
307 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
308 AddUsersToWorkList(I); // Add all modified instrs to worklist
310 I.replaceAllUsesWith(V);
313 // If we are replacing the instruction with itself, this must be in a
314 // segment of unreachable code, so just clobber the instruction.
315 I.replaceAllUsesWith(UndefValue::get(I.getType()));
320 // EraseInstFromFunction - When dealing with an instruction that has side
321 // effects or produces a void value, we can't rely on DCE to delete the
322 // instruction. Instead, visit methods should return the value returned by
324 Instruction *EraseInstFromFunction(Instruction &I) {
325 assert(I.use_empty() && "Cannot erase instruction that is used!");
326 AddUsesToWorkList(I);
327 RemoveFromWorkList(&I);
329 return 0; // Don't do anything with FI
332 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
333 APInt &KnownOne, unsigned Depth = 0) const {
334 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
337 bool MaskedValueIsZero(Value *V, const APInt &Mask,
338 unsigned Depth = 0) const {
339 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
341 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
342 return llvm::ComputeNumSignBits(Op, TD, Depth);
347 /// SimplifyCommutative - This performs a few simplifications for
348 /// commutative operators.
349 bool SimplifyCommutative(BinaryOperator &I);
351 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
352 /// most-complex to least-complex order.
353 bool SimplifyCompare(CmpInst &I);
355 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
356 /// based on the demanded bits.
357 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
358 APInt& KnownZero, APInt& KnownOne,
360 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
361 APInt& KnownZero, APInt& KnownOne,
364 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
365 /// SimplifyDemandedBits knows about. See if the instruction has any
366 /// properties that allow us to simplify its operands.
367 bool SimplifyDemandedInstructionBits(Instruction &Inst);
369 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
370 APInt& UndefElts, unsigned Depth = 0);
372 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
373 // PHI node as operand #0, see if we can fold the instruction into the PHI
374 // (which is only possible if all operands to the PHI are constants).
375 Instruction *FoldOpIntoPhi(Instruction &I);
377 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
378 // operator and they all are only used by the PHI, PHI together their
379 // inputs, and do the operation once, to the result of the PHI.
380 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
381 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
382 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
385 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
386 ConstantInt *AndRHS, BinaryOperator &TheAnd);
388 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
389 bool isSub, Instruction &I);
390 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
391 bool isSigned, bool Inside, Instruction &IB);
392 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
393 Instruction *MatchBSwap(BinaryOperator &I);
394 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
395 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
396 Instruction *SimplifyMemSet(MemSetInst *MI);
399 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
401 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
402 unsigned CastOpc, int &NumCastsRemoved);
403 unsigned GetOrEnforceKnownAlignment(Value *V,
404 unsigned PrefAlign = 0);
409 char InstCombiner::ID = 0;
410 static RegisterPass<InstCombiner>
411 X("instcombine", "Combine redundant instructions");
413 // getComplexity: Assign a complexity or rank value to LLVM Values...
414 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
415 static unsigned getComplexity(Value *V) {
416 if (isa<Instruction>(V)) {
417 if (BinaryOperator::isNeg(V) ||
418 BinaryOperator::isFNeg(V) ||
419 BinaryOperator::isNot(V))
423 if (isa<Argument>(V)) return 3;
424 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
427 // isOnlyUse - Return true if this instruction will be deleted if we stop using
429 static bool isOnlyUse(Value *V) {
430 return V->hasOneUse() || isa<Constant>(V);
433 // getPromotedType - Return the specified type promoted as it would be to pass
434 // though a va_arg area...
435 static const Type *getPromotedType(const Type *Ty) {
436 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
437 if (ITy->getBitWidth() < 32)
438 return Type::getInt32Ty(Ty->getContext());
443 /// getBitCastOperand - If the specified operand is a CastInst, a constant
444 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
445 /// operand value, otherwise return null.
446 static Value *getBitCastOperand(Value *V) {
447 if (Operator *O = dyn_cast<Operator>(V)) {
448 if (O->getOpcode() == Instruction::BitCast)
449 return O->getOperand(0);
450 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
451 if (GEP->hasAllZeroIndices())
452 return GEP->getPointerOperand();
457 /// This function is a wrapper around CastInst::isEliminableCastPair. It
458 /// simply extracts arguments and returns what that function returns.
459 static Instruction::CastOps
460 isEliminableCastPair(
461 const CastInst *CI, ///< The first cast instruction
462 unsigned opcode, ///< The opcode of the second cast instruction
463 const Type *DstTy, ///< The target type for the second cast instruction
464 TargetData *TD ///< The target data for pointer size
467 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
468 const Type *MidTy = CI->getType(); // B from above
470 // Get the opcodes of the two Cast instructions
471 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
472 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
474 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
476 TD ? TD->getIntPtrType(CI->getContext()) : 0);
478 // We don't want to form an inttoptr or ptrtoint that converts to an integer
479 // type that differs from the pointer size.
480 if ((Res == Instruction::IntToPtr &&
481 (!TD || SrcTy != TD->getIntPtrType(CI->getContext()))) ||
482 (Res == Instruction::PtrToInt &&
483 (!TD || DstTy != TD->getIntPtrType(CI->getContext()))))
486 return Instruction::CastOps(Res);
489 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
490 /// in any code being generated. It does not require codegen if V is simple
491 /// enough or if the cast can be folded into other casts.
492 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
493 const Type *Ty, TargetData *TD) {
494 if (V->getType() == Ty || isa<Constant>(V)) return false;
496 // If this is another cast that can be eliminated, it isn't codegen either.
497 if (const CastInst *CI = dyn_cast<CastInst>(V))
498 if (isEliminableCastPair(CI, opcode, Ty, TD))
503 // SimplifyCommutative - This performs a few simplifications for commutative
506 // 1. Order operands such that they are listed from right (least complex) to
507 // left (most complex). This puts constants before unary operators before
510 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
511 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
513 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
514 bool Changed = false;
515 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
516 Changed = !I.swapOperands();
518 if (!I.isAssociative()) return Changed;
519 Instruction::BinaryOps Opcode = I.getOpcode();
520 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
521 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
522 if (isa<Constant>(I.getOperand(1))) {
523 Constant *Folded = ConstantExpr::get(I.getOpcode(),
524 cast<Constant>(I.getOperand(1)),
525 cast<Constant>(Op->getOperand(1)));
526 I.setOperand(0, Op->getOperand(0));
527 I.setOperand(1, Folded);
529 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
530 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
531 isOnlyUse(Op) && isOnlyUse(Op1)) {
532 Constant *C1 = cast<Constant>(Op->getOperand(1));
533 Constant *C2 = cast<Constant>(Op1->getOperand(1));
535 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
536 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
537 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
541 I.setOperand(0, New);
542 I.setOperand(1, Folded);
549 /// SimplifyCompare - For a CmpInst this function just orders the operands
550 /// so that theyare listed from right (least complex) to left (most complex).
551 /// This puts constants before unary operators before binary operators.
552 bool InstCombiner::SimplifyCompare(CmpInst &I) {
553 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
556 // Compare instructions are not associative so there's nothing else we can do.
560 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
561 // if the LHS is a constant zero (which is the 'negate' form).
563 static inline Value *dyn_castNegVal(Value *V) {
564 if (BinaryOperator::isNeg(V))
565 return BinaryOperator::getNegArgument(V);
567 // Constants can be considered to be negated values if they can be folded.
568 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
569 return ConstantExpr::getNeg(C);
571 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
572 if (C->getType()->getElementType()->isInteger())
573 return ConstantExpr::getNeg(C);
578 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
579 // instruction if the LHS is a constant negative zero (which is the 'negate'
582 static inline Value *dyn_castFNegVal(Value *V) {
583 if (BinaryOperator::isFNeg(V))
584 return BinaryOperator::getFNegArgument(V);
586 // Constants can be considered to be negated values if they can be folded.
587 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
588 return ConstantExpr::getFNeg(C);
590 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
591 if (C->getType()->getElementType()->isFloatingPoint())
592 return ConstantExpr::getFNeg(C);
597 static inline Value *dyn_castNotVal(Value *V) {
598 if (BinaryOperator::isNot(V))
599 return BinaryOperator::getNotArgument(V);
601 // Constants can be considered to be not'ed values...
602 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
603 return ConstantInt::get(C->getType(), ~C->getValue());
607 // dyn_castFoldableMul - If this value is a multiply that can be folded into
608 // other computations (because it has a constant operand), return the
609 // non-constant operand of the multiply, and set CST to point to the multiplier.
610 // Otherwise, return null.
612 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
613 if (V->hasOneUse() && V->getType()->isInteger())
614 if (Instruction *I = dyn_cast<Instruction>(V)) {
615 if (I->getOpcode() == Instruction::Mul)
616 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
617 return I->getOperand(0);
618 if (I->getOpcode() == Instruction::Shl)
619 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
620 // The multiplier is really 1 << CST.
621 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
622 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
623 CST = ConstantInt::get(V->getType()->getContext(),
624 APInt(BitWidth, 1).shl(CSTVal));
625 return I->getOperand(0);
631 /// AddOne - Add one to a ConstantInt
632 static Constant *AddOne(Constant *C) {
633 return ConstantExpr::getAdd(C,
634 ConstantInt::get(C->getType(), 1));
636 /// SubOne - Subtract one from a ConstantInt
637 static Constant *SubOne(ConstantInt *C) {
638 return ConstantExpr::getSub(C,
639 ConstantInt::get(C->getType(), 1));
641 /// MultiplyOverflows - True if the multiply can not be expressed in an int
643 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
644 uint32_t W = C1->getBitWidth();
645 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
654 APInt MulExt = LHSExt * RHSExt;
657 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
658 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
659 return MulExt.slt(Min) || MulExt.sgt(Max);
661 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
665 /// ShrinkDemandedConstant - Check to see if the specified operand of the
666 /// specified instruction is a constant integer. If so, check to see if there
667 /// are any bits set in the constant that are not demanded. If so, shrink the
668 /// constant and return true.
669 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
671 assert(I && "No instruction?");
672 assert(OpNo < I->getNumOperands() && "Operand index too large");
674 // If the operand is not a constant integer, nothing to do.
675 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
676 if (!OpC) return false;
678 // If there are no bits set that aren't demanded, nothing to do.
679 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
680 if ((~Demanded & OpC->getValue()) == 0)
683 // This instruction is producing bits that are not demanded. Shrink the RHS.
684 Demanded &= OpC->getValue();
685 I->setOperand(OpNo, ConstantInt::get(OpC->getType(), Demanded));
689 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
690 // set of known zero and one bits, compute the maximum and minimum values that
691 // could have the specified known zero and known one bits, returning them in
693 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
694 const APInt& KnownOne,
695 APInt& Min, APInt& Max) {
696 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
697 KnownZero.getBitWidth() == Min.getBitWidth() &&
698 KnownZero.getBitWidth() == Max.getBitWidth() &&
699 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
700 APInt UnknownBits = ~(KnownZero|KnownOne);
702 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
703 // bit if it is unknown.
705 Max = KnownOne|UnknownBits;
707 if (UnknownBits.isNegative()) { // Sign bit is unknown
708 Min.set(Min.getBitWidth()-1);
709 Max.clear(Max.getBitWidth()-1);
713 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
714 // a set of known zero and one bits, compute the maximum and minimum values that
715 // could have the specified known zero and known one bits, returning them in
717 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
718 const APInt &KnownOne,
719 APInt &Min, APInt &Max) {
720 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
721 KnownZero.getBitWidth() == Min.getBitWidth() &&
722 KnownZero.getBitWidth() == Max.getBitWidth() &&
723 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
724 APInt UnknownBits = ~(KnownZero|KnownOne);
726 // The minimum value is when the unknown bits are all zeros.
728 // The maximum value is when the unknown bits are all ones.
729 Max = KnownOne|UnknownBits;
732 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
733 /// SimplifyDemandedBits knows about. See if the instruction has any
734 /// properties that allow us to simplify its operands.
735 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
736 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
737 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
738 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
740 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
741 KnownZero, KnownOne, 0);
742 if (V == 0) return false;
743 if (V == &Inst) return true;
744 ReplaceInstUsesWith(Inst, V);
748 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
749 /// specified instruction operand if possible, updating it in place. It returns
750 /// true if it made any change and false otherwise.
751 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
752 APInt &KnownZero, APInt &KnownOne,
754 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
755 KnownZero, KnownOne, Depth);
756 if (NewVal == 0) return false;
762 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
763 /// value based on the demanded bits. When this function is called, it is known
764 /// that only the bits set in DemandedMask of the result of V are ever used
765 /// downstream. Consequently, depending on the mask and V, it may be possible
766 /// to replace V with a constant or one of its operands. In such cases, this
767 /// function does the replacement and returns true. In all other cases, it
768 /// returns false after analyzing the expression and setting KnownOne and known
769 /// to be one in the expression. KnownZero contains all the bits that are known
770 /// to be zero in the expression. These are provided to potentially allow the
771 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
772 /// the expression. KnownOne and KnownZero always follow the invariant that
773 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
774 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
775 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
776 /// and KnownOne must all be the same.
778 /// This returns null if it did not change anything and it permits no
779 /// simplification. This returns V itself if it did some simplification of V's
780 /// operands based on the information about what bits are demanded. This returns
781 /// some other non-null value if it found out that V is equal to another value
782 /// in the context where the specified bits are demanded, but not for all users.
783 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
784 APInt &KnownZero, APInt &KnownOne,
786 assert(V != 0 && "Null pointer of Value???");
787 assert(Depth <= 6 && "Limit Search Depth");
788 uint32_t BitWidth = DemandedMask.getBitWidth();
789 const Type *VTy = V->getType();
790 assert((TD || !isa<PointerType>(VTy)) &&
791 "SimplifyDemandedBits needs to know bit widths!");
792 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
793 (!VTy->isIntOrIntVector() ||
794 VTy->getScalarSizeInBits() == BitWidth) &&
795 KnownZero.getBitWidth() == BitWidth &&
796 KnownOne.getBitWidth() == BitWidth &&
797 "Value *V, DemandedMask, KnownZero and KnownOne "
798 "must have same BitWidth");
799 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
800 // We know all of the bits for a constant!
801 KnownOne = CI->getValue() & DemandedMask;
802 KnownZero = ~KnownOne & DemandedMask;
805 if (isa<ConstantPointerNull>(V)) {
806 // We know all of the bits for a constant!
808 KnownZero = DemandedMask;
814 if (DemandedMask == 0) { // Not demanding any bits from V.
815 if (isa<UndefValue>(V))
817 return UndefValue::get(VTy);
820 if (Depth == 6) // Limit search depth.
823 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
824 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
826 Instruction *I = dyn_cast<Instruction>(V);
828 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
829 return 0; // Only analyze instructions.
832 // If there are multiple uses of this value and we aren't at the root, then
833 // we can't do any simplifications of the operands, because DemandedMask
834 // only reflects the bits demanded by *one* of the users.
835 if (Depth != 0 && !I->hasOneUse()) {
836 // Despite the fact that we can't simplify this instruction in all User's
837 // context, we can at least compute the knownzero/knownone bits, and we can
838 // do simplifications that apply to *just* the one user if we know that
839 // this instruction has a simpler value in that context.
840 if (I->getOpcode() == Instruction::And) {
841 // If either the LHS or the RHS are Zero, the result is zero.
842 ComputeMaskedBits(I->getOperand(1), DemandedMask,
843 RHSKnownZero, RHSKnownOne, Depth+1);
844 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
845 LHSKnownZero, LHSKnownOne, Depth+1);
847 // If all of the demanded bits are known 1 on one side, return the other.
848 // These bits cannot contribute to the result of the 'and' in this
850 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
851 (DemandedMask & ~LHSKnownZero))
852 return I->getOperand(0);
853 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
854 (DemandedMask & ~RHSKnownZero))
855 return I->getOperand(1);
857 // If all of the demanded bits in the inputs are known zeros, return zero.
858 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
859 return Constant::getNullValue(VTy);
861 } else if (I->getOpcode() == Instruction::Or) {
862 // We can simplify (X|Y) -> X or Y in the user's context if we know that
863 // only bits from X or Y are demanded.
865 // If either the LHS or the RHS are One, the result is One.
866 ComputeMaskedBits(I->getOperand(1), DemandedMask,
867 RHSKnownZero, RHSKnownOne, Depth+1);
868 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
869 LHSKnownZero, LHSKnownOne, Depth+1);
871 // If all of the demanded bits are known zero on one side, return the
872 // other. These bits cannot contribute to the result of the 'or' in this
874 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
875 (DemandedMask & ~LHSKnownOne))
876 return I->getOperand(0);
877 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
878 (DemandedMask & ~RHSKnownOne))
879 return I->getOperand(1);
881 // If all of the potentially set bits on one side are known to be set on
882 // the other side, just use the 'other' side.
883 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
884 (DemandedMask & (~RHSKnownZero)))
885 return I->getOperand(0);
886 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
887 (DemandedMask & (~LHSKnownZero)))
888 return I->getOperand(1);
891 // Compute the KnownZero/KnownOne bits to simplify things downstream.
892 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
896 // If this is the root being simplified, allow it to have multiple uses,
897 // just set the DemandedMask to all bits so that we can try to simplify the
898 // operands. This allows visitTruncInst (for example) to simplify the
899 // operand of a trunc without duplicating all the logic below.
900 if (Depth == 0 && !V->hasOneUse())
901 DemandedMask = APInt::getAllOnesValue(BitWidth);
903 switch (I->getOpcode()) {
905 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
907 case Instruction::And:
908 // If either the LHS or the RHS are Zero, the result is zero.
909 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
910 RHSKnownZero, RHSKnownOne, Depth+1) ||
911 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
912 LHSKnownZero, LHSKnownOne, Depth+1))
914 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
915 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
917 // If all of the demanded bits are known 1 on one side, return the other.
918 // These bits cannot contribute to the result of the 'and'.
919 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
920 (DemandedMask & ~LHSKnownZero))
921 return I->getOperand(0);
922 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
923 (DemandedMask & ~RHSKnownZero))
924 return I->getOperand(1);
926 // If all of the demanded bits in the inputs are known zeros, return zero.
927 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
928 return Constant::getNullValue(VTy);
930 // If the RHS is a constant, see if we can simplify it.
931 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
934 // Output known-1 bits are only known if set in both the LHS & RHS.
935 RHSKnownOne &= LHSKnownOne;
936 // Output known-0 are known to be clear if zero in either the LHS | RHS.
937 RHSKnownZero |= LHSKnownZero;
939 case Instruction::Or:
940 // If either the LHS or the RHS are One, the result is One.
941 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
942 RHSKnownZero, RHSKnownOne, Depth+1) ||
943 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
944 LHSKnownZero, LHSKnownOne, Depth+1))
946 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
947 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
949 // If all of the demanded bits are known zero on one side, return the other.
950 // These bits cannot contribute to the result of the 'or'.
951 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
952 (DemandedMask & ~LHSKnownOne))
953 return I->getOperand(0);
954 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
955 (DemandedMask & ~RHSKnownOne))
956 return I->getOperand(1);
958 // If all of the potentially set bits on one side are known to be set on
959 // the other side, just use the 'other' side.
960 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
961 (DemandedMask & (~RHSKnownZero)))
962 return I->getOperand(0);
963 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
964 (DemandedMask & (~LHSKnownZero)))
965 return I->getOperand(1);
967 // If the RHS is a constant, see if we can simplify it.
968 if (ShrinkDemandedConstant(I, 1, DemandedMask))
971 // Output known-0 bits are only known if clear in both the LHS & RHS.
972 RHSKnownZero &= LHSKnownZero;
973 // Output known-1 are known to be set if set in either the LHS | RHS.
974 RHSKnownOne |= LHSKnownOne;
976 case Instruction::Xor: {
977 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
978 RHSKnownZero, RHSKnownOne, Depth+1) ||
979 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
980 LHSKnownZero, LHSKnownOne, Depth+1))
982 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
983 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
985 // If all of the demanded bits are known zero on one side, return the other.
986 // These bits cannot contribute to the result of the 'xor'.
987 if ((DemandedMask & RHSKnownZero) == DemandedMask)
988 return I->getOperand(0);
989 if ((DemandedMask & LHSKnownZero) == DemandedMask)
990 return I->getOperand(1);
992 // Output known-0 bits are known if clear or set in both the LHS & RHS.
993 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
994 (RHSKnownOne & LHSKnownOne);
995 // Output known-1 are known to be set if set in only one of the LHS, RHS.
996 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
997 (RHSKnownOne & LHSKnownZero);
999 // If all of the demanded bits are known to be zero on one side or the
1000 // other, turn this into an *inclusive* or.
1001 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1002 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1004 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1006 return InsertNewInstBefore(Or, *I);
1009 // If all of the demanded bits on one side are known, and all of the set
1010 // bits on that side are also known to be set on the other side, turn this
1011 // into an AND, as we know the bits will be cleared.
1012 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1013 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1015 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1016 Constant *AndC = Constant::getIntegerValue(VTy,
1017 ~RHSKnownOne & DemandedMask);
1019 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1020 return InsertNewInstBefore(And, *I);
1024 // If the RHS is a constant, see if we can simplify it.
1025 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1026 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1029 RHSKnownZero = KnownZeroOut;
1030 RHSKnownOne = KnownOneOut;
1033 case Instruction::Select:
1034 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1035 RHSKnownZero, RHSKnownOne, Depth+1) ||
1036 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1037 LHSKnownZero, LHSKnownOne, Depth+1))
1039 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1040 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1042 // If the operands are constants, see if we can simplify them.
1043 if (ShrinkDemandedConstant(I, 1, DemandedMask) ||
1044 ShrinkDemandedConstant(I, 2, DemandedMask))
1047 // Only known if known in both the LHS and RHS.
1048 RHSKnownOne &= LHSKnownOne;
1049 RHSKnownZero &= LHSKnownZero;
1051 case Instruction::Trunc: {
1052 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1053 DemandedMask.zext(truncBf);
1054 RHSKnownZero.zext(truncBf);
1055 RHSKnownOne.zext(truncBf);
1056 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1057 RHSKnownZero, RHSKnownOne, Depth+1))
1059 DemandedMask.trunc(BitWidth);
1060 RHSKnownZero.trunc(BitWidth);
1061 RHSKnownOne.trunc(BitWidth);
1062 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1065 case Instruction::BitCast:
1066 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1067 return false; // vector->int or fp->int?
1069 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1070 if (const VectorType *SrcVTy =
1071 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1072 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1073 // Don't touch a bitcast between vectors of different element counts.
1076 // Don't touch a scalar-to-vector bitcast.
1078 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1079 // Don't touch a vector-to-scalar bitcast.
1082 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1083 RHSKnownZero, RHSKnownOne, Depth+1))
1085 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1087 case Instruction::ZExt: {
1088 // Compute the bits in the result that are not present in the input.
1089 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1091 DemandedMask.trunc(SrcBitWidth);
1092 RHSKnownZero.trunc(SrcBitWidth);
1093 RHSKnownOne.trunc(SrcBitWidth);
1094 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1095 RHSKnownZero, RHSKnownOne, Depth+1))
1097 DemandedMask.zext(BitWidth);
1098 RHSKnownZero.zext(BitWidth);
1099 RHSKnownOne.zext(BitWidth);
1100 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1101 // The top bits are known to be zero.
1102 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1105 case Instruction::SExt: {
1106 // Compute the bits in the result that are not present in the input.
1107 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1109 APInt InputDemandedBits = DemandedMask &
1110 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1112 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1113 // If any of the sign extended bits are demanded, we know that the sign
1115 if ((NewBits & DemandedMask) != 0)
1116 InputDemandedBits.set(SrcBitWidth-1);
1118 InputDemandedBits.trunc(SrcBitWidth);
1119 RHSKnownZero.trunc(SrcBitWidth);
1120 RHSKnownOne.trunc(SrcBitWidth);
1121 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1122 RHSKnownZero, RHSKnownOne, Depth+1))
1124 InputDemandedBits.zext(BitWidth);
1125 RHSKnownZero.zext(BitWidth);
1126 RHSKnownOne.zext(BitWidth);
1127 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1129 // If the sign bit of the input is known set or clear, then we know the
1130 // top bits of the result.
1132 // If the input sign bit is known zero, or if the NewBits are not demanded
1133 // convert this into a zero extension.
1134 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1135 // Convert to ZExt cast
1136 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1137 return InsertNewInstBefore(NewCast, *I);
1138 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1139 RHSKnownOne |= NewBits;
1143 case Instruction::Add: {
1144 // Figure out what the input bits are. If the top bits of the and result
1145 // are not demanded, then the add doesn't demand them from its input
1147 unsigned NLZ = DemandedMask.countLeadingZeros();
1149 // If there is a constant on the RHS, there are a variety of xformations
1151 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1152 // If null, this should be simplified elsewhere. Some of the xforms here
1153 // won't work if the RHS is zero.
1157 // If the top bit of the output is demanded, demand everything from the
1158 // input. Otherwise, we demand all the input bits except NLZ top bits.
1159 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1161 // Find information about known zero/one bits in the input.
1162 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1163 LHSKnownZero, LHSKnownOne, Depth+1))
1166 // If the RHS of the add has bits set that can't affect the input, reduce
1168 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1171 // Avoid excess work.
1172 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1175 // Turn it into OR if input bits are zero.
1176 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1178 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1180 return InsertNewInstBefore(Or, *I);
1183 // We can say something about the output known-zero and known-one bits,
1184 // depending on potential carries from the input constant and the
1185 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1186 // bits set and the RHS constant is 0x01001, then we know we have a known
1187 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1189 // To compute this, we first compute the potential carry bits. These are
1190 // the bits which may be modified. I'm not aware of a better way to do
1192 const APInt &RHSVal = RHS->getValue();
1193 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1195 // Now that we know which bits have carries, compute the known-1/0 sets.
1197 // Bits are known one if they are known zero in one operand and one in the
1198 // other, and there is no input carry.
1199 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1200 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1202 // Bits are known zero if they are known zero in both operands and there
1203 // is no input carry.
1204 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1206 // If the high-bits of this ADD are not demanded, then it does not demand
1207 // the high bits of its LHS or RHS.
1208 if (DemandedMask[BitWidth-1] == 0) {
1209 // Right fill the mask of bits for this ADD to demand the most
1210 // significant bit and all those below it.
1211 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1212 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1213 LHSKnownZero, LHSKnownOne, Depth+1) ||
1214 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1215 LHSKnownZero, LHSKnownOne, Depth+1))
1221 case Instruction::Sub:
1222 // If the high-bits of this SUB are not demanded, then it does not demand
1223 // the high bits of its LHS or RHS.
1224 if (DemandedMask[BitWidth-1] == 0) {
1225 // Right fill the mask of bits for this SUB to demand the most
1226 // significant bit and all those below it.
1227 uint32_t NLZ = DemandedMask.countLeadingZeros();
1228 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1229 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1230 LHSKnownZero, LHSKnownOne, Depth+1) ||
1231 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1232 LHSKnownZero, LHSKnownOne, Depth+1))
1235 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1236 // the known zeros and ones.
1237 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1239 case Instruction::Shl:
1240 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1241 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1242 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1243 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1244 RHSKnownZero, RHSKnownOne, Depth+1))
1246 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1247 RHSKnownZero <<= ShiftAmt;
1248 RHSKnownOne <<= ShiftAmt;
1249 // low bits known zero.
1251 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1254 case Instruction::LShr:
1255 // For a logical shift right
1256 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1257 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1259 // Unsigned shift right.
1260 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1261 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1262 RHSKnownZero, RHSKnownOne, Depth+1))
1264 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1265 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1266 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1268 // Compute the new bits that are at the top now.
1269 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1270 RHSKnownZero |= HighBits; // high bits known zero.
1274 case Instruction::AShr:
1275 // If this is an arithmetic shift right and only the low-bit is set, we can
1276 // always convert this into a logical shr, even if the shift amount is
1277 // variable. The low bit of the shift cannot be an input sign bit unless
1278 // the shift amount is >= the size of the datatype, which is undefined.
1279 if (DemandedMask == 1) {
1280 // Perform the logical shift right.
1281 Instruction *NewVal = BinaryOperator::CreateLShr(
1282 I->getOperand(0), I->getOperand(1), I->getName());
1283 return InsertNewInstBefore(NewVal, *I);
1286 // If the sign bit is the only bit demanded by this ashr, then there is no
1287 // need to do it, the shift doesn't change the high bit.
1288 if (DemandedMask.isSignBit())
1289 return I->getOperand(0);
1291 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1292 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1294 // Signed shift right.
1295 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1296 // If any of the "high bits" are demanded, we should set the sign bit as
1298 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1299 DemandedMaskIn.set(BitWidth-1);
1300 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1301 RHSKnownZero, RHSKnownOne, Depth+1))
1303 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1304 // Compute the new bits that are at the top now.
1305 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1306 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1307 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1309 // Handle the sign bits.
1310 APInt SignBit(APInt::getSignBit(BitWidth));
1311 // Adjust to where it is now in the mask.
1312 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1314 // If the input sign bit is known to be zero, or if none of the top bits
1315 // are demanded, turn this into an unsigned shift right.
1316 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1317 (HighBits & ~DemandedMask) == HighBits) {
1318 // Perform the logical shift right.
1319 Instruction *NewVal = BinaryOperator::CreateLShr(
1320 I->getOperand(0), SA, I->getName());
1321 return InsertNewInstBefore(NewVal, *I);
1322 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1323 RHSKnownOne |= HighBits;
1327 case Instruction::SRem:
1328 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1329 APInt RA = Rem->getValue().abs();
1330 if (RA.isPowerOf2()) {
1331 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1332 return I->getOperand(0);
1334 APInt LowBits = RA - 1;
1335 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1336 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1337 LHSKnownZero, LHSKnownOne, Depth+1))
1340 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1341 LHSKnownZero |= ~LowBits;
1343 KnownZero |= LHSKnownZero & DemandedMask;
1345 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1349 case Instruction::URem: {
1350 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1351 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1352 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1353 KnownZero2, KnownOne2, Depth+1) ||
1354 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1355 KnownZero2, KnownOne2, Depth+1))
1358 unsigned Leaders = KnownZero2.countLeadingOnes();
1359 Leaders = std::max(Leaders,
1360 KnownZero2.countLeadingOnes());
1361 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1364 case Instruction::Call:
1365 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1366 switch (II->getIntrinsicID()) {
1368 case Intrinsic::bswap: {
1369 // If the only bits demanded come from one byte of the bswap result,
1370 // just shift the input byte into position to eliminate the bswap.
1371 unsigned NLZ = DemandedMask.countLeadingZeros();
1372 unsigned NTZ = DemandedMask.countTrailingZeros();
1374 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1375 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1376 // have 14 leading zeros, round to 8.
1379 // If we need exactly one byte, we can do this transformation.
1380 if (BitWidth-NLZ-NTZ == 8) {
1381 unsigned ResultBit = NTZ;
1382 unsigned InputBit = BitWidth-NTZ-8;
1384 // Replace this with either a left or right shift to get the byte into
1386 Instruction *NewVal;
1387 if (InputBit > ResultBit)
1388 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1389 ConstantInt::get(I->getType(), InputBit-ResultBit));
1391 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1392 ConstantInt::get(I->getType(), ResultBit-InputBit));
1393 NewVal->takeName(I);
1394 return InsertNewInstBefore(NewVal, *I);
1397 // TODO: Could compute known zero/one bits based on the input.
1402 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1406 // If the client is only demanding bits that we know, return the known
1408 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1409 return Constant::getIntegerValue(VTy, RHSKnownOne);
1414 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1415 /// any number of elements. DemandedElts contains the set of elements that are
1416 /// actually used by the caller. This method analyzes which elements of the
1417 /// operand are undef and returns that information in UndefElts.
1419 /// If the information about demanded elements can be used to simplify the
1420 /// operation, the operation is simplified, then the resultant value is
1421 /// returned. This returns null if no change was made.
1422 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1425 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1426 APInt EltMask(APInt::getAllOnesValue(VWidth));
1427 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1429 if (isa<UndefValue>(V)) {
1430 // If the entire vector is undefined, just return this info.
1431 UndefElts = EltMask;
1433 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1434 UndefElts = EltMask;
1435 return UndefValue::get(V->getType());
1439 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1440 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1441 Constant *Undef = UndefValue::get(EltTy);
1443 std::vector<Constant*> Elts;
1444 for (unsigned i = 0; i != VWidth; ++i)
1445 if (!DemandedElts[i]) { // If not demanded, set to undef.
1446 Elts.push_back(Undef);
1448 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1449 Elts.push_back(Undef);
1451 } else { // Otherwise, defined.
1452 Elts.push_back(CP->getOperand(i));
1455 // If we changed the constant, return it.
1456 Constant *NewCP = ConstantVector::get(Elts);
1457 return NewCP != CP ? NewCP : 0;
1458 } else if (isa<ConstantAggregateZero>(V)) {
1459 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1462 // Check if this is identity. If so, return 0 since we are not simplifying
1464 if (DemandedElts == ((1ULL << VWidth) -1))
1467 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1468 Constant *Zero = Constant::getNullValue(EltTy);
1469 Constant *Undef = UndefValue::get(EltTy);
1470 std::vector<Constant*> Elts;
1471 for (unsigned i = 0; i != VWidth; ++i) {
1472 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1473 Elts.push_back(Elt);
1475 UndefElts = DemandedElts ^ EltMask;
1476 return ConstantVector::get(Elts);
1479 // Limit search depth.
1483 // If multiple users are using the root value, procede with
1484 // simplification conservatively assuming that all elements
1486 if (!V->hasOneUse()) {
1487 // Quit if we find multiple users of a non-root value though.
1488 // They'll be handled when it's their turn to be visited by
1489 // the main instcombine process.
1491 // TODO: Just compute the UndefElts information recursively.
1494 // Conservatively assume that all elements are needed.
1495 DemandedElts = EltMask;
1498 Instruction *I = dyn_cast<Instruction>(V);
1499 if (!I) return 0; // Only analyze instructions.
1501 bool MadeChange = false;
1502 APInt UndefElts2(VWidth, 0);
1504 switch (I->getOpcode()) {
1507 case Instruction::InsertElement: {
1508 // If this is a variable index, we don't know which element it overwrites.
1509 // demand exactly the same input as we produce.
1510 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1512 // Note that we can't propagate undef elt info, because we don't know
1513 // which elt is getting updated.
1514 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1515 UndefElts2, Depth+1);
1516 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1520 // If this is inserting an element that isn't demanded, remove this
1522 unsigned IdxNo = Idx->getZExtValue();
1523 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1524 return AddSoonDeadInstToWorklist(*I, 0);
1526 // Otherwise, the element inserted overwrites whatever was there, so the
1527 // input demanded set is simpler than the output set.
1528 APInt DemandedElts2 = DemandedElts;
1529 DemandedElts2.clear(IdxNo);
1530 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1531 UndefElts, Depth+1);
1532 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1534 // The inserted element is defined.
1535 UndefElts.clear(IdxNo);
1538 case Instruction::ShuffleVector: {
1539 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1540 uint64_t LHSVWidth =
1541 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1542 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1543 for (unsigned i = 0; i < VWidth; i++) {
1544 if (DemandedElts[i]) {
1545 unsigned MaskVal = Shuffle->getMaskValue(i);
1546 if (MaskVal != -1u) {
1547 assert(MaskVal < LHSVWidth * 2 &&
1548 "shufflevector mask index out of range!");
1549 if (MaskVal < LHSVWidth)
1550 LeftDemanded.set(MaskVal);
1552 RightDemanded.set(MaskVal - LHSVWidth);
1557 APInt UndefElts4(LHSVWidth, 0);
1558 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1559 UndefElts4, Depth+1);
1560 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1562 APInt UndefElts3(LHSVWidth, 0);
1563 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1564 UndefElts3, Depth+1);
1565 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1567 bool NewUndefElts = false;
1568 for (unsigned i = 0; i < VWidth; i++) {
1569 unsigned MaskVal = Shuffle->getMaskValue(i);
1570 if (MaskVal == -1u) {
1572 } else if (MaskVal < LHSVWidth) {
1573 if (UndefElts4[MaskVal]) {
1574 NewUndefElts = true;
1578 if (UndefElts3[MaskVal - LHSVWidth]) {
1579 NewUndefElts = true;
1586 // Add additional discovered undefs.
1587 std::vector<Constant*> Elts;
1588 for (unsigned i = 0; i < VWidth; ++i) {
1590 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
1592 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context),
1593 Shuffle->getMaskValue(i)));
1595 I->setOperand(2, ConstantVector::get(Elts));
1600 case Instruction::BitCast: {
1601 // Vector->vector casts only.
1602 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1604 unsigned InVWidth = VTy->getNumElements();
1605 APInt InputDemandedElts(InVWidth, 0);
1608 if (VWidth == InVWidth) {
1609 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1610 // elements as are demanded of us.
1612 InputDemandedElts = DemandedElts;
1613 } else if (VWidth > InVWidth) {
1617 // If there are more elements in the result than there are in the source,
1618 // then an input element is live if any of the corresponding output
1619 // elements are live.
1620 Ratio = VWidth/InVWidth;
1621 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1622 if (DemandedElts[OutIdx])
1623 InputDemandedElts.set(OutIdx/Ratio);
1629 // If there are more elements in the source than there are in the result,
1630 // then an input element is live if the corresponding output element is
1632 Ratio = InVWidth/VWidth;
1633 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1634 if (DemandedElts[InIdx/Ratio])
1635 InputDemandedElts.set(InIdx);
1638 // div/rem demand all inputs, because they don't want divide by zero.
1639 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1640 UndefElts2, Depth+1);
1642 I->setOperand(0, TmpV);
1646 UndefElts = UndefElts2;
1647 if (VWidth > InVWidth) {
1648 llvm_unreachable("Unimp");
1649 // If there are more elements in the result than there are in the source,
1650 // then an output element is undef if the corresponding input element is
1652 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1653 if (UndefElts2[OutIdx/Ratio])
1654 UndefElts.set(OutIdx);
1655 } else if (VWidth < InVWidth) {
1656 llvm_unreachable("Unimp");
1657 // If there are more elements in the source than there are in the result,
1658 // then a result element is undef if all of the corresponding input
1659 // elements are undef.
1660 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1661 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1662 if (!UndefElts2[InIdx]) // Not undef?
1663 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1667 case Instruction::And:
1668 case Instruction::Or:
1669 case Instruction::Xor:
1670 case Instruction::Add:
1671 case Instruction::Sub:
1672 case Instruction::Mul:
1673 // div/rem demand all inputs, because they don't want divide by zero.
1674 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1675 UndefElts, Depth+1);
1676 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1677 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1678 UndefElts2, Depth+1);
1679 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1681 // Output elements are undefined if both are undefined. Consider things
1682 // like undef&0. The result is known zero, not undef.
1683 UndefElts &= UndefElts2;
1686 case Instruction::Call: {
1687 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1689 switch (II->getIntrinsicID()) {
1692 // Binary vector operations that work column-wise. A dest element is a
1693 // function of the corresponding input elements from the two inputs.
1694 case Intrinsic::x86_sse_sub_ss:
1695 case Intrinsic::x86_sse_mul_ss:
1696 case Intrinsic::x86_sse_min_ss:
1697 case Intrinsic::x86_sse_max_ss:
1698 case Intrinsic::x86_sse2_sub_sd:
1699 case Intrinsic::x86_sse2_mul_sd:
1700 case Intrinsic::x86_sse2_min_sd:
1701 case Intrinsic::x86_sse2_max_sd:
1702 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1703 UndefElts, Depth+1);
1704 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1705 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1706 UndefElts2, Depth+1);
1707 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1709 // If only the low elt is demanded and this is a scalarizable intrinsic,
1710 // scalarize it now.
1711 if (DemandedElts == 1) {
1712 switch (II->getIntrinsicID()) {
1714 case Intrinsic::x86_sse_sub_ss:
1715 case Intrinsic::x86_sse_mul_ss:
1716 case Intrinsic::x86_sse2_sub_sd:
1717 case Intrinsic::x86_sse2_mul_sd:
1718 // TODO: Lower MIN/MAX/ABS/etc
1719 Value *LHS = II->getOperand(1);
1720 Value *RHS = II->getOperand(2);
1721 // Extract the element as scalars.
1722 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1723 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1724 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1725 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), "tmp"), *II);
1727 switch (II->getIntrinsicID()) {
1728 default: llvm_unreachable("Case stmts out of sync!");
1729 case Intrinsic::x86_sse_sub_ss:
1730 case Intrinsic::x86_sse2_sub_sd:
1731 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1732 II->getName()), *II);
1734 case Intrinsic::x86_sse_mul_ss:
1735 case Intrinsic::x86_sse2_mul_sd:
1736 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1737 II->getName()), *II);
1742 InsertElementInst::Create(
1743 UndefValue::get(II->getType()), TmpV,
1744 ConstantInt::get(Type::getInt32Ty(*Context), 0U, false), II->getName());
1745 InsertNewInstBefore(New, *II);
1746 AddSoonDeadInstToWorklist(*II, 0);
1751 // Output elements are undefined if both are undefined. Consider things
1752 // like undef&0. The result is known zero, not undef.
1753 UndefElts &= UndefElts2;
1759 return MadeChange ? I : 0;
1763 /// AssociativeOpt - Perform an optimization on an associative operator. This
1764 /// function is designed to check a chain of associative operators for a
1765 /// potential to apply a certain optimization. Since the optimization may be
1766 /// applicable if the expression was reassociated, this checks the chain, then
1767 /// reassociates the expression as necessary to expose the optimization
1768 /// opportunity. This makes use of a special Functor, which must define
1769 /// 'shouldApply' and 'apply' methods.
1771 template<typename Functor>
1772 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1773 unsigned Opcode = Root.getOpcode();
1774 Value *LHS = Root.getOperand(0);
1776 // Quick check, see if the immediate LHS matches...
1777 if (F.shouldApply(LHS))
1778 return F.apply(Root);
1780 // Otherwise, if the LHS is not of the same opcode as the root, return.
1781 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1782 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1783 // Should we apply this transform to the RHS?
1784 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1786 // If not to the RHS, check to see if we should apply to the LHS...
1787 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1788 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1792 // If the functor wants to apply the optimization to the RHS of LHSI,
1793 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1795 // Now all of the instructions are in the current basic block, go ahead
1796 // and perform the reassociation.
1797 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1799 // First move the selected RHS to the LHS of the root...
1800 Root.setOperand(0, LHSI->getOperand(1));
1802 // Make what used to be the LHS of the root be the user of the root...
1803 Value *ExtraOperand = TmpLHSI->getOperand(1);
1804 if (&Root == TmpLHSI) {
1805 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1808 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1809 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1810 BasicBlock::iterator ARI = &Root; ++ARI;
1811 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1814 // Now propagate the ExtraOperand down the chain of instructions until we
1816 while (TmpLHSI != LHSI) {
1817 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1818 // Move the instruction to immediately before the chain we are
1819 // constructing to avoid breaking dominance properties.
1820 NextLHSI->moveBefore(ARI);
1823 Value *NextOp = NextLHSI->getOperand(1);
1824 NextLHSI->setOperand(1, ExtraOperand);
1826 ExtraOperand = NextOp;
1829 // Now that the instructions are reassociated, have the functor perform
1830 // the transformation...
1831 return F.apply(Root);
1834 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1841 // AddRHS - Implements: X + X --> X << 1
1844 explicit AddRHS(Value *rhs) : RHS(rhs) {}
1845 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1846 Instruction *apply(BinaryOperator &Add) const {
1847 return BinaryOperator::CreateShl(Add.getOperand(0),
1848 ConstantInt::get(Add.getType(), 1));
1852 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1854 struct AddMaskingAnd {
1856 explicit AddMaskingAnd(Constant *c) : C2(c) {}
1857 bool shouldApply(Value *LHS) const {
1859 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1860 ConstantExpr::getAnd(C1, C2)->isNullValue();
1862 Instruction *apply(BinaryOperator &Add) const {
1863 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1869 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1871 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1872 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1875 // Figure out if the constant is the left or the right argument.
1876 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1877 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1879 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1881 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1882 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1885 Value *Op0 = SO, *Op1 = ConstOperand;
1887 std::swap(Op0, Op1);
1889 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1890 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1891 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1892 New = CmpInst::Create(CI->getOpcode(), CI->getPredicate(),
1893 Op0, Op1, SO->getName()+".cmp");
1895 llvm_unreachable("Unknown binary instruction type!");
1897 return IC->InsertNewInstBefore(New, I);
1900 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1901 // constant as the other operand, try to fold the binary operator into the
1902 // select arguments. This also works for Cast instructions, which obviously do
1903 // not have a second operand.
1904 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1906 // Don't modify shared select instructions
1907 if (!SI->hasOneUse()) return 0;
1908 Value *TV = SI->getOperand(1);
1909 Value *FV = SI->getOperand(2);
1911 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1912 // Bool selects with constant operands can be folded to logical ops.
1913 if (SI->getType() == Type::getInt1Ty(*IC->getContext())) return 0;
1915 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1916 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1918 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1925 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1926 /// node as operand #0, see if we can fold the instruction into the PHI (which
1927 /// is only possible if all operands to the PHI are constants).
1928 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1929 PHINode *PN = cast<PHINode>(I.getOperand(0));
1930 unsigned NumPHIValues = PN->getNumIncomingValues();
1931 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1933 // Check to see if all of the operands of the PHI are constants. If there is
1934 // one non-constant value, remember the BB it is. If there is more than one
1935 // or if *it* is a PHI, bail out.
1936 BasicBlock *NonConstBB = 0;
1937 for (unsigned i = 0; i != NumPHIValues; ++i)
1938 if (!isa<Constant>(PN->getIncomingValue(i))) {
1939 if (NonConstBB) return 0; // More than one non-const value.
1940 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1941 NonConstBB = PN->getIncomingBlock(i);
1943 // If the incoming non-constant value is in I's block, we have an infinite
1945 if (NonConstBB == I.getParent())
1949 // If there is exactly one non-constant value, we can insert a copy of the
1950 // operation in that block. However, if this is a critical edge, we would be
1951 // inserting the computation one some other paths (e.g. inside a loop). Only
1952 // do this if the pred block is unconditionally branching into the phi block.
1954 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1955 if (!BI || !BI->isUnconditional()) return 0;
1958 // Okay, we can do the transformation: create the new PHI node.
1959 PHINode *NewPN = PHINode::Create(I.getType(), "");
1960 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1961 InsertNewInstBefore(NewPN, *PN);
1962 NewPN->takeName(PN);
1964 // Next, add all of the operands to the PHI.
1965 if (I.getNumOperands() == 2) {
1966 Constant *C = cast<Constant>(I.getOperand(1));
1967 for (unsigned i = 0; i != NumPHIValues; ++i) {
1969 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1970 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1971 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1973 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1975 assert(PN->getIncomingBlock(i) == NonConstBB);
1976 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1977 InV = BinaryOperator::Create(BO->getOpcode(),
1978 PN->getIncomingValue(i), C, "phitmp",
1979 NonConstBB->getTerminator());
1980 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1981 InV = CmpInst::Create(CI->getOpcode(),
1983 PN->getIncomingValue(i), C, "phitmp",
1984 NonConstBB->getTerminator());
1986 llvm_unreachable("Unknown binop!");
1988 AddToWorkList(cast<Instruction>(InV));
1990 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1993 CastInst *CI = cast<CastInst>(&I);
1994 const Type *RetTy = CI->getType();
1995 for (unsigned i = 0; i != NumPHIValues; ++i) {
1997 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1998 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2000 assert(PN->getIncomingBlock(i) == NonConstBB);
2001 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2002 I.getType(), "phitmp",
2003 NonConstBB->getTerminator());
2004 AddToWorkList(cast<Instruction>(InV));
2006 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2009 return ReplaceInstUsesWith(I, NewPN);
2013 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2014 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2015 /// This basically requires proving that the add in the original type would not
2016 /// overflow to change the sign bit or have a carry out.
2017 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2018 // There are different heuristics we can use for this. Here are some simple
2021 // Add has the property that adding any two 2's complement numbers can only
2022 // have one carry bit which can change a sign. As such, if LHS and RHS each
2023 // have at least two sign bits, we know that the addition of the two values will
2024 // sign extend fine.
2025 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2029 // If one of the operands only has one non-zero bit, and if the other operand
2030 // has a known-zero bit in a more significant place than it (not including the
2031 // sign bit) the ripple may go up to and fill the zero, but won't change the
2032 // sign. For example, (X & ~4) + 1.
2040 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2041 bool Changed = SimplifyCommutative(I);
2042 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2044 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2045 // X + undef -> undef
2046 if (isa<UndefValue>(RHS))
2047 return ReplaceInstUsesWith(I, RHS);
2050 if (RHSC->isNullValue())
2051 return ReplaceInstUsesWith(I, LHS);
2053 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2054 // X + (signbit) --> X ^ signbit
2055 const APInt& Val = CI->getValue();
2056 uint32_t BitWidth = Val.getBitWidth();
2057 if (Val == APInt::getSignBit(BitWidth))
2058 return BinaryOperator::CreateXor(LHS, RHS);
2060 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2061 // (X & 254)+1 -> (X&254)|1
2062 if (SimplifyDemandedInstructionBits(I))
2065 // zext(bool) + C -> bool ? C + 1 : C
2066 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2067 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2068 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2071 if (isa<PHINode>(LHS))
2072 if (Instruction *NV = FoldOpIntoPhi(I))
2075 ConstantInt *XorRHS = 0;
2077 if (isa<ConstantInt>(RHSC) &&
2078 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2079 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2080 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2082 uint32_t Size = TySizeBits / 2;
2083 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2084 APInt CFF80Val(-C0080Val);
2086 if (TySizeBits > Size) {
2087 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2088 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2089 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2090 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2091 // This is a sign extend if the top bits are known zero.
2092 if (!MaskedValueIsZero(XorLHS,
2093 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2094 Size = 0; // Not a sign ext, but can't be any others either.
2099 C0080Val = APIntOps::lshr(C0080Val, Size);
2100 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2101 } while (Size >= 1);
2103 // FIXME: This shouldn't be necessary. When the backends can handle types
2104 // with funny bit widths then this switch statement should be removed. It
2105 // is just here to get the size of the "middle" type back up to something
2106 // that the back ends can handle.
2107 const Type *MiddleType = 0;
2110 case 32: MiddleType = Type::getInt32Ty(*Context); break;
2111 case 16: MiddleType = Type::getInt16Ty(*Context); break;
2112 case 8: MiddleType = Type::getInt8Ty(*Context); break;
2115 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2116 InsertNewInstBefore(NewTrunc, I);
2117 return new SExtInst(NewTrunc, I.getType(), I.getName());
2122 if (I.getType() == Type::getInt1Ty(*Context))
2123 return BinaryOperator::CreateXor(LHS, RHS);
2126 if (I.getType()->isInteger()) {
2127 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2130 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2131 if (RHSI->getOpcode() == Instruction::Sub)
2132 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2133 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2135 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2136 if (LHSI->getOpcode() == Instruction::Sub)
2137 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2138 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2143 // -A + -B --> -(A + B)
2144 if (Value *LHSV = dyn_castNegVal(LHS)) {
2145 if (LHS->getType()->isIntOrIntVector()) {
2146 if (Value *RHSV = dyn_castNegVal(RHS)) {
2147 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2148 InsertNewInstBefore(NewAdd, I);
2149 return BinaryOperator::CreateNeg(NewAdd);
2153 return BinaryOperator::CreateSub(RHS, LHSV);
2157 if (!isa<Constant>(RHS))
2158 if (Value *V = dyn_castNegVal(RHS))
2159 return BinaryOperator::CreateSub(LHS, V);
2163 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2164 if (X == RHS) // X*C + X --> X * (C+1)
2165 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2167 // X*C1 + X*C2 --> X * (C1+C2)
2169 if (X == dyn_castFoldableMul(RHS, C1))
2170 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2173 // X + X*C --> X * (C+1)
2174 if (dyn_castFoldableMul(RHS, C2) == LHS)
2175 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2177 // X + ~X --> -1 since ~X = -X-1
2178 if (dyn_castNotVal(LHS) == RHS ||
2179 dyn_castNotVal(RHS) == LHS)
2180 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2183 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2184 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2185 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2188 // A+B --> A|B iff A and B have no bits set in common.
2189 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2190 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2191 APInt LHSKnownOne(IT->getBitWidth(), 0);
2192 APInt LHSKnownZero(IT->getBitWidth(), 0);
2193 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2194 if (LHSKnownZero != 0) {
2195 APInt RHSKnownOne(IT->getBitWidth(), 0);
2196 APInt RHSKnownZero(IT->getBitWidth(), 0);
2197 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2199 // No bits in common -> bitwise or.
2200 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2201 return BinaryOperator::CreateOr(LHS, RHS);
2205 // W*X + Y*Z --> W * (X+Z) iff W == Y
2206 if (I.getType()->isIntOrIntVector()) {
2207 Value *W, *X, *Y, *Z;
2208 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2209 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2213 } else if (Y == X) {
2215 } else if (X == Z) {
2222 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2223 LHS->getName()), I);
2224 return BinaryOperator::CreateMul(W, NewAdd);
2229 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2231 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2232 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2234 // (X & FF00) + xx00 -> (X+xx00) & FF00
2235 if (LHS->hasOneUse() &&
2236 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2237 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2238 if (Anded == CRHS) {
2239 // See if all bits from the first bit set in the Add RHS up are included
2240 // in the mask. First, get the rightmost bit.
2241 const APInt& AddRHSV = CRHS->getValue();
2243 // Form a mask of all bits from the lowest bit added through the top.
2244 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2246 // See if the and mask includes all of these bits.
2247 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2249 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2250 // Okay, the xform is safe. Insert the new add pronto.
2251 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2252 LHS->getName()), I);
2253 return BinaryOperator::CreateAnd(NewAdd, C2);
2258 // Try to fold constant add into select arguments.
2259 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2260 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2264 // add (select X 0 (sub n A)) A --> select X A n
2266 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2269 SI = dyn_cast<SelectInst>(RHS);
2272 if (SI && SI->hasOneUse()) {
2273 Value *TV = SI->getTrueValue();
2274 Value *FV = SI->getFalseValue();
2277 // Can we fold the add into the argument of the select?
2278 // We check both true and false select arguments for a matching subtract.
2279 if (match(FV, m_Zero()) &&
2280 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2281 // Fold the add into the true select value.
2282 return SelectInst::Create(SI->getCondition(), N, A);
2283 if (match(TV, m_Zero()) &&
2284 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2285 // Fold the add into the false select value.
2286 return SelectInst::Create(SI->getCondition(), A, N);
2290 // Check for (add (sext x), y), see if we can merge this into an
2291 // integer add followed by a sext.
2292 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2293 // (add (sext x), cst) --> (sext (add x, cst'))
2294 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2296 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2297 if (LHSConv->hasOneUse() &&
2298 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2299 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2300 // Insert the new, smaller add.
2301 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2303 InsertNewInstBefore(NewAdd, I);
2304 return new SExtInst(NewAdd, I.getType());
2308 // (add (sext x), (sext y)) --> (sext (add int x, y))
2309 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2310 // Only do this if x/y have the same type, if at last one of them has a
2311 // single use (so we don't increase the number of sexts), and if the
2312 // integer add will not overflow.
2313 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2314 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2315 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2316 RHSConv->getOperand(0))) {
2317 // Insert the new integer add.
2318 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2319 RHSConv->getOperand(0),
2321 InsertNewInstBefore(NewAdd, I);
2322 return new SExtInst(NewAdd, I.getType());
2327 return Changed ? &I : 0;
2330 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2331 bool Changed = SimplifyCommutative(I);
2332 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2334 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2336 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2337 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2338 (I.getType())->getValueAPF()))
2339 return ReplaceInstUsesWith(I, LHS);
2342 if (isa<PHINode>(LHS))
2343 if (Instruction *NV = FoldOpIntoPhi(I))
2348 // -A + -B --> -(A + B)
2349 if (Value *LHSV = dyn_castFNegVal(LHS))
2350 return BinaryOperator::CreateFSub(RHS, LHSV);
2353 if (!isa<Constant>(RHS))
2354 if (Value *V = dyn_castFNegVal(RHS))
2355 return BinaryOperator::CreateFSub(LHS, V);
2357 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2358 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2359 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2360 return ReplaceInstUsesWith(I, LHS);
2362 // Check for (add double (sitofp x), y), see if we can merge this into an
2363 // integer add followed by a promotion.
2364 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2365 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2366 // ... if the constant fits in the integer value. This is useful for things
2367 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2368 // requires a constant pool load, and generally allows the add to be better
2370 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2372 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2373 if (LHSConv->hasOneUse() &&
2374 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2375 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2376 // Insert the new integer add.
2377 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2379 InsertNewInstBefore(NewAdd, I);
2380 return new SIToFPInst(NewAdd, I.getType());
2384 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2385 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2386 // Only do this if x/y have the same type, if at last one of them has a
2387 // single use (so we don't increase the number of int->fp conversions),
2388 // and if the integer add will not overflow.
2389 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2390 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2391 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2392 RHSConv->getOperand(0))) {
2393 // Insert the new integer add.
2394 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2395 RHSConv->getOperand(0),
2397 InsertNewInstBefore(NewAdd, I);
2398 return new SIToFPInst(NewAdd, I.getType());
2403 return Changed ? &I : 0;
2406 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2407 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2409 if (Op0 == Op1) // sub X, X -> 0
2410 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2412 // If this is a 'B = x-(-A)', change to B = x+A...
2413 if (Value *V = dyn_castNegVal(Op1))
2414 return BinaryOperator::CreateAdd(Op0, V);
2416 if (isa<UndefValue>(Op0))
2417 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2418 if (isa<UndefValue>(Op1))
2419 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2421 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2422 // Replace (-1 - A) with (~A)...
2423 if (C->isAllOnesValue())
2424 return BinaryOperator::CreateNot(Op1);
2426 // C - ~X == X + (1+C)
2428 if (match(Op1, m_Not(m_Value(X))))
2429 return BinaryOperator::CreateAdd(X, AddOne(C));
2431 // -(X >>u 31) -> (X >>s 31)
2432 // -(X >>s 31) -> (X >>u 31)
2434 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2435 if (SI->getOpcode() == Instruction::LShr) {
2436 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2437 // Check to see if we are shifting out everything but the sign bit.
2438 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2439 SI->getType()->getPrimitiveSizeInBits()-1) {
2440 // Ok, the transformation is safe. Insert AShr.
2441 return BinaryOperator::Create(Instruction::AShr,
2442 SI->getOperand(0), CU, SI->getName());
2446 else if (SI->getOpcode() == Instruction::AShr) {
2447 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2448 // Check to see if we are shifting out everything but the sign bit.
2449 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2450 SI->getType()->getPrimitiveSizeInBits()-1) {
2451 // Ok, the transformation is safe. Insert LShr.
2452 return BinaryOperator::CreateLShr(
2453 SI->getOperand(0), CU, SI->getName());
2460 // Try to fold constant sub into select arguments.
2461 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2462 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2465 // C - zext(bool) -> bool ? C - 1 : C
2466 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2467 if (ZI->getSrcTy() == Type::getInt1Ty(*Context))
2468 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2471 if (I.getType() == Type::getInt1Ty(*Context))
2472 return BinaryOperator::CreateXor(Op0, Op1);
2474 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2475 if (Op1I->getOpcode() == Instruction::Add) {
2476 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2477 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2479 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2480 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2482 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2483 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2484 // C1-(X+C2) --> (C1-C2)-X
2485 return BinaryOperator::CreateSub(
2486 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2490 if (Op1I->hasOneUse()) {
2491 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2492 // is not used by anyone else...
2494 if (Op1I->getOpcode() == Instruction::Sub) {
2495 // Swap the two operands of the subexpr...
2496 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2497 Op1I->setOperand(0, IIOp1);
2498 Op1I->setOperand(1, IIOp0);
2500 // Create the new top level add instruction...
2501 return BinaryOperator::CreateAdd(Op0, Op1);
2504 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2506 if (Op1I->getOpcode() == Instruction::And &&
2507 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2508 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2511 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2512 return BinaryOperator::CreateAnd(Op0, NewNot);
2515 // 0 - (X sdiv C) -> (X sdiv -C)
2516 if (Op1I->getOpcode() == Instruction::SDiv)
2517 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2519 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2520 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2521 ConstantExpr::getNeg(DivRHS));
2523 // X - X*C --> X * (1-C)
2524 ConstantInt *C2 = 0;
2525 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2527 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2529 return BinaryOperator::CreateMul(Op0, CP1);
2534 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2535 if (Op0I->getOpcode() == Instruction::Add) {
2536 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2537 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2538 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2539 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2540 } else if (Op0I->getOpcode() == Instruction::Sub) {
2541 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2542 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2548 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2549 if (X == Op1) // X*C - X --> X * (C-1)
2550 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2552 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2553 if (X == dyn_castFoldableMul(Op1, C2))
2554 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2559 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2560 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2562 // If this is a 'B = x-(-A)', change to B = x+A...
2563 if (Value *V = dyn_castFNegVal(Op1))
2564 return BinaryOperator::CreateFAdd(Op0, V);
2566 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2567 if (Op1I->getOpcode() == Instruction::FAdd) {
2568 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2569 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
2571 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2572 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
2580 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2581 /// comparison only checks the sign bit. If it only checks the sign bit, set
2582 /// TrueIfSigned if the result of the comparison is true when the input value is
2584 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2585 bool &TrueIfSigned) {
2587 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2588 TrueIfSigned = true;
2589 return RHS->isZero();
2590 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2591 TrueIfSigned = true;
2592 return RHS->isAllOnesValue();
2593 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2594 TrueIfSigned = false;
2595 return RHS->isAllOnesValue();
2596 case ICmpInst::ICMP_UGT:
2597 // True if LHS u> RHS and RHS == high-bit-mask - 1
2598 TrueIfSigned = true;
2599 return RHS->getValue() ==
2600 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2601 case ICmpInst::ICMP_UGE:
2602 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2603 TrueIfSigned = true;
2604 return RHS->getValue().isSignBit();
2610 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2611 bool Changed = SimplifyCommutative(I);
2612 Value *Op0 = I.getOperand(0);
2614 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2615 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2617 // Simplify mul instructions with a constant RHS...
2618 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2619 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2621 // ((X << C1)*C2) == (X * (C2 << C1))
2622 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2623 if (SI->getOpcode() == Instruction::Shl)
2624 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2625 return BinaryOperator::CreateMul(SI->getOperand(0),
2626 ConstantExpr::getShl(CI, ShOp));
2629 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2630 if (CI->equalsInt(1)) // X * 1 == X
2631 return ReplaceInstUsesWith(I, Op0);
2632 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2633 return BinaryOperator::CreateNeg(Op0, I.getName());
2635 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2636 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2637 return BinaryOperator::CreateShl(Op0,
2638 ConstantInt::get(Op0->getType(), Val.logBase2()));
2640 } else if (isa<VectorType>(Op1->getType())) {
2641 if (Op1->isNullValue())
2642 return ReplaceInstUsesWith(I, Op1);
2644 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2645 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2646 return BinaryOperator::CreateNeg(Op0, I.getName());
2648 // As above, vector X*splat(1.0) -> X in all defined cases.
2649 if (Constant *Splat = Op1V->getSplatValue()) {
2650 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2651 if (CI->equalsInt(1))
2652 return ReplaceInstUsesWith(I, Op0);
2657 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2658 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2659 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2660 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2661 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2663 InsertNewInstBefore(Add, I);
2664 Value *C1C2 = ConstantExpr::getMul(Op1,
2665 cast<Constant>(Op0I->getOperand(1)));
2666 return BinaryOperator::CreateAdd(Add, C1C2);
2670 // Try to fold constant mul into select arguments.
2671 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2672 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2675 if (isa<PHINode>(Op0))
2676 if (Instruction *NV = FoldOpIntoPhi(I))
2680 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2681 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2682 return BinaryOperator::CreateMul(Op0v, Op1v);
2684 // (X / Y) * Y = X - (X % Y)
2685 // (X / Y) * -Y = (X % Y) - X
2687 Value *Op1 = I.getOperand(1);
2688 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2690 (BO->getOpcode() != Instruction::UDiv &&
2691 BO->getOpcode() != Instruction::SDiv)) {
2693 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2695 Value *Neg = dyn_castNegVal(Op1);
2696 if (BO && BO->hasOneUse() &&
2697 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2698 (BO->getOpcode() == Instruction::UDiv ||
2699 BO->getOpcode() == Instruction::SDiv)) {
2700 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2702 // If the division is exact, X % Y is zero.
2703 if (SDivOperator *SDiv = dyn_cast<SDivOperator>(BO))
2704 if (SDiv->isExact()) {
2706 return ReplaceInstUsesWith(I, Op0BO);
2708 return BinaryOperator::CreateNeg(Op0BO);
2712 if (BO->getOpcode() == Instruction::UDiv)
2713 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2715 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2717 InsertNewInstBefore(Rem, I);
2721 return BinaryOperator::CreateSub(Op0BO, Rem);
2723 return BinaryOperator::CreateSub(Rem, Op0BO);
2727 if (I.getType() == Type::getInt1Ty(*Context))
2728 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2730 // If one of the operands of the multiply is a cast from a boolean value, then
2731 // we know the bool is either zero or one, so this is a 'masking' multiply.
2732 // See if we can simplify things based on how the boolean was originally
2734 CastInst *BoolCast = 0;
2735 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2736 if (CI->getOperand(0)->getType() == Type::getInt1Ty(*Context))
2739 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2740 if (CI->getOperand(0)->getType() == Type::getInt1Ty(*Context))
2743 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2744 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2745 const Type *SCOpTy = SCIOp0->getType();
2748 // If the icmp is true iff the sign bit of X is set, then convert this
2749 // multiply into a shift/and combination.
2750 if (isa<ConstantInt>(SCIOp1) &&
2751 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2753 // Shift the X value right to turn it into "all signbits".
2754 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2755 SCOpTy->getPrimitiveSizeInBits()-1);
2757 InsertNewInstBefore(
2758 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2759 BoolCast->getOperand(0)->getName()+
2762 // If the multiply type is not the same as the source type, sign extend
2763 // or truncate to the multiply type.
2764 if (I.getType() != V->getType()) {
2765 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2766 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2767 Instruction::CastOps opcode =
2768 (SrcBits == DstBits ? Instruction::BitCast :
2769 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2770 V = InsertCastBefore(opcode, V, I.getType(), I);
2773 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2774 return BinaryOperator::CreateAnd(V, OtherOp);
2779 return Changed ? &I : 0;
2782 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2783 bool Changed = SimplifyCommutative(I);
2784 Value *Op0 = I.getOperand(0);
2786 // Simplify mul instructions with a constant RHS...
2787 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2788 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2789 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2790 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2791 if (Op1F->isExactlyValue(1.0))
2792 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2793 } else if (isa<VectorType>(Op1->getType())) {
2794 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2795 // As above, vector X*splat(1.0) -> X in all defined cases.
2796 if (Constant *Splat = Op1V->getSplatValue()) {
2797 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2798 if (F->isExactlyValue(1.0))
2799 return ReplaceInstUsesWith(I, Op0);
2804 // Try to fold constant mul into select arguments.
2805 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2806 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2809 if (isa<PHINode>(Op0))
2810 if (Instruction *NV = FoldOpIntoPhi(I))
2814 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
2815 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1)))
2816 return BinaryOperator::CreateFMul(Op0v, Op1v);
2818 return Changed ? &I : 0;
2821 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2823 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2824 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2826 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2827 int NonNullOperand = -1;
2828 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2829 if (ST->isNullValue())
2831 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2832 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2833 if (ST->isNullValue())
2836 if (NonNullOperand == -1)
2839 Value *SelectCond = SI->getOperand(0);
2841 // Change the div/rem to use 'Y' instead of the select.
2842 I.setOperand(1, SI->getOperand(NonNullOperand));
2844 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2845 // problem. However, the select, or the condition of the select may have
2846 // multiple uses. Based on our knowledge that the operand must be non-zero,
2847 // propagate the known value for the select into other uses of it, and
2848 // propagate a known value of the condition into its other users.
2850 // If the select and condition only have a single use, don't bother with this,
2852 if (SI->use_empty() && SelectCond->hasOneUse())
2855 // Scan the current block backward, looking for other uses of SI.
2856 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2858 while (BBI != BBFront) {
2860 // If we found a call to a function, we can't assume it will return, so
2861 // information from below it cannot be propagated above it.
2862 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2865 // Replace uses of the select or its condition with the known values.
2866 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2869 *I = SI->getOperand(NonNullOperand);
2871 } else if (*I == SelectCond) {
2872 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
2873 ConstantInt::getFalse(*Context);
2878 // If we past the instruction, quit looking for it.
2881 if (&*BBI == SelectCond)
2884 // If we ran out of things to eliminate, break out of the loop.
2885 if (SelectCond == 0 && SI == 0)
2893 /// This function implements the transforms on div instructions that work
2894 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2895 /// used by the visitors to those instructions.
2896 /// @brief Transforms common to all three div instructions
2897 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2898 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2900 // undef / X -> 0 for integer.
2901 // undef / X -> undef for FP (the undef could be a snan).
2902 if (isa<UndefValue>(Op0)) {
2903 if (Op0->getType()->isFPOrFPVector())
2904 return ReplaceInstUsesWith(I, Op0);
2905 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2908 // X / undef -> undef
2909 if (isa<UndefValue>(Op1))
2910 return ReplaceInstUsesWith(I, Op1);
2915 /// This function implements the transforms common to both integer division
2916 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2917 /// division instructions.
2918 /// @brief Common integer divide transforms
2919 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2920 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2922 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2924 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2925 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
2926 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2927 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2930 Constant *CI = ConstantInt::get(I.getType(), 1);
2931 return ReplaceInstUsesWith(I, CI);
2934 if (Instruction *Common = commonDivTransforms(I))
2937 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2938 // This does not apply for fdiv.
2939 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2942 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2944 if (RHS->equalsInt(1))
2945 return ReplaceInstUsesWith(I, Op0);
2947 // (X / C1) / C2 -> X / (C1*C2)
2948 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2949 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2950 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2951 if (MultiplyOverflows(RHS, LHSRHS,
2952 I.getOpcode()==Instruction::SDiv))
2953 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2955 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2956 ConstantExpr::getMul(RHS, LHSRHS));
2959 if (!RHS->isZero()) { // avoid X udiv 0
2960 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2961 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2963 if (isa<PHINode>(Op0))
2964 if (Instruction *NV = FoldOpIntoPhi(I))
2969 // 0 / X == 0, we don't need to preserve faults!
2970 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2971 if (LHS->equalsInt(0))
2972 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2974 // It can't be division by zero, hence it must be division by one.
2975 if (I.getType() == Type::getInt1Ty(*Context))
2976 return ReplaceInstUsesWith(I, Op0);
2978 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2979 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2982 return ReplaceInstUsesWith(I, Op0);
2988 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2989 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2991 // Handle the integer div common cases
2992 if (Instruction *Common = commonIDivTransforms(I))
2995 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2996 // X udiv C^2 -> X >> C
2997 // Check to see if this is an unsigned division with an exact power of 2,
2998 // if so, convert to a right shift.
2999 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3000 return BinaryOperator::CreateLShr(Op0,
3001 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3003 // X udiv C, where C >= signbit
3004 if (C->getValue().isNegative()) {
3005 Value *IC = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_ULT, Op0, C),
3007 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3008 ConstantInt::get(I.getType(), 1));
3012 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3013 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3014 if (RHSI->getOpcode() == Instruction::Shl &&
3015 isa<ConstantInt>(RHSI->getOperand(0))) {
3016 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3017 if (C1.isPowerOf2()) {
3018 Value *N = RHSI->getOperand(1);
3019 const Type *NTy = N->getType();
3020 if (uint32_t C2 = C1.logBase2()) {
3021 Constant *C2V = ConstantInt::get(NTy, C2);
3022 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3024 return BinaryOperator::CreateLShr(Op0, N);
3029 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3030 // where C1&C2 are powers of two.
3031 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3032 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3033 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3034 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3035 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3036 // Compute the shift amounts
3037 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3038 // Construct the "on true" case of the select
3039 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3040 Instruction *TSI = BinaryOperator::CreateLShr(
3041 Op0, TC, SI->getName()+".t");
3042 TSI = InsertNewInstBefore(TSI, I);
3044 // Construct the "on false" case of the select
3045 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3046 Instruction *FSI = BinaryOperator::CreateLShr(
3047 Op0, FC, SI->getName()+".f");
3048 FSI = InsertNewInstBefore(FSI, I);
3050 // construct the select instruction and return it.
3051 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3057 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3058 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3060 // Handle the integer div common cases
3061 if (Instruction *Common = commonIDivTransforms(I))
3064 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3066 if (RHS->isAllOnesValue())
3067 return BinaryOperator::CreateNeg(Op0);
3069 // sdiv X, C --> ashr X, log2(C)
3070 if (cast<SDivOperator>(&I)->isExact() &&
3071 RHS->getValue().isNonNegative() &&
3072 RHS->getValue().isPowerOf2()) {
3073 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3074 RHS->getValue().exactLogBase2());
3075 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3078 // -X/C --> X/-C provided the negation doesn't overflow.
3079 if (SubOperator *Sub = dyn_cast<SubOperator>(Op0))
3080 if (isa<Constant>(Sub->getOperand(0)) &&
3081 cast<Constant>(Sub->getOperand(0))->isNullValue() &&
3082 Sub->hasNoSignedWrap())
3083 return BinaryOperator::CreateSDiv(Sub->getOperand(1),
3084 ConstantExpr::getNeg(RHS));
3087 // If the sign bits of both operands are zero (i.e. we can prove they are
3088 // unsigned inputs), turn this into a udiv.
3089 if (I.getType()->isInteger()) {
3090 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3091 if (MaskedValueIsZero(Op0, Mask)) {
3092 if (MaskedValueIsZero(Op1, Mask)) {
3093 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3094 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3096 ConstantInt *ShiftedInt;
3097 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3098 ShiftedInt->getValue().isPowerOf2()) {
3099 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3100 // Safe because the only negative value (1 << Y) can take on is
3101 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3102 // the sign bit set.
3103 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3111 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3112 return commonDivTransforms(I);
3115 /// This function implements the transforms on rem instructions that work
3116 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3117 /// is used by the visitors to those instructions.
3118 /// @brief Transforms common to all three rem instructions
3119 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3120 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3122 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3123 if (I.getType()->isFPOrFPVector())
3124 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3125 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3127 if (isa<UndefValue>(Op1))
3128 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3130 // Handle cases involving: rem X, (select Cond, Y, Z)
3131 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3137 /// This function implements the transforms common to both integer remainder
3138 /// instructions (urem and srem). It is called by the visitors to those integer
3139 /// remainder instructions.
3140 /// @brief Common integer remainder transforms
3141 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3142 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3144 if (Instruction *common = commonRemTransforms(I))
3147 // 0 % X == 0 for integer, we don't need to preserve faults!
3148 if (Constant *LHS = dyn_cast<Constant>(Op0))
3149 if (LHS->isNullValue())
3150 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3152 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3153 // X % 0 == undef, we don't need to preserve faults!
3154 if (RHS->equalsInt(0))
3155 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3157 if (RHS->equalsInt(1)) // X % 1 == 0
3158 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3160 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3161 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3162 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3164 } else if (isa<PHINode>(Op0I)) {
3165 if (Instruction *NV = FoldOpIntoPhi(I))
3169 // See if we can fold away this rem instruction.
3170 if (SimplifyDemandedInstructionBits(I))
3178 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3179 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3181 if (Instruction *common = commonIRemTransforms(I))
3184 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3185 // X urem C^2 -> X and C
3186 // Check to see if this is an unsigned remainder with an exact power of 2,
3187 // if so, convert to a bitwise and.
3188 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3189 if (C->getValue().isPowerOf2())
3190 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3193 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3194 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3195 if (RHSI->getOpcode() == Instruction::Shl &&
3196 isa<ConstantInt>(RHSI->getOperand(0))) {
3197 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3198 Constant *N1 = Constant::getAllOnesValue(I.getType());
3199 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3201 return BinaryOperator::CreateAnd(Op0, Add);
3206 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3207 // where C1&C2 are powers of two.
3208 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3209 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3210 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3211 // STO == 0 and SFO == 0 handled above.
3212 if ((STO->getValue().isPowerOf2()) &&
3213 (SFO->getValue().isPowerOf2())) {
3214 Value *TrueAnd = InsertNewInstBefore(
3215 BinaryOperator::CreateAnd(Op0, SubOne(STO),
3216 SI->getName()+".t"), I);
3217 Value *FalseAnd = InsertNewInstBefore(
3218 BinaryOperator::CreateAnd(Op0, SubOne(SFO),
3219 SI->getName()+".f"), I);
3220 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3228 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3229 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3231 // Handle the integer rem common cases
3232 if (Instruction *common = commonIRemTransforms(I))
3235 if (Value *RHSNeg = dyn_castNegVal(Op1))
3236 if (!isa<Constant>(RHSNeg) ||
3237 (isa<ConstantInt>(RHSNeg) &&
3238 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3240 AddUsesToWorkList(I);
3241 I.setOperand(1, RHSNeg);
3245 // If the sign bits of both operands are zero (i.e. we can prove they are
3246 // unsigned inputs), turn this into a urem.
3247 if (I.getType()->isInteger()) {
3248 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3249 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3250 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3251 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3255 // If it's a constant vector, flip any negative values positive.
3256 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3257 unsigned VWidth = RHSV->getNumOperands();
3259 bool hasNegative = false;
3260 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3261 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3262 if (RHS->getValue().isNegative())
3266 std::vector<Constant *> Elts(VWidth);
3267 for (unsigned i = 0; i != VWidth; ++i) {
3268 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3269 if (RHS->getValue().isNegative())
3270 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3276 Constant *NewRHSV = ConstantVector::get(Elts);
3277 if (NewRHSV != RHSV) {
3278 AddUsesToWorkList(I);
3279 I.setOperand(1, NewRHSV);
3288 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3289 return commonRemTransforms(I);
3292 // isOneBitSet - Return true if there is exactly one bit set in the specified
3294 static bool isOneBitSet(const ConstantInt *CI) {
3295 return CI->getValue().isPowerOf2();
3298 // isHighOnes - Return true if the constant is of the form 1+0+.
3299 // This is the same as lowones(~X).
3300 static bool isHighOnes(const ConstantInt *CI) {
3301 return (~CI->getValue() + 1).isPowerOf2();
3304 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3305 /// are carefully arranged to allow folding of expressions such as:
3307 /// (A < B) | (A > B) --> (A != B)
3309 /// Note that this is only valid if the first and second predicates have the
3310 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3312 /// Three bits are used to represent the condition, as follows:
3317 /// <=> Value Definition
3318 /// 000 0 Always false
3325 /// 111 7 Always true
3327 static unsigned getICmpCode(const ICmpInst *ICI) {
3328 switch (ICI->getPredicate()) {
3330 case ICmpInst::ICMP_UGT: return 1; // 001
3331 case ICmpInst::ICMP_SGT: return 1; // 001
3332 case ICmpInst::ICMP_EQ: return 2; // 010
3333 case ICmpInst::ICMP_UGE: return 3; // 011
3334 case ICmpInst::ICMP_SGE: return 3; // 011
3335 case ICmpInst::ICMP_ULT: return 4; // 100
3336 case ICmpInst::ICMP_SLT: return 4; // 100
3337 case ICmpInst::ICMP_NE: return 5; // 101
3338 case ICmpInst::ICMP_ULE: return 6; // 110
3339 case ICmpInst::ICMP_SLE: return 6; // 110
3342 llvm_unreachable("Invalid ICmp predicate!");
3347 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3348 /// predicate into a three bit mask. It also returns whether it is an ordered
3349 /// predicate by reference.
3350 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3353 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3354 case FCmpInst::FCMP_UNO: return 0; // 000
3355 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3356 case FCmpInst::FCMP_UGT: return 1; // 001
3357 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3358 case FCmpInst::FCMP_UEQ: return 2; // 010
3359 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3360 case FCmpInst::FCMP_UGE: return 3; // 011
3361 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3362 case FCmpInst::FCMP_ULT: return 4; // 100
3363 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3364 case FCmpInst::FCMP_UNE: return 5; // 101
3365 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3366 case FCmpInst::FCMP_ULE: return 6; // 110
3369 // Not expecting FCMP_FALSE and FCMP_TRUE;
3370 llvm_unreachable("Unexpected FCmp predicate!");
3375 /// getICmpValue - This is the complement of getICmpCode, which turns an
3376 /// opcode and two operands into either a constant true or false, or a brand
3377 /// new ICmp instruction. The sign is passed in to determine which kind
3378 /// of predicate to use in the new icmp instruction.
3379 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3380 LLVMContext *Context) {
3382 default: llvm_unreachable("Illegal ICmp code!");
3383 case 0: return ConstantInt::getFalse(*Context);
3386 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3388 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3389 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3392 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3394 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3397 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3399 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3400 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3403 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3405 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3406 case 7: return ConstantInt::getTrue(*Context);
3410 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3411 /// opcode and two operands into either a FCmp instruction. isordered is passed
3412 /// in to determine which kind of predicate to use in the new fcmp instruction.
3413 static Value *getFCmpValue(bool isordered, unsigned code,
3414 Value *LHS, Value *RHS, LLVMContext *Context) {
3416 default: llvm_unreachable("Illegal FCmp code!");
3419 return new FCmpInst(FCmpInst::FCMP_ORD, LHS, RHS);
3421 return new FCmpInst(FCmpInst::FCMP_UNO, LHS, RHS);
3424 return new FCmpInst(FCmpInst::FCMP_OGT, LHS, RHS);
3426 return new FCmpInst(FCmpInst::FCMP_UGT, LHS, RHS);
3429 return new FCmpInst(FCmpInst::FCMP_OEQ, LHS, RHS);
3431 return new FCmpInst(FCmpInst::FCMP_UEQ, LHS, RHS);
3434 return new FCmpInst(FCmpInst::FCMP_OGE, LHS, RHS);
3436 return new FCmpInst(FCmpInst::FCMP_UGE, LHS, RHS);
3439 return new FCmpInst(FCmpInst::FCMP_OLT, LHS, RHS);
3441 return new FCmpInst(FCmpInst::FCMP_ULT, LHS, RHS);
3444 return new FCmpInst(FCmpInst::FCMP_ONE, LHS, RHS);
3446 return new FCmpInst(FCmpInst::FCMP_UNE, LHS, RHS);
3449 return new FCmpInst(FCmpInst::FCMP_OLE, LHS, RHS);
3451 return new FCmpInst(FCmpInst::FCMP_ULE, LHS, RHS);
3452 case 7: return ConstantInt::getTrue(*Context);
3456 /// PredicatesFoldable - Return true if both predicates match sign or if at
3457 /// least one of them is an equality comparison (which is signless).
3458 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3459 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3460 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3461 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3465 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3466 struct FoldICmpLogical {
3469 ICmpInst::Predicate pred;
3470 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3471 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3472 pred(ICI->getPredicate()) {}
3473 bool shouldApply(Value *V) const {
3474 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3475 if (PredicatesFoldable(pred, ICI->getPredicate()))
3476 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3477 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3480 Instruction *apply(Instruction &Log) const {
3481 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3482 if (ICI->getOperand(0) != LHS) {
3483 assert(ICI->getOperand(1) == LHS);
3484 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3487 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3488 unsigned LHSCode = getICmpCode(ICI);
3489 unsigned RHSCode = getICmpCode(RHSICI);
3491 switch (Log.getOpcode()) {
3492 case Instruction::And: Code = LHSCode & RHSCode; break;
3493 case Instruction::Or: Code = LHSCode | RHSCode; break;
3494 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3495 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3498 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3499 ICmpInst::isSignedPredicate(ICI->getPredicate());
3501 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3502 if (Instruction *I = dyn_cast<Instruction>(RV))
3504 // Otherwise, it's a constant boolean value...
3505 return IC.ReplaceInstUsesWith(Log, RV);
3508 } // end anonymous namespace
3510 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3511 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3512 // guaranteed to be a binary operator.
3513 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3515 ConstantInt *AndRHS,
3516 BinaryOperator &TheAnd) {
3517 Value *X = Op->getOperand(0);
3518 Constant *Together = 0;
3520 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3522 switch (Op->getOpcode()) {
3523 case Instruction::Xor:
3524 if (Op->hasOneUse()) {
3525 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3526 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3527 InsertNewInstBefore(And, TheAnd);
3529 return BinaryOperator::CreateXor(And, Together);
3532 case Instruction::Or:
3533 if (Together == AndRHS) // (X | C) & C --> C
3534 return ReplaceInstUsesWith(TheAnd, AndRHS);
3536 if (Op->hasOneUse() && Together != OpRHS) {
3537 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3538 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3539 InsertNewInstBefore(Or, TheAnd);
3541 return BinaryOperator::CreateAnd(Or, AndRHS);
3544 case Instruction::Add:
3545 if (Op->hasOneUse()) {
3546 // Adding a one to a single bit bit-field should be turned into an XOR
3547 // of the bit. First thing to check is to see if this AND is with a
3548 // single bit constant.
3549 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3551 // If there is only one bit set...
3552 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3553 // Ok, at this point, we know that we are masking the result of the
3554 // ADD down to exactly one bit. If the constant we are adding has
3555 // no bits set below this bit, then we can eliminate the ADD.
3556 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3558 // Check to see if any bits below the one bit set in AndRHSV are set.
3559 if ((AddRHS & (AndRHSV-1)) == 0) {
3560 // If not, the only thing that can effect the output of the AND is
3561 // the bit specified by AndRHSV. If that bit is set, the effect of
3562 // the XOR is to toggle the bit. If it is clear, then the ADD has
3564 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3565 TheAnd.setOperand(0, X);
3568 // Pull the XOR out of the AND.
3569 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3570 InsertNewInstBefore(NewAnd, TheAnd);
3571 NewAnd->takeName(Op);
3572 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3579 case Instruction::Shl: {
3580 // We know that the AND will not produce any of the bits shifted in, so if
3581 // the anded constant includes them, clear them now!
3583 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3584 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3585 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3586 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3588 if (CI->getValue() == ShlMask) {
3589 // Masking out bits that the shift already masks
3590 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3591 } else if (CI != AndRHS) { // Reducing bits set in and.
3592 TheAnd.setOperand(1, CI);
3597 case Instruction::LShr:
3599 // We know that the AND will not produce any of the bits shifted in, so if
3600 // the anded constant includes them, clear them now! This only applies to
3601 // unsigned shifts, because a signed shr may bring in set bits!
3603 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3604 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3605 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3606 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3608 if (CI->getValue() == ShrMask) {
3609 // Masking out bits that the shift already masks.
3610 return ReplaceInstUsesWith(TheAnd, Op);
3611 } else if (CI != AndRHS) {
3612 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3617 case Instruction::AShr:
3619 // See if this is shifting in some sign extension, then masking it out
3621 if (Op->hasOneUse()) {
3622 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3623 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3624 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3625 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3626 if (C == AndRHS) { // Masking out bits shifted in.
3627 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3628 // Make the argument unsigned.
3629 Value *ShVal = Op->getOperand(0);
3630 ShVal = InsertNewInstBefore(
3631 BinaryOperator::CreateLShr(ShVal, OpRHS,
3632 Op->getName()), TheAnd);
3633 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3642 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3643 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3644 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3645 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3646 /// insert new instructions.
3647 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3648 bool isSigned, bool Inside,
3650 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3651 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3652 "Lo is not <= Hi in range emission code!");
3655 if (Lo == Hi) // Trivially false.
3656 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3658 // V >= Min && V < Hi --> V < Hi
3659 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3660 ICmpInst::Predicate pred = (isSigned ?
3661 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3662 return new ICmpInst(pred, V, Hi);
3665 // Emit V-Lo <u Hi-Lo
3666 Constant *NegLo = ConstantExpr::getNeg(Lo);
3667 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3668 InsertNewInstBefore(Add, IB);
3669 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3670 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3673 if (Lo == Hi) // Trivially true.
3674 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3676 // V < Min || V >= Hi -> V > Hi-1
3677 Hi = SubOne(cast<ConstantInt>(Hi));
3678 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3679 ICmpInst::Predicate pred = (isSigned ?
3680 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3681 return new ICmpInst(pred, V, Hi);
3684 // Emit V-Lo >u Hi-1-Lo
3685 // Note that Hi has already had one subtracted from it, above.
3686 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3687 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3688 InsertNewInstBefore(Add, IB);
3689 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3690 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3693 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3694 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3695 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3696 // not, since all 1s are not contiguous.
3697 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3698 const APInt& V = Val->getValue();
3699 uint32_t BitWidth = Val->getType()->getBitWidth();
3700 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3702 // look for the first zero bit after the run of ones
3703 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3704 // look for the first non-zero bit
3705 ME = V.getActiveBits();
3709 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3710 /// where isSub determines whether the operator is a sub. If we can fold one of
3711 /// the following xforms:
3713 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3714 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3715 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3717 /// return (A +/- B).
3719 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3720 ConstantInt *Mask, bool isSub,
3722 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3723 if (!LHSI || LHSI->getNumOperands() != 2 ||
3724 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3726 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3728 switch (LHSI->getOpcode()) {
3730 case Instruction::And:
3731 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3732 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3733 if ((Mask->getValue().countLeadingZeros() +
3734 Mask->getValue().countPopulation()) ==
3735 Mask->getValue().getBitWidth())
3738 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3739 // part, we don't need any explicit masks to take them out of A. If that
3740 // is all N is, ignore it.
3741 uint32_t MB = 0, ME = 0;
3742 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3743 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3744 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3745 if (MaskedValueIsZero(RHS, Mask))
3750 case Instruction::Or:
3751 case Instruction::Xor:
3752 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3753 if ((Mask->getValue().countLeadingZeros() +
3754 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3755 && ConstantExpr::getAnd(N, Mask)->isNullValue())
3762 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3764 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3765 return InsertNewInstBefore(New, I);
3768 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3769 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3770 ICmpInst *LHS, ICmpInst *RHS) {
3772 ConstantInt *LHSCst, *RHSCst;
3773 ICmpInst::Predicate LHSCC, RHSCC;
3775 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3776 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3777 m_ConstantInt(LHSCst))) ||
3778 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3779 m_ConstantInt(RHSCst))))
3782 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3783 // where C is a power of 2
3784 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3785 LHSCst->getValue().isPowerOf2()) {
3786 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3787 InsertNewInstBefore(NewOr, I);
3788 return new ICmpInst(LHSCC, NewOr, LHSCst);
3791 // From here on, we only handle:
3792 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3793 if (Val != Val2) return 0;
3795 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3796 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3797 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3798 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3799 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3802 // We can't fold (ugt x, C) & (sgt x, C2).
3803 if (!PredicatesFoldable(LHSCC, RHSCC))
3806 // Ensure that the larger constant is on the RHS.
3808 if (ICmpInst::isSignedPredicate(LHSCC) ||
3809 (ICmpInst::isEquality(LHSCC) &&
3810 ICmpInst::isSignedPredicate(RHSCC)))
3811 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3813 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3816 std::swap(LHS, RHS);
3817 std::swap(LHSCst, RHSCst);
3818 std::swap(LHSCC, RHSCC);
3821 // At this point, we know we have have two icmp instructions
3822 // comparing a value against two constants and and'ing the result
3823 // together. Because of the above check, we know that we only have
3824 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3825 // (from the FoldICmpLogical check above), that the two constants
3826 // are not equal and that the larger constant is on the RHS
3827 assert(LHSCst != RHSCst && "Compares not folded above?");
3830 default: llvm_unreachable("Unknown integer condition code!");
3831 case ICmpInst::ICMP_EQ:
3833 default: llvm_unreachable("Unknown integer condition code!");
3834 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3835 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3836 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3837 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3838 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3839 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3840 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3841 return ReplaceInstUsesWith(I, LHS);
3843 case ICmpInst::ICMP_NE:
3845 default: llvm_unreachable("Unknown integer condition code!");
3846 case ICmpInst::ICMP_ULT:
3847 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3848 return new ICmpInst(ICmpInst::ICMP_ULT, Val, LHSCst);
3849 break; // (X != 13 & X u< 15) -> no change
3850 case ICmpInst::ICMP_SLT:
3851 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3852 return new ICmpInst(ICmpInst::ICMP_SLT, Val, LHSCst);
3853 break; // (X != 13 & X s< 15) -> no change
3854 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3855 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3856 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3857 return ReplaceInstUsesWith(I, RHS);
3858 case ICmpInst::ICMP_NE:
3859 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3860 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3861 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3862 Val->getName()+".off");
3863 InsertNewInstBefore(Add, I);
3864 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3865 ConstantInt::get(Add->getType(), 1));
3867 break; // (X != 13 & X != 15) -> no change
3870 case ICmpInst::ICMP_ULT:
3872 default: llvm_unreachable("Unknown integer condition code!");
3873 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3874 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3875 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3876 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3878 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3879 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3880 return ReplaceInstUsesWith(I, LHS);
3881 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3885 case ICmpInst::ICMP_SLT:
3887 default: llvm_unreachable("Unknown integer condition code!");
3888 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3889 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3890 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3891 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3893 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3894 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3895 return ReplaceInstUsesWith(I, LHS);
3896 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3900 case ICmpInst::ICMP_UGT:
3902 default: llvm_unreachable("Unknown integer condition code!");
3903 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3904 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3905 return ReplaceInstUsesWith(I, RHS);
3906 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3908 case ICmpInst::ICMP_NE:
3909 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3910 return new ICmpInst(LHSCC, Val, RHSCst);
3911 break; // (X u> 13 & X != 15) -> no change
3912 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3913 return InsertRangeTest(Val, AddOne(LHSCst),
3914 RHSCst, false, true, I);
3915 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3919 case ICmpInst::ICMP_SGT:
3921 default: llvm_unreachable("Unknown integer condition code!");
3922 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3923 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3924 return ReplaceInstUsesWith(I, RHS);
3925 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3927 case ICmpInst::ICMP_NE:
3928 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3929 return new ICmpInst(LHSCC, Val, RHSCst);
3930 break; // (X s> 13 & X != 15) -> no change
3931 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3932 return InsertRangeTest(Val, AddOne(LHSCst),
3933 RHSCst, true, true, I);
3934 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3943 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
3946 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3947 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3948 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3949 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3950 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3951 // If either of the constants are nans, then the whole thing returns
3953 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3954 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3955 return new FCmpInst(FCmpInst::FCMP_ORD,
3956 LHS->getOperand(0), RHS->getOperand(0));
3959 // Handle vector zeros. This occurs because the canonical form of
3960 // "fcmp ord x,x" is "fcmp ord x, 0".
3961 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
3962 isa<ConstantAggregateZero>(RHS->getOperand(1)))
3963 return new FCmpInst(FCmpInst::FCMP_ORD,
3964 LHS->getOperand(0), RHS->getOperand(0));
3968 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
3969 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
3970 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
3973 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
3974 // Swap RHS operands to match LHS.
3975 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
3976 std::swap(Op1LHS, Op1RHS);
3979 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
3980 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
3982 return new FCmpInst((FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
3984 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
3985 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3986 if (Op0CC == FCmpInst::FCMP_TRUE)
3987 return ReplaceInstUsesWith(I, RHS);
3988 if (Op1CC == FCmpInst::FCMP_TRUE)
3989 return ReplaceInstUsesWith(I, LHS);
3993 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
3994 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
3996 std::swap(LHS, RHS);
3997 std::swap(Op0Pred, Op1Pred);
3998 std::swap(Op0Ordered, Op1Ordered);
4001 // uno && ueq -> uno && (uno || eq) -> ueq
4002 // ord && olt -> ord && (ord && lt) -> olt
4003 if (Op0Ordered == Op1Ordered)
4004 return ReplaceInstUsesWith(I, RHS);
4006 // uno && oeq -> uno && (ord && eq) -> false
4007 // uno && ord -> false
4009 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4010 // ord && ueq -> ord && (uno || eq) -> oeq
4011 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4012 Op0LHS, Op0RHS, Context));
4020 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4021 bool Changed = SimplifyCommutative(I);
4022 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4024 if (isa<UndefValue>(Op1)) // X & undef -> 0
4025 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4029 return ReplaceInstUsesWith(I, Op1);
4031 // See if we can simplify any instructions used by the instruction whose sole
4032 // purpose is to compute bits we don't care about.
4033 if (SimplifyDemandedInstructionBits(I))
4035 if (isa<VectorType>(I.getType())) {
4036 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4037 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4038 return ReplaceInstUsesWith(I, I.getOperand(0));
4039 } else if (isa<ConstantAggregateZero>(Op1)) {
4040 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4044 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4045 const APInt& AndRHSMask = AndRHS->getValue();
4046 APInt NotAndRHS(~AndRHSMask);
4048 // Optimize a variety of ((val OP C1) & C2) combinations...
4049 if (isa<BinaryOperator>(Op0)) {
4050 Instruction *Op0I = cast<Instruction>(Op0);
4051 Value *Op0LHS = Op0I->getOperand(0);
4052 Value *Op0RHS = Op0I->getOperand(1);
4053 switch (Op0I->getOpcode()) {
4054 case Instruction::Xor:
4055 case Instruction::Or:
4056 // If the mask is only needed on one incoming arm, push it up.
4057 if (Op0I->hasOneUse()) {
4058 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4059 // Not masking anything out for the LHS, move to RHS.
4060 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4061 Op0RHS->getName()+".masked");
4062 InsertNewInstBefore(NewRHS, I);
4063 return BinaryOperator::Create(
4064 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4066 if (!isa<Constant>(Op0RHS) &&
4067 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4068 // Not masking anything out for the RHS, move to LHS.
4069 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4070 Op0LHS->getName()+".masked");
4071 InsertNewInstBefore(NewLHS, I);
4072 return BinaryOperator::Create(
4073 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4078 case Instruction::Add:
4079 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4080 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4081 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4082 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4083 return BinaryOperator::CreateAnd(V, AndRHS);
4084 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4085 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4088 case Instruction::Sub:
4089 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4090 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4091 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4092 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4093 return BinaryOperator::CreateAnd(V, AndRHS);
4095 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4096 // has 1's for all bits that the subtraction with A might affect.
4097 if (Op0I->hasOneUse()) {
4098 uint32_t BitWidth = AndRHSMask.getBitWidth();
4099 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4100 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4102 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4103 if (!(A && A->isZero()) && // avoid infinite recursion.
4104 MaskedValueIsZero(Op0LHS, Mask)) {
4105 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
4106 InsertNewInstBefore(NewNeg, I);
4107 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4112 case Instruction::Shl:
4113 case Instruction::LShr:
4114 // (1 << x) & 1 --> zext(x == 0)
4115 // (1 >> x) & 1 --> zext(x == 0)
4116 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4117 Instruction *NewICmp = new ICmpInst(ICmpInst::ICMP_EQ,
4118 Op0RHS, Constant::getNullValue(I.getType()));
4119 InsertNewInstBefore(NewICmp, I);
4120 return new ZExtInst(NewICmp, I.getType());
4125 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4126 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4128 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4129 // If this is an integer truncation or change from signed-to-unsigned, and
4130 // if the source is an and/or with immediate, transform it. This
4131 // frequently occurs for bitfield accesses.
4132 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4133 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4134 CastOp->getNumOperands() == 2)
4135 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4136 if (CastOp->getOpcode() == Instruction::And) {
4137 // Change: and (cast (and X, C1) to T), C2
4138 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4139 // This will fold the two constants together, which may allow
4140 // other simplifications.
4141 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4142 CastOp->getOperand(0), I.getType(),
4143 CastOp->getName()+".shrunk");
4144 NewCast = InsertNewInstBefore(NewCast, I);
4145 // trunc_or_bitcast(C1)&C2
4147 ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4148 C3 = ConstantExpr::getAnd(C3, AndRHS);
4149 return BinaryOperator::CreateAnd(NewCast, C3);
4150 } else if (CastOp->getOpcode() == Instruction::Or) {
4151 // Change: and (cast (or X, C1) to T), C2
4152 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4154 ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4155 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4157 return ReplaceInstUsesWith(I, AndRHS);
4163 // Try to fold constant and into select arguments.
4164 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4165 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4167 if (isa<PHINode>(Op0))
4168 if (Instruction *NV = FoldOpIntoPhi(I))
4172 Value *Op0NotVal = dyn_castNotVal(Op0);
4173 Value *Op1NotVal = dyn_castNotVal(Op1);
4175 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4176 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4178 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4179 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4180 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4181 I.getName()+".demorgan");
4182 InsertNewInstBefore(Or, I);
4183 return BinaryOperator::CreateNot(Or);
4187 Value *A = 0, *B = 0, *C = 0, *D = 0;
4188 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4189 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4190 return ReplaceInstUsesWith(I, Op1);
4192 // (A|B) & ~(A&B) -> A^B
4193 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4194 if ((A == C && B == D) || (A == D && B == C))
4195 return BinaryOperator::CreateXor(A, B);
4199 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4200 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4201 return ReplaceInstUsesWith(I, Op0);
4203 // ~(A&B) & (A|B) -> A^B
4204 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4205 if ((A == C && B == D) || (A == D && B == C))
4206 return BinaryOperator::CreateXor(A, B);
4210 if (Op0->hasOneUse() &&
4211 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4212 if (A == Op1) { // (A^B)&A -> A&(A^B)
4213 I.swapOperands(); // Simplify below
4214 std::swap(Op0, Op1);
4215 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4216 cast<BinaryOperator>(Op0)->swapOperands();
4217 I.swapOperands(); // Simplify below
4218 std::swap(Op0, Op1);
4222 if (Op1->hasOneUse() &&
4223 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4224 if (B == Op0) { // B&(A^B) -> B&(B^A)
4225 cast<BinaryOperator>(Op1)->swapOperands();
4228 if (A == Op0) { // A&(A^B) -> A & ~B
4229 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4230 InsertNewInstBefore(NotB, I);
4231 return BinaryOperator::CreateAnd(A, NotB);
4235 // (A&((~A)|B)) -> A&B
4236 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4237 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4238 return BinaryOperator::CreateAnd(A, Op1);
4239 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4240 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4241 return BinaryOperator::CreateAnd(A, Op0);
4244 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4245 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4246 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4249 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4250 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4254 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4255 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4256 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4257 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4258 const Type *SrcTy = Op0C->getOperand(0)->getType();
4259 if (SrcTy == Op1C->getOperand(0)->getType() &&
4260 SrcTy->isIntOrIntVector() &&
4261 // Only do this if the casts both really cause code to be generated.
4262 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4264 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4266 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4267 Op1C->getOperand(0),
4269 InsertNewInstBefore(NewOp, I);
4270 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4274 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4275 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4276 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4277 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4278 SI0->getOperand(1) == SI1->getOperand(1) &&
4279 (SI0->hasOneUse() || SI1->hasOneUse())) {
4280 Instruction *NewOp =
4281 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4283 SI0->getName()), I);
4284 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4285 SI1->getOperand(1));
4289 // If and'ing two fcmp, try combine them into one.
4290 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4291 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4292 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4296 return Changed ? &I : 0;
4299 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4300 /// capable of providing pieces of a bswap. The subexpression provides pieces
4301 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4302 /// the expression came from the corresponding "byte swapped" byte in some other
4303 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4304 /// we know that the expression deposits the low byte of %X into the high byte
4305 /// of the bswap result and that all other bytes are zero. This expression is
4306 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4309 /// This function returns true if the match was unsuccessful and false if so.
4310 /// On entry to the function the "OverallLeftShift" is a signed integer value
4311 /// indicating the number of bytes that the subexpression is later shifted. For
4312 /// example, if the expression is later right shifted by 16 bits, the
4313 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4314 /// byte of ByteValues is actually being set.
4316 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4317 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4318 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4319 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4320 /// always in the local (OverallLeftShift) coordinate space.
4322 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4323 SmallVector<Value*, 8> &ByteValues) {
4324 if (Instruction *I = dyn_cast<Instruction>(V)) {
4325 // If this is an or instruction, it may be an inner node of the bswap.
4326 if (I->getOpcode() == Instruction::Or) {
4327 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4329 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4333 // If this is a logical shift by a constant multiple of 8, recurse with
4334 // OverallLeftShift and ByteMask adjusted.
4335 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4337 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4338 // Ensure the shift amount is defined and of a byte value.
4339 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4342 unsigned ByteShift = ShAmt >> 3;
4343 if (I->getOpcode() == Instruction::Shl) {
4344 // X << 2 -> collect(X, +2)
4345 OverallLeftShift += ByteShift;
4346 ByteMask >>= ByteShift;
4348 // X >>u 2 -> collect(X, -2)
4349 OverallLeftShift -= ByteShift;
4350 ByteMask <<= ByteShift;
4351 ByteMask &= (~0U >> (32-ByteValues.size()));
4354 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4355 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4357 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4361 // If this is a logical 'and' with a mask that clears bytes, clear the
4362 // corresponding bytes in ByteMask.
4363 if (I->getOpcode() == Instruction::And &&
4364 isa<ConstantInt>(I->getOperand(1))) {
4365 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4366 unsigned NumBytes = ByteValues.size();
4367 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4368 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4370 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4371 // If this byte is masked out by a later operation, we don't care what
4373 if ((ByteMask & (1 << i)) == 0)
4376 // If the AndMask is all zeros for this byte, clear the bit.
4377 APInt MaskB = AndMask & Byte;
4379 ByteMask &= ~(1U << i);
4383 // If the AndMask is not all ones for this byte, it's not a bytezap.
4387 // Otherwise, this byte is kept.
4390 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4395 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4396 // the input value to the bswap. Some observations: 1) if more than one byte
4397 // is demanded from this input, then it could not be successfully assembled
4398 // into a byteswap. At least one of the two bytes would not be aligned with
4399 // their ultimate destination.
4400 if (!isPowerOf2_32(ByteMask)) return true;
4401 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4403 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4404 // is demanded, it needs to go into byte 0 of the result. This means that the
4405 // byte needs to be shifted until it lands in the right byte bucket. The
4406 // shift amount depends on the position: if the byte is coming from the high
4407 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4408 // low part, it must be shifted left.
4409 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4410 if (InputByteNo < ByteValues.size()/2) {
4411 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4414 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4418 // If the destination byte value is already defined, the values are or'd
4419 // together, which isn't a bswap (unless it's an or of the same bits).
4420 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4422 ByteValues[DestByteNo] = V;
4426 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4427 /// If so, insert the new bswap intrinsic and return it.
4428 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4429 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4430 if (!ITy || ITy->getBitWidth() % 16 ||
4431 // ByteMask only allows up to 32-byte values.
4432 ITy->getBitWidth() > 32*8)
4433 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4435 /// ByteValues - For each byte of the result, we keep track of which value
4436 /// defines each byte.
4437 SmallVector<Value*, 8> ByteValues;
4438 ByteValues.resize(ITy->getBitWidth()/8);
4440 // Try to find all the pieces corresponding to the bswap.
4441 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4442 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4445 // Check to see if all of the bytes come from the same value.
4446 Value *V = ByteValues[0];
4447 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4449 // Check to make sure that all of the bytes come from the same value.
4450 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4451 if (ByteValues[i] != V)
4453 const Type *Tys[] = { ITy };
4454 Module *M = I.getParent()->getParent()->getParent();
4455 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4456 return CallInst::Create(F, V);
4459 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4460 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4461 /// we can simplify this expression to "cond ? C : D or B".
4462 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4464 LLVMContext *Context) {
4465 // If A is not a select of -1/0, this cannot match.
4467 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4470 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4471 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4472 return SelectInst::Create(Cond, C, B);
4473 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4474 return SelectInst::Create(Cond, C, B);
4475 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4476 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4477 return SelectInst::Create(Cond, C, D);
4478 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4479 return SelectInst::Create(Cond, C, D);
4483 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4484 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4485 ICmpInst *LHS, ICmpInst *RHS) {
4487 ConstantInt *LHSCst, *RHSCst;
4488 ICmpInst::Predicate LHSCC, RHSCC;
4490 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4491 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4492 m_ConstantInt(LHSCst))) ||
4493 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4494 m_ConstantInt(RHSCst))))
4497 // From here on, we only handle:
4498 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4499 if (Val != Val2) return 0;
4501 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4502 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4503 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4504 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4505 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4508 // We can't fold (ugt x, C) | (sgt x, C2).
4509 if (!PredicatesFoldable(LHSCC, RHSCC))
4512 // Ensure that the larger constant is on the RHS.
4514 if (ICmpInst::isSignedPredicate(LHSCC) ||
4515 (ICmpInst::isEquality(LHSCC) &&
4516 ICmpInst::isSignedPredicate(RHSCC)))
4517 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4519 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4522 std::swap(LHS, RHS);
4523 std::swap(LHSCst, RHSCst);
4524 std::swap(LHSCC, RHSCC);
4527 // At this point, we know we have have two icmp instructions
4528 // comparing a value against two constants and or'ing the result
4529 // together. Because of the above check, we know that we only have
4530 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4531 // FoldICmpLogical check above), that the two constants are not
4533 assert(LHSCst != RHSCst && "Compares not folded above?");
4536 default: llvm_unreachable("Unknown integer condition code!");
4537 case ICmpInst::ICMP_EQ:
4539 default: llvm_unreachable("Unknown integer condition code!");
4540 case ICmpInst::ICMP_EQ:
4541 if (LHSCst == SubOne(RHSCst)) {
4542 // (X == 13 | X == 14) -> X-13 <u 2
4543 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4544 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4545 Val->getName()+".off");
4546 InsertNewInstBefore(Add, I);
4547 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4548 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4550 break; // (X == 13 | X == 15) -> no change
4551 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4552 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4554 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4555 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4556 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4557 return ReplaceInstUsesWith(I, RHS);
4560 case ICmpInst::ICMP_NE:
4562 default: llvm_unreachable("Unknown integer condition code!");
4563 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4564 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4565 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4566 return ReplaceInstUsesWith(I, LHS);
4567 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4568 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4569 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4570 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4573 case ICmpInst::ICMP_ULT:
4575 default: llvm_unreachable("Unknown integer condition code!");
4576 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4578 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4579 // If RHSCst is [us]MAXINT, it is always false. Not handling
4580 // this can cause overflow.
4581 if (RHSCst->isMaxValue(false))
4582 return ReplaceInstUsesWith(I, LHS);
4583 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4585 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4587 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4588 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4589 return ReplaceInstUsesWith(I, RHS);
4590 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4594 case ICmpInst::ICMP_SLT:
4596 default: llvm_unreachable("Unknown integer condition code!");
4597 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4599 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4600 // If RHSCst is [us]MAXINT, it is always false. Not handling
4601 // this can cause overflow.
4602 if (RHSCst->isMaxValue(true))
4603 return ReplaceInstUsesWith(I, LHS);
4604 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4606 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4608 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4609 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4610 return ReplaceInstUsesWith(I, RHS);
4611 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4615 case ICmpInst::ICMP_UGT:
4617 default: llvm_unreachable("Unknown integer condition code!");
4618 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4619 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4620 return ReplaceInstUsesWith(I, LHS);
4621 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4623 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4624 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4625 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4626 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4630 case ICmpInst::ICMP_SGT:
4632 default: llvm_unreachable("Unknown integer condition code!");
4633 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4634 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4635 return ReplaceInstUsesWith(I, LHS);
4636 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4638 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4639 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4640 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4641 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4649 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
4651 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4652 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4653 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4654 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4655 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4656 // If either of the constants are nans, then the whole thing returns
4658 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4659 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4661 // Otherwise, no need to compare the two constants, compare the
4663 return new FCmpInst(FCmpInst::FCMP_UNO,
4664 LHS->getOperand(0), RHS->getOperand(0));
4667 // Handle vector zeros. This occurs because the canonical form of
4668 // "fcmp uno x,x" is "fcmp uno x, 0".
4669 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4670 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4671 return new FCmpInst(FCmpInst::FCMP_UNO,
4672 LHS->getOperand(0), RHS->getOperand(0));
4677 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4678 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4679 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4681 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4682 // Swap RHS operands to match LHS.
4683 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4684 std::swap(Op1LHS, Op1RHS);
4686 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4687 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4689 return new FCmpInst((FCmpInst::Predicate)Op0CC,
4691 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
4692 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4693 if (Op0CC == FCmpInst::FCMP_FALSE)
4694 return ReplaceInstUsesWith(I, RHS);
4695 if (Op1CC == FCmpInst::FCMP_FALSE)
4696 return ReplaceInstUsesWith(I, LHS);
4699 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4700 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4701 if (Op0Ordered == Op1Ordered) {
4702 // If both are ordered or unordered, return a new fcmp with
4703 // or'ed predicates.
4704 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4705 Op0LHS, Op0RHS, Context);
4706 if (Instruction *I = dyn_cast<Instruction>(RV))
4708 // Otherwise, it's a constant boolean value...
4709 return ReplaceInstUsesWith(I, RV);
4715 /// FoldOrWithConstants - This helper function folds:
4717 /// ((A | B) & C1) | (B & C2)
4723 /// when the XOR of the two constants is "all ones" (-1).
4724 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4725 Value *A, Value *B, Value *C) {
4726 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4730 ConstantInt *CI2 = 0;
4731 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4733 APInt Xor = CI1->getValue() ^ CI2->getValue();
4734 if (!Xor.isAllOnesValue()) return 0;
4736 if (V1 == A || V1 == B) {
4737 Instruction *NewOp =
4738 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4739 return BinaryOperator::CreateOr(NewOp, V1);
4745 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4746 bool Changed = SimplifyCommutative(I);
4747 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4749 if (isa<UndefValue>(Op1)) // X | undef -> -1
4750 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4754 return ReplaceInstUsesWith(I, Op0);
4756 // See if we can simplify any instructions used by the instruction whose sole
4757 // purpose is to compute bits we don't care about.
4758 if (SimplifyDemandedInstructionBits(I))
4760 if (isa<VectorType>(I.getType())) {
4761 if (isa<ConstantAggregateZero>(Op1)) {
4762 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4763 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4764 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4765 return ReplaceInstUsesWith(I, I.getOperand(1));
4770 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4771 ConstantInt *C1 = 0; Value *X = 0;
4772 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4773 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
4775 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4776 InsertNewInstBefore(Or, I);
4778 return BinaryOperator::CreateAnd(Or,
4779 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
4782 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4783 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
4785 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4786 InsertNewInstBefore(Or, I);
4788 return BinaryOperator::CreateXor(Or,
4789 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
4792 // Try to fold constant and into select arguments.
4793 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4794 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4796 if (isa<PHINode>(Op0))
4797 if (Instruction *NV = FoldOpIntoPhi(I))
4801 Value *A = 0, *B = 0;
4802 ConstantInt *C1 = 0, *C2 = 0;
4804 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4805 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4806 return ReplaceInstUsesWith(I, Op1);
4807 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4808 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4809 return ReplaceInstUsesWith(I, Op0);
4811 // (A | B) | C and A | (B | C) -> bswap if possible.
4812 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4813 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4814 match(Op1, m_Or(m_Value(), m_Value())) ||
4815 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4816 match(Op1, m_Shift(m_Value(), m_Value())))) {
4817 if (Instruction *BSwap = MatchBSwap(I))
4821 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4822 if (Op0->hasOneUse() &&
4823 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4824 MaskedValueIsZero(Op1, C1->getValue())) {
4825 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4826 InsertNewInstBefore(NOr, I);
4828 return BinaryOperator::CreateXor(NOr, C1);
4831 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4832 if (Op1->hasOneUse() &&
4833 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4834 MaskedValueIsZero(Op0, C1->getValue())) {
4835 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4836 InsertNewInstBefore(NOr, I);
4838 return BinaryOperator::CreateXor(NOr, C1);
4842 Value *C = 0, *D = 0;
4843 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4844 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4845 Value *V1 = 0, *V2 = 0, *V3 = 0;
4846 C1 = dyn_cast<ConstantInt>(C);
4847 C2 = dyn_cast<ConstantInt>(D);
4848 if (C1 && C2) { // (A & C1)|(B & C2)
4849 // If we have: ((V + N) & C1) | (V & C2)
4850 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4851 // replace with V+N.
4852 if (C1->getValue() == ~C2->getValue()) {
4853 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4854 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4855 // Add commutes, try both ways.
4856 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4857 return ReplaceInstUsesWith(I, A);
4858 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4859 return ReplaceInstUsesWith(I, A);
4861 // Or commutes, try both ways.
4862 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4863 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4864 // Add commutes, try both ways.
4865 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4866 return ReplaceInstUsesWith(I, B);
4867 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4868 return ReplaceInstUsesWith(I, B);
4871 V1 = 0; V2 = 0; V3 = 0;
4874 // Check to see if we have any common things being and'ed. If so, find the
4875 // terms for V1 & (V2|V3).
4876 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4877 if (A == B) // (A & C)|(A & D) == A & (C|D)
4878 V1 = A, V2 = C, V3 = D;
4879 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4880 V1 = A, V2 = B, V3 = C;
4881 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4882 V1 = C, V2 = A, V3 = D;
4883 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4884 V1 = C, V2 = A, V3 = B;
4888 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4889 return BinaryOperator::CreateAnd(V1, Or);
4893 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4894 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4896 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4898 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4900 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4903 // ((A&~B)|(~A&B)) -> A^B
4904 if ((match(C, m_Not(m_Specific(D))) &&
4905 match(B, m_Not(m_Specific(A)))))
4906 return BinaryOperator::CreateXor(A, D);
4907 // ((~B&A)|(~A&B)) -> A^B
4908 if ((match(A, m_Not(m_Specific(D))) &&
4909 match(B, m_Not(m_Specific(C)))))
4910 return BinaryOperator::CreateXor(C, D);
4911 // ((A&~B)|(B&~A)) -> A^B
4912 if ((match(C, m_Not(m_Specific(B))) &&
4913 match(D, m_Not(m_Specific(A)))))
4914 return BinaryOperator::CreateXor(A, B);
4915 // ((~B&A)|(B&~A)) -> A^B
4916 if ((match(A, m_Not(m_Specific(B))) &&
4917 match(D, m_Not(m_Specific(C)))))
4918 return BinaryOperator::CreateXor(C, B);
4921 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4922 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4923 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4924 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4925 SI0->getOperand(1) == SI1->getOperand(1) &&
4926 (SI0->hasOneUse() || SI1->hasOneUse())) {
4927 Instruction *NewOp =
4928 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4930 SI0->getName()), I);
4931 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4932 SI1->getOperand(1));
4936 // ((A|B)&1)|(B&-2) -> (A&1) | B
4937 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4938 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4939 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4940 if (Ret) return Ret;
4942 // (B&-2)|((A|B)&1) -> (A&1) | B
4943 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4944 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4945 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4946 if (Ret) return Ret;
4949 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4950 if (A == Op1) // ~A | A == -1
4951 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4955 // Note, A is still live here!
4956 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4958 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4960 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4961 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4962 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4963 I.getName()+".demorgan"), I);
4964 return BinaryOperator::CreateNot(And);
4968 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4969 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4970 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4973 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4974 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4978 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4979 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4980 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4981 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4982 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4983 !isa<ICmpInst>(Op1C->getOperand(0))) {
4984 const Type *SrcTy = Op0C->getOperand(0)->getType();
4985 if (SrcTy == Op1C->getOperand(0)->getType() &&
4986 SrcTy->isIntOrIntVector() &&
4987 // Only do this if the casts both really cause code to be
4989 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4991 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4993 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4994 Op1C->getOperand(0),
4996 InsertNewInstBefore(NewOp, I);
4997 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5004 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
5005 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
5006 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
5007 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
5011 return Changed ? &I : 0;
5016 // XorSelf - Implements: X ^ X --> 0
5019 XorSelf(Value *rhs) : RHS(rhs) {}
5020 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5021 Instruction *apply(BinaryOperator &Xor) const {
5028 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5029 bool Changed = SimplifyCommutative(I);
5030 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5032 if (isa<UndefValue>(Op1)) {
5033 if (isa<UndefValue>(Op0))
5034 // Handle undef ^ undef -> 0 special case. This is a common
5036 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5037 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5040 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5041 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
5042 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5043 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5046 // See if we can simplify any instructions used by the instruction whose sole
5047 // purpose is to compute bits we don't care about.
5048 if (SimplifyDemandedInstructionBits(I))
5050 if (isa<VectorType>(I.getType()))
5051 if (isa<ConstantAggregateZero>(Op1))
5052 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5054 // Is this a ~ operation?
5055 if (Value *NotOp = dyn_castNotVal(&I)) {
5056 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5057 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5058 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5059 if (Op0I->getOpcode() == Instruction::And ||
5060 Op0I->getOpcode() == Instruction::Or) {
5061 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
5062 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5064 BinaryOperator::CreateNot(Op0I->getOperand(1),
5065 Op0I->getOperand(1)->getName()+".not");
5066 InsertNewInstBefore(NotY, I);
5067 if (Op0I->getOpcode() == Instruction::And)
5068 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5070 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5077 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5078 if (RHS == ConstantInt::getTrue(*Context) && Op0->hasOneUse()) {
5079 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5080 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5081 return new ICmpInst(ICI->getInversePredicate(),
5082 ICI->getOperand(0), ICI->getOperand(1));
5084 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5085 return new FCmpInst(FCI->getInversePredicate(),
5086 FCI->getOperand(0), FCI->getOperand(1));
5089 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5090 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5091 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5092 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5093 Instruction::CastOps Opcode = Op0C->getOpcode();
5094 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
5095 if (RHS == ConstantExpr::getCast(Opcode,
5096 ConstantInt::getTrue(*Context),
5097 Op0C->getDestTy())) {
5098 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
5099 CI->getOpcode(), CI->getInversePredicate(),
5100 CI->getOperand(0), CI->getOperand(1)), I);
5101 NewCI->takeName(CI);
5102 return CastInst::Create(Opcode, NewCI, Op0C->getType());
5109 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5110 // ~(c-X) == X-c-1 == X+(-c-1)
5111 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5112 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5113 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5114 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5115 ConstantInt::get(I.getType(), 1));
5116 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5119 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5120 if (Op0I->getOpcode() == Instruction::Add) {
5121 // ~(X-c) --> (-c-1)-X
5122 if (RHS->isAllOnesValue()) {
5123 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5124 return BinaryOperator::CreateSub(
5125 ConstantExpr::getSub(NegOp0CI,
5126 ConstantInt::get(I.getType(), 1)),
5127 Op0I->getOperand(0));
5128 } else if (RHS->getValue().isSignBit()) {
5129 // (X + C) ^ signbit -> (X + C + signbit)
5130 Constant *C = ConstantInt::get(*Context,
5131 RHS->getValue() + Op0CI->getValue());
5132 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5135 } else if (Op0I->getOpcode() == Instruction::Or) {
5136 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5137 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5138 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5139 // Anything in both C1 and C2 is known to be zero, remove it from
5141 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5142 NewRHS = ConstantExpr::getAnd(NewRHS,
5143 ConstantExpr::getNot(CommonBits));
5144 AddToWorkList(Op0I);
5145 I.setOperand(0, Op0I->getOperand(0));
5146 I.setOperand(1, NewRHS);
5153 // Try to fold constant and into select arguments.
5154 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5155 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5157 if (isa<PHINode>(Op0))
5158 if (Instruction *NV = FoldOpIntoPhi(I))
5162 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5164 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5166 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5168 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5171 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5174 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5175 if (A == Op0) { // B^(B|A) == (A|B)^B
5176 Op1I->swapOperands();
5178 std::swap(Op0, Op1);
5179 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5180 I.swapOperands(); // Simplified below.
5181 std::swap(Op0, Op1);
5183 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5184 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5185 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5186 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5187 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5189 if (A == Op0) { // A^(A&B) -> A^(B&A)
5190 Op1I->swapOperands();
5193 if (B == Op0) { // A^(B&A) -> (B&A)^A
5194 I.swapOperands(); // Simplified below.
5195 std::swap(Op0, Op1);
5200 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5203 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5204 Op0I->hasOneUse()) {
5205 if (A == Op1) // (B|A)^B == (A|B)^B
5207 if (B == Op1) { // (A|B)^B == A & ~B
5209 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5210 return BinaryOperator::CreateAnd(A, NotB);
5212 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5213 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5214 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5215 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5216 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5218 if (A == Op1) // (A&B)^A -> (B&A)^A
5220 if (B == Op1 && // (B&A)^A == ~B & A
5221 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5223 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5224 return BinaryOperator::CreateAnd(N, Op1);
5229 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5230 if (Op0I && Op1I && Op0I->isShift() &&
5231 Op0I->getOpcode() == Op1I->getOpcode() &&
5232 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5233 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5234 Instruction *NewOp =
5235 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5236 Op1I->getOperand(0),
5237 Op0I->getName()), I);
5238 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5239 Op1I->getOperand(1));
5243 Value *A, *B, *C, *D;
5244 // (A & B)^(A | B) -> A ^ B
5245 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5246 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5247 if ((A == C && B == D) || (A == D && B == C))
5248 return BinaryOperator::CreateXor(A, B);
5250 // (A | B)^(A & B) -> A ^ B
5251 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5252 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5253 if ((A == C && B == D) || (A == D && B == C))
5254 return BinaryOperator::CreateXor(A, B);
5258 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5259 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5260 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5261 // (X & Y)^(X & Y) -> (Y^Z) & X
5262 Value *X = 0, *Y = 0, *Z = 0;
5264 X = A, Y = B, Z = D;
5266 X = A, Y = B, Z = C;
5268 X = B, Y = A, Z = D;
5270 X = B, Y = A, Z = C;
5273 Instruction *NewOp =
5274 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5275 return BinaryOperator::CreateAnd(NewOp, X);
5280 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5281 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5282 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5285 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5286 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5287 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5288 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5289 const Type *SrcTy = Op0C->getOperand(0)->getType();
5290 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5291 // Only do this if the casts both really cause code to be generated.
5292 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5294 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5296 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5297 Op1C->getOperand(0),
5299 InsertNewInstBefore(NewOp, I);
5300 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5305 return Changed ? &I : 0;
5308 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5309 LLVMContext *Context) {
5310 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5313 static bool HasAddOverflow(ConstantInt *Result,
5314 ConstantInt *In1, ConstantInt *In2,
5317 if (In2->getValue().isNegative())
5318 return Result->getValue().sgt(In1->getValue());
5320 return Result->getValue().slt(In1->getValue());
5322 return Result->getValue().ult(In1->getValue());
5325 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5326 /// overflowed for this type.
5327 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5328 Constant *In2, LLVMContext *Context,
5329 bool IsSigned = false) {
5330 Result = ConstantExpr::getAdd(In1, In2);
5332 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5333 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5334 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5335 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5336 ExtractElement(In1, Idx, Context),
5337 ExtractElement(In2, Idx, Context),
5344 return HasAddOverflow(cast<ConstantInt>(Result),
5345 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5349 static bool HasSubOverflow(ConstantInt *Result,
5350 ConstantInt *In1, ConstantInt *In2,
5353 if (In2->getValue().isNegative())
5354 return Result->getValue().slt(In1->getValue());
5356 return Result->getValue().sgt(In1->getValue());
5358 return Result->getValue().ugt(In1->getValue());
5361 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5362 /// overflowed for this type.
5363 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5364 Constant *In2, LLVMContext *Context,
5365 bool IsSigned = false) {
5366 Result = ConstantExpr::getSub(In1, In2);
5368 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5369 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5370 Constant *Idx = ConstantInt::get(Type::getInt32Ty(*Context), i);
5371 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5372 ExtractElement(In1, Idx, Context),
5373 ExtractElement(In2, Idx, Context),
5380 return HasSubOverflow(cast<ConstantInt>(Result),
5381 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5385 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5386 /// code necessary to compute the offset from the base pointer (without adding
5387 /// in the base pointer). Return the result as a signed integer of intptr size.
5388 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5389 TargetData &TD = *IC.getTargetData();
5390 gep_type_iterator GTI = gep_type_begin(GEP);
5391 const Type *IntPtrTy = TD.getIntPtrType(I.getContext());
5392 LLVMContext *Context = IC.getContext();
5393 Value *Result = Constant::getNullValue(IntPtrTy);
5395 // Build a mask for high order bits.
5396 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5397 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5399 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5402 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5403 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5404 if (OpC->isZero()) continue;
5406 // Handle a struct index, which adds its field offset to the pointer.
5407 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5408 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5410 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5412 ConstantInt::get(*Context,
5413 RC->getValue() + APInt(IntPtrWidth, Size));
5415 Result = IC.InsertNewInstBefore(
5416 BinaryOperator::CreateAdd(Result,
5417 ConstantInt::get(IntPtrTy, Size),
5418 GEP->getName()+".offs"), I);
5422 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5424 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5425 Scale = ConstantExpr::getMul(OC, Scale);
5426 if (Constant *RC = dyn_cast<Constant>(Result))
5427 Result = ConstantExpr::getAdd(RC, Scale);
5429 // Emit an add instruction.
5430 Result = IC.InsertNewInstBefore(
5431 BinaryOperator::CreateAdd(Result, Scale,
5432 GEP->getName()+".offs"), I);
5436 // Convert to correct type.
5437 if (Op->getType() != IntPtrTy) {
5438 if (Constant *OpC = dyn_cast<Constant>(Op))
5439 Op = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true);
5441 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5443 Op->getName()+".c"), I);
5446 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5447 if (Constant *OpC = dyn_cast<Constant>(Op))
5448 Op = ConstantExpr::getMul(OpC, Scale);
5449 else // We'll let instcombine(mul) convert this to a shl if possible.
5450 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5451 GEP->getName()+".idx"), I);
5454 // Emit an add instruction.
5455 if (isa<Constant>(Op) && isa<Constant>(Result))
5456 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5457 cast<Constant>(Result));
5459 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5460 GEP->getName()+".offs"), I);
5466 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5467 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5468 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5469 /// be complex, and scales are involved. The above expression would also be
5470 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5471 /// This later form is less amenable to optimization though, and we are allowed
5472 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5474 /// If we can't emit an optimized form for this expression, this returns null.
5476 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5478 TargetData &TD = *IC.getTargetData();
5479 gep_type_iterator GTI = gep_type_begin(GEP);
5481 // Check to see if this gep only has a single variable index. If so, and if
5482 // any constant indices are a multiple of its scale, then we can compute this
5483 // in terms of the scale of the variable index. For example, if the GEP
5484 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5485 // because the expression will cross zero at the same point.
5486 unsigned i, e = GEP->getNumOperands();
5488 for (i = 1; i != e; ++i, ++GTI) {
5489 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5490 // Compute the aggregate offset of constant indices.
5491 if (CI->isZero()) continue;
5493 // Handle a struct index, which adds its field offset to the pointer.
5494 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5495 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5497 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5498 Offset += Size*CI->getSExtValue();
5501 // Found our variable index.
5506 // If there are no variable indices, we must have a constant offset, just
5507 // evaluate it the general way.
5508 if (i == e) return 0;
5510 Value *VariableIdx = GEP->getOperand(i);
5511 // Determine the scale factor of the variable element. For example, this is
5512 // 4 if the variable index is into an array of i32.
5513 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5515 // Verify that there are no other variable indices. If so, emit the hard way.
5516 for (++i, ++GTI; i != e; ++i, ++GTI) {
5517 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5520 // Compute the aggregate offset of constant indices.
5521 if (CI->isZero()) continue;
5523 // Handle a struct index, which adds its field offset to the pointer.
5524 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5525 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5527 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5528 Offset += Size*CI->getSExtValue();
5532 // Okay, we know we have a single variable index, which must be a
5533 // pointer/array/vector index. If there is no offset, life is simple, return
5535 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5537 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5538 // we don't need to bother extending: the extension won't affect where the
5539 // computation crosses zero.
5540 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5541 VariableIdx = new TruncInst(VariableIdx,
5542 TD.getIntPtrType(VariableIdx->getContext()),
5543 VariableIdx->getName(), &I);
5547 // Otherwise, there is an index. The computation we will do will be modulo
5548 // the pointer size, so get it.
5549 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5551 Offset &= PtrSizeMask;
5552 VariableScale &= PtrSizeMask;
5554 // To do this transformation, any constant index must be a multiple of the
5555 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5556 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5557 // multiple of the variable scale.
5558 int64_t NewOffs = Offset / (int64_t)VariableScale;
5559 if (Offset != NewOffs*(int64_t)VariableScale)
5562 // Okay, we can do this evaluation. Start by converting the index to intptr.
5563 const Type *IntPtrTy = TD.getIntPtrType(VariableIdx->getContext());
5564 if (VariableIdx->getType() != IntPtrTy)
5565 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5567 VariableIdx->getName(), &I);
5568 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5569 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5573 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5574 /// else. At this point we know that the GEP is on the LHS of the comparison.
5575 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5576 ICmpInst::Predicate Cond,
5578 // Look through bitcasts.
5579 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5580 RHS = BCI->getOperand(0);
5582 Value *PtrBase = GEPLHS->getOperand(0);
5583 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5584 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5585 // This transformation (ignoring the base and scales) is valid because we
5586 // know pointers can't overflow since the gep is inbounds. See if we can
5587 // output an optimized form.
5588 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5590 // If not, synthesize the offset the hard way.
5592 Offset = EmitGEPOffset(GEPLHS, I, *this);
5593 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5594 Constant::getNullValue(Offset->getType()));
5595 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5596 // If the base pointers are different, but the indices are the same, just
5597 // compare the base pointer.
5598 if (PtrBase != GEPRHS->getOperand(0)) {
5599 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5600 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5601 GEPRHS->getOperand(0)->getType();
5603 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5604 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5605 IndicesTheSame = false;
5609 // If all indices are the same, just compare the base pointers.
5611 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5612 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5614 // Otherwise, the base pointers are different and the indices are
5615 // different, bail out.
5619 // If one of the GEPs has all zero indices, recurse.
5620 bool AllZeros = true;
5621 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5622 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5623 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5628 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5629 ICmpInst::getSwappedPredicate(Cond), I);
5631 // If the other GEP has all zero indices, recurse.
5633 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5634 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5635 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5640 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5642 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5643 // If the GEPs only differ by one index, compare it.
5644 unsigned NumDifferences = 0; // Keep track of # differences.
5645 unsigned DiffOperand = 0; // The operand that differs.
5646 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5647 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5648 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5649 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5650 // Irreconcilable differences.
5654 if (NumDifferences++) break;
5659 if (NumDifferences == 0) // SAME GEP?
5660 return ReplaceInstUsesWith(I, // No comparison is needed here.
5661 ConstantInt::get(Type::getInt1Ty(*Context),
5662 ICmpInst::isTrueWhenEqual(Cond)));
5664 else if (NumDifferences == 1) {
5665 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5666 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5667 // Make sure we do a signed comparison here.
5668 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5672 // Only lower this if the icmp is the only user of the GEP or if we expect
5673 // the result to fold to a constant!
5675 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5676 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5677 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5678 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5679 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5680 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5686 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5688 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5691 if (!isa<ConstantFP>(RHSC)) return 0;
5692 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5694 // Get the width of the mantissa. We don't want to hack on conversions that
5695 // might lose information from the integer, e.g. "i64 -> float"
5696 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5697 if (MantissaWidth == -1) return 0; // Unknown.
5699 // Check to see that the input is converted from an integer type that is small
5700 // enough that preserves all bits. TODO: check here for "known" sign bits.
5701 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5702 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5704 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5705 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5709 // If the conversion would lose info, don't hack on this.
5710 if ((int)InputSize > MantissaWidth)
5713 // Otherwise, we can potentially simplify the comparison. We know that it
5714 // will always come through as an integer value and we know the constant is
5715 // not a NAN (it would have been previously simplified).
5716 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5718 ICmpInst::Predicate Pred;
5719 switch (I.getPredicate()) {
5720 default: llvm_unreachable("Unexpected predicate!");
5721 case FCmpInst::FCMP_UEQ:
5722 case FCmpInst::FCMP_OEQ:
5723 Pred = ICmpInst::ICMP_EQ;
5725 case FCmpInst::FCMP_UGT:
5726 case FCmpInst::FCMP_OGT:
5727 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5729 case FCmpInst::FCMP_UGE:
5730 case FCmpInst::FCMP_OGE:
5731 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5733 case FCmpInst::FCMP_ULT:
5734 case FCmpInst::FCMP_OLT:
5735 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5737 case FCmpInst::FCMP_ULE:
5738 case FCmpInst::FCMP_OLE:
5739 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5741 case FCmpInst::FCMP_UNE:
5742 case FCmpInst::FCMP_ONE:
5743 Pred = ICmpInst::ICMP_NE;
5745 case FCmpInst::FCMP_ORD:
5746 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5747 case FCmpInst::FCMP_UNO:
5748 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5751 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5753 // Now we know that the APFloat is a normal number, zero or inf.
5755 // See if the FP constant is too large for the integer. For example,
5756 // comparing an i8 to 300.0.
5757 unsigned IntWidth = IntTy->getScalarSizeInBits();
5760 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5761 // and large values.
5762 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5763 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5764 APFloat::rmNearestTiesToEven);
5765 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5766 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5767 Pred == ICmpInst::ICMP_SLE)
5768 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5769 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5772 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5773 // +INF and large values.
5774 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5775 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5776 APFloat::rmNearestTiesToEven);
5777 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5778 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5779 Pred == ICmpInst::ICMP_ULE)
5780 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5781 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5786 // See if the RHS value is < SignedMin.
5787 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5788 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5789 APFloat::rmNearestTiesToEven);
5790 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5791 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5792 Pred == ICmpInst::ICMP_SGE)
5793 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5794 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5798 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5799 // [0, UMAX], but it may still be fractional. See if it is fractional by
5800 // casting the FP value to the integer value and back, checking for equality.
5801 // Don't do this for zero, because -0.0 is not fractional.
5802 Constant *RHSInt = LHSUnsigned
5803 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5804 : ConstantExpr::getFPToSI(RHSC, IntTy);
5805 if (!RHS.isZero()) {
5806 bool Equal = LHSUnsigned
5807 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5808 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5810 // If we had a comparison against a fractional value, we have to adjust
5811 // the compare predicate and sometimes the value. RHSC is rounded towards
5812 // zero at this point.
5814 default: llvm_unreachable("Unexpected integer comparison!");
5815 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5816 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5817 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5818 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5819 case ICmpInst::ICMP_ULE:
5820 // (float)int <= 4.4 --> int <= 4
5821 // (float)int <= -4.4 --> false
5822 if (RHS.isNegative())
5823 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5825 case ICmpInst::ICMP_SLE:
5826 // (float)int <= 4.4 --> int <= 4
5827 // (float)int <= -4.4 --> int < -4
5828 if (RHS.isNegative())
5829 Pred = ICmpInst::ICMP_SLT;
5831 case ICmpInst::ICMP_ULT:
5832 // (float)int < -4.4 --> false
5833 // (float)int < 4.4 --> int <= 4
5834 if (RHS.isNegative())
5835 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5836 Pred = ICmpInst::ICMP_ULE;
5838 case ICmpInst::ICMP_SLT:
5839 // (float)int < -4.4 --> int < -4
5840 // (float)int < 4.4 --> int <= 4
5841 if (!RHS.isNegative())
5842 Pred = ICmpInst::ICMP_SLE;
5844 case ICmpInst::ICMP_UGT:
5845 // (float)int > 4.4 --> int > 4
5846 // (float)int > -4.4 --> true
5847 if (RHS.isNegative())
5848 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5850 case ICmpInst::ICMP_SGT:
5851 // (float)int > 4.4 --> int > 4
5852 // (float)int > -4.4 --> int >= -4
5853 if (RHS.isNegative())
5854 Pred = ICmpInst::ICMP_SGE;
5856 case ICmpInst::ICMP_UGE:
5857 // (float)int >= -4.4 --> true
5858 // (float)int >= 4.4 --> int > 4
5859 if (!RHS.isNegative())
5860 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5861 Pred = ICmpInst::ICMP_UGT;
5863 case ICmpInst::ICMP_SGE:
5864 // (float)int >= -4.4 --> int >= -4
5865 // (float)int >= 4.4 --> int > 4
5866 if (!RHS.isNegative())
5867 Pred = ICmpInst::ICMP_SGT;
5873 // Lower this FP comparison into an appropriate integer version of the
5875 return new ICmpInst(Pred, LHSI->getOperand(0), RHSInt);
5878 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5879 bool Changed = SimplifyCompare(I);
5880 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5882 // Fold trivial predicates.
5883 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5884 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5885 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5886 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5888 // Simplify 'fcmp pred X, X'
5890 switch (I.getPredicate()) {
5891 default: llvm_unreachable("Unknown predicate!");
5892 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5893 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5894 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5895 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5896 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5897 case FCmpInst::FCMP_OLT: // True if ordered and less than
5898 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5899 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5901 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5902 case FCmpInst::FCMP_ULT: // True if unordered or less than
5903 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5904 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5905 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5906 I.setPredicate(FCmpInst::FCMP_UNO);
5907 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5910 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5911 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5912 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5913 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5914 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5915 I.setPredicate(FCmpInst::FCMP_ORD);
5916 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5921 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5922 return ReplaceInstUsesWith(I, UndefValue::get(Type::getInt1Ty(*Context)));
5924 // Handle fcmp with constant RHS
5925 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5926 // If the constant is a nan, see if we can fold the comparison based on it.
5927 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5928 if (CFP->getValueAPF().isNaN()) {
5929 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5930 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5931 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5932 "Comparison must be either ordered or unordered!");
5933 // True if unordered.
5934 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5938 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5939 switch (LHSI->getOpcode()) {
5940 case Instruction::PHI:
5941 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5942 // block. If in the same block, we're encouraging jump threading. If
5943 // not, we are just pessimizing the code by making an i1 phi.
5944 if (LHSI->getParent() == I.getParent())
5945 if (Instruction *NV = FoldOpIntoPhi(I))
5948 case Instruction::SIToFP:
5949 case Instruction::UIToFP:
5950 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5953 case Instruction::Select:
5954 // If either operand of the select is a constant, we can fold the
5955 // comparison into the select arms, which will cause one to be
5956 // constant folded and the select turned into a bitwise or.
5957 Value *Op1 = 0, *Op2 = 0;
5958 if (LHSI->hasOneUse()) {
5959 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5960 // Fold the known value into the constant operand.
5961 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5962 // Insert a new FCmp of the other select operand.
5963 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5964 LHSI->getOperand(2), RHSC,
5966 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5967 // Fold the known value into the constant operand.
5968 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5969 // Insert a new FCmp of the other select operand.
5970 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5971 LHSI->getOperand(1), RHSC,
5977 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5982 return Changed ? &I : 0;
5985 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5986 bool Changed = SimplifyCompare(I);
5987 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5988 const Type *Ty = Op0->getType();
5992 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
5993 I.isTrueWhenEqual()));
5995 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5996 return ReplaceInstUsesWith(I, UndefValue::get(Type::getInt1Ty(*Context)));
5998 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5999 // addresses never equal each other! We already know that Op0 != Op1.
6000 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
6001 isa<ConstantPointerNull>(Op0)) &&
6002 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
6003 isa<ConstantPointerNull>(Op1)))
6004 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
6005 !I.isTrueWhenEqual()));
6007 // icmp's with boolean values can always be turned into bitwise operations
6008 if (Ty == Type::getInt1Ty(*Context)) {
6009 switch (I.getPredicate()) {
6010 default: llvm_unreachable("Invalid icmp instruction!");
6011 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
6012 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
6013 InsertNewInstBefore(Xor, I);
6014 return BinaryOperator::CreateNot(Xor);
6016 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6017 return BinaryOperator::CreateXor(Op0, Op1);
6019 case ICmpInst::ICMP_UGT:
6020 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6022 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6023 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
6024 InsertNewInstBefore(Not, I);
6025 return BinaryOperator::CreateAnd(Not, Op1);
6027 case ICmpInst::ICMP_SGT:
6028 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6030 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6031 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
6032 InsertNewInstBefore(Not, I);
6033 return BinaryOperator::CreateAnd(Not, Op0);
6035 case ICmpInst::ICMP_UGE:
6036 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6038 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6039 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
6040 InsertNewInstBefore(Not, I);
6041 return BinaryOperator::CreateOr(Not, Op1);
6043 case ICmpInst::ICMP_SGE:
6044 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6046 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6047 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
6048 InsertNewInstBefore(Not, I);
6049 return BinaryOperator::CreateOr(Not, Op0);
6054 unsigned BitWidth = 0;
6056 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6057 else if (Ty->isIntOrIntVector())
6058 BitWidth = Ty->getScalarSizeInBits();
6060 bool isSignBit = false;
6062 // See if we are doing a comparison with a constant.
6063 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6064 Value *A = 0, *B = 0;
6066 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6067 if (I.isEquality() && CI->isNullValue() &&
6068 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
6069 // (icmp cond A B) if cond is equality
6070 return new ICmpInst(I.getPredicate(), A, B);
6073 // If we have an icmp le or icmp ge instruction, turn it into the
6074 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6075 // them being folded in the code below.
6076 switch (I.getPredicate()) {
6078 case ICmpInst::ICMP_ULE:
6079 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6080 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6081 return new ICmpInst(ICmpInst::ICMP_ULT, Op0,
6083 case ICmpInst::ICMP_SLE:
6084 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6085 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6086 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6088 case ICmpInst::ICMP_UGE:
6089 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6090 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6091 return new ICmpInst(ICmpInst::ICMP_UGT, Op0,
6093 case ICmpInst::ICMP_SGE:
6094 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6095 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6096 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6100 // If this comparison is a normal comparison, it demands all
6101 // bits, if it is a sign bit comparison, it only demands the sign bit.
6103 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6106 // See if we can fold the comparison based on range information we can get
6107 // by checking whether bits are known to be zero or one in the input.
6108 if (BitWidth != 0) {
6109 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6110 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6112 if (SimplifyDemandedBits(I.getOperandUse(0),
6113 isSignBit ? APInt::getSignBit(BitWidth)
6114 : APInt::getAllOnesValue(BitWidth),
6115 Op0KnownZero, Op0KnownOne, 0))
6117 if (SimplifyDemandedBits(I.getOperandUse(1),
6118 APInt::getAllOnesValue(BitWidth),
6119 Op1KnownZero, Op1KnownOne, 0))
6122 // Given the known and unknown bits, compute a range that the LHS could be
6123 // in. Compute the Min, Max and RHS values based on the known bits. For the
6124 // EQ and NE we use unsigned values.
6125 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6126 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6127 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6128 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6130 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6133 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6135 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6139 // If Min and Max are known to be the same, then SimplifyDemandedBits
6140 // figured out that the LHS is a constant. Just constant fold this now so
6141 // that code below can assume that Min != Max.
6142 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6143 return new ICmpInst(I.getPredicate(),
6144 ConstantInt::get(*Context, Op0Min), Op1);
6145 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6146 return new ICmpInst(I.getPredicate(), Op0,
6147 ConstantInt::get(*Context, Op1Min));
6149 // Based on the range information we know about the LHS, see if we can
6150 // simplify this comparison. For example, (x&4) < 8 is always true.
6151 switch (I.getPredicate()) {
6152 default: llvm_unreachable("Unknown icmp opcode!");
6153 case ICmpInst::ICMP_EQ:
6154 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6155 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6157 case ICmpInst::ICMP_NE:
6158 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6159 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6161 case ICmpInst::ICMP_ULT:
6162 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6163 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6164 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6165 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6166 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6167 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6168 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6169 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6170 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6173 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6174 if (CI->isMinValue(true))
6175 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
6176 Constant::getAllOnesValue(Op0->getType()));
6179 case ICmpInst::ICMP_UGT:
6180 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6181 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6182 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6183 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6185 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6186 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6187 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6188 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6189 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6192 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6193 if (CI->isMaxValue(true))
6194 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
6195 Constant::getNullValue(Op0->getType()));
6198 case ICmpInst::ICMP_SLT:
6199 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6200 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6201 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6202 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6203 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6204 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6205 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6206 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6207 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6211 case ICmpInst::ICMP_SGT:
6212 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6213 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6214 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6215 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6217 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6218 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
6219 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6220 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6221 return new ICmpInst(ICmpInst::ICMP_EQ, Op0,
6225 case ICmpInst::ICMP_SGE:
6226 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6227 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6228 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6229 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6230 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6232 case ICmpInst::ICMP_SLE:
6233 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6234 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6235 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6236 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6237 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6239 case ICmpInst::ICMP_UGE:
6240 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6241 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6242 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6243 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6244 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6246 case ICmpInst::ICMP_ULE:
6247 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6248 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6249 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6250 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6251 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6255 // Turn a signed comparison into an unsigned one if both operands
6256 // are known to have the same sign.
6257 if (I.isSignedPredicate() &&
6258 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6259 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6260 return new ICmpInst(I.getUnsignedPredicate(), Op0, Op1);
6263 // Test if the ICmpInst instruction is used exclusively by a select as
6264 // part of a minimum or maximum operation. If so, refrain from doing
6265 // any other folding. This helps out other analyses which understand
6266 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6267 // and CodeGen. And in this case, at least one of the comparison
6268 // operands has at least one user besides the compare (the select),
6269 // which would often largely negate the benefit of folding anyway.
6271 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6272 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6273 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6276 // See if we are doing a comparison between a constant and an instruction that
6277 // can be folded into the comparison.
6278 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6279 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6280 // instruction, see if that instruction also has constants so that the
6281 // instruction can be folded into the icmp
6282 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6283 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6287 // Handle icmp with constant (but not simple integer constant) RHS
6288 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6289 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6290 switch (LHSI->getOpcode()) {
6291 case Instruction::GetElementPtr:
6292 if (RHSC->isNullValue()) {
6293 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6294 bool isAllZeros = true;
6295 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6296 if (!isa<Constant>(LHSI->getOperand(i)) ||
6297 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6302 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
6303 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6307 case Instruction::PHI:
6308 // Only fold icmp into the PHI if the phi and fcmp are in the same
6309 // block. If in the same block, we're encouraging jump threading. If
6310 // not, we are just pessimizing the code by making an i1 phi.
6311 if (LHSI->getParent() == I.getParent())
6312 if (Instruction *NV = FoldOpIntoPhi(I))
6315 case Instruction::Select: {
6316 // If either operand of the select is a constant, we can fold the
6317 // comparison into the select arms, which will cause one to be
6318 // constant folded and the select turned into a bitwise or.
6319 Value *Op1 = 0, *Op2 = 0;
6320 if (LHSI->hasOneUse()) {
6321 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6322 // Fold the known value into the constant operand.
6323 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6324 // Insert a new ICmp of the other select operand.
6325 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6326 LHSI->getOperand(2), RHSC,
6328 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6329 // Fold the known value into the constant operand.
6330 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6331 // Insert a new ICmp of the other select operand.
6332 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
6333 LHSI->getOperand(1), RHSC,
6339 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6342 case Instruction::Malloc:
6343 // If we have (malloc != null), and if the malloc has a single use, we
6344 // can assume it is successful and remove the malloc.
6345 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6346 AddToWorkList(LHSI);
6347 return ReplaceInstUsesWith(I, ConstantInt::get(Type::getInt1Ty(*Context),
6348 !I.isTrueWhenEqual()));
6354 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6355 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6356 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6358 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6359 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6360 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6363 // Test to see if the operands of the icmp are casted versions of other
6364 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6366 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6367 if (isa<PointerType>(Op0->getType()) &&
6368 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6369 // We keep moving the cast from the left operand over to the right
6370 // operand, where it can often be eliminated completely.
6371 Op0 = CI->getOperand(0);
6373 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6374 // so eliminate it as well.
6375 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6376 Op1 = CI2->getOperand(0);
6378 // If Op1 is a constant, we can fold the cast into the constant.
6379 if (Op0->getType() != Op1->getType()) {
6380 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6381 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6383 // Otherwise, cast the RHS right before the icmp
6384 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6387 return new ICmpInst(I.getPredicate(), Op0, Op1);
6391 if (isa<CastInst>(Op0)) {
6392 // Handle the special case of: icmp (cast bool to X), <cst>
6393 // This comes up when you have code like
6396 // For generality, we handle any zero-extension of any operand comparison
6397 // with a constant or another cast from the same type.
6398 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6399 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6403 // See if it's the same type of instruction on the left and right.
6404 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6405 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6406 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6407 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6408 switch (Op0I->getOpcode()) {
6410 case Instruction::Add:
6411 case Instruction::Sub:
6412 case Instruction::Xor:
6413 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6414 return new ICmpInst(I.getPredicate(), Op0I->getOperand(0),
6415 Op1I->getOperand(0));
6416 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6417 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6418 if (CI->getValue().isSignBit()) {
6419 ICmpInst::Predicate Pred = I.isSignedPredicate()
6420 ? I.getUnsignedPredicate()
6421 : I.getSignedPredicate();
6422 return new ICmpInst(Pred, Op0I->getOperand(0),
6423 Op1I->getOperand(0));
6426 if (CI->getValue().isMaxSignedValue()) {
6427 ICmpInst::Predicate Pred = I.isSignedPredicate()
6428 ? I.getUnsignedPredicate()
6429 : I.getSignedPredicate();
6430 Pred = I.getSwappedPredicate(Pred);
6431 return new ICmpInst(Pred, Op0I->getOperand(0),
6432 Op1I->getOperand(0));
6436 case Instruction::Mul:
6437 if (!I.isEquality())
6440 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6441 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6442 // Mask = -1 >> count-trailing-zeros(Cst).
6443 if (!CI->isZero() && !CI->isOne()) {
6444 const APInt &AP = CI->getValue();
6445 ConstantInt *Mask = ConstantInt::get(*Context,
6446 APInt::getLowBitsSet(AP.getBitWidth(),
6448 AP.countTrailingZeros()));
6449 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6451 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6453 InsertNewInstBefore(And1, I);
6454 InsertNewInstBefore(And2, I);
6455 return new ICmpInst(I.getPredicate(), And1, And2);
6464 // ~x < ~y --> y < x
6466 if (match(Op0, m_Not(m_Value(A))) &&
6467 match(Op1, m_Not(m_Value(B))))
6468 return new ICmpInst(I.getPredicate(), B, A);
6471 if (I.isEquality()) {
6472 Value *A, *B, *C, *D;
6474 // -x == -y --> x == y
6475 if (match(Op0, m_Neg(m_Value(A))) &&
6476 match(Op1, m_Neg(m_Value(B))))
6477 return new ICmpInst(I.getPredicate(), A, B);
6479 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6480 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6481 Value *OtherVal = A == Op1 ? B : A;
6482 return new ICmpInst(I.getPredicate(), OtherVal,
6483 Constant::getNullValue(A->getType()));
6486 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6487 // A^c1 == C^c2 --> A == C^(c1^c2)
6488 ConstantInt *C1, *C2;
6489 if (match(B, m_ConstantInt(C1)) &&
6490 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6492 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6493 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6494 return new ICmpInst(I.getPredicate(), A,
6495 InsertNewInstBefore(Xor, I));
6498 // A^B == A^D -> B == D
6499 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
6500 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
6501 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
6502 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
6506 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6507 (A == Op0 || B == Op0)) {
6508 // A == (A^B) -> B == 0
6509 Value *OtherVal = A == Op0 ? B : A;
6510 return new ICmpInst(I.getPredicate(), OtherVal,
6511 Constant::getNullValue(A->getType()));
6514 // (A-B) == A -> B == 0
6515 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6516 return new ICmpInst(I.getPredicate(), B,
6517 Constant::getNullValue(B->getType()));
6519 // A == (A-B) -> B == 0
6520 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6521 return new ICmpInst(I.getPredicate(), B,
6522 Constant::getNullValue(B->getType()));
6524 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6525 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6526 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6527 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6528 Value *X = 0, *Y = 0, *Z = 0;
6531 X = B; Y = D; Z = A;
6532 } else if (A == D) {
6533 X = B; Y = C; Z = A;
6534 } else if (B == C) {
6535 X = A; Y = D; Z = B;
6536 } else if (B == D) {
6537 X = A; Y = C; Z = B;
6540 if (X) { // Build (X^Y) & Z
6541 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6542 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6543 I.setOperand(0, Op1);
6544 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6549 return Changed ? &I : 0;
6553 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6554 /// and CmpRHS are both known to be integer constants.
6555 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6556 ConstantInt *DivRHS) {
6557 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6558 const APInt &CmpRHSV = CmpRHS->getValue();
6560 // FIXME: If the operand types don't match the type of the divide
6561 // then don't attempt this transform. The code below doesn't have the
6562 // logic to deal with a signed divide and an unsigned compare (and
6563 // vice versa). This is because (x /s C1) <s C2 produces different
6564 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6565 // (x /u C1) <u C2. Simply casting the operands and result won't
6566 // work. :( The if statement below tests that condition and bails
6568 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6569 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6571 if (DivRHS->isZero())
6572 return 0; // The ProdOV computation fails on divide by zero.
6573 if (DivIsSigned && DivRHS->isAllOnesValue())
6574 return 0; // The overflow computation also screws up here
6575 if (DivRHS->isOne())
6576 return 0; // Not worth bothering, and eliminates some funny cases
6579 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6580 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6581 // C2 (CI). By solving for X we can turn this into a range check
6582 // instead of computing a divide.
6583 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6585 // Determine if the product overflows by seeing if the product is
6586 // not equal to the divide. Make sure we do the same kind of divide
6587 // as in the LHS instruction that we're folding.
6588 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6589 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6591 // Get the ICmp opcode
6592 ICmpInst::Predicate Pred = ICI.getPredicate();
6594 // Figure out the interval that is being checked. For example, a comparison
6595 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6596 // Compute this interval based on the constants involved and the signedness of
6597 // the compare/divide. This computes a half-open interval, keeping track of
6598 // whether either value in the interval overflows. After analysis each
6599 // overflow variable is set to 0 if it's corresponding bound variable is valid
6600 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6601 int LoOverflow = 0, HiOverflow = 0;
6602 Constant *LoBound = 0, *HiBound = 0;
6604 if (!DivIsSigned) { // udiv
6605 // e.g. X/5 op 3 --> [15, 20)
6607 HiOverflow = LoOverflow = ProdOV;
6609 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6610 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6611 if (CmpRHSV == 0) { // (X / pos) op 0
6612 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6613 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6615 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6616 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6617 HiOverflow = LoOverflow = ProdOV;
6619 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6620 } else { // (X / pos) op neg
6621 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6622 HiBound = AddOne(Prod);
6623 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6625 ConstantInt* DivNeg =
6626 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6627 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6631 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6632 if (CmpRHSV == 0) { // (X / neg) op 0
6633 // e.g. X/-5 op 0 --> [-4, 5)
6634 LoBound = AddOne(DivRHS);
6635 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6636 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6637 HiOverflow = 1; // [INTMIN+1, overflow)
6638 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6640 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6641 // e.g. X/-5 op 3 --> [-19, -14)
6642 HiBound = AddOne(Prod);
6643 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6645 LoOverflow = AddWithOverflow(LoBound, HiBound,
6646 DivRHS, Context, true) ? -1 : 0;
6647 } else { // (X / neg) op neg
6648 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6649 LoOverflow = HiOverflow = ProdOV;
6651 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6654 // Dividing by a negative swaps the condition. LT <-> GT
6655 Pred = ICmpInst::getSwappedPredicate(Pred);
6658 Value *X = DivI->getOperand(0);
6660 default: llvm_unreachable("Unhandled icmp opcode!");
6661 case ICmpInst::ICMP_EQ:
6662 if (LoOverflow && HiOverflow)
6663 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6664 else if (HiOverflow)
6665 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6666 ICmpInst::ICMP_UGE, X, LoBound);
6667 else if (LoOverflow)
6668 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6669 ICmpInst::ICMP_ULT, X, HiBound);
6671 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6672 case ICmpInst::ICMP_NE:
6673 if (LoOverflow && HiOverflow)
6674 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6675 else if (HiOverflow)
6676 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
6677 ICmpInst::ICMP_ULT, X, LoBound);
6678 else if (LoOverflow)
6679 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
6680 ICmpInst::ICMP_UGE, X, HiBound);
6682 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6683 case ICmpInst::ICMP_ULT:
6684 case ICmpInst::ICMP_SLT:
6685 if (LoOverflow == +1) // Low bound is greater than input range.
6686 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6687 if (LoOverflow == -1) // Low bound is less than input range.
6688 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6689 return new ICmpInst(Pred, X, LoBound);
6690 case ICmpInst::ICMP_UGT:
6691 case ICmpInst::ICMP_SGT:
6692 if (HiOverflow == +1) // High bound greater than input range.
6693 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6694 else if (HiOverflow == -1) // High bound less than input range.
6695 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6696 if (Pred == ICmpInst::ICMP_UGT)
6697 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
6699 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
6704 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6706 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6709 const APInt &RHSV = RHS->getValue();
6711 switch (LHSI->getOpcode()) {
6712 case Instruction::Trunc:
6713 if (ICI.isEquality() && LHSI->hasOneUse()) {
6714 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6715 // of the high bits truncated out of x are known.
6716 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6717 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6718 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6719 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6720 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6722 // If all the high bits are known, we can do this xform.
6723 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6724 // Pull in the high bits from known-ones set.
6725 APInt NewRHS(RHS->getValue());
6726 NewRHS.zext(SrcBits);
6728 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6729 ConstantInt::get(*Context, NewRHS));
6734 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6735 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6736 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6738 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6739 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6740 Value *CompareVal = LHSI->getOperand(0);
6742 // If the sign bit of the XorCST is not set, there is no change to
6743 // the operation, just stop using the Xor.
6744 if (!XorCST->getValue().isNegative()) {
6745 ICI.setOperand(0, CompareVal);
6746 AddToWorkList(LHSI);
6750 // Was the old condition true if the operand is positive?
6751 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6753 // If so, the new one isn't.
6754 isTrueIfPositive ^= true;
6756 if (isTrueIfPositive)
6757 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal,
6760 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal,
6764 if (LHSI->hasOneUse()) {
6765 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6766 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6767 const APInt &SignBit = XorCST->getValue();
6768 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6769 ? ICI.getUnsignedPredicate()
6770 : ICI.getSignedPredicate();
6771 return new ICmpInst(Pred, LHSI->getOperand(0),
6772 ConstantInt::get(*Context, RHSV ^ SignBit));
6775 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6776 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6777 const APInt &NotSignBit = XorCST->getValue();
6778 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6779 ? ICI.getUnsignedPredicate()
6780 : ICI.getSignedPredicate();
6781 Pred = ICI.getSwappedPredicate(Pred);
6782 return new ICmpInst(Pred, LHSI->getOperand(0),
6783 ConstantInt::get(*Context, RHSV ^ NotSignBit));
6788 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6789 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6790 LHSI->getOperand(0)->hasOneUse()) {
6791 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6793 // If the LHS is an AND of a truncating cast, we can widen the
6794 // and/compare to be the input width without changing the value
6795 // produced, eliminating a cast.
6796 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6797 // We can do this transformation if either the AND constant does not
6798 // have its sign bit set or if it is an equality comparison.
6799 // Extending a relational comparison when we're checking the sign
6800 // bit would not work.
6801 if (Cast->hasOneUse() &&
6802 (ICI.isEquality() ||
6803 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6805 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6806 APInt NewCST = AndCST->getValue();
6807 NewCST.zext(BitWidth);
6809 NewCI.zext(BitWidth);
6810 Instruction *NewAnd =
6811 BinaryOperator::CreateAnd(Cast->getOperand(0),
6812 ConstantInt::get(*Context, NewCST), LHSI->getName());
6813 InsertNewInstBefore(NewAnd, ICI);
6814 return new ICmpInst(ICI.getPredicate(), NewAnd,
6815 ConstantInt::get(*Context, NewCI));
6819 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6820 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6821 // happens a LOT in code produced by the C front-end, for bitfield
6823 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6824 if (Shift && !Shift->isShift())
6828 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6829 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6830 const Type *AndTy = AndCST->getType(); // Type of the and.
6832 // We can fold this as long as we can't shift unknown bits
6833 // into the mask. This can only happen with signed shift
6834 // rights, as they sign-extend.
6836 bool CanFold = Shift->isLogicalShift();
6838 // To test for the bad case of the signed shr, see if any
6839 // of the bits shifted in could be tested after the mask.
6840 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6841 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6843 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6844 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6845 AndCST->getValue()) == 0)
6851 if (Shift->getOpcode() == Instruction::Shl)
6852 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6854 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6856 // Check to see if we are shifting out any of the bits being
6858 if (ConstantExpr::get(Shift->getOpcode(),
6859 NewCst, ShAmt) != RHS) {
6860 // If we shifted bits out, the fold is not going to work out.
6861 // As a special case, check to see if this means that the
6862 // result is always true or false now.
6863 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6864 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6865 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6866 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6868 ICI.setOperand(1, NewCst);
6869 Constant *NewAndCST;
6870 if (Shift->getOpcode() == Instruction::Shl)
6871 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6873 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6874 LHSI->setOperand(1, NewAndCST);
6875 LHSI->setOperand(0, Shift->getOperand(0));
6876 AddToWorkList(Shift); // Shift is dead.
6877 AddUsesToWorkList(ICI);
6883 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6884 // preferable because it allows the C<<Y expression to be hoisted out
6885 // of a loop if Y is invariant and X is not.
6886 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6887 ICI.isEquality() && !Shift->isArithmeticShift() &&
6888 !isa<Constant>(Shift->getOperand(0))) {
6891 if (Shift->getOpcode() == Instruction::LShr) {
6892 NS = BinaryOperator::CreateShl(AndCST,
6893 Shift->getOperand(1), "tmp");
6895 // Insert a logical shift.
6896 NS = BinaryOperator::CreateLShr(AndCST,
6897 Shift->getOperand(1), "tmp");
6899 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6901 // Compute X & (C << Y).
6902 Instruction *NewAnd =
6903 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6904 InsertNewInstBefore(NewAnd, ICI);
6906 ICI.setOperand(0, NewAnd);
6912 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6913 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6916 uint32_t TypeBits = RHSV.getBitWidth();
6918 // Check that the shift amount is in range. If not, don't perform
6919 // undefined shifts. When the shift is visited it will be
6921 if (ShAmt->uge(TypeBits))
6924 if (ICI.isEquality()) {
6925 // If we are comparing against bits always shifted out, the
6926 // comparison cannot succeed.
6928 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
6930 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6931 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6932 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6933 return ReplaceInstUsesWith(ICI, Cst);
6936 if (LHSI->hasOneUse()) {
6937 // Otherwise strength reduce the shift into an and.
6938 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6940 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
6941 TypeBits-ShAmtVal));
6944 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6945 Mask, LHSI->getName()+".mask");
6946 Value *And = InsertNewInstBefore(AndI, ICI);
6947 return new ICmpInst(ICI.getPredicate(), And,
6948 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
6952 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6953 bool TrueIfSigned = false;
6954 if (LHSI->hasOneUse() &&
6955 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6956 // (X << 31) <s 0 --> (X&1) != 0
6957 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
6958 (TypeBits-ShAmt->getZExtValue()-1));
6960 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6961 Mask, LHSI->getName()+".mask");
6962 Value *And = InsertNewInstBefore(AndI, ICI);
6964 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6965 And, Constant::getNullValue(And->getType()));
6970 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6971 case Instruction::AShr: {
6972 // Only handle equality comparisons of shift-by-constant.
6973 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6974 if (!ShAmt || !ICI.isEquality()) break;
6976 // Check that the shift amount is in range. If not, don't perform
6977 // undefined shifts. When the shift is visited it will be
6979 uint32_t TypeBits = RHSV.getBitWidth();
6980 if (ShAmt->uge(TypeBits))
6983 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6985 // If we are comparing against bits always shifted out, the
6986 // comparison cannot succeed.
6987 APInt Comp = RHSV << ShAmtVal;
6988 if (LHSI->getOpcode() == Instruction::LShr)
6989 Comp = Comp.lshr(ShAmtVal);
6991 Comp = Comp.ashr(ShAmtVal);
6993 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6994 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6995 Constant *Cst = ConstantInt::get(Type::getInt1Ty(*Context), IsICMP_NE);
6996 return ReplaceInstUsesWith(ICI, Cst);
6999 // Otherwise, check to see if the bits shifted out are known to be zero.
7000 // If so, we can compare against the unshifted value:
7001 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
7002 if (LHSI->hasOneUse() &&
7003 MaskedValueIsZero(LHSI->getOperand(0),
7004 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
7005 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
7006 ConstantExpr::getShl(RHS, ShAmt));
7009 if (LHSI->hasOneUse()) {
7010 // Otherwise strength reduce the shift into an and.
7011 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7012 Constant *Mask = ConstantInt::get(*Context, Val);
7015 BinaryOperator::CreateAnd(LHSI->getOperand(0),
7016 Mask, LHSI->getName()+".mask");
7017 Value *And = InsertNewInstBefore(AndI, ICI);
7018 return new ICmpInst(ICI.getPredicate(), And,
7019 ConstantExpr::getShl(RHS, ShAmt));
7024 case Instruction::SDiv:
7025 case Instruction::UDiv:
7026 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7027 // Fold this div into the comparison, producing a range check.
7028 // Determine, based on the divide type, what the range is being
7029 // checked. If there is an overflow on the low or high side, remember
7030 // it, otherwise compute the range [low, hi) bounding the new value.
7031 // See: InsertRangeTest above for the kinds of replacements possible.
7032 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7033 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7038 case Instruction::Add:
7039 // Fold: icmp pred (add, X, C1), C2
7041 if (!ICI.isEquality()) {
7042 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7044 const APInt &LHSV = LHSC->getValue();
7046 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7049 if (ICI.isSignedPredicate()) {
7050 if (CR.getLower().isSignBit()) {
7051 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7052 ConstantInt::get(*Context, CR.getUpper()));
7053 } else if (CR.getUpper().isSignBit()) {
7054 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7055 ConstantInt::get(*Context, CR.getLower()));
7058 if (CR.getLower().isMinValue()) {
7059 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7060 ConstantInt::get(*Context, CR.getUpper()));
7061 } else if (CR.getUpper().isMinValue()) {
7062 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7063 ConstantInt::get(*Context, CR.getLower()));
7070 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7071 if (ICI.isEquality()) {
7072 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7074 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7075 // the second operand is a constant, simplify a bit.
7076 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7077 switch (BO->getOpcode()) {
7078 case Instruction::SRem:
7079 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7080 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7081 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7082 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7083 Instruction *NewRem =
7084 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
7086 InsertNewInstBefore(NewRem, ICI);
7087 return new ICmpInst(ICI.getPredicate(), NewRem,
7088 Constant::getNullValue(BO->getType()));
7092 case Instruction::Add:
7093 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7094 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7095 if (BO->hasOneUse())
7096 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7097 ConstantExpr::getSub(RHS, BOp1C));
7098 } else if (RHSV == 0) {
7099 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7100 // efficiently invertible, or if the add has just this one use.
7101 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7103 if (Value *NegVal = dyn_castNegVal(BOp1))
7104 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
7105 else if (Value *NegVal = dyn_castNegVal(BOp0))
7106 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
7107 else if (BO->hasOneUse()) {
7108 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
7109 InsertNewInstBefore(Neg, ICI);
7111 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
7115 case Instruction::Xor:
7116 // For the xor case, we can xor two constants together, eliminating
7117 // the explicit xor.
7118 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7119 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7120 ConstantExpr::getXor(RHS, BOC));
7123 case Instruction::Sub:
7124 // Replace (([sub|xor] A, B) != 0) with (A != B)
7126 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
7130 case Instruction::Or:
7131 // If bits are being or'd in that are not present in the constant we
7132 // are comparing against, then the comparison could never succeed!
7133 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7134 Constant *NotCI = ConstantExpr::getNot(RHS);
7135 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7136 return ReplaceInstUsesWith(ICI,
7137 ConstantInt::get(Type::getInt1Ty(*Context),
7142 case Instruction::And:
7143 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7144 // If bits are being compared against that are and'd out, then the
7145 // comparison can never succeed!
7146 if ((RHSV & ~BOC->getValue()) != 0)
7147 return ReplaceInstUsesWith(ICI,
7148 ConstantInt::get(Type::getInt1Ty(*Context),
7151 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7152 if (RHS == BOC && RHSV.isPowerOf2())
7153 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
7154 ICmpInst::ICMP_NE, LHSI,
7155 Constant::getNullValue(RHS->getType()));
7157 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7158 if (BOC->getValue().isSignBit()) {
7159 Value *X = BO->getOperand(0);
7160 Constant *Zero = Constant::getNullValue(X->getType());
7161 ICmpInst::Predicate pred = isICMP_NE ?
7162 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7163 return new ICmpInst(pred, X, Zero);
7166 // ((X & ~7) == 0) --> X < 8
7167 if (RHSV == 0 && isHighOnes(BOC)) {
7168 Value *X = BO->getOperand(0);
7169 Constant *NegX = ConstantExpr::getNeg(BOC);
7170 ICmpInst::Predicate pred = isICMP_NE ?
7171 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7172 return new ICmpInst(pred, X, NegX);
7177 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7178 // Handle icmp {eq|ne} <intrinsic>, intcst.
7179 if (II->getIntrinsicID() == Intrinsic::bswap) {
7181 ICI.setOperand(0, II->getOperand(1));
7182 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7190 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7191 /// We only handle extending casts so far.
7193 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7194 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7195 Value *LHSCIOp = LHSCI->getOperand(0);
7196 const Type *SrcTy = LHSCIOp->getType();
7197 const Type *DestTy = LHSCI->getType();
7200 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7201 // integer type is the same size as the pointer type.
7202 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7203 TD->getPointerSizeInBits() ==
7204 cast<IntegerType>(DestTy)->getBitWidth()) {
7206 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7207 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7208 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7209 RHSOp = RHSC->getOperand(0);
7210 // If the pointer types don't match, insert a bitcast.
7211 if (LHSCIOp->getType() != RHSOp->getType())
7212 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
7216 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
7219 // The code below only handles extension cast instructions, so far.
7221 if (LHSCI->getOpcode() != Instruction::ZExt &&
7222 LHSCI->getOpcode() != Instruction::SExt)
7225 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7226 bool isSignedCmp = ICI.isSignedPredicate();
7228 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7229 // Not an extension from the same type?
7230 RHSCIOp = CI->getOperand(0);
7231 if (RHSCIOp->getType() != LHSCIOp->getType())
7234 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7235 // and the other is a zext), then we can't handle this.
7236 if (CI->getOpcode() != LHSCI->getOpcode())
7239 // Deal with equality cases early.
7240 if (ICI.isEquality())
7241 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7243 // A signed comparison of sign extended values simplifies into a
7244 // signed comparison.
7245 if (isSignedCmp && isSignedExt)
7246 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
7248 // The other three cases all fold into an unsigned comparison.
7249 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7252 // If we aren't dealing with a constant on the RHS, exit early
7253 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7257 // Compute the constant that would happen if we truncated to SrcTy then
7258 // reextended to DestTy.
7259 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7260 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7263 // If the re-extended constant didn't change...
7265 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7266 // For example, we might have:
7267 // %A = sext i16 %X to i32
7268 // %B = icmp ugt i32 %A, 1330
7269 // It is incorrect to transform this into
7270 // %B = icmp ugt i16 %X, 1330
7271 // because %A may have negative value.
7273 // However, we allow this when the compare is EQ/NE, because they are
7275 if (isSignedExt == isSignedCmp || ICI.isEquality())
7276 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
7280 // The re-extended constant changed so the constant cannot be represented
7281 // in the shorter type. Consequently, we cannot emit a simple comparison.
7283 // First, handle some easy cases. We know the result cannot be equal at this
7284 // point so handle the ICI.isEquality() cases
7285 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7286 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7287 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7288 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7290 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7291 // should have been folded away previously and not enter in here.
7294 // We're performing a signed comparison.
7295 if (cast<ConstantInt>(CI)->getValue().isNegative())
7296 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7298 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7300 // We're performing an unsigned comparison.
7302 // We're performing an unsigned comp with a sign extended value.
7303 // This is true if the input is >= 0. [aka >s -1]
7304 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7305 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT,
7306 LHSCIOp, NegOne, ICI.getName()), ICI);
7308 // Unsigned extend & unsigned compare -> always true.
7309 Result = ConstantInt::getTrue(*Context);
7313 // Finally, return the value computed.
7314 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7315 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7316 return ReplaceInstUsesWith(ICI, Result);
7318 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7319 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7320 "ICmp should be folded!");
7321 if (Constant *CI = dyn_cast<Constant>(Result))
7322 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7323 return BinaryOperator::CreateNot(Result);
7326 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7327 return commonShiftTransforms(I);
7330 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7331 return commonShiftTransforms(I);
7334 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7335 if (Instruction *R = commonShiftTransforms(I))
7338 Value *Op0 = I.getOperand(0);
7340 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7341 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7342 if (CSI->isAllOnesValue())
7343 return ReplaceInstUsesWith(I, CSI);
7345 // See if we can turn a signed shr into an unsigned shr.
7346 if (MaskedValueIsZero(Op0,
7347 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7348 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7350 // Arithmetic shifting an all-sign-bit value is a no-op.
7351 unsigned NumSignBits = ComputeNumSignBits(Op0);
7352 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7353 return ReplaceInstUsesWith(I, Op0);
7358 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7359 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7360 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7362 // shl X, 0 == X and shr X, 0 == X
7363 // shl 0, X == 0 and shr 0, X == 0
7364 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7365 Op0 == Constant::getNullValue(Op0->getType()))
7366 return ReplaceInstUsesWith(I, Op0);
7368 if (isa<UndefValue>(Op0)) {
7369 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7370 return ReplaceInstUsesWith(I, Op0);
7371 else // undef << X -> 0, undef >>u X -> 0
7372 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7374 if (isa<UndefValue>(Op1)) {
7375 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7376 return ReplaceInstUsesWith(I, Op0);
7377 else // X << undef, X >>u undef -> 0
7378 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7381 // See if we can fold away this shift.
7382 if (SimplifyDemandedInstructionBits(I))
7385 // Try to fold constant and into select arguments.
7386 if (isa<Constant>(Op0))
7387 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7388 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7391 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7392 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7397 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7398 BinaryOperator &I) {
7399 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7401 // See if we can simplify any instructions used by the instruction whose sole
7402 // purpose is to compute bits we don't care about.
7403 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7405 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7408 if (Op1->uge(TypeBits)) {
7409 if (I.getOpcode() != Instruction::AShr)
7410 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7412 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7417 // ((X*C1) << C2) == (X * (C1 << C2))
7418 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7419 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7420 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7421 return BinaryOperator::CreateMul(BO->getOperand(0),
7422 ConstantExpr::getShl(BOOp, Op1));
7424 // Try to fold constant and into select arguments.
7425 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7426 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7428 if (isa<PHINode>(Op0))
7429 if (Instruction *NV = FoldOpIntoPhi(I))
7432 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7433 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7434 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7435 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7436 // place. Don't try to do this transformation in this case. Also, we
7437 // require that the input operand is a shift-by-constant so that we have
7438 // confidence that the shifts will get folded together. We could do this
7439 // xform in more cases, but it is unlikely to be profitable.
7440 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7441 isa<ConstantInt>(TrOp->getOperand(1))) {
7442 // Okay, we'll do this xform. Make the shift of shift.
7443 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7444 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7446 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7448 // For logical shifts, the truncation has the effect of making the high
7449 // part of the register be zeros. Emulate this by inserting an AND to
7450 // clear the top bits as needed. This 'and' will usually be zapped by
7451 // other xforms later if dead.
7452 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7453 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7454 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7456 // The mask we constructed says what the trunc would do if occurring
7457 // between the shifts. We want to know the effect *after* the second
7458 // shift. We know that it is a logical shift by a constant, so adjust the
7459 // mask as appropriate.
7460 if (I.getOpcode() == Instruction::Shl)
7461 MaskV <<= Op1->getZExtValue();
7463 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7464 MaskV = MaskV.lshr(Op1->getZExtValue());
7468 BinaryOperator::CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7470 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7472 // Return the value truncated to the interesting size.
7473 return new TruncInst(And, I.getType());
7477 if (Op0->hasOneUse()) {
7478 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7479 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7482 switch (Op0BO->getOpcode()) {
7484 case Instruction::Add:
7485 case Instruction::And:
7486 case Instruction::Or:
7487 case Instruction::Xor: {
7488 // These operators commute.
7489 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7490 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7491 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7493 Instruction *YS = BinaryOperator::CreateShl(
7494 Op0BO->getOperand(0), Op1,
7496 InsertNewInstBefore(YS, I); // (Y << C)
7498 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7499 Op0BO->getOperand(1)->getName());
7500 InsertNewInstBefore(X, I); // (X + (Y << C))
7501 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7502 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7503 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7506 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7507 Value *Op0BOOp1 = Op0BO->getOperand(1);
7508 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7510 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7511 m_ConstantInt(CC))) &&
7512 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7513 Instruction *YS = BinaryOperator::CreateShl(
7514 Op0BO->getOperand(0), Op1,
7516 InsertNewInstBefore(YS, I); // (Y << C)
7518 BinaryOperator::CreateAnd(V1,
7519 ConstantExpr::getShl(CC, Op1),
7520 V1->getName()+".mask");
7521 InsertNewInstBefore(XM, I); // X & (CC << C)
7523 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7528 case Instruction::Sub: {
7529 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7530 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7531 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7532 m_Specific(Op1)))) {
7533 Instruction *YS = BinaryOperator::CreateShl(
7534 Op0BO->getOperand(1), Op1,
7536 InsertNewInstBefore(YS, I); // (Y << C)
7538 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7539 Op0BO->getOperand(0)->getName());
7540 InsertNewInstBefore(X, I); // (X + (Y << C))
7541 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7542 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7543 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7546 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7547 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7548 match(Op0BO->getOperand(0),
7549 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7550 m_ConstantInt(CC))) && V2 == Op1 &&
7551 cast<BinaryOperator>(Op0BO->getOperand(0))
7552 ->getOperand(0)->hasOneUse()) {
7553 Instruction *YS = BinaryOperator::CreateShl(
7554 Op0BO->getOperand(1), Op1,
7556 InsertNewInstBefore(YS, I); // (Y << C)
7558 BinaryOperator::CreateAnd(V1,
7559 ConstantExpr::getShl(CC, Op1),
7560 V1->getName()+".mask");
7561 InsertNewInstBefore(XM, I); // X & (CC << C)
7563 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7571 // If the operand is an bitwise operator with a constant RHS, and the
7572 // shift is the only use, we can pull it out of the shift.
7573 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7574 bool isValid = true; // Valid only for And, Or, Xor
7575 bool highBitSet = false; // Transform if high bit of constant set?
7577 switch (Op0BO->getOpcode()) {
7578 default: isValid = false; break; // Do not perform transform!
7579 case Instruction::Add:
7580 isValid = isLeftShift;
7582 case Instruction::Or:
7583 case Instruction::Xor:
7586 case Instruction::And:
7591 // If this is a signed shift right, and the high bit is modified
7592 // by the logical operation, do not perform the transformation.
7593 // The highBitSet boolean indicates the value of the high bit of
7594 // the constant which would cause it to be modified for this
7597 if (isValid && I.getOpcode() == Instruction::AShr)
7598 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7601 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7603 Instruction *NewShift =
7604 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7605 InsertNewInstBefore(NewShift, I);
7606 NewShift->takeName(Op0BO);
7608 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7615 // Find out if this is a shift of a shift by a constant.
7616 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7617 if (ShiftOp && !ShiftOp->isShift())
7620 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7621 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7622 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7623 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7624 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7625 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7626 Value *X = ShiftOp->getOperand(0);
7628 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7630 const IntegerType *Ty = cast<IntegerType>(I.getType());
7632 // Check for (X << c1) << c2 and (X >> c1) >> c2
7633 if (I.getOpcode() == ShiftOp->getOpcode()) {
7634 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7636 if (AmtSum >= TypeBits) {
7637 if (I.getOpcode() != Instruction::AShr)
7638 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7639 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7642 return BinaryOperator::Create(I.getOpcode(), X,
7643 ConstantInt::get(Ty, AmtSum));
7644 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7645 I.getOpcode() == Instruction::AShr) {
7646 if (AmtSum >= TypeBits)
7647 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7649 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7650 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7651 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7652 I.getOpcode() == Instruction::LShr) {
7653 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7654 if (AmtSum >= TypeBits)
7655 AmtSum = TypeBits-1;
7657 Instruction *Shift =
7658 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7659 InsertNewInstBefore(Shift, I);
7661 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7662 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7665 // Okay, if we get here, one shift must be left, and the other shift must be
7666 // right. See if the amounts are equal.
7667 if (ShiftAmt1 == ShiftAmt2) {
7668 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7669 if (I.getOpcode() == Instruction::Shl) {
7670 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7671 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7673 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7674 if (I.getOpcode() == Instruction::LShr) {
7675 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7676 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7678 // We can simplify ((X << C) >>s C) into a trunc + sext.
7679 // NOTE: we could do this for any C, but that would make 'unusual' integer
7680 // types. For now, just stick to ones well-supported by the code
7682 const Type *SExtType = 0;
7683 switch (Ty->getBitWidth() - ShiftAmt1) {
7690 SExtType = IntegerType::get(*Context, Ty->getBitWidth() - ShiftAmt1);
7695 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7696 InsertNewInstBefore(NewTrunc, I);
7697 return new SExtInst(NewTrunc, Ty);
7699 // Otherwise, we can't handle it yet.
7700 } else if (ShiftAmt1 < ShiftAmt2) {
7701 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7703 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7704 if (I.getOpcode() == Instruction::Shl) {
7705 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7706 ShiftOp->getOpcode() == Instruction::AShr);
7707 Instruction *Shift =
7708 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7709 InsertNewInstBefore(Shift, I);
7711 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7712 return BinaryOperator::CreateAnd(Shift,
7713 ConstantInt::get(*Context, Mask));
7716 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7717 if (I.getOpcode() == Instruction::LShr) {
7718 assert(ShiftOp->getOpcode() == Instruction::Shl);
7719 Instruction *Shift =
7720 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7721 InsertNewInstBefore(Shift, I);
7723 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7724 return BinaryOperator::CreateAnd(Shift,
7725 ConstantInt::get(*Context, Mask));
7728 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7730 assert(ShiftAmt2 < ShiftAmt1);
7731 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7733 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7734 if (I.getOpcode() == Instruction::Shl) {
7735 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7736 ShiftOp->getOpcode() == Instruction::AShr);
7737 Instruction *Shift =
7738 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7739 ConstantInt::get(Ty, ShiftDiff));
7740 InsertNewInstBefore(Shift, I);
7742 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7743 return BinaryOperator::CreateAnd(Shift,
7744 ConstantInt::get(*Context, Mask));
7747 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7748 if (I.getOpcode() == Instruction::LShr) {
7749 assert(ShiftOp->getOpcode() == Instruction::Shl);
7750 Instruction *Shift =
7751 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7752 InsertNewInstBefore(Shift, I);
7754 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7755 return BinaryOperator::CreateAnd(Shift,
7756 ConstantInt::get(*Context, Mask));
7759 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7766 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7767 /// expression. If so, decompose it, returning some value X, such that Val is
7770 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7771 int &Offset, LLVMContext *Context) {
7772 assert(Val->getType() == Type::getInt32Ty(*Context) && "Unexpected allocation size type!");
7773 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7774 Offset = CI->getZExtValue();
7776 return ConstantInt::get(Type::getInt32Ty(*Context), 0);
7777 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7778 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7779 if (I->getOpcode() == Instruction::Shl) {
7780 // This is a value scaled by '1 << the shift amt'.
7781 Scale = 1U << RHS->getZExtValue();
7783 return I->getOperand(0);
7784 } else if (I->getOpcode() == Instruction::Mul) {
7785 // This value is scaled by 'RHS'.
7786 Scale = RHS->getZExtValue();
7788 return I->getOperand(0);
7789 } else if (I->getOpcode() == Instruction::Add) {
7790 // We have X+C. Check to see if we really have (X*C2)+C1,
7791 // where C1 is divisible by C2.
7794 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7796 Offset += RHS->getZExtValue();
7803 // Otherwise, we can't look past this.
7810 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7811 /// try to eliminate the cast by moving the type information into the alloc.
7812 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7813 AllocationInst &AI) {
7814 const PointerType *PTy = cast<PointerType>(CI.getType());
7816 // Remove any uses of AI that are dead.
7817 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7819 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7820 Instruction *User = cast<Instruction>(*UI++);
7821 if (isInstructionTriviallyDead(User)) {
7822 while (UI != E && *UI == User)
7823 ++UI; // If this instruction uses AI more than once, don't break UI.
7826 DEBUG(errs() << "IC: DCE: " << *User << '\n');
7827 EraseInstFromFunction(*User);
7831 // This requires TargetData to get the alloca alignment and size information.
7834 // Get the type really allocated and the type casted to.
7835 const Type *AllocElTy = AI.getAllocatedType();
7836 const Type *CastElTy = PTy->getElementType();
7837 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7839 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7840 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7841 if (CastElTyAlign < AllocElTyAlign) return 0;
7843 // If the allocation has multiple uses, only promote it if we are strictly
7844 // increasing the alignment of the resultant allocation. If we keep it the
7845 // same, we open the door to infinite loops of various kinds. (A reference
7846 // from a dbg.declare doesn't count as a use for this purpose.)
7847 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7848 CastElTyAlign == AllocElTyAlign) return 0;
7850 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7851 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7852 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7854 // See if we can satisfy the modulus by pulling a scale out of the array
7856 unsigned ArraySizeScale;
7858 Value *NumElements = // See if the array size is a decomposable linear expr.
7859 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7860 ArrayOffset, Context);
7862 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7864 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7865 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7867 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7872 // If the allocation size is constant, form a constant mul expression
7873 Amt = ConstantInt::get(Type::getInt32Ty(*Context), Scale);
7874 if (isa<ConstantInt>(NumElements))
7875 Amt = ConstantExpr::getMul(cast<ConstantInt>(NumElements),
7876 cast<ConstantInt>(Amt));
7877 // otherwise multiply the amount and the number of elements
7879 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7880 Amt = InsertNewInstBefore(Tmp, AI);
7884 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7885 Value *Off = ConstantInt::get(Type::getInt32Ty(*Context), Offset, true);
7886 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7887 Amt = InsertNewInstBefore(Tmp, AI);
7890 AllocationInst *New;
7891 if (isa<MallocInst>(AI))
7892 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7894 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7895 InsertNewInstBefore(New, AI);
7898 // If the allocation has one real use plus a dbg.declare, just remove the
7900 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7901 EraseInstFromFunction(*DI);
7903 // If the allocation has multiple real uses, insert a cast and change all
7904 // things that used it to use the new cast. This will also hack on CI, but it
7906 else if (!AI.hasOneUse()) {
7907 AddUsesToWorkList(AI);
7908 // New is the allocation instruction, pointer typed. AI is the original
7909 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7910 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7911 InsertNewInstBefore(NewCast, AI);
7912 AI.replaceAllUsesWith(NewCast);
7914 return ReplaceInstUsesWith(CI, New);
7917 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7918 /// and return it as type Ty without inserting any new casts and without
7919 /// changing the computed value. This is used by code that tries to decide
7920 /// whether promoting or shrinking integer operations to wider or smaller types
7921 /// will allow us to eliminate a truncate or extend.
7923 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7924 /// extension operation if Ty is larger.
7926 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7927 /// should return true if trunc(V) can be computed by computing V in the smaller
7928 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7929 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7930 /// efficiently truncated.
7932 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7933 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7934 /// the final result.
7935 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7937 int &NumCastsRemoved){
7938 // We can always evaluate constants in another type.
7939 if (isa<Constant>(V))
7942 Instruction *I = dyn_cast<Instruction>(V);
7943 if (!I) return false;
7945 const Type *OrigTy = V->getType();
7947 // If this is an extension or truncate, we can often eliminate it.
7948 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7949 // If this is a cast from the destination type, we can trivially eliminate
7950 // it, and this will remove a cast overall.
7951 if (I->getOperand(0)->getType() == Ty) {
7952 // If the first operand is itself a cast, and is eliminable, do not count
7953 // this as an eliminable cast. We would prefer to eliminate those two
7955 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7961 // We can't extend or shrink something that has multiple uses: doing so would
7962 // require duplicating the instruction in general, which isn't profitable.
7963 if (!I->hasOneUse()) return false;
7965 unsigned Opc = I->getOpcode();
7967 case Instruction::Add:
7968 case Instruction::Sub:
7969 case Instruction::Mul:
7970 case Instruction::And:
7971 case Instruction::Or:
7972 case Instruction::Xor:
7973 // These operators can all arbitrarily be extended or truncated.
7974 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7976 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7979 case Instruction::UDiv:
7980 case Instruction::URem: {
7981 // UDiv and URem can be truncated if all the truncated bits are zero.
7982 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7983 uint32_t BitWidth = Ty->getScalarSizeInBits();
7984 if (BitWidth < OrigBitWidth) {
7985 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7986 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7987 MaskedValueIsZero(I->getOperand(1), Mask)) {
7988 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7990 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7996 case Instruction::Shl:
7997 // If we are truncating the result of this SHL, and if it's a shift of a
7998 // constant amount, we can always perform a SHL in a smaller type.
7999 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8000 uint32_t BitWidth = Ty->getScalarSizeInBits();
8001 if (BitWidth < OrigTy->getScalarSizeInBits() &&
8002 CI->getLimitedValue(BitWidth) < BitWidth)
8003 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8007 case Instruction::LShr:
8008 // If this is a truncate of a logical shr, we can truncate it to a smaller
8009 // lshr iff we know that the bits we would otherwise be shifting in are
8011 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8012 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8013 uint32_t BitWidth = Ty->getScalarSizeInBits();
8014 if (BitWidth < OrigBitWidth &&
8015 MaskedValueIsZero(I->getOperand(0),
8016 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
8017 CI->getLimitedValue(BitWidth) < BitWidth) {
8018 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8023 case Instruction::ZExt:
8024 case Instruction::SExt:
8025 case Instruction::Trunc:
8026 // If this is the same kind of case as our original (e.g. zext+zext), we
8027 // can safely replace it. Note that replacing it does not reduce the number
8028 // of casts in the input.
8032 // sext (zext ty1), ty2 -> zext ty2
8033 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8036 case Instruction::Select: {
8037 SelectInst *SI = cast<SelectInst>(I);
8038 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8040 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8043 case Instruction::PHI: {
8044 // We can change a phi if we can change all operands.
8045 PHINode *PN = cast<PHINode>(I);
8046 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8047 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8053 // TODO: Can handle more cases here.
8060 /// EvaluateInDifferentType - Given an expression that
8061 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8062 /// evaluate the expression.
8063 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8065 if (Constant *C = dyn_cast<Constant>(V))
8066 return ConstantExpr::getIntegerCast(C, Ty,
8067 isSigned /*Sext or ZExt*/);
8069 // Otherwise, it must be an instruction.
8070 Instruction *I = cast<Instruction>(V);
8071 Instruction *Res = 0;
8072 unsigned Opc = I->getOpcode();
8074 case Instruction::Add:
8075 case Instruction::Sub:
8076 case Instruction::Mul:
8077 case Instruction::And:
8078 case Instruction::Or:
8079 case Instruction::Xor:
8080 case Instruction::AShr:
8081 case Instruction::LShr:
8082 case Instruction::Shl:
8083 case Instruction::UDiv:
8084 case Instruction::URem: {
8085 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8086 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8087 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8090 case Instruction::Trunc:
8091 case Instruction::ZExt:
8092 case Instruction::SExt:
8093 // If the source type of the cast is the type we're trying for then we can
8094 // just return the source. There's no need to insert it because it is not
8096 if (I->getOperand(0)->getType() == Ty)
8097 return I->getOperand(0);
8099 // Otherwise, must be the same type of cast, so just reinsert a new one.
8100 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8103 case Instruction::Select: {
8104 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8105 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8106 Res = SelectInst::Create(I->getOperand(0), True, False);
8109 case Instruction::PHI: {
8110 PHINode *OPN = cast<PHINode>(I);
8111 PHINode *NPN = PHINode::Create(Ty);
8112 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8113 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8114 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8120 // TODO: Can handle more cases here.
8121 llvm_unreachable("Unreachable!");
8126 return InsertNewInstBefore(Res, *I);
8129 /// @brief Implement the transforms common to all CastInst visitors.
8130 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8131 Value *Src = CI.getOperand(0);
8133 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8134 // eliminate it now.
8135 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8136 if (Instruction::CastOps opc =
8137 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8138 // The first cast (CSrc) is eliminable so we need to fix up or replace
8139 // the second cast (CI). CSrc will then have a good chance of being dead.
8140 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8144 // If we are casting a select then fold the cast into the select
8145 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8146 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8149 // If we are casting a PHI then fold the cast into the PHI
8150 if (isa<PHINode>(Src))
8151 if (Instruction *NV = FoldOpIntoPhi(CI))
8157 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8158 /// or not there is a sequence of GEP indices into the type that will land us at
8159 /// the specified offset. If so, fill them into NewIndices and return the
8160 /// resultant element type, otherwise return null.
8161 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8162 SmallVectorImpl<Value*> &NewIndices,
8163 const TargetData *TD,
8164 LLVMContext *Context) {
8166 if (!Ty->isSized()) return 0;
8168 // Start with the index over the outer type. Note that the type size
8169 // might be zero (even if the offset isn't zero) if the indexed type
8170 // is something like [0 x {int, int}]
8171 const Type *IntPtrTy = TD->getIntPtrType(*Context);
8172 int64_t FirstIdx = 0;
8173 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8174 FirstIdx = Offset/TySize;
8175 Offset -= FirstIdx*TySize;
8177 // Handle hosts where % returns negative instead of values [0..TySize).
8181 assert(Offset >= 0);
8183 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8186 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8188 // Index into the types. If we fail, set OrigBase to null.
8190 // Indexing into tail padding between struct/array elements.
8191 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8194 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8195 const StructLayout *SL = TD->getStructLayout(STy);
8196 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8197 "Offset must stay within the indexed type");
8199 unsigned Elt = SL->getElementContainingOffset(Offset);
8200 NewIndices.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Elt));
8202 Offset -= SL->getElementOffset(Elt);
8203 Ty = STy->getElementType(Elt);
8204 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8205 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8206 assert(EltSize && "Cannot index into a zero-sized array");
8207 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8209 Ty = AT->getElementType();
8211 // Otherwise, we can't index into the middle of this atomic type, bail.
8219 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8220 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8221 Value *Src = CI.getOperand(0);
8223 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8224 // If casting the result of a getelementptr instruction with no offset, turn
8225 // this into a cast of the original pointer!
8226 if (GEP->hasAllZeroIndices()) {
8227 // Changing the cast operand is usually not a good idea but it is safe
8228 // here because the pointer operand is being replaced with another
8229 // pointer operand so the opcode doesn't need to change.
8231 CI.setOperand(0, GEP->getOperand(0));
8235 // If the GEP has a single use, and the base pointer is a bitcast, and the
8236 // GEP computes a constant offset, see if we can convert these three
8237 // instructions into fewer. This typically happens with unions and other
8238 // non-type-safe code.
8239 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8240 if (GEP->hasAllConstantIndices()) {
8241 // We are guaranteed to get a constant from EmitGEPOffset.
8242 ConstantInt *OffsetV =
8243 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8244 int64_t Offset = OffsetV->getSExtValue();
8246 // Get the base pointer input of the bitcast, and the type it points to.
8247 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8248 const Type *GEPIdxTy =
8249 cast<PointerType>(OrigBase->getType())->getElementType();
8250 SmallVector<Value*, 8> NewIndices;
8251 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8252 // If we were able to index down into an element, create the GEP
8253 // and bitcast the result. This eliminates one bitcast, potentially
8255 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
8257 NewIndices.end(), "");
8258 InsertNewInstBefore(NGEP, CI);
8259 NGEP->takeName(GEP);
8260 if (cast<GEPOperator>(GEP)->isInBounds())
8261 cast<GEPOperator>(NGEP)->setIsInBounds(true);
8263 if (isa<BitCastInst>(CI))
8264 return new BitCastInst(NGEP, CI.getType());
8265 assert(isa<PtrToIntInst>(CI));
8266 return new PtrToIntInst(NGEP, CI.getType());
8272 return commonCastTransforms(CI);
8275 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8276 /// type like i42. We don't want to introduce operations on random non-legal
8277 /// integer types where they don't already exist in the code. In the future,
8278 /// we should consider making this based off target-data, so that 32-bit targets
8279 /// won't get i64 operations etc.
8280 static bool isSafeIntegerType(const Type *Ty) {
8281 switch (Ty->getPrimitiveSizeInBits()) {
8292 /// commonIntCastTransforms - This function implements the common transforms
8293 /// for trunc, zext, and sext.
8294 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8295 if (Instruction *Result = commonCastTransforms(CI))
8298 Value *Src = CI.getOperand(0);
8299 const Type *SrcTy = Src->getType();
8300 const Type *DestTy = CI.getType();
8301 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8302 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8304 // See if we can simplify any instructions used by the LHS whose sole
8305 // purpose is to compute bits we don't care about.
8306 if (SimplifyDemandedInstructionBits(CI))
8309 // If the source isn't an instruction or has more than one use then we
8310 // can't do anything more.
8311 Instruction *SrcI = dyn_cast<Instruction>(Src);
8312 if (!SrcI || !Src->hasOneUse())
8315 // Attempt to propagate the cast into the instruction for int->int casts.
8316 int NumCastsRemoved = 0;
8317 // Only do this if the dest type is a simple type, don't convert the
8318 // expression tree to something weird like i93 unless the source is also
8320 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8321 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8322 CanEvaluateInDifferentType(SrcI, DestTy,
8323 CI.getOpcode(), NumCastsRemoved)) {
8324 // If this cast is a truncate, evaluting in a different type always
8325 // eliminates the cast, so it is always a win. If this is a zero-extension,
8326 // we need to do an AND to maintain the clear top-part of the computation,
8327 // so we require that the input have eliminated at least one cast. If this
8328 // is a sign extension, we insert two new casts (to do the extension) so we
8329 // require that two casts have been eliminated.
8330 bool DoXForm = false;
8331 bool JustReplace = false;
8332 switch (CI.getOpcode()) {
8334 // All the others use floating point so we shouldn't actually
8335 // get here because of the check above.
8336 llvm_unreachable("Unknown cast type");
8337 case Instruction::Trunc:
8340 case Instruction::ZExt: {
8341 DoXForm = NumCastsRemoved >= 1;
8342 if (!DoXForm && 0) {
8343 // If it's unnecessary to issue an AND to clear the high bits, it's
8344 // always profitable to do this xform.
8345 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8346 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8347 if (MaskedValueIsZero(TryRes, Mask))
8348 return ReplaceInstUsesWith(CI, TryRes);
8350 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8351 if (TryI->use_empty())
8352 EraseInstFromFunction(*TryI);
8356 case Instruction::SExt: {
8357 DoXForm = NumCastsRemoved >= 2;
8358 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8359 // If we do not have to emit the truncate + sext pair, then it's always
8360 // profitable to do this xform.
8362 // It's not safe to eliminate the trunc + sext pair if one of the
8363 // eliminated cast is a truncate. e.g.
8364 // t2 = trunc i32 t1 to i16
8365 // t3 = sext i16 t2 to i32
8368 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8369 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8370 if (NumSignBits > (DestBitSize - SrcBitSize))
8371 return ReplaceInstUsesWith(CI, TryRes);
8373 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8374 if (TryI->use_empty())
8375 EraseInstFromFunction(*TryI);
8382 DEBUG(errs() << "ICE: EvaluateInDifferentType converting expression type"
8383 " to avoid cast: " << CI);
8384 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8385 CI.getOpcode() == Instruction::SExt);
8387 // Just replace this cast with the result.
8388 return ReplaceInstUsesWith(CI, Res);
8390 assert(Res->getType() == DestTy);
8391 switch (CI.getOpcode()) {
8392 default: llvm_unreachable("Unknown cast type!");
8393 case Instruction::Trunc:
8394 // Just replace this cast with the result.
8395 return ReplaceInstUsesWith(CI, Res);
8396 case Instruction::ZExt: {
8397 assert(SrcBitSize < DestBitSize && "Not a zext?");
8399 // If the high bits are already zero, just replace this cast with the
8401 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8402 if (MaskedValueIsZero(Res, Mask))
8403 return ReplaceInstUsesWith(CI, Res);
8405 // We need to emit an AND to clear the high bits.
8406 Constant *C = ConstantInt::get(*Context,
8407 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8408 return BinaryOperator::CreateAnd(Res, C);
8410 case Instruction::SExt: {
8411 // If the high bits are already filled with sign bit, just replace this
8412 // cast with the result.
8413 unsigned NumSignBits = ComputeNumSignBits(Res);
8414 if (NumSignBits > (DestBitSize - SrcBitSize))
8415 return ReplaceInstUsesWith(CI, Res);
8417 // We need to emit a cast to truncate, then a cast to sext.
8418 return CastInst::Create(Instruction::SExt,
8419 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8426 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8427 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8429 switch (SrcI->getOpcode()) {
8430 case Instruction::Add:
8431 case Instruction::Mul:
8432 case Instruction::And:
8433 case Instruction::Or:
8434 case Instruction::Xor:
8435 // If we are discarding information, rewrite.
8436 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8437 // Don't insert two casts unless at least one can be eliminated.
8438 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8439 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8440 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8441 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8442 return BinaryOperator::Create(
8443 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8447 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8448 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8449 SrcI->getOpcode() == Instruction::Xor &&
8450 Op1 == ConstantInt::getTrue(*Context) &&
8451 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8452 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8453 return BinaryOperator::CreateXor(New,
8454 ConstantInt::get(CI.getType(), 1));
8458 case Instruction::Shl: {
8459 // Canonicalize trunc inside shl, if we can.
8460 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8461 if (CI && DestBitSize < SrcBitSize &&
8462 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8463 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8464 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8465 return BinaryOperator::CreateShl(Op0c, Op1c);
8473 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8474 if (Instruction *Result = commonIntCastTransforms(CI))
8477 Value *Src = CI.getOperand(0);
8478 const Type *Ty = CI.getType();
8479 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8480 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8482 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8483 if (DestBitWidth == 1) {
8484 Constant *One = ConstantInt::get(Src->getType(), 1);
8485 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8486 Value *Zero = Constant::getNullValue(Src->getType());
8487 return new ICmpInst(ICmpInst::ICMP_NE, Src, Zero);
8490 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8491 ConstantInt *ShAmtV = 0;
8493 if (Src->hasOneUse() &&
8494 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8495 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8497 // Get a mask for the bits shifting in.
8498 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8499 if (MaskedValueIsZero(ShiftOp, Mask)) {
8500 if (ShAmt >= DestBitWidth) // All zeros.
8501 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8503 // Okay, we can shrink this. Truncate the input, then return a new
8505 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8506 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8507 return BinaryOperator::CreateLShr(V1, V2);
8514 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8515 /// in order to eliminate the icmp.
8516 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8518 // If we are just checking for a icmp eq of a single bit and zext'ing it
8519 // to an integer, then shift the bit to the appropriate place and then
8520 // cast to integer to avoid the comparison.
8521 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8522 const APInt &Op1CV = Op1C->getValue();
8524 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8525 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8526 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8527 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8528 if (!DoXform) return ICI;
8530 Value *In = ICI->getOperand(0);
8531 Value *Sh = ConstantInt::get(In->getType(),
8532 In->getType()->getScalarSizeInBits()-1);
8533 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8534 In->getName()+".lobit"),
8536 if (In->getType() != CI.getType())
8537 In = CastInst::CreateIntegerCast(In, CI.getType(),
8538 false/*ZExt*/, "tmp", &CI);
8540 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8541 Constant *One = ConstantInt::get(In->getType(), 1);
8542 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8543 In->getName()+".not"),
8547 return ReplaceInstUsesWith(CI, In);
8552 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8553 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8554 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8555 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8556 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8557 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8558 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8559 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8560 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8561 // This only works for EQ and NE
8562 ICI->isEquality()) {
8563 // If Op1C some other power of two, convert:
8564 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8565 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8566 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8567 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8569 APInt KnownZeroMask(~KnownZero);
8570 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8571 if (!DoXform) return ICI;
8573 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8574 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8575 // (X&4) == 2 --> false
8576 // (X&4) != 2 --> true
8577 Constant *Res = ConstantInt::get(Type::getInt1Ty(*Context), isNE);
8578 Res = ConstantExpr::getZExt(Res, CI.getType());
8579 return ReplaceInstUsesWith(CI, Res);
8582 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8583 Value *In = ICI->getOperand(0);
8585 // Perform a logical shr by shiftamt.
8586 // Insert the shift to put the result in the low bit.
8587 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8588 ConstantInt::get(In->getType(), ShiftAmt),
8589 In->getName()+".lobit"), CI);
8592 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8593 Constant *One = ConstantInt::get(In->getType(), 1);
8594 In = BinaryOperator::CreateXor(In, One, "tmp");
8595 InsertNewInstBefore(cast<Instruction>(In), CI);
8598 if (CI.getType() == In->getType())
8599 return ReplaceInstUsesWith(CI, In);
8601 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8609 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8610 // If one of the common conversion will work ..
8611 if (Instruction *Result = commonIntCastTransforms(CI))
8614 Value *Src = CI.getOperand(0);
8616 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8617 // types and if the sizes are just right we can convert this into a logical
8618 // 'and' which will be much cheaper than the pair of casts.
8619 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8620 // Get the sizes of the types involved. We know that the intermediate type
8621 // will be smaller than A or C, but don't know the relation between A and C.
8622 Value *A = CSrc->getOperand(0);
8623 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8624 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8625 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8626 // If we're actually extending zero bits, then if
8627 // SrcSize < DstSize: zext(a & mask)
8628 // SrcSize == DstSize: a & mask
8629 // SrcSize > DstSize: trunc(a) & mask
8630 if (SrcSize < DstSize) {
8631 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8632 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8634 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8635 InsertNewInstBefore(And, CI);
8636 return new ZExtInst(And, CI.getType());
8637 } else if (SrcSize == DstSize) {
8638 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8639 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8641 } else if (SrcSize > DstSize) {
8642 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8643 InsertNewInstBefore(Trunc, CI);
8644 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8645 return BinaryOperator::CreateAnd(Trunc,
8646 ConstantInt::get(Trunc->getType(),
8651 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8652 return transformZExtICmp(ICI, CI);
8654 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8655 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8656 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8657 // of the (zext icmp) will be transformed.
8658 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8659 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8660 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8661 (transformZExtICmp(LHS, CI, false) ||
8662 transformZExtICmp(RHS, CI, false))) {
8663 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8664 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8665 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8669 // zext(trunc(t) & C) -> (t & zext(C)).
8670 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8671 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8672 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8673 Value *TI0 = TI->getOperand(0);
8674 if (TI0->getType() == CI.getType())
8676 BinaryOperator::CreateAnd(TI0,
8677 ConstantExpr::getZExt(C, CI.getType()));
8680 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8681 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8682 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8683 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8684 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8685 And->getOperand(1) == C)
8686 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8687 Value *TI0 = TI->getOperand(0);
8688 if (TI0->getType() == CI.getType()) {
8689 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8690 Instruction *NewAnd = BinaryOperator::CreateAnd(TI0, ZC, "tmp");
8691 InsertNewInstBefore(NewAnd, *And);
8692 return BinaryOperator::CreateXor(NewAnd, ZC);
8699 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8700 if (Instruction *I = commonIntCastTransforms(CI))
8703 Value *Src = CI.getOperand(0);
8705 // Canonicalize sign-extend from i1 to a select.
8706 if (Src->getType() == Type::getInt1Ty(*Context))
8707 return SelectInst::Create(Src,
8708 Constant::getAllOnesValue(CI.getType()),
8709 Constant::getNullValue(CI.getType()));
8711 // See if the value being truncated is already sign extended. If so, just
8712 // eliminate the trunc/sext pair.
8713 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8714 Value *Op = cast<User>(Src)->getOperand(0);
8715 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8716 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8717 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8718 unsigned NumSignBits = ComputeNumSignBits(Op);
8720 if (OpBits == DestBits) {
8721 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8722 // bits, it is already ready.
8723 if (NumSignBits > DestBits-MidBits)
8724 return ReplaceInstUsesWith(CI, Op);
8725 } else if (OpBits < DestBits) {
8726 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8727 // bits, just sext from i32.
8728 if (NumSignBits > OpBits-MidBits)
8729 return new SExtInst(Op, CI.getType(), "tmp");
8731 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8732 // bits, just truncate to i32.
8733 if (NumSignBits > OpBits-MidBits)
8734 return new TruncInst(Op, CI.getType(), "tmp");
8738 // If the input is a shl/ashr pair of a same constant, then this is a sign
8739 // extension from a smaller value. If we could trust arbitrary bitwidth
8740 // integers, we could turn this into a truncate to the smaller bit and then
8741 // use a sext for the whole extension. Since we don't, look deeper and check
8742 // for a truncate. If the source and dest are the same type, eliminate the
8743 // trunc and extend and just do shifts. For example, turn:
8744 // %a = trunc i32 %i to i8
8745 // %b = shl i8 %a, 6
8746 // %c = ashr i8 %b, 6
8747 // %d = sext i8 %c to i32
8749 // %a = shl i32 %i, 30
8750 // %d = ashr i32 %a, 30
8752 ConstantInt *BA = 0, *CA = 0;
8753 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8754 m_ConstantInt(CA))) &&
8755 BA == CA && isa<TruncInst>(A)) {
8756 Value *I = cast<TruncInst>(A)->getOperand(0);
8757 if (I->getType() == CI.getType()) {
8758 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8759 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8760 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8761 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8762 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8764 return BinaryOperator::CreateAShr(I, ShAmtV);
8771 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8772 /// in the specified FP type without changing its value.
8773 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8774 LLVMContext *Context) {
8776 APFloat F = CFP->getValueAPF();
8777 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8779 return ConstantFP::get(*Context, F);
8783 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8784 /// through it until we get the source value.
8785 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8786 if (Instruction *I = dyn_cast<Instruction>(V))
8787 if (I->getOpcode() == Instruction::FPExt)
8788 return LookThroughFPExtensions(I->getOperand(0), Context);
8790 // If this value is a constant, return the constant in the smallest FP type
8791 // that can accurately represent it. This allows us to turn
8792 // (float)((double)X+2.0) into x+2.0f.
8793 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8794 if (CFP->getType() == Type::getPPC_FP128Ty(*Context))
8795 return V; // No constant folding of this.
8796 // See if the value can be truncated to float and then reextended.
8797 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8799 if (CFP->getType() == Type::getDoubleTy(*Context))
8800 return V; // Won't shrink.
8801 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8803 // Don't try to shrink to various long double types.
8809 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8810 if (Instruction *I = commonCastTransforms(CI))
8813 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8814 // smaller than the destination type, we can eliminate the truncate by doing
8815 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8816 // many builtins (sqrt, etc).
8817 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8818 if (OpI && OpI->hasOneUse()) {
8819 switch (OpI->getOpcode()) {
8821 case Instruction::FAdd:
8822 case Instruction::FSub:
8823 case Instruction::FMul:
8824 case Instruction::FDiv:
8825 case Instruction::FRem:
8826 const Type *SrcTy = OpI->getType();
8827 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8828 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8829 if (LHSTrunc->getType() != SrcTy &&
8830 RHSTrunc->getType() != SrcTy) {
8831 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8832 // If the source types were both smaller than the destination type of
8833 // the cast, do this xform.
8834 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8835 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8836 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8838 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8840 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8849 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8850 return commonCastTransforms(CI);
8853 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8854 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8856 return commonCastTransforms(FI);
8858 // fptoui(uitofp(X)) --> X
8859 // fptoui(sitofp(X)) --> X
8860 // This is safe if the intermediate type has enough bits in its mantissa to
8861 // accurately represent all values of X. For example, do not do this with
8862 // i64->float->i64. This is also safe for sitofp case, because any negative
8863 // 'X' value would cause an undefined result for the fptoui.
8864 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8865 OpI->getOperand(0)->getType() == FI.getType() &&
8866 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8867 OpI->getType()->getFPMantissaWidth())
8868 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8870 return commonCastTransforms(FI);
8873 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8874 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8876 return commonCastTransforms(FI);
8878 // fptosi(sitofp(X)) --> X
8879 // fptosi(uitofp(X)) --> X
8880 // This is safe if the intermediate type has enough bits in its mantissa to
8881 // accurately represent all values of X. For example, do not do this with
8882 // i64->float->i64. This is also safe for sitofp case, because any negative
8883 // 'X' value would cause an undefined result for the fptoui.
8884 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8885 OpI->getOperand(0)->getType() == FI.getType() &&
8886 (int)FI.getType()->getScalarSizeInBits() <=
8887 OpI->getType()->getFPMantissaWidth())
8888 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8890 return commonCastTransforms(FI);
8893 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8894 return commonCastTransforms(CI);
8897 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8898 return commonCastTransforms(CI);
8901 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8902 // If the destination integer type is smaller than the intptr_t type for
8903 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8904 // trunc to be exposed to other transforms. Don't do this for extending
8905 // ptrtoint's, because we don't know if the target sign or zero extends its
8908 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8909 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8910 TD->getIntPtrType(CI.getContext()),
8912 return new TruncInst(P, CI.getType());
8915 return commonPointerCastTransforms(CI);
8918 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8919 // If the source integer type is larger than the intptr_t type for
8920 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8921 // allows the trunc to be exposed to other transforms. Don't do this for
8922 // extending inttoptr's, because we don't know if the target sign or zero
8923 // extends to pointers.
8925 CI.getOperand(0)->getType()->getScalarSizeInBits() >
8926 TD->getPointerSizeInBits()) {
8927 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8928 TD->getIntPtrType(CI.getContext()),
8930 return new IntToPtrInst(P, CI.getType());
8933 if (Instruction *I = commonCastTransforms(CI))
8939 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8940 // If the operands are integer typed then apply the integer transforms,
8941 // otherwise just apply the common ones.
8942 Value *Src = CI.getOperand(0);
8943 const Type *SrcTy = Src->getType();
8944 const Type *DestTy = CI.getType();
8946 if (isa<PointerType>(SrcTy)) {
8947 if (Instruction *I = commonPointerCastTransforms(CI))
8950 if (Instruction *Result = commonCastTransforms(CI))
8955 // Get rid of casts from one type to the same type. These are useless and can
8956 // be replaced by the operand.
8957 if (DestTy == Src->getType())
8958 return ReplaceInstUsesWith(CI, Src);
8960 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8961 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8962 const Type *DstElTy = DstPTy->getElementType();
8963 const Type *SrcElTy = SrcPTy->getElementType();
8965 // If the address spaces don't match, don't eliminate the bitcast, which is
8966 // required for changing types.
8967 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8970 // If we are casting a malloc or alloca to a pointer to a type of the same
8971 // size, rewrite the allocation instruction to allocate the "right" type.
8972 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8973 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8976 // If the source and destination are pointers, and this cast is equivalent
8977 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8978 // This can enhance SROA and other transforms that want type-safe pointers.
8979 Constant *ZeroUInt = Constant::getNullValue(Type::getInt32Ty(*Context));
8980 unsigned NumZeros = 0;
8981 while (SrcElTy != DstElTy &&
8982 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8983 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8984 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8988 // If we found a path from the src to dest, create the getelementptr now.
8989 if (SrcElTy == DstElTy) {
8990 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8991 Instruction *GEP = GetElementPtrInst::Create(Src,
8992 Idxs.begin(), Idxs.end(), "",
8993 ((Instruction*) NULL));
8994 cast<GEPOperator>(GEP)->setIsInBounds(true);
8999 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
9000 if (DestVTy->getNumElements() == 1) {
9001 if (!isa<VectorType>(SrcTy)) {
9002 Value *Elem = InsertCastBefore(Instruction::BitCast, Src,
9003 DestVTy->getElementType(), CI);
9004 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
9005 Constant::getNullValue(Type::getInt32Ty(*Context)));
9007 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
9011 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
9012 if (SrcVTy->getNumElements() == 1) {
9013 if (!isa<VectorType>(DestTy)) {
9015 ExtractElementInst::Create(Src, Constant::getNullValue(Type::getInt32Ty(*Context)));
9016 InsertNewInstBefore(Elem, CI);
9017 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
9022 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9023 if (SVI->hasOneUse()) {
9024 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9025 // a bitconvert to a vector with the same # elts.
9026 if (isa<VectorType>(DestTy) &&
9027 cast<VectorType>(DestTy)->getNumElements() ==
9028 SVI->getType()->getNumElements() &&
9029 SVI->getType()->getNumElements() ==
9030 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9032 // If either of the operands is a cast from CI.getType(), then
9033 // evaluating the shuffle in the casted destination's type will allow
9034 // us to eliminate at least one cast.
9035 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9036 Tmp->getOperand(0)->getType() == DestTy) ||
9037 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9038 Tmp->getOperand(0)->getType() == DestTy)) {
9039 Value *LHS = InsertCastBefore(Instruction::BitCast,
9040 SVI->getOperand(0), DestTy, CI);
9041 Value *RHS = InsertCastBefore(Instruction::BitCast,
9042 SVI->getOperand(1), DestTy, CI);
9043 // Return a new shuffle vector. Use the same element ID's, as we
9044 // know the vector types match #elts.
9045 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9053 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9055 /// %D = select %cond, %C, %A
9057 /// %C = select %cond, %B, 0
9060 /// Assuming that the specified instruction is an operand to the select, return
9061 /// a bitmask indicating which operands of this instruction are foldable if they
9062 /// equal the other incoming value of the select.
9064 static unsigned GetSelectFoldableOperands(Instruction *I) {
9065 switch (I->getOpcode()) {
9066 case Instruction::Add:
9067 case Instruction::Mul:
9068 case Instruction::And:
9069 case Instruction::Or:
9070 case Instruction::Xor:
9071 return 3; // Can fold through either operand.
9072 case Instruction::Sub: // Can only fold on the amount subtracted.
9073 case Instruction::Shl: // Can only fold on the shift amount.
9074 case Instruction::LShr:
9075 case Instruction::AShr:
9078 return 0; // Cannot fold
9082 /// GetSelectFoldableConstant - For the same transformation as the previous
9083 /// function, return the identity constant that goes into the select.
9084 static Constant *GetSelectFoldableConstant(Instruction *I,
9085 LLVMContext *Context) {
9086 switch (I->getOpcode()) {
9087 default: llvm_unreachable("This cannot happen!");
9088 case Instruction::Add:
9089 case Instruction::Sub:
9090 case Instruction::Or:
9091 case Instruction::Xor:
9092 case Instruction::Shl:
9093 case Instruction::LShr:
9094 case Instruction::AShr:
9095 return Constant::getNullValue(I->getType());
9096 case Instruction::And:
9097 return Constant::getAllOnesValue(I->getType());
9098 case Instruction::Mul:
9099 return ConstantInt::get(I->getType(), 1);
9103 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9104 /// have the same opcode and only one use each. Try to simplify this.
9105 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9107 if (TI->getNumOperands() == 1) {
9108 // If this is a non-volatile load or a cast from the same type,
9111 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9114 return 0; // unknown unary op.
9117 // Fold this by inserting a select from the input values.
9118 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9119 FI->getOperand(0), SI.getName()+".v");
9120 InsertNewInstBefore(NewSI, SI);
9121 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9125 // Only handle binary operators here.
9126 if (!isa<BinaryOperator>(TI))
9129 // Figure out if the operations have any operands in common.
9130 Value *MatchOp, *OtherOpT, *OtherOpF;
9132 if (TI->getOperand(0) == FI->getOperand(0)) {
9133 MatchOp = TI->getOperand(0);
9134 OtherOpT = TI->getOperand(1);
9135 OtherOpF = FI->getOperand(1);
9136 MatchIsOpZero = true;
9137 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9138 MatchOp = TI->getOperand(1);
9139 OtherOpT = TI->getOperand(0);
9140 OtherOpF = FI->getOperand(0);
9141 MatchIsOpZero = false;
9142 } else if (!TI->isCommutative()) {
9144 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9145 MatchOp = TI->getOperand(0);
9146 OtherOpT = TI->getOperand(1);
9147 OtherOpF = FI->getOperand(0);
9148 MatchIsOpZero = true;
9149 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9150 MatchOp = TI->getOperand(1);
9151 OtherOpT = TI->getOperand(0);
9152 OtherOpF = FI->getOperand(1);
9153 MatchIsOpZero = true;
9158 // If we reach here, they do have operations in common.
9159 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9160 OtherOpF, SI.getName()+".v");
9161 InsertNewInstBefore(NewSI, SI);
9163 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9165 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9167 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9169 llvm_unreachable("Shouldn't get here");
9173 static bool isSelect01(Constant *C1, Constant *C2) {
9174 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9177 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9180 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9183 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9184 /// facilitate further optimization.
9185 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9187 // See the comment above GetSelectFoldableOperands for a description of the
9188 // transformation we are doing here.
9189 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9190 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9191 !isa<Constant>(FalseVal)) {
9192 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9193 unsigned OpToFold = 0;
9194 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9196 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9201 Constant *C = GetSelectFoldableConstant(TVI, Context);
9202 Value *OOp = TVI->getOperand(2-OpToFold);
9203 // Avoid creating select between 2 constants unless it's selecting
9205 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9206 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9207 InsertNewInstBefore(NewSel, SI);
9208 NewSel->takeName(TVI);
9209 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9210 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9211 llvm_unreachable("Unknown instruction!!");
9218 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9219 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9220 !isa<Constant>(TrueVal)) {
9221 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9222 unsigned OpToFold = 0;
9223 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9225 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9230 Constant *C = GetSelectFoldableConstant(FVI, Context);
9231 Value *OOp = FVI->getOperand(2-OpToFold);
9232 // Avoid creating select between 2 constants unless it's selecting
9234 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9235 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9236 InsertNewInstBefore(NewSel, SI);
9237 NewSel->takeName(FVI);
9238 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9239 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9240 llvm_unreachable("Unknown instruction!!");
9250 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9251 /// ICmpInst as its first operand.
9253 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9255 bool Changed = false;
9256 ICmpInst::Predicate Pred = ICI->getPredicate();
9257 Value *CmpLHS = ICI->getOperand(0);
9258 Value *CmpRHS = ICI->getOperand(1);
9259 Value *TrueVal = SI.getTrueValue();
9260 Value *FalseVal = SI.getFalseValue();
9262 // Check cases where the comparison is with a constant that
9263 // can be adjusted to fit the min/max idiom. We may edit ICI in
9264 // place here, so make sure the select is the only user.
9265 if (ICI->hasOneUse())
9266 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9269 case ICmpInst::ICMP_ULT:
9270 case ICmpInst::ICMP_SLT: {
9271 // X < MIN ? T : F --> F
9272 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9273 return ReplaceInstUsesWith(SI, FalseVal);
9274 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9275 Constant *AdjustedRHS = SubOne(CI);
9276 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9277 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9278 Pred = ICmpInst::getSwappedPredicate(Pred);
9279 CmpRHS = AdjustedRHS;
9280 std::swap(FalseVal, TrueVal);
9281 ICI->setPredicate(Pred);
9282 ICI->setOperand(1, CmpRHS);
9283 SI.setOperand(1, TrueVal);
9284 SI.setOperand(2, FalseVal);
9289 case ICmpInst::ICMP_UGT:
9290 case ICmpInst::ICMP_SGT: {
9291 // X > MAX ? T : F --> F
9292 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9293 return ReplaceInstUsesWith(SI, FalseVal);
9294 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9295 Constant *AdjustedRHS = AddOne(CI);
9296 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9297 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9298 Pred = ICmpInst::getSwappedPredicate(Pred);
9299 CmpRHS = AdjustedRHS;
9300 std::swap(FalseVal, TrueVal);
9301 ICI->setPredicate(Pred);
9302 ICI->setOperand(1, CmpRHS);
9303 SI.setOperand(1, TrueVal);
9304 SI.setOperand(2, FalseVal);
9311 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9312 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9313 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9314 if (match(TrueVal, m_ConstantInt<-1>()) &&
9315 match(FalseVal, m_ConstantInt<0>()))
9316 Pred = ICI->getPredicate();
9317 else if (match(TrueVal, m_ConstantInt<0>()) &&
9318 match(FalseVal, m_ConstantInt<-1>()))
9319 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9321 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9322 // If we are just checking for a icmp eq of a single bit and zext'ing it
9323 // to an integer, then shift the bit to the appropriate place and then
9324 // cast to integer to avoid the comparison.
9325 const APInt &Op1CV = CI->getValue();
9327 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9328 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9329 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9330 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9331 Value *In = ICI->getOperand(0);
9332 Value *Sh = ConstantInt::get(In->getType(),
9333 In->getType()->getScalarSizeInBits()-1);
9334 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9335 In->getName()+".lobit"),
9337 if (In->getType() != SI.getType())
9338 In = CastInst::CreateIntegerCast(In, SI.getType(),
9339 true/*SExt*/, "tmp", ICI);
9341 if (Pred == ICmpInst::ICMP_SGT)
9342 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9343 In->getName()+".not"), *ICI);
9345 return ReplaceInstUsesWith(SI, In);
9350 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9351 // Transform (X == Y) ? X : Y -> Y
9352 if (Pred == ICmpInst::ICMP_EQ)
9353 return ReplaceInstUsesWith(SI, FalseVal);
9354 // Transform (X != Y) ? X : Y -> X
9355 if (Pred == ICmpInst::ICMP_NE)
9356 return ReplaceInstUsesWith(SI, TrueVal);
9357 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9359 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9360 // Transform (X == Y) ? Y : X -> X
9361 if (Pred == ICmpInst::ICMP_EQ)
9362 return ReplaceInstUsesWith(SI, FalseVal);
9363 // Transform (X != Y) ? Y : X -> Y
9364 if (Pred == ICmpInst::ICMP_NE)
9365 return ReplaceInstUsesWith(SI, TrueVal);
9366 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9369 /// NOTE: if we wanted to, this is where to detect integer ABS
9371 return Changed ? &SI : 0;
9374 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9375 Value *CondVal = SI.getCondition();
9376 Value *TrueVal = SI.getTrueValue();
9377 Value *FalseVal = SI.getFalseValue();
9379 // select true, X, Y -> X
9380 // select false, X, Y -> Y
9381 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9382 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9384 // select C, X, X -> X
9385 if (TrueVal == FalseVal)
9386 return ReplaceInstUsesWith(SI, TrueVal);
9388 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9389 return ReplaceInstUsesWith(SI, FalseVal);
9390 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9391 return ReplaceInstUsesWith(SI, TrueVal);
9392 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9393 if (isa<Constant>(TrueVal))
9394 return ReplaceInstUsesWith(SI, TrueVal);
9396 return ReplaceInstUsesWith(SI, FalseVal);
9399 if (SI.getType() == Type::getInt1Ty(*Context)) {
9400 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9401 if (C->getZExtValue()) {
9402 // Change: A = select B, true, C --> A = or B, C
9403 return BinaryOperator::CreateOr(CondVal, FalseVal);
9405 // Change: A = select B, false, C --> A = and !B, C
9407 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9408 "not."+CondVal->getName()), SI);
9409 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9411 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9412 if (C->getZExtValue() == false) {
9413 // Change: A = select B, C, false --> A = and B, C
9414 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9416 // Change: A = select B, C, true --> A = or !B, C
9418 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9419 "not."+CondVal->getName()), SI);
9420 return BinaryOperator::CreateOr(NotCond, TrueVal);
9424 // select a, b, a -> a&b
9425 // select a, a, b -> a|b
9426 if (CondVal == TrueVal)
9427 return BinaryOperator::CreateOr(CondVal, FalseVal);
9428 else if (CondVal == FalseVal)
9429 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9432 // Selecting between two integer constants?
9433 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9434 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9435 // select C, 1, 0 -> zext C to int
9436 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9437 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9438 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9439 // select C, 0, 1 -> zext !C to int
9441 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9442 "not."+CondVal->getName()), SI);
9443 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9446 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9447 // If one of the constants is zero (we know they can't both be) and we
9448 // have an icmp instruction with zero, and we have an 'and' with the
9449 // non-constant value, eliminate this whole mess. This corresponds to
9450 // cases like this: ((X & 27) ? 27 : 0)
9451 if (TrueValC->isZero() || FalseValC->isZero())
9452 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9453 cast<Constant>(IC->getOperand(1))->isNullValue())
9454 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9455 if (ICA->getOpcode() == Instruction::And &&
9456 isa<ConstantInt>(ICA->getOperand(1)) &&
9457 (ICA->getOperand(1) == TrueValC ||
9458 ICA->getOperand(1) == FalseValC) &&
9459 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9460 // Okay, now we know that everything is set up, we just don't
9461 // know whether we have a icmp_ne or icmp_eq and whether the
9462 // true or false val is the zero.
9463 bool ShouldNotVal = !TrueValC->isZero();
9464 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9467 V = InsertNewInstBefore(BinaryOperator::Create(
9468 Instruction::Xor, V, ICA->getOperand(1)), SI);
9469 return ReplaceInstUsesWith(SI, V);
9474 // See if we are selecting two values based on a comparison of the two values.
9475 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9476 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9477 // Transform (X == Y) ? X : Y -> Y
9478 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9479 // This is not safe in general for floating point:
9480 // consider X== -0, Y== +0.
9481 // It becomes safe if either operand is a nonzero constant.
9482 ConstantFP *CFPt, *CFPf;
9483 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9484 !CFPt->getValueAPF().isZero()) ||
9485 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9486 !CFPf->getValueAPF().isZero()))
9487 return ReplaceInstUsesWith(SI, FalseVal);
9489 // Transform (X != Y) ? X : Y -> X
9490 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9491 return ReplaceInstUsesWith(SI, TrueVal);
9492 // NOTE: if we wanted to, this is where to detect MIN/MAX
9494 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9495 // Transform (X == Y) ? Y : X -> X
9496 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9497 // This is not safe in general for floating point:
9498 // consider X== -0, Y== +0.
9499 // It becomes safe if either operand is a nonzero constant.
9500 ConstantFP *CFPt, *CFPf;
9501 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9502 !CFPt->getValueAPF().isZero()) ||
9503 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9504 !CFPf->getValueAPF().isZero()))
9505 return ReplaceInstUsesWith(SI, FalseVal);
9507 // Transform (X != Y) ? Y : X -> Y
9508 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9509 return ReplaceInstUsesWith(SI, TrueVal);
9510 // NOTE: if we wanted to, this is where to detect MIN/MAX
9512 // NOTE: if we wanted to, this is where to detect ABS
9515 // See if we are selecting two values based on a comparison of the two values.
9516 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9517 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9520 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9521 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9522 if (TI->hasOneUse() && FI->hasOneUse()) {
9523 Instruction *AddOp = 0, *SubOp = 0;
9525 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9526 if (TI->getOpcode() == FI->getOpcode())
9527 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9530 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9531 // even legal for FP.
9532 if ((TI->getOpcode() == Instruction::Sub &&
9533 FI->getOpcode() == Instruction::Add) ||
9534 (TI->getOpcode() == Instruction::FSub &&
9535 FI->getOpcode() == Instruction::FAdd)) {
9536 AddOp = FI; SubOp = TI;
9537 } else if ((FI->getOpcode() == Instruction::Sub &&
9538 TI->getOpcode() == Instruction::Add) ||
9539 (FI->getOpcode() == Instruction::FSub &&
9540 TI->getOpcode() == Instruction::FAdd)) {
9541 AddOp = TI; SubOp = FI;
9545 Value *OtherAddOp = 0;
9546 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9547 OtherAddOp = AddOp->getOperand(1);
9548 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9549 OtherAddOp = AddOp->getOperand(0);
9553 // So at this point we know we have (Y -> OtherAddOp):
9554 // select C, (add X, Y), (sub X, Z)
9555 Value *NegVal; // Compute -Z
9556 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9557 NegVal = ConstantExpr::getNeg(C);
9559 NegVal = InsertNewInstBefore(
9560 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9564 Value *NewTrueOp = OtherAddOp;
9565 Value *NewFalseOp = NegVal;
9567 std::swap(NewTrueOp, NewFalseOp);
9568 Instruction *NewSel =
9569 SelectInst::Create(CondVal, NewTrueOp,
9570 NewFalseOp, SI.getName() + ".p");
9572 NewSel = InsertNewInstBefore(NewSel, SI);
9573 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9578 // See if we can fold the select into one of our operands.
9579 if (SI.getType()->isInteger()) {
9580 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9585 if (BinaryOperator::isNot(CondVal)) {
9586 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9587 SI.setOperand(1, FalseVal);
9588 SI.setOperand(2, TrueVal);
9595 /// EnforceKnownAlignment - If the specified pointer points to an object that
9596 /// we control, modify the object's alignment to PrefAlign. This isn't
9597 /// often possible though. If alignment is important, a more reliable approach
9598 /// is to simply align all global variables and allocation instructions to
9599 /// their preferred alignment from the beginning.
9601 static unsigned EnforceKnownAlignment(Value *V,
9602 unsigned Align, unsigned PrefAlign) {
9604 User *U = dyn_cast<User>(V);
9605 if (!U) return Align;
9607 switch (Operator::getOpcode(U)) {
9609 case Instruction::BitCast:
9610 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9611 case Instruction::GetElementPtr: {
9612 // If all indexes are zero, it is just the alignment of the base pointer.
9613 bool AllZeroOperands = true;
9614 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9615 if (!isa<Constant>(*i) ||
9616 !cast<Constant>(*i)->isNullValue()) {
9617 AllZeroOperands = false;
9621 if (AllZeroOperands) {
9622 // Treat this like a bitcast.
9623 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9629 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9630 // If there is a large requested alignment and we can, bump up the alignment
9632 if (!GV->isDeclaration()) {
9633 if (GV->getAlignment() >= PrefAlign)
9634 Align = GV->getAlignment();
9636 GV->setAlignment(PrefAlign);
9640 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9641 // If there is a requested alignment and if this is an alloca, round up. We
9642 // don't do this for malloc, because some systems can't respect the request.
9643 if (isa<AllocaInst>(AI)) {
9644 if (AI->getAlignment() >= PrefAlign)
9645 Align = AI->getAlignment();
9647 AI->setAlignment(PrefAlign);
9656 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9657 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9658 /// and it is more than the alignment of the ultimate object, see if we can
9659 /// increase the alignment of the ultimate object, making this check succeed.
9660 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9661 unsigned PrefAlign) {
9662 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9663 sizeof(PrefAlign) * CHAR_BIT;
9664 APInt Mask = APInt::getAllOnesValue(BitWidth);
9665 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9666 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9667 unsigned TrailZ = KnownZero.countTrailingOnes();
9668 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9670 if (PrefAlign > Align)
9671 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9673 // We don't need to make any adjustment.
9677 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9678 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9679 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9680 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9681 unsigned CopyAlign = MI->getAlignment();
9683 if (CopyAlign < MinAlign) {
9684 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9689 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9691 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9692 if (MemOpLength == 0) return 0;
9694 // Source and destination pointer types are always "i8*" for intrinsic. See
9695 // if the size is something we can handle with a single primitive load/store.
9696 // A single load+store correctly handles overlapping memory in the memmove
9698 unsigned Size = MemOpLength->getZExtValue();
9699 if (Size == 0) return MI; // Delete this mem transfer.
9701 if (Size > 8 || (Size&(Size-1)))
9702 return 0; // If not 1/2/4/8 bytes, exit.
9704 // Use an integer load+store unless we can find something better.
9706 PointerType::getUnqual(IntegerType::get(*Context, Size<<3));
9708 // Memcpy forces the use of i8* for the source and destination. That means
9709 // that if you're using memcpy to move one double around, you'll get a cast
9710 // from double* to i8*. We'd much rather use a double load+store rather than
9711 // an i64 load+store, here because this improves the odds that the source or
9712 // dest address will be promotable. See if we can find a better type than the
9713 // integer datatype.
9714 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9715 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9716 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9717 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9718 // down through these levels if so.
9719 while (!SrcETy->isSingleValueType()) {
9720 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9721 if (STy->getNumElements() == 1)
9722 SrcETy = STy->getElementType(0);
9725 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9726 if (ATy->getNumElements() == 1)
9727 SrcETy = ATy->getElementType();
9734 if (SrcETy->isSingleValueType())
9735 NewPtrTy = PointerType::getUnqual(SrcETy);
9740 // If the memcpy/memmove provides better alignment info than we can
9742 SrcAlign = std::max(SrcAlign, CopyAlign);
9743 DstAlign = std::max(DstAlign, CopyAlign);
9745 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9746 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9747 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9748 InsertNewInstBefore(L, *MI);
9749 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9751 // Set the size of the copy to 0, it will be deleted on the next iteration.
9752 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9756 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9757 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9758 if (MI->getAlignment() < Alignment) {
9759 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9764 // Extract the length and alignment and fill if they are constant.
9765 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9766 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9767 if (!LenC || !FillC || FillC->getType() != Type::getInt8Ty(*Context))
9769 uint64_t Len = LenC->getZExtValue();
9770 Alignment = MI->getAlignment();
9772 // If the length is zero, this is a no-op
9773 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9775 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9776 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9777 const Type *ITy = IntegerType::get(*Context, Len*8); // n=1 -> i8.
9779 Value *Dest = MI->getDest();
9780 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9782 // Alignment 0 is identity for alignment 1 for memset, but not store.
9783 if (Alignment == 0) Alignment = 1;
9785 // Extract the fill value and store.
9786 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9787 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
9788 Dest, false, Alignment), *MI);
9790 // Set the size of the copy to 0, it will be deleted on the next iteration.
9791 MI->setLength(Constant::getNullValue(LenC->getType()));
9799 /// visitCallInst - CallInst simplification. This mostly only handles folding
9800 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9801 /// the heavy lifting.
9803 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9804 // If the caller function is nounwind, mark the call as nounwind, even if the
9806 if (CI.getParent()->getParent()->doesNotThrow() &&
9807 !CI.doesNotThrow()) {
9808 CI.setDoesNotThrow();
9814 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9815 if (!II) return visitCallSite(&CI);
9817 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9819 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9820 bool Changed = false;
9822 // memmove/cpy/set of zero bytes is a noop.
9823 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9824 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9826 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9827 if (CI->getZExtValue() == 1) {
9828 // Replace the instruction with just byte operations. We would
9829 // transform other cases to loads/stores, but we don't know if
9830 // alignment is sufficient.
9834 // If we have a memmove and the source operation is a constant global,
9835 // then the source and dest pointers can't alias, so we can change this
9836 // into a call to memcpy.
9837 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9838 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9839 if (GVSrc->isConstant()) {
9840 Module *M = CI.getParent()->getParent()->getParent();
9841 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9843 Tys[0] = CI.getOperand(3)->getType();
9845 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9849 // memmove(x,x,size) -> noop.
9850 if (MMI->getSource() == MMI->getDest())
9851 return EraseInstFromFunction(CI);
9854 // If we can determine a pointer alignment that is bigger than currently
9855 // set, update the alignment.
9856 if (isa<MemTransferInst>(MI)) {
9857 if (Instruction *I = SimplifyMemTransfer(MI))
9859 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9860 if (Instruction *I = SimplifyMemSet(MSI))
9864 if (Changed) return II;
9867 switch (II->getIntrinsicID()) {
9869 case Intrinsic::bswap:
9870 // bswap(bswap(x)) -> x
9871 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9872 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9873 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9875 case Intrinsic::ppc_altivec_lvx:
9876 case Intrinsic::ppc_altivec_lvxl:
9877 case Intrinsic::x86_sse_loadu_ps:
9878 case Intrinsic::x86_sse2_loadu_pd:
9879 case Intrinsic::x86_sse2_loadu_dq:
9880 // Turn PPC lvx -> load if the pointer is known aligned.
9881 // Turn X86 loadups -> load if the pointer is known aligned.
9882 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9883 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9884 PointerType::getUnqual(II->getType()),
9886 return new LoadInst(Ptr);
9889 case Intrinsic::ppc_altivec_stvx:
9890 case Intrinsic::ppc_altivec_stvxl:
9891 // Turn stvx -> store if the pointer is known aligned.
9892 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9893 const Type *OpPtrTy =
9894 PointerType::getUnqual(II->getOperand(1)->getType());
9895 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9896 return new StoreInst(II->getOperand(1), Ptr);
9899 case Intrinsic::x86_sse_storeu_ps:
9900 case Intrinsic::x86_sse2_storeu_pd:
9901 case Intrinsic::x86_sse2_storeu_dq:
9902 // Turn X86 storeu -> store if the pointer is known aligned.
9903 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9904 const Type *OpPtrTy =
9905 PointerType::getUnqual(II->getOperand(2)->getType());
9906 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9907 return new StoreInst(II->getOperand(2), Ptr);
9911 case Intrinsic::x86_sse_cvttss2si: {
9912 // These intrinsics only demands the 0th element of its input vector. If
9913 // we can simplify the input based on that, do so now.
9915 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9916 APInt DemandedElts(VWidth, 1);
9917 APInt UndefElts(VWidth, 0);
9918 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9920 II->setOperand(1, V);
9926 case Intrinsic::ppc_altivec_vperm:
9927 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9928 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9929 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9931 // Check that all of the elements are integer constants or undefs.
9932 bool AllEltsOk = true;
9933 for (unsigned i = 0; i != 16; ++i) {
9934 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9935 !isa<UndefValue>(Mask->getOperand(i))) {
9942 // Cast the input vectors to byte vectors.
9943 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9944 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9945 Value *Result = UndefValue::get(Op0->getType());
9947 // Only extract each element once.
9948 Value *ExtractedElts[32];
9949 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9951 for (unsigned i = 0; i != 16; ++i) {
9952 if (isa<UndefValue>(Mask->getOperand(i)))
9954 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9955 Idx &= 31; // Match the hardware behavior.
9957 if (ExtractedElts[Idx] == 0) {
9959 ExtractElementInst::Create(Idx < 16 ? Op0 : Op1,
9960 ConstantInt::get(Type::getInt32Ty(*Context), Idx&15, false), "tmp");
9961 InsertNewInstBefore(Elt, CI);
9962 ExtractedElts[Idx] = Elt;
9965 // Insert this value into the result vector.
9966 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9967 ConstantInt::get(Type::getInt32Ty(*Context), i, false),
9969 InsertNewInstBefore(cast<Instruction>(Result), CI);
9971 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9976 case Intrinsic::stackrestore: {
9977 // If the save is right next to the restore, remove the restore. This can
9978 // happen when variable allocas are DCE'd.
9979 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9980 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9981 BasicBlock::iterator BI = SS;
9983 return EraseInstFromFunction(CI);
9987 // Scan down this block to see if there is another stack restore in the
9988 // same block without an intervening call/alloca.
9989 BasicBlock::iterator BI = II;
9990 TerminatorInst *TI = II->getParent()->getTerminator();
9991 bool CannotRemove = false;
9992 for (++BI; &*BI != TI; ++BI) {
9993 if (isa<AllocaInst>(BI)) {
9994 CannotRemove = true;
9997 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9998 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9999 // If there is a stackrestore below this one, remove this one.
10000 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10001 return EraseInstFromFunction(CI);
10002 // Otherwise, ignore the intrinsic.
10004 // If we found a non-intrinsic call, we can't remove the stack
10006 CannotRemove = true;
10012 // If the stack restore is in a return/unwind block and if there are no
10013 // allocas or calls between the restore and the return, nuke the restore.
10014 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10015 return EraseInstFromFunction(CI);
10020 return visitCallSite(II);
10023 // InvokeInst simplification
10025 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10026 return visitCallSite(&II);
10029 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10030 /// passed through the varargs area, we can eliminate the use of the cast.
10031 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10032 const CastInst * const CI,
10033 const TargetData * const TD,
10035 if (!CI->isLosslessCast())
10038 // The size of ByVal arguments is derived from the type, so we
10039 // can't change to a type with a different size. If the size were
10040 // passed explicitly we could avoid this check.
10041 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10044 const Type* SrcTy =
10045 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10046 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10047 if (!SrcTy->isSized() || !DstTy->isSized())
10049 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10054 // visitCallSite - Improvements for call and invoke instructions.
10056 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10057 bool Changed = false;
10059 // If the callee is a constexpr cast of a function, attempt to move the cast
10060 // to the arguments of the call/invoke.
10061 if (transformConstExprCastCall(CS)) return 0;
10063 Value *Callee = CS.getCalledValue();
10065 if (Function *CalleeF = dyn_cast<Function>(Callee))
10066 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10067 Instruction *OldCall = CS.getInstruction();
10068 // If the call and callee calling conventions don't match, this call must
10069 // be unreachable, as the call is undefined.
10070 new StoreInst(ConstantInt::getTrue(*Context),
10071 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))),
10073 if (!OldCall->use_empty())
10074 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
10075 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10076 return EraseInstFromFunction(*OldCall);
10080 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10081 // This instruction is not reachable, just remove it. We insert a store to
10082 // undef so that we know that this code is not reachable, despite the fact
10083 // that we can't modify the CFG here.
10084 new StoreInst(ConstantInt::getTrue(*Context),
10085 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))),
10086 CS.getInstruction());
10088 if (!CS.getInstruction()->use_empty())
10089 CS.getInstruction()->
10090 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
10092 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10093 // Don't break the CFG, insert a dummy cond branch.
10094 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10095 ConstantInt::getTrue(*Context), II);
10097 return EraseInstFromFunction(*CS.getInstruction());
10100 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10101 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10102 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10103 return transformCallThroughTrampoline(CS);
10105 const PointerType *PTy = cast<PointerType>(Callee->getType());
10106 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10107 if (FTy->isVarArg()) {
10108 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10109 // See if we can optimize any arguments passed through the varargs area of
10111 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10112 E = CS.arg_end(); I != E; ++I, ++ix) {
10113 CastInst *CI = dyn_cast<CastInst>(*I);
10114 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10115 *I = CI->getOperand(0);
10121 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10122 // Inline asm calls cannot throw - mark them 'nounwind'.
10123 CS.setDoesNotThrow();
10127 return Changed ? CS.getInstruction() : 0;
10130 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10131 // attempt to move the cast to the arguments of the call/invoke.
10133 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10134 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10135 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10136 if (CE->getOpcode() != Instruction::BitCast ||
10137 !isa<Function>(CE->getOperand(0)))
10139 Function *Callee = cast<Function>(CE->getOperand(0));
10140 Instruction *Caller = CS.getInstruction();
10141 const AttrListPtr &CallerPAL = CS.getAttributes();
10143 // Okay, this is a cast from a function to a different type. Unless doing so
10144 // would cause a type conversion of one of our arguments, change this call to
10145 // be a direct call with arguments casted to the appropriate types.
10147 const FunctionType *FT = Callee->getFunctionType();
10148 const Type *OldRetTy = Caller->getType();
10149 const Type *NewRetTy = FT->getReturnType();
10151 if (isa<StructType>(NewRetTy))
10152 return false; // TODO: Handle multiple return values.
10154 // Check to see if we are changing the return type...
10155 if (OldRetTy != NewRetTy) {
10156 if (Callee->isDeclaration() &&
10157 // Conversion is ok if changing from one pointer type to another or from
10158 // a pointer to an integer of the same size.
10159 !((isa<PointerType>(OldRetTy) || !TD ||
10160 OldRetTy == TD->getIntPtrType(Caller->getContext())) &&
10161 (isa<PointerType>(NewRetTy) || !TD ||
10162 NewRetTy == TD->getIntPtrType(Caller->getContext()))))
10163 return false; // Cannot transform this return value.
10165 if (!Caller->use_empty() &&
10166 // void -> non-void is handled specially
10167 NewRetTy != Type::getVoidTy(*Context) && !CastInst::isCastable(NewRetTy, OldRetTy))
10168 return false; // Cannot transform this return value.
10170 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10171 Attributes RAttrs = CallerPAL.getRetAttributes();
10172 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10173 return false; // Attribute not compatible with transformed value.
10176 // If the callsite is an invoke instruction, and the return value is used by
10177 // a PHI node in a successor, we cannot change the return type of the call
10178 // because there is no place to put the cast instruction (without breaking
10179 // the critical edge). Bail out in this case.
10180 if (!Caller->use_empty())
10181 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10182 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10184 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10185 if (PN->getParent() == II->getNormalDest() ||
10186 PN->getParent() == II->getUnwindDest())
10190 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10191 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10193 CallSite::arg_iterator AI = CS.arg_begin();
10194 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10195 const Type *ParamTy = FT->getParamType(i);
10196 const Type *ActTy = (*AI)->getType();
10198 if (!CastInst::isCastable(ActTy, ParamTy))
10199 return false; // Cannot transform this parameter value.
10201 if (CallerPAL.getParamAttributes(i + 1)
10202 & Attribute::typeIncompatible(ParamTy))
10203 return false; // Attribute not compatible with transformed value.
10205 // Converting from one pointer type to another or between a pointer and an
10206 // integer of the same size is safe even if we do not have a body.
10207 bool isConvertible = ActTy == ParamTy ||
10208 (TD && ((isa<PointerType>(ParamTy) ||
10209 ParamTy == TD->getIntPtrType(Caller->getContext())) &&
10210 (isa<PointerType>(ActTy) ||
10211 ActTy == TD->getIntPtrType(Caller->getContext()))));
10212 if (Callee->isDeclaration() && !isConvertible) return false;
10215 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10216 Callee->isDeclaration())
10217 return false; // Do not delete arguments unless we have a function body.
10219 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10220 !CallerPAL.isEmpty())
10221 // In this case we have more arguments than the new function type, but we
10222 // won't be dropping them. Check that these extra arguments have attributes
10223 // that are compatible with being a vararg call argument.
10224 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10225 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10227 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10228 if (PAttrs & Attribute::VarArgsIncompatible)
10232 // Okay, we decided that this is a safe thing to do: go ahead and start
10233 // inserting cast instructions as necessary...
10234 std::vector<Value*> Args;
10235 Args.reserve(NumActualArgs);
10236 SmallVector<AttributeWithIndex, 8> attrVec;
10237 attrVec.reserve(NumCommonArgs);
10239 // Get any return attributes.
10240 Attributes RAttrs = CallerPAL.getRetAttributes();
10242 // If the return value is not being used, the type may not be compatible
10243 // with the existing attributes. Wipe out any problematic attributes.
10244 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10246 // Add the new return attributes.
10248 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10250 AI = CS.arg_begin();
10251 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10252 const Type *ParamTy = FT->getParamType(i);
10253 if ((*AI)->getType() == ParamTy) {
10254 Args.push_back(*AI);
10256 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10257 false, ParamTy, false);
10258 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
10259 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
10262 // Add any parameter attributes.
10263 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10264 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10267 // If the function takes more arguments than the call was taking, add them
10269 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10270 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10272 // If we are removing arguments to the function, emit an obnoxious warning...
10273 if (FT->getNumParams() < NumActualArgs) {
10274 if (!FT->isVarArg()) {
10275 errs() << "WARNING: While resolving call to function '"
10276 << Callee->getName() << "' arguments were dropped!\n";
10278 // Add all of the arguments in their promoted form to the arg list...
10279 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10280 const Type *PTy = getPromotedType((*AI)->getType());
10281 if (PTy != (*AI)->getType()) {
10282 // Must promote to pass through va_arg area!
10283 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
10285 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
10286 InsertNewInstBefore(Cast, *Caller);
10287 Args.push_back(Cast);
10289 Args.push_back(*AI);
10292 // Add any parameter attributes.
10293 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10294 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10299 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10300 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10302 if (NewRetTy == Type::getVoidTy(*Context))
10303 Caller->setName(""); // Void type should not have a name.
10305 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10309 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10310 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10311 Args.begin(), Args.end(),
10312 Caller->getName(), Caller);
10313 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10314 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10316 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10317 Caller->getName(), Caller);
10318 CallInst *CI = cast<CallInst>(Caller);
10319 if (CI->isTailCall())
10320 cast<CallInst>(NC)->setTailCall();
10321 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10322 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10325 // Insert a cast of the return type as necessary.
10327 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10328 if (NV->getType() != Type::getVoidTy(*Context)) {
10329 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10331 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10333 // If this is an invoke instruction, we should insert it after the first
10334 // non-phi, instruction in the normal successor block.
10335 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10336 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10337 InsertNewInstBefore(NC, *I);
10339 // Otherwise, it's a call, just insert cast right after the call instr
10340 InsertNewInstBefore(NC, *Caller);
10342 AddUsersToWorkList(*Caller);
10344 NV = UndefValue::get(Caller->getType());
10348 if (Caller->getType() != Type::getVoidTy(*Context) && !Caller->use_empty())
10349 Caller->replaceAllUsesWith(NV);
10350 Caller->eraseFromParent();
10351 RemoveFromWorkList(Caller);
10355 // transformCallThroughTrampoline - Turn a call to a function created by the
10356 // init_trampoline intrinsic into a direct call to the underlying function.
10358 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10359 Value *Callee = CS.getCalledValue();
10360 const PointerType *PTy = cast<PointerType>(Callee->getType());
10361 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10362 const AttrListPtr &Attrs = CS.getAttributes();
10364 // If the call already has the 'nest' attribute somewhere then give up -
10365 // otherwise 'nest' would occur twice after splicing in the chain.
10366 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10369 IntrinsicInst *Tramp =
10370 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10372 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10373 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10374 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10376 const AttrListPtr &NestAttrs = NestF->getAttributes();
10377 if (!NestAttrs.isEmpty()) {
10378 unsigned NestIdx = 1;
10379 const Type *NestTy = 0;
10380 Attributes NestAttr = Attribute::None;
10382 // Look for a parameter marked with the 'nest' attribute.
10383 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10384 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10385 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10386 // Record the parameter type and any other attributes.
10388 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10393 Instruction *Caller = CS.getInstruction();
10394 std::vector<Value*> NewArgs;
10395 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10397 SmallVector<AttributeWithIndex, 8> NewAttrs;
10398 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10400 // Insert the nest argument into the call argument list, which may
10401 // mean appending it. Likewise for attributes.
10403 // Add any result attributes.
10404 if (Attributes Attr = Attrs.getRetAttributes())
10405 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10409 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10411 if (Idx == NestIdx) {
10412 // Add the chain argument and attributes.
10413 Value *NestVal = Tramp->getOperand(3);
10414 if (NestVal->getType() != NestTy)
10415 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10416 NewArgs.push_back(NestVal);
10417 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10423 // Add the original argument and attributes.
10424 NewArgs.push_back(*I);
10425 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10427 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10433 // Add any function attributes.
10434 if (Attributes Attr = Attrs.getFnAttributes())
10435 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10437 // The trampoline may have been bitcast to a bogus type (FTy).
10438 // Handle this by synthesizing a new function type, equal to FTy
10439 // with the chain parameter inserted.
10441 std::vector<const Type*> NewTypes;
10442 NewTypes.reserve(FTy->getNumParams()+1);
10444 // Insert the chain's type into the list of parameter types, which may
10445 // mean appending it.
10448 FunctionType::param_iterator I = FTy->param_begin(),
10449 E = FTy->param_end();
10452 if (Idx == NestIdx)
10453 // Add the chain's type.
10454 NewTypes.push_back(NestTy);
10459 // Add the original type.
10460 NewTypes.push_back(*I);
10466 // Replace the trampoline call with a direct call. Let the generic
10467 // code sort out any function type mismatches.
10468 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10470 Constant *NewCallee =
10471 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10472 NestF : ConstantExpr::getBitCast(NestF,
10473 PointerType::getUnqual(NewFTy));
10474 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10477 Instruction *NewCaller;
10478 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10479 NewCaller = InvokeInst::Create(NewCallee,
10480 II->getNormalDest(), II->getUnwindDest(),
10481 NewArgs.begin(), NewArgs.end(),
10482 Caller->getName(), Caller);
10483 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10484 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10486 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10487 Caller->getName(), Caller);
10488 if (cast<CallInst>(Caller)->isTailCall())
10489 cast<CallInst>(NewCaller)->setTailCall();
10490 cast<CallInst>(NewCaller)->
10491 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10492 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10494 if (Caller->getType() != Type::getVoidTy(*Context) && !Caller->use_empty())
10495 Caller->replaceAllUsesWith(NewCaller);
10496 Caller->eraseFromParent();
10497 RemoveFromWorkList(Caller);
10502 // Replace the trampoline call with a direct call. Since there is no 'nest'
10503 // parameter, there is no need to adjust the argument list. Let the generic
10504 // code sort out any function type mismatches.
10505 Constant *NewCallee =
10506 NestF->getType() == PTy ? NestF :
10507 ConstantExpr::getBitCast(NestF, PTy);
10508 CS.setCalledFunction(NewCallee);
10509 return CS.getInstruction();
10512 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10513 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10514 /// and a single binop.
10515 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10516 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10517 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10518 unsigned Opc = FirstInst->getOpcode();
10519 Value *LHSVal = FirstInst->getOperand(0);
10520 Value *RHSVal = FirstInst->getOperand(1);
10522 const Type *LHSType = LHSVal->getType();
10523 const Type *RHSType = RHSVal->getType();
10525 // Scan to see if all operands are the same opcode, all have one use, and all
10526 // kill their operands (i.e. the operands have one use).
10527 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10528 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10529 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10530 // Verify type of the LHS matches so we don't fold cmp's of different
10531 // types or GEP's with different index types.
10532 I->getOperand(0)->getType() != LHSType ||
10533 I->getOperand(1)->getType() != RHSType)
10536 // If they are CmpInst instructions, check their predicates
10537 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10538 if (cast<CmpInst>(I)->getPredicate() !=
10539 cast<CmpInst>(FirstInst)->getPredicate())
10542 // Keep track of which operand needs a phi node.
10543 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10544 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10547 // Otherwise, this is safe to transform!
10549 Value *InLHS = FirstInst->getOperand(0);
10550 Value *InRHS = FirstInst->getOperand(1);
10551 PHINode *NewLHS = 0, *NewRHS = 0;
10553 NewLHS = PHINode::Create(LHSType,
10554 FirstInst->getOperand(0)->getName() + ".pn");
10555 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10556 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10557 InsertNewInstBefore(NewLHS, PN);
10562 NewRHS = PHINode::Create(RHSType,
10563 FirstInst->getOperand(1)->getName() + ".pn");
10564 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10565 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10566 InsertNewInstBefore(NewRHS, PN);
10570 // Add all operands to the new PHIs.
10571 if (NewLHS || NewRHS) {
10572 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10573 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10575 Value *NewInLHS = InInst->getOperand(0);
10576 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10579 Value *NewInRHS = InInst->getOperand(1);
10580 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10585 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10586 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10587 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10588 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10592 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10593 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10595 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10596 FirstInst->op_end());
10597 // This is true if all GEP bases are allocas and if all indices into them are
10599 bool AllBasePointersAreAllocas = true;
10601 // Scan to see if all operands are the same opcode, all have one use, and all
10602 // kill their operands (i.e. the operands have one use).
10603 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10604 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10605 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10606 GEP->getNumOperands() != FirstInst->getNumOperands())
10609 // Keep track of whether or not all GEPs are of alloca pointers.
10610 if (AllBasePointersAreAllocas &&
10611 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10612 !GEP->hasAllConstantIndices()))
10613 AllBasePointersAreAllocas = false;
10615 // Compare the operand lists.
10616 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10617 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10620 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10621 // if one of the PHIs has a constant for the index. The index may be
10622 // substantially cheaper to compute for the constants, so making it a
10623 // variable index could pessimize the path. This also handles the case
10624 // for struct indices, which must always be constant.
10625 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10626 isa<ConstantInt>(GEP->getOperand(op)))
10629 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10631 FixedOperands[op] = 0; // Needs a PHI.
10635 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10636 // bother doing this transformation. At best, this will just save a bit of
10637 // offset calculation, but all the predecessors will have to materialize the
10638 // stack address into a register anyway. We'd actually rather *clone* the
10639 // load up into the predecessors so that we have a load of a gep of an alloca,
10640 // which can usually all be folded into the load.
10641 if (AllBasePointersAreAllocas)
10644 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10645 // that is variable.
10646 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10648 bool HasAnyPHIs = false;
10649 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10650 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10651 Value *FirstOp = FirstInst->getOperand(i);
10652 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10653 FirstOp->getName()+".pn");
10654 InsertNewInstBefore(NewPN, PN);
10656 NewPN->reserveOperandSpace(e);
10657 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10658 OperandPhis[i] = NewPN;
10659 FixedOperands[i] = NewPN;
10664 // Add all operands to the new PHIs.
10666 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10667 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10668 BasicBlock *InBB = PN.getIncomingBlock(i);
10670 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10671 if (PHINode *OpPhi = OperandPhis[op])
10672 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10676 Value *Base = FixedOperands[0];
10677 GetElementPtrInst *GEP =
10678 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10679 FixedOperands.end());
10680 if (cast<GEPOperator>(FirstInst)->isInBounds())
10681 cast<GEPOperator>(GEP)->setIsInBounds(true);
10686 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10687 /// sink the load out of the block that defines it. This means that it must be
10688 /// obvious the value of the load is not changed from the point of the load to
10689 /// the end of the block it is in.
10691 /// Finally, it is safe, but not profitable, to sink a load targetting a
10692 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10694 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10695 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10697 for (++BBI; BBI != E; ++BBI)
10698 if (BBI->mayWriteToMemory())
10701 // Check for non-address taken alloca. If not address-taken already, it isn't
10702 // profitable to do this xform.
10703 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10704 bool isAddressTaken = false;
10705 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10707 if (isa<LoadInst>(UI)) continue;
10708 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10709 // If storing TO the alloca, then the address isn't taken.
10710 if (SI->getOperand(1) == AI) continue;
10712 isAddressTaken = true;
10716 if (!isAddressTaken && AI->isStaticAlloca())
10720 // If this load is a load from a GEP with a constant offset from an alloca,
10721 // then we don't want to sink it. In its present form, it will be
10722 // load [constant stack offset]. Sinking it will cause us to have to
10723 // materialize the stack addresses in each predecessor in a register only to
10724 // do a shared load from register in the successor.
10725 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10726 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10727 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10734 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10735 // operator and they all are only used by the PHI, PHI together their
10736 // inputs, and do the operation once, to the result of the PHI.
10737 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10738 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10740 // Scan the instruction, looking for input operations that can be folded away.
10741 // If all input operands to the phi are the same instruction (e.g. a cast from
10742 // the same type or "+42") we can pull the operation through the PHI, reducing
10743 // code size and simplifying code.
10744 Constant *ConstantOp = 0;
10745 const Type *CastSrcTy = 0;
10746 bool isVolatile = false;
10747 if (isa<CastInst>(FirstInst)) {
10748 CastSrcTy = FirstInst->getOperand(0)->getType();
10749 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10750 // Can fold binop, compare or shift here if the RHS is a constant,
10751 // otherwise call FoldPHIArgBinOpIntoPHI.
10752 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10753 if (ConstantOp == 0)
10754 return FoldPHIArgBinOpIntoPHI(PN);
10755 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10756 isVolatile = LI->isVolatile();
10757 // We can't sink the load if the loaded value could be modified between the
10758 // load and the PHI.
10759 if (LI->getParent() != PN.getIncomingBlock(0) ||
10760 !isSafeAndProfitableToSinkLoad(LI))
10763 // If the PHI is of volatile loads and the load block has multiple
10764 // successors, sinking it would remove a load of the volatile value from
10765 // the path through the other successor.
10767 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10770 } else if (isa<GetElementPtrInst>(FirstInst)) {
10771 return FoldPHIArgGEPIntoPHI(PN);
10773 return 0; // Cannot fold this operation.
10776 // Check to see if all arguments are the same operation.
10777 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10778 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10779 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10780 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10783 if (I->getOperand(0)->getType() != CastSrcTy)
10784 return 0; // Cast operation must match.
10785 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10786 // We can't sink the load if the loaded value could be modified between
10787 // the load and the PHI.
10788 if (LI->isVolatile() != isVolatile ||
10789 LI->getParent() != PN.getIncomingBlock(i) ||
10790 !isSafeAndProfitableToSinkLoad(LI))
10793 // If the PHI is of volatile loads and the load block has multiple
10794 // successors, sinking it would remove a load of the volatile value from
10795 // the path through the other successor.
10797 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10800 } else if (I->getOperand(1) != ConstantOp) {
10805 // Okay, they are all the same operation. Create a new PHI node of the
10806 // correct type, and PHI together all of the LHS's of the instructions.
10807 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10808 PN.getName()+".in");
10809 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10811 Value *InVal = FirstInst->getOperand(0);
10812 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10814 // Add all operands to the new PHI.
10815 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10816 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10817 if (NewInVal != InVal)
10819 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10824 // The new PHI unions all of the same values together. This is really
10825 // common, so we handle it intelligently here for compile-time speed.
10829 InsertNewInstBefore(NewPN, PN);
10833 // Insert and return the new operation.
10834 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10835 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10836 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10837 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10838 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10839 return CmpInst::Create(CIOp->getOpcode(), CIOp->getPredicate(),
10840 PhiVal, ConstantOp);
10841 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10843 // If this was a volatile load that we are merging, make sure to loop through
10844 // and mark all the input loads as non-volatile. If we don't do this, we will
10845 // insert a new volatile load and the old ones will not be deletable.
10847 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10848 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10850 return new LoadInst(PhiVal, "", isVolatile);
10853 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10855 static bool DeadPHICycle(PHINode *PN,
10856 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10857 if (PN->use_empty()) return true;
10858 if (!PN->hasOneUse()) return false;
10860 // Remember this node, and if we find the cycle, return.
10861 if (!PotentiallyDeadPHIs.insert(PN))
10864 // Don't scan crazily complex things.
10865 if (PotentiallyDeadPHIs.size() == 16)
10868 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10869 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10874 /// PHIsEqualValue - Return true if this phi node is always equal to
10875 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10876 /// z = some value; x = phi (y, z); y = phi (x, z)
10877 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10878 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10879 // See if we already saw this PHI node.
10880 if (!ValueEqualPHIs.insert(PN))
10883 // Don't scan crazily complex things.
10884 if (ValueEqualPHIs.size() == 16)
10887 // Scan the operands to see if they are either phi nodes or are equal to
10889 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10890 Value *Op = PN->getIncomingValue(i);
10891 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10892 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10894 } else if (Op != NonPhiInVal)
10902 // PHINode simplification
10904 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10905 // If LCSSA is around, don't mess with Phi nodes
10906 if (MustPreserveLCSSA) return 0;
10908 if (Value *V = PN.hasConstantValue())
10909 return ReplaceInstUsesWith(PN, V);
10911 // If all PHI operands are the same operation, pull them through the PHI,
10912 // reducing code size.
10913 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10914 isa<Instruction>(PN.getIncomingValue(1)) &&
10915 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10916 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10917 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10918 // than themselves more than once.
10919 PN.getIncomingValue(0)->hasOneUse())
10920 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10923 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10924 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10925 // PHI)... break the cycle.
10926 if (PN.hasOneUse()) {
10927 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10928 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10929 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10930 PotentiallyDeadPHIs.insert(&PN);
10931 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10932 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10935 // If this phi has a single use, and if that use just computes a value for
10936 // the next iteration of a loop, delete the phi. This occurs with unused
10937 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10938 // common case here is good because the only other things that catch this
10939 // are induction variable analysis (sometimes) and ADCE, which is only run
10941 if (PHIUser->hasOneUse() &&
10942 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10943 PHIUser->use_back() == &PN) {
10944 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10948 // We sometimes end up with phi cycles that non-obviously end up being the
10949 // same value, for example:
10950 // z = some value; x = phi (y, z); y = phi (x, z)
10951 // where the phi nodes don't necessarily need to be in the same block. Do a
10952 // quick check to see if the PHI node only contains a single non-phi value, if
10953 // so, scan to see if the phi cycle is actually equal to that value.
10955 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10956 // Scan for the first non-phi operand.
10957 while (InValNo != NumOperandVals &&
10958 isa<PHINode>(PN.getIncomingValue(InValNo)))
10961 if (InValNo != NumOperandVals) {
10962 Value *NonPhiInVal = PN.getOperand(InValNo);
10964 // Scan the rest of the operands to see if there are any conflicts, if so
10965 // there is no need to recursively scan other phis.
10966 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10967 Value *OpVal = PN.getIncomingValue(InValNo);
10968 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10972 // If we scanned over all operands, then we have one unique value plus
10973 // phi values. Scan PHI nodes to see if they all merge in each other or
10975 if (InValNo == NumOperandVals) {
10976 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10977 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10978 return ReplaceInstUsesWith(PN, NonPhiInVal);
10985 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10986 Instruction *InsertPoint,
10987 InstCombiner *IC) {
10988 unsigned PtrSize = DTy->getScalarSizeInBits();
10989 unsigned VTySize = V->getType()->getScalarSizeInBits();
10990 // We must cast correctly to the pointer type. Ensure that we
10991 // sign extend the integer value if it is smaller as this is
10992 // used for address computation.
10993 Instruction::CastOps opcode =
10994 (VTySize < PtrSize ? Instruction::SExt :
10995 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10996 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
11000 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
11001 Value *PtrOp = GEP.getOperand(0);
11002 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
11003 // If so, eliminate the noop.
11004 if (GEP.getNumOperands() == 1)
11005 return ReplaceInstUsesWith(GEP, PtrOp);
11007 if (isa<UndefValue>(GEP.getOperand(0)))
11008 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
11010 bool HasZeroPointerIndex = false;
11011 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
11012 HasZeroPointerIndex = C->isNullValue();
11014 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11015 return ReplaceInstUsesWith(GEP, PtrOp);
11017 // Eliminate unneeded casts for indices.
11019 bool MadeChange = false;
11020 unsigned PtrSize = TD->getPointerSizeInBits();
11022 gep_type_iterator GTI = gep_type_begin(GEP);
11023 for (User::op_iterator I = GEP.op_begin() + 1, E = GEP.op_end();
11024 I != E; ++I, ++GTI) {
11025 if (!isa<SequentialType>(*GTI)) continue;
11027 // If we are using a wider index than needed for this platform, shrink it
11028 // to what we need. If narrower, sign-extend it to what we need. This
11029 // explicit cast can make subsequent optimizations more obvious.
11030 unsigned OpBits = cast<IntegerType>((*I)->getType())->getBitWidth();
11032 if (OpBits == PtrSize)
11035 Instruction::CastOps Opc =
11036 OpBits > PtrSize ? Instruction::Trunc : Instruction::SExt;
11037 *I = InsertCastBefore(Opc, *I, TD->getIntPtrType(GEP.getContext()), GEP);
11040 if (MadeChange) return &GEP;
11043 // Combine Indices - If the source pointer to this getelementptr instruction
11044 // is a getelementptr instruction, combine the indices of the two
11045 // getelementptr instructions into a single instruction.
11047 SmallVector<Value*, 8> SrcGEPOperands;
11048 bool BothInBounds = cast<GEPOperator>(&GEP)->isInBounds();
11049 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
11050 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
11051 if (!Src->isInBounds())
11052 BothInBounds = false;
11055 if (!SrcGEPOperands.empty()) {
11056 // Note that if our source is a gep chain itself that we wait for that
11057 // chain to be resolved before we perform this transformation. This
11058 // avoids us creating a TON of code in some cases.
11060 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
11061 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
11062 return 0; // Wait until our source is folded to completion.
11064 SmallVector<Value*, 8> Indices;
11066 // Find out whether the last index in the source GEP is a sequential idx.
11067 bool EndsWithSequential = false;
11068 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
11069 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
11070 EndsWithSequential = !isa<StructType>(*I);
11072 // Can we combine the two pointer arithmetics offsets?
11073 if (EndsWithSequential) {
11074 // Replace: gep (gep %P, long B), long A, ...
11075 // With: T = long A+B; gep %P, T, ...
11077 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
11078 if (SO1 == Constant::getNullValue(SO1->getType())) {
11080 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
11083 // If they aren't the same type, convert both to an integer of the
11084 // target's pointer size.
11085 if (SO1->getType() != GO1->getType()) {
11086 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
11088 ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
11089 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
11091 ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
11093 unsigned PS = TD->getPointerSizeInBits();
11094 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
11095 // Convert GO1 to SO1's type.
11096 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
11098 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
11099 // Convert SO1 to GO1's type.
11100 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
11102 const Type *PT = TD->getIntPtrType(GEP.getContext());
11103 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
11104 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
11108 if (isa<Constant>(SO1) && isa<Constant>(GO1))
11109 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
11111 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11112 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
11116 // Recycle the GEP we already have if possible.
11117 if (SrcGEPOperands.size() == 2) {
11118 GEP.setOperand(0, SrcGEPOperands[0]);
11119 GEP.setOperand(1, Sum);
11122 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11123 SrcGEPOperands.end()-1);
11124 Indices.push_back(Sum);
11125 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
11126 } else if (isa<Constant>(*GEP.idx_begin()) &&
11127 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11128 SrcGEPOperands.size() != 1) {
11129 // Otherwise we can do the fold if the first index of the GEP is a zero
11130 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11131 SrcGEPOperands.end());
11132 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
11135 if (!Indices.empty()) {
11136 GetElementPtrInst *NewGEP =
11137 GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
11138 Indices.end(), GEP.getName());
11140 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11144 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
11145 if (!isa<PointerType>(X->getType())) {
11146 // Not interesting. Source pointer must be a cast from pointer.
11147 } else if (HasZeroPointerIndex) {
11148 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11149 // into : GEP [10 x i8]* X, i32 0, ...
11151 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11152 // into : GEP i8* X, ...
11154 // This occurs when the program declares an array extern like "int X[];"
11155 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11156 const PointerType *XTy = cast<PointerType>(X->getType());
11157 if (const ArrayType *CATy =
11158 dyn_cast<ArrayType>(CPTy->getElementType())) {
11159 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11160 if (CATy->getElementType() == XTy->getElementType()) {
11161 // -> GEP i8* X, ...
11162 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11163 GetElementPtrInst *NewGEP =
11164 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11166 if (cast<GEPOperator>(&GEP)->isInBounds())
11167 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11169 } else if (const ArrayType *XATy =
11170 dyn_cast<ArrayType>(XTy->getElementType())) {
11171 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11172 if (CATy->getElementType() == XATy->getElementType()) {
11173 // -> GEP [10 x i8]* X, i32 0, ...
11174 // At this point, we know that the cast source type is a pointer
11175 // to an array of the same type as the destination pointer
11176 // array. Because the array type is never stepped over (there
11177 // is a leading zero) we can fold the cast into this GEP.
11178 GEP.setOperand(0, X);
11183 } else if (GEP.getNumOperands() == 2) {
11184 // Transform things like:
11185 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11186 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11187 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11188 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11189 if (TD && isa<ArrayType>(SrcElTy) &&
11190 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11191 TD->getTypeAllocSize(ResElTy)) {
11193 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11194 Idx[1] = GEP.getOperand(1);
11195 GetElementPtrInst *NewGEP =
11196 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11197 if (cast<GEPOperator>(&GEP)->isInBounds())
11198 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11199 Value *V = InsertNewInstBefore(NewGEP, GEP);
11200 // V and GEP are both pointer types --> BitCast
11201 return new BitCastInst(V, GEP.getType());
11204 // Transform things like:
11205 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11206 // (where tmp = 8*tmp2) into:
11207 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11209 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::getInt8Ty(*Context)) {
11210 uint64_t ArrayEltSize =
11211 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11213 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11214 // allow either a mul, shift, or constant here.
11216 ConstantInt *Scale = 0;
11217 if (ArrayEltSize == 1) {
11218 NewIdx = GEP.getOperand(1);
11220 ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11221 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11222 NewIdx = ConstantInt::get(CI->getType(), 1);
11224 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11225 if (Inst->getOpcode() == Instruction::Shl &&
11226 isa<ConstantInt>(Inst->getOperand(1))) {
11227 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11228 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11229 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11231 NewIdx = Inst->getOperand(0);
11232 } else if (Inst->getOpcode() == Instruction::Mul &&
11233 isa<ConstantInt>(Inst->getOperand(1))) {
11234 Scale = cast<ConstantInt>(Inst->getOperand(1));
11235 NewIdx = Inst->getOperand(0);
11239 // If the index will be to exactly the right offset with the scale taken
11240 // out, perform the transformation. Note, we don't know whether Scale is
11241 // signed or not. We'll use unsigned version of division/modulo
11242 // operation after making sure Scale doesn't have the sign bit set.
11243 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11244 Scale->getZExtValue() % ArrayEltSize == 0) {
11245 Scale = ConstantInt::get(Scale->getType(),
11246 Scale->getZExtValue() / ArrayEltSize);
11247 if (Scale->getZExtValue() != 1) {
11249 ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11251 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
11252 NewIdx = InsertNewInstBefore(Sc, GEP);
11255 // Insert the new GEP instruction.
11257 Idx[0] = Constant::getNullValue(Type::getInt32Ty(*Context));
11259 Instruction *NewGEP =
11260 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11261 if (cast<GEPOperator>(&GEP)->isInBounds())
11262 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11263 NewGEP = InsertNewInstBefore(NewGEP, GEP);
11264 // The NewGEP must be pointer typed, so must the old one -> BitCast
11265 return new BitCastInst(NewGEP, GEP.getType());
11271 /// See if we can simplify:
11272 /// X = bitcast A to B*
11273 /// Y = gep X, <...constant indices...>
11274 /// into a gep of the original struct. This is important for SROA and alias
11275 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11276 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11278 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11279 // Determine how much the GEP moves the pointer. We are guaranteed to get
11280 // a constant back from EmitGEPOffset.
11281 ConstantInt *OffsetV =
11282 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11283 int64_t Offset = OffsetV->getSExtValue();
11285 // If this GEP instruction doesn't move the pointer, just replace the GEP
11286 // with a bitcast of the real input to the dest type.
11288 // If the bitcast is of an allocation, and the allocation will be
11289 // converted to match the type of the cast, don't touch this.
11290 if (isa<AllocationInst>(BCI->getOperand(0))) {
11291 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11292 if (Instruction *I = visitBitCast(*BCI)) {
11295 BCI->getParent()->getInstList().insert(BCI, I);
11296 ReplaceInstUsesWith(*BCI, I);
11301 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11304 // Otherwise, if the offset is non-zero, we need to find out if there is a
11305 // field at Offset in 'A's type. If so, we can pull the cast through the
11307 SmallVector<Value*, 8> NewIndices;
11309 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11310 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11311 Instruction *NGEP =
11312 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11314 if (NGEP->getType() == GEP.getType()) return NGEP;
11315 if (cast<GEPOperator>(&GEP)->isInBounds())
11316 cast<GEPOperator>(NGEP)->setIsInBounds(true);
11317 InsertNewInstBefore(NGEP, GEP);
11318 NGEP->takeName(&GEP);
11319 return new BitCastInst(NGEP, GEP.getType());
11327 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11328 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11329 if (AI.isArrayAllocation()) { // Check C != 1
11330 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11331 const Type *NewTy =
11332 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11333 AllocationInst *New = 0;
11335 // Create and insert the replacement instruction...
11336 if (isa<MallocInst>(AI))
11337 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11339 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11340 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11343 InsertNewInstBefore(New, AI);
11345 // Scan to the end of the allocation instructions, to skip over a block of
11346 // allocas if possible...also skip interleaved debug info
11348 BasicBlock::iterator It = New;
11349 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11351 // Now that I is pointing to the first non-allocation-inst in the block,
11352 // insert our getelementptr instruction...
11354 Value *NullIdx = Constant::getNullValue(Type::getInt32Ty(*Context));
11358 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11359 New->getName()+".sub", It);
11360 cast<GEPOperator>(V)->setIsInBounds(true);
11362 // Now make everything use the getelementptr instead of the original
11364 return ReplaceInstUsesWith(AI, V);
11365 } else if (isa<UndefValue>(AI.getArraySize())) {
11366 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11370 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11371 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11372 // Note that we only do this for alloca's, because malloc should allocate
11373 // and return a unique pointer, even for a zero byte allocation.
11374 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11375 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11377 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11378 if (AI.getAlignment() == 0)
11379 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11385 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11386 Value *Op = FI.getOperand(0);
11388 // free undef -> unreachable.
11389 if (isa<UndefValue>(Op)) {
11390 // Insert a new store to null because we cannot modify the CFG here.
11391 new StoreInst(ConstantInt::getTrue(*Context),
11392 UndefValue::get(PointerType::getUnqual(Type::getInt1Ty(*Context))), &FI);
11393 return EraseInstFromFunction(FI);
11396 // If we have 'free null' delete the instruction. This can happen in stl code
11397 // when lots of inlining happens.
11398 if (isa<ConstantPointerNull>(Op))
11399 return EraseInstFromFunction(FI);
11401 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11402 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11403 FI.setOperand(0, CI->getOperand(0));
11407 // Change free (gep X, 0,0,0,0) into free(X)
11408 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11409 if (GEPI->hasAllZeroIndices()) {
11410 AddToWorkList(GEPI);
11411 FI.setOperand(0, GEPI->getOperand(0));
11416 // Change free(malloc) into nothing, if the malloc has a single use.
11417 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11418 if (MI->hasOneUse()) {
11419 EraseInstFromFunction(FI);
11420 return EraseInstFromFunction(*MI);
11427 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11428 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11429 const TargetData *TD) {
11430 User *CI = cast<User>(LI.getOperand(0));
11431 Value *CastOp = CI->getOperand(0);
11432 LLVMContext *Context = IC.getContext();
11435 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11436 // Instead of loading constant c string, use corresponding integer value
11437 // directly if string length is small enough.
11439 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11440 unsigned len = Str.length();
11441 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11442 unsigned numBits = Ty->getPrimitiveSizeInBits();
11443 // Replace LI with immediate integer store.
11444 if ((numBits >> 3) == len + 1) {
11445 APInt StrVal(numBits, 0);
11446 APInt SingleChar(numBits, 0);
11447 if (TD->isLittleEndian()) {
11448 for (signed i = len-1; i >= 0; i--) {
11449 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11450 StrVal = (StrVal << 8) | SingleChar;
11453 for (unsigned i = 0; i < len; i++) {
11454 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11455 StrVal = (StrVal << 8) | SingleChar;
11457 // Append NULL at the end.
11459 StrVal = (StrVal << 8) | SingleChar;
11461 Value *NL = ConstantInt::get(*Context, StrVal);
11462 return IC.ReplaceInstUsesWith(LI, NL);
11468 const PointerType *DestTy = cast<PointerType>(CI->getType());
11469 const Type *DestPTy = DestTy->getElementType();
11470 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11472 // If the address spaces don't match, don't eliminate the cast.
11473 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11476 const Type *SrcPTy = SrcTy->getElementType();
11478 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11479 isa<VectorType>(DestPTy)) {
11480 // If the source is an array, the code below will not succeed. Check to
11481 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11483 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11484 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11485 if (ASrcTy->getNumElements() != 0) {
11487 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::getInt32Ty(*Context));
11488 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11489 SrcTy = cast<PointerType>(CastOp->getType());
11490 SrcPTy = SrcTy->getElementType();
11493 if (IC.getTargetData() &&
11494 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11495 isa<VectorType>(SrcPTy)) &&
11496 // Do not allow turning this into a load of an integer, which is then
11497 // casted to a pointer, this pessimizes pointer analysis a lot.
11498 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11499 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
11500 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
11502 // Okay, we are casting from one integer or pointer type to another of
11503 // the same size. Instead of casting the pointer before the load, cast
11504 // the result of the loaded value.
11505 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11507 LI.isVolatile()),LI);
11508 // Now cast the result of the load.
11509 return new BitCastInst(NewLoad, LI.getType());
11516 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11517 Value *Op = LI.getOperand(0);
11519 // Attempt to improve the alignment.
11521 unsigned KnownAlign =
11522 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11524 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11525 LI.getAlignment()))
11526 LI.setAlignment(KnownAlign);
11529 // load (cast X) --> cast (load X) iff safe
11530 if (isa<CastInst>(Op))
11531 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11534 // None of the following transforms are legal for volatile loads.
11535 if (LI.isVolatile()) return 0;
11537 // Do really simple store-to-load forwarding and load CSE, to catch cases
11538 // where there are several consequtive memory accesses to the same location,
11539 // separated by a few arithmetic operations.
11540 BasicBlock::iterator BBI = &LI;
11541 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11542 return ReplaceInstUsesWith(LI, AvailableVal);
11544 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11545 const Value *GEPI0 = GEPI->getOperand(0);
11546 // TODO: Consider a target hook for valid address spaces for this xform.
11547 if (isa<ConstantPointerNull>(GEPI0) &&
11548 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11549 // Insert a new store to null instruction before the load to indicate
11550 // that this code is not reachable. We do this instead of inserting
11551 // an unreachable instruction directly because we cannot modify the
11553 new StoreInst(UndefValue::get(LI.getType()),
11554 Constant::getNullValue(Op->getType()), &LI);
11555 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11559 if (Constant *C = dyn_cast<Constant>(Op)) {
11560 // load null/undef -> undef
11561 // TODO: Consider a target hook for valid address spaces for this xform.
11562 if (isa<UndefValue>(C) || (C->isNullValue() &&
11563 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11564 // Insert a new store to null instruction before the load to indicate that
11565 // this code is not reachable. We do this instead of inserting an
11566 // unreachable instruction directly because we cannot modify the CFG.
11567 new StoreInst(UndefValue::get(LI.getType()),
11568 Constant::getNullValue(Op->getType()), &LI);
11569 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11572 // Instcombine load (constant global) into the value loaded.
11573 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11574 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11575 return ReplaceInstUsesWith(LI, GV->getInitializer());
11577 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11578 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11579 if (CE->getOpcode() == Instruction::GetElementPtr) {
11580 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11581 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11583 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11585 return ReplaceInstUsesWith(LI, V);
11586 if (CE->getOperand(0)->isNullValue()) {
11587 // Insert a new store to null instruction before the load to indicate
11588 // that this code is not reachable. We do this instead of inserting
11589 // an unreachable instruction directly because we cannot modify the
11591 new StoreInst(UndefValue::get(LI.getType()),
11592 Constant::getNullValue(Op->getType()), &LI);
11593 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11596 } else if (CE->isCast()) {
11597 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11603 // If this load comes from anywhere in a constant global, and if the global
11604 // is all undef or zero, we know what it loads.
11605 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11606 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11607 if (GV->getInitializer()->isNullValue())
11608 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11609 else if (isa<UndefValue>(GV->getInitializer()))
11610 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11614 if (Op->hasOneUse()) {
11615 // Change select and PHI nodes to select values instead of addresses: this
11616 // helps alias analysis out a lot, allows many others simplifications, and
11617 // exposes redundancy in the code.
11619 // Note that we cannot do the transformation unless we know that the
11620 // introduced loads cannot trap! Something like this is valid as long as
11621 // the condition is always false: load (select bool %C, int* null, int* %G),
11622 // but it would not be valid if we transformed it to load from null
11623 // unconditionally.
11625 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11626 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11627 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11628 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11629 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11630 SI->getOperand(1)->getName()+".val"), LI);
11631 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11632 SI->getOperand(2)->getName()+".val"), LI);
11633 return SelectInst::Create(SI->getCondition(), V1, V2);
11636 // load (select (cond, null, P)) -> load P
11637 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11638 if (C->isNullValue()) {
11639 LI.setOperand(0, SI->getOperand(2));
11643 // load (select (cond, P, null)) -> load P
11644 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11645 if (C->isNullValue()) {
11646 LI.setOperand(0, SI->getOperand(1));
11654 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11655 /// when possible. This makes it generally easy to do alias analysis and/or
11656 /// SROA/mem2reg of the memory object.
11657 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11658 User *CI = cast<User>(SI.getOperand(1));
11659 Value *CastOp = CI->getOperand(0);
11661 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11662 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11663 if (SrcTy == 0) return 0;
11665 const Type *SrcPTy = SrcTy->getElementType();
11667 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11670 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11671 /// to its first element. This allows us to handle things like:
11672 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11673 /// on 32-bit hosts.
11674 SmallVector<Value*, 4> NewGEPIndices;
11676 // If the source is an array, the code below will not succeed. Check to
11677 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11679 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11680 // Index through pointer.
11681 Constant *Zero = Constant::getNullValue(Type::getInt32Ty(*IC.getContext()));
11682 NewGEPIndices.push_back(Zero);
11685 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11686 if (!STy->getNumElements()) /* Struct can be empty {} */
11688 NewGEPIndices.push_back(Zero);
11689 SrcPTy = STy->getElementType(0);
11690 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11691 NewGEPIndices.push_back(Zero);
11692 SrcPTy = ATy->getElementType();
11698 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11701 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11704 // If the pointers point into different address spaces or if they point to
11705 // values with different sizes, we can't do the transformation.
11706 if (!IC.getTargetData() ||
11707 SrcTy->getAddressSpace() !=
11708 cast<PointerType>(CI->getType())->getAddressSpace() ||
11709 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
11710 IC.getTargetData()->getTypeSizeInBits(DestPTy))
11713 // Okay, we are casting from one integer or pointer type to another of
11714 // the same size. Instead of casting the pointer before
11715 // the store, cast the value to be stored.
11717 Value *SIOp0 = SI.getOperand(0);
11718 Instruction::CastOps opcode = Instruction::BitCast;
11719 const Type* CastSrcTy = SIOp0->getType();
11720 const Type* CastDstTy = SrcPTy;
11721 if (isa<PointerType>(CastDstTy)) {
11722 if (CastSrcTy->isInteger())
11723 opcode = Instruction::IntToPtr;
11724 } else if (isa<IntegerType>(CastDstTy)) {
11725 if (isa<PointerType>(SIOp0->getType()))
11726 opcode = Instruction::PtrToInt;
11729 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11730 // emit a GEP to index into its first field.
11731 if (!NewGEPIndices.empty()) {
11732 if (Constant *C = dyn_cast<Constant>(CastOp))
11733 CastOp = ConstantExpr::getGetElementPtr(C, &NewGEPIndices[0],
11734 NewGEPIndices.size());
11736 CastOp = IC.InsertNewInstBefore(
11737 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11738 NewGEPIndices.end()), SI);
11739 cast<GEPOperator>(CastOp)->setIsInBounds(true);
11742 if (Constant *C = dyn_cast<Constant>(SIOp0))
11743 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11745 NewCast = IC.InsertNewInstBefore(
11746 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11748 return new StoreInst(NewCast, CastOp);
11751 /// equivalentAddressValues - Test if A and B will obviously have the same
11752 /// value. This includes recognizing that %t0 and %t1 will have the same
11753 /// value in code like this:
11754 /// %t0 = getelementptr \@a, 0, 3
11755 /// store i32 0, i32* %t0
11756 /// %t1 = getelementptr \@a, 0, 3
11757 /// %t2 = load i32* %t1
11759 static bool equivalentAddressValues(Value *A, Value *B) {
11760 // Test if the values are trivially equivalent.
11761 if (A == B) return true;
11763 // Test if the values come form identical arithmetic instructions.
11764 // This uses isIdenticalToWhenDefined instead of isIdenticalTo because
11765 // its only used to compare two uses within the same basic block, which
11766 // means that they'll always either have the same value or one of them
11767 // will have an undefined value.
11768 if (isa<BinaryOperator>(A) ||
11769 isa<CastInst>(A) ||
11771 isa<GetElementPtrInst>(A))
11772 if (Instruction *BI = dyn_cast<Instruction>(B))
11773 if (cast<Instruction>(A)->isIdenticalToWhenDefined(BI))
11776 // Otherwise they may not be equivalent.
11780 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11781 // return the llvm.dbg.declare.
11782 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11783 if (!V->hasNUses(2))
11785 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11787 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11789 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11790 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11797 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11798 Value *Val = SI.getOperand(0);
11799 Value *Ptr = SI.getOperand(1);
11801 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11802 EraseInstFromFunction(SI);
11807 // If the RHS is an alloca with a single use, zapify the store, making the
11809 // If the RHS is an alloca with a two uses, the other one being a
11810 // llvm.dbg.declare, zapify the store and the declare, making the
11811 // alloca dead. We must do this to prevent declare's from affecting
11813 if (!SI.isVolatile()) {
11814 if (Ptr->hasOneUse()) {
11815 if (isa<AllocaInst>(Ptr)) {
11816 EraseInstFromFunction(SI);
11820 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11821 if (isa<AllocaInst>(GEP->getOperand(0))) {
11822 if (GEP->getOperand(0)->hasOneUse()) {
11823 EraseInstFromFunction(SI);
11827 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11828 EraseInstFromFunction(*DI);
11829 EraseInstFromFunction(SI);
11836 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11837 EraseInstFromFunction(*DI);
11838 EraseInstFromFunction(SI);
11844 // Attempt to improve the alignment.
11846 unsigned KnownAlign =
11847 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11849 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11850 SI.getAlignment()))
11851 SI.setAlignment(KnownAlign);
11854 // Do really simple DSE, to catch cases where there are several consecutive
11855 // stores to the same location, separated by a few arithmetic operations. This
11856 // situation often occurs with bitfield accesses.
11857 BasicBlock::iterator BBI = &SI;
11858 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11861 // Don't count debug info directives, lest they affect codegen,
11862 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11863 // It is necessary for correctness to skip those that feed into a
11864 // llvm.dbg.declare, as these are not present when debugging is off.
11865 if (isa<DbgInfoIntrinsic>(BBI) ||
11866 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11871 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11872 // Prev store isn't volatile, and stores to the same location?
11873 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11874 SI.getOperand(1))) {
11877 EraseInstFromFunction(*PrevSI);
11883 // If this is a load, we have to stop. However, if the loaded value is from
11884 // the pointer we're loading and is producing the pointer we're storing,
11885 // then *this* store is dead (X = load P; store X -> P).
11886 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11887 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11888 !SI.isVolatile()) {
11889 EraseInstFromFunction(SI);
11893 // Otherwise, this is a load from some other location. Stores before it
11894 // may not be dead.
11898 // Don't skip over loads or things that can modify memory.
11899 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11904 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11906 // store X, null -> turns into 'unreachable' in SimplifyCFG
11907 if (isa<ConstantPointerNull>(Ptr) &&
11908 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11909 if (!isa<UndefValue>(Val)) {
11910 SI.setOperand(0, UndefValue::get(Val->getType()));
11911 if (Instruction *U = dyn_cast<Instruction>(Val))
11912 AddToWorkList(U); // Dropped a use.
11915 return 0; // Do not modify these!
11918 // store undef, Ptr -> noop
11919 if (isa<UndefValue>(Val)) {
11920 EraseInstFromFunction(SI);
11925 // If the pointer destination is a cast, see if we can fold the cast into the
11927 if (isa<CastInst>(Ptr))
11928 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11930 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11932 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11936 // If this store is the last instruction in the basic block (possibly
11937 // excepting debug info instructions and the pointer bitcasts that feed
11938 // into them), and if the block ends with an unconditional branch, try
11939 // to move it to the successor block.
11943 } while (isa<DbgInfoIntrinsic>(BBI) ||
11944 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11945 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11946 if (BI->isUnconditional())
11947 if (SimplifyStoreAtEndOfBlock(SI))
11948 return 0; // xform done!
11953 /// SimplifyStoreAtEndOfBlock - Turn things like:
11954 /// if () { *P = v1; } else { *P = v2 }
11955 /// into a phi node with a store in the successor.
11957 /// Simplify things like:
11958 /// *P = v1; if () { *P = v2; }
11959 /// into a phi node with a store in the successor.
11961 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11962 BasicBlock *StoreBB = SI.getParent();
11964 // Check to see if the successor block has exactly two incoming edges. If
11965 // so, see if the other predecessor contains a store to the same location.
11966 // if so, insert a PHI node (if needed) and move the stores down.
11967 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11969 // Determine whether Dest has exactly two predecessors and, if so, compute
11970 // the other predecessor.
11971 pred_iterator PI = pred_begin(DestBB);
11972 BasicBlock *OtherBB = 0;
11973 if (*PI != StoreBB)
11976 if (PI == pred_end(DestBB))
11979 if (*PI != StoreBB) {
11984 if (++PI != pred_end(DestBB))
11987 // Bail out if all the relevant blocks aren't distinct (this can happen,
11988 // for example, if SI is in an infinite loop)
11989 if (StoreBB == DestBB || OtherBB == DestBB)
11992 // Verify that the other block ends in a branch and is not otherwise empty.
11993 BasicBlock::iterator BBI = OtherBB->getTerminator();
11994 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
11995 if (!OtherBr || BBI == OtherBB->begin())
11998 // If the other block ends in an unconditional branch, check for the 'if then
11999 // else' case. there is an instruction before the branch.
12000 StoreInst *OtherStore = 0;
12001 if (OtherBr->isUnconditional()) {
12003 // Skip over debugging info.
12004 while (isa<DbgInfoIntrinsic>(BBI) ||
12005 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12006 if (BBI==OtherBB->begin())
12010 // If this isn't a store, or isn't a store to the same location, bail out.
12011 OtherStore = dyn_cast<StoreInst>(BBI);
12012 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
12015 // Otherwise, the other block ended with a conditional branch. If one of the
12016 // destinations is StoreBB, then we have the if/then case.
12017 if (OtherBr->getSuccessor(0) != StoreBB &&
12018 OtherBr->getSuccessor(1) != StoreBB)
12021 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12022 // if/then triangle. See if there is a store to the same ptr as SI that
12023 // lives in OtherBB.
12025 // Check to see if we find the matching store.
12026 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12027 if (OtherStore->getOperand(1) != SI.getOperand(1))
12031 // If we find something that may be using or overwriting the stored
12032 // value, or if we run out of instructions, we can't do the xform.
12033 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12034 BBI == OtherBB->begin())
12038 // In order to eliminate the store in OtherBr, we have to
12039 // make sure nothing reads or overwrites the stored value in
12041 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12042 // FIXME: This should really be AA driven.
12043 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12048 // Insert a PHI node now if we need it.
12049 Value *MergedVal = OtherStore->getOperand(0);
12050 if (MergedVal != SI.getOperand(0)) {
12051 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12052 PN->reserveOperandSpace(2);
12053 PN->addIncoming(SI.getOperand(0), SI.getParent());
12054 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12055 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12058 // Advance to a place where it is safe to insert the new store and
12060 BBI = DestBB->getFirstNonPHI();
12061 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12062 OtherStore->isVolatile()), *BBI);
12064 // Nuke the old stores.
12065 EraseInstFromFunction(SI);
12066 EraseInstFromFunction(*OtherStore);
12072 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12073 // Change br (not X), label True, label False to: br X, label False, True
12075 BasicBlock *TrueDest;
12076 BasicBlock *FalseDest;
12077 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
12078 !isa<Constant>(X)) {
12079 // Swap Destinations and condition...
12080 BI.setCondition(X);
12081 BI.setSuccessor(0, FalseDest);
12082 BI.setSuccessor(1, TrueDest);
12086 // Cannonicalize fcmp_one -> fcmp_oeq
12087 FCmpInst::Predicate FPred; Value *Y;
12088 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12089 TrueDest, FalseDest)))
12090 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12091 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
12092 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
12093 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
12094 Instruction *NewSCC = new FCmpInst(I, NewPred, X, Y, "");
12095 NewSCC->takeName(I);
12096 // Swap Destinations and condition...
12097 BI.setCondition(NewSCC);
12098 BI.setSuccessor(0, FalseDest);
12099 BI.setSuccessor(1, TrueDest);
12100 RemoveFromWorkList(I);
12101 I->eraseFromParent();
12102 AddToWorkList(NewSCC);
12106 // Cannonicalize icmp_ne -> icmp_eq
12107 ICmpInst::Predicate IPred;
12108 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12109 TrueDest, FalseDest)))
12110 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12111 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12112 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
12113 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
12114 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
12115 Instruction *NewSCC = new ICmpInst(I, NewPred, X, Y, "");
12116 NewSCC->takeName(I);
12117 // Swap Destinations and condition...
12118 BI.setCondition(NewSCC);
12119 BI.setSuccessor(0, FalseDest);
12120 BI.setSuccessor(1, TrueDest);
12121 RemoveFromWorkList(I);
12122 I->eraseFromParent();;
12123 AddToWorkList(NewSCC);
12130 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12131 Value *Cond = SI.getCondition();
12132 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12133 if (I->getOpcode() == Instruction::Add)
12134 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12135 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12136 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12138 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
12140 SI.setOperand(0, I->getOperand(0));
12148 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12149 Value *Agg = EV.getAggregateOperand();
12151 if (!EV.hasIndices())
12152 return ReplaceInstUsesWith(EV, Agg);
12154 if (Constant *C = dyn_cast<Constant>(Agg)) {
12155 if (isa<UndefValue>(C))
12156 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12158 if (isa<ConstantAggregateZero>(C))
12159 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12161 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12162 // Extract the element indexed by the first index out of the constant
12163 Value *V = C->getOperand(*EV.idx_begin());
12164 if (EV.getNumIndices() > 1)
12165 // Extract the remaining indices out of the constant indexed by the
12167 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12169 return ReplaceInstUsesWith(EV, V);
12171 return 0; // Can't handle other constants
12173 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12174 // We're extracting from an insertvalue instruction, compare the indices
12175 const unsigned *exti, *exte, *insi, *inse;
12176 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12177 exte = EV.idx_end(), inse = IV->idx_end();
12178 exti != exte && insi != inse;
12180 if (*insi != *exti)
12181 // The insert and extract both reference distinctly different elements.
12182 // This means the extract is not influenced by the insert, and we can
12183 // replace the aggregate operand of the extract with the aggregate
12184 // operand of the insert. i.e., replace
12185 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12186 // %E = extractvalue { i32, { i32 } } %I, 0
12188 // %E = extractvalue { i32, { i32 } } %A, 0
12189 return ExtractValueInst::Create(IV->getAggregateOperand(),
12190 EV.idx_begin(), EV.idx_end());
12192 if (exti == exte && insi == inse)
12193 // Both iterators are at the end: Index lists are identical. Replace
12194 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12195 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12197 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12198 if (exti == exte) {
12199 // The extract list is a prefix of the insert list. i.e. replace
12200 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12201 // %E = extractvalue { i32, { i32 } } %I, 1
12203 // %X = extractvalue { i32, { i32 } } %A, 1
12204 // %E = insertvalue { i32 } %X, i32 42, 0
12205 // by switching the order of the insert and extract (though the
12206 // insertvalue should be left in, since it may have other uses).
12207 Value *NewEV = InsertNewInstBefore(
12208 ExtractValueInst::Create(IV->getAggregateOperand(),
12209 EV.idx_begin(), EV.idx_end()),
12211 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12215 // The insert list is a prefix of the extract list
12216 // We can simply remove the common indices from the extract and make it
12217 // operate on the inserted value instead of the insertvalue result.
12219 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12220 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12222 // %E extractvalue { i32 } { i32 42 }, 0
12223 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12226 // Can't simplify extracts from other values. Note that nested extracts are
12227 // already simplified implicitely by the above (extract ( extract (insert) )
12228 // will be translated into extract ( insert ( extract ) ) first and then just
12229 // the value inserted, if appropriate).
12233 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12234 /// is to leave as a vector operation.
12235 static bool CheapToScalarize(Value *V, bool isConstant) {
12236 if (isa<ConstantAggregateZero>(V))
12238 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12239 if (isConstant) return true;
12240 // If all elts are the same, we can extract.
12241 Constant *Op0 = C->getOperand(0);
12242 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12243 if (C->getOperand(i) != Op0)
12247 Instruction *I = dyn_cast<Instruction>(V);
12248 if (!I) return false;
12250 // Insert element gets simplified to the inserted element or is deleted if
12251 // this is constant idx extract element and its a constant idx insertelt.
12252 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12253 isa<ConstantInt>(I->getOperand(2)))
12255 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12257 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12258 if (BO->hasOneUse() &&
12259 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12260 CheapToScalarize(BO->getOperand(1), isConstant)))
12262 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12263 if (CI->hasOneUse() &&
12264 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12265 CheapToScalarize(CI->getOperand(1), isConstant)))
12271 /// Read and decode a shufflevector mask.
12273 /// It turns undef elements into values that are larger than the number of
12274 /// elements in the input.
12275 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12276 unsigned NElts = SVI->getType()->getNumElements();
12277 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12278 return std::vector<unsigned>(NElts, 0);
12279 if (isa<UndefValue>(SVI->getOperand(2)))
12280 return std::vector<unsigned>(NElts, 2*NElts);
12282 std::vector<unsigned> Result;
12283 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12284 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12285 if (isa<UndefValue>(*i))
12286 Result.push_back(NElts*2); // undef -> 8
12288 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12292 /// FindScalarElement - Given a vector and an element number, see if the scalar
12293 /// value is already around as a register, for example if it were inserted then
12294 /// extracted from the vector.
12295 static Value *FindScalarElement(Value *V, unsigned EltNo,
12296 LLVMContext *Context) {
12297 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12298 const VectorType *PTy = cast<VectorType>(V->getType());
12299 unsigned Width = PTy->getNumElements();
12300 if (EltNo >= Width) // Out of range access.
12301 return UndefValue::get(PTy->getElementType());
12303 if (isa<UndefValue>(V))
12304 return UndefValue::get(PTy->getElementType());
12305 else if (isa<ConstantAggregateZero>(V))
12306 return Constant::getNullValue(PTy->getElementType());
12307 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12308 return CP->getOperand(EltNo);
12309 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12310 // If this is an insert to a variable element, we don't know what it is.
12311 if (!isa<ConstantInt>(III->getOperand(2)))
12313 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12315 // If this is an insert to the element we are looking for, return the
12317 if (EltNo == IIElt)
12318 return III->getOperand(1);
12320 // Otherwise, the insertelement doesn't modify the value, recurse on its
12322 return FindScalarElement(III->getOperand(0), EltNo, Context);
12323 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12324 unsigned LHSWidth =
12325 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12326 unsigned InEl = getShuffleMask(SVI)[EltNo];
12327 if (InEl < LHSWidth)
12328 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12329 else if (InEl < LHSWidth*2)
12330 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12332 return UndefValue::get(PTy->getElementType());
12335 // Otherwise, we don't know.
12339 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12340 // If vector val is undef, replace extract with scalar undef.
12341 if (isa<UndefValue>(EI.getOperand(0)))
12342 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12344 // If vector val is constant 0, replace extract with scalar 0.
12345 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12346 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12348 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12349 // If vector val is constant with all elements the same, replace EI with
12350 // that element. When the elements are not identical, we cannot replace yet
12351 // (we do that below, but only when the index is constant).
12352 Constant *op0 = C->getOperand(0);
12353 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12354 if (C->getOperand(i) != op0) {
12359 return ReplaceInstUsesWith(EI, op0);
12362 // If extracting a specified index from the vector, see if we can recursively
12363 // find a previously computed scalar that was inserted into the vector.
12364 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12365 unsigned IndexVal = IdxC->getZExtValue();
12366 unsigned VectorWidth =
12367 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12369 // If this is extracting an invalid index, turn this into undef, to avoid
12370 // crashing the code below.
12371 if (IndexVal >= VectorWidth)
12372 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12374 // This instruction only demands the single element from the input vector.
12375 // If the input vector has a single use, simplify it based on this use
12377 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12378 APInt UndefElts(VectorWidth, 0);
12379 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12380 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12381 DemandedMask, UndefElts)) {
12382 EI.setOperand(0, V);
12387 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12388 return ReplaceInstUsesWith(EI, Elt);
12390 // If the this extractelement is directly using a bitcast from a vector of
12391 // the same number of elements, see if we can find the source element from
12392 // it. In this case, we will end up needing to bitcast the scalars.
12393 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12394 if (const VectorType *VT =
12395 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12396 if (VT->getNumElements() == VectorWidth)
12397 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12398 IndexVal, Context))
12399 return new BitCastInst(Elt, EI.getType());
12403 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12404 if (I->hasOneUse()) {
12405 // Push extractelement into predecessor operation if legal and
12406 // profitable to do so
12407 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12408 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12409 if (CheapToScalarize(BO, isConstantElt)) {
12410 ExtractElementInst *newEI0 =
12411 ExtractElementInst::Create(BO->getOperand(0), EI.getOperand(1),
12412 EI.getName()+".lhs");
12413 ExtractElementInst *newEI1 =
12414 ExtractElementInst::Create(BO->getOperand(1), EI.getOperand(1),
12415 EI.getName()+".rhs");
12416 InsertNewInstBefore(newEI0, EI);
12417 InsertNewInstBefore(newEI1, EI);
12418 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12420 } else if (isa<LoadInst>(I)) {
12422 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12423 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12424 PointerType::get(EI.getType(), AS),*I);
12425 GetElementPtrInst *GEP =
12426 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12427 cast<GEPOperator>(GEP)->setIsInBounds(true);
12428 InsertNewInstBefore(GEP, *I);
12429 LoadInst* Load = new LoadInst(GEP, "tmp");
12430 InsertNewInstBefore(Load, *I);
12431 return ReplaceInstUsesWith(EI, Load);
12434 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12435 // Extracting the inserted element?
12436 if (IE->getOperand(2) == EI.getOperand(1))
12437 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12438 // If the inserted and extracted elements are constants, they must not
12439 // be the same value, extract from the pre-inserted value instead.
12440 if (isa<Constant>(IE->getOperand(2)) &&
12441 isa<Constant>(EI.getOperand(1))) {
12442 AddUsesToWorkList(EI);
12443 EI.setOperand(0, IE->getOperand(0));
12446 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12447 // If this is extracting an element from a shufflevector, figure out where
12448 // it came from and extract from the appropriate input element instead.
12449 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12450 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12452 unsigned LHSWidth =
12453 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12455 if (SrcIdx < LHSWidth)
12456 Src = SVI->getOperand(0);
12457 else if (SrcIdx < LHSWidth*2) {
12458 SrcIdx -= LHSWidth;
12459 Src = SVI->getOperand(1);
12461 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12463 return ExtractElementInst::Create(Src,
12464 ConstantInt::get(Type::getInt32Ty(*Context), SrcIdx, false));
12467 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12472 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12473 /// elements from either LHS or RHS, return the shuffle mask and true.
12474 /// Otherwise, return false.
12475 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12476 std::vector<Constant*> &Mask,
12477 LLVMContext *Context) {
12478 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12479 "Invalid CollectSingleShuffleElements");
12480 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12482 if (isa<UndefValue>(V)) {
12483 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12485 } else if (V == LHS) {
12486 for (unsigned i = 0; i != NumElts; ++i)
12487 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12489 } else if (V == RHS) {
12490 for (unsigned i = 0; i != NumElts; ++i)
12491 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i+NumElts));
12493 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12494 // If this is an insert of an extract from some other vector, include it.
12495 Value *VecOp = IEI->getOperand(0);
12496 Value *ScalarOp = IEI->getOperand(1);
12497 Value *IdxOp = IEI->getOperand(2);
12499 if (!isa<ConstantInt>(IdxOp))
12501 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12503 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12504 // Okay, we can handle this if the vector we are insertinting into is
12505 // transitively ok.
12506 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12507 // If so, update the mask to reflect the inserted undef.
12508 Mask[InsertedIdx] = UndefValue::get(Type::getInt32Ty(*Context));
12511 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12512 if (isa<ConstantInt>(EI->getOperand(1)) &&
12513 EI->getOperand(0)->getType() == V->getType()) {
12514 unsigned ExtractedIdx =
12515 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12517 // This must be extracting from either LHS or RHS.
12518 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12519 // Okay, we can handle this if the vector we are insertinting into is
12520 // transitively ok.
12521 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12522 // If so, update the mask to reflect the inserted value.
12523 if (EI->getOperand(0) == LHS) {
12524 Mask[InsertedIdx % NumElts] =
12525 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12527 assert(EI->getOperand(0) == RHS);
12528 Mask[InsertedIdx % NumElts] =
12529 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx+NumElts);
12538 // TODO: Handle shufflevector here!
12543 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12544 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12545 /// that computes V and the LHS value of the shuffle.
12546 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12547 Value *&RHS, LLVMContext *Context) {
12548 assert(isa<VectorType>(V->getType()) &&
12549 (RHS == 0 || V->getType() == RHS->getType()) &&
12550 "Invalid shuffle!");
12551 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12553 if (isa<UndefValue>(V)) {
12554 Mask.assign(NumElts, UndefValue::get(Type::getInt32Ty(*Context)));
12556 } else if (isa<ConstantAggregateZero>(V)) {
12557 Mask.assign(NumElts, ConstantInt::get(Type::getInt32Ty(*Context), 0));
12559 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12560 // If this is an insert of an extract from some other vector, include it.
12561 Value *VecOp = IEI->getOperand(0);
12562 Value *ScalarOp = IEI->getOperand(1);
12563 Value *IdxOp = IEI->getOperand(2);
12565 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12566 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12567 EI->getOperand(0)->getType() == V->getType()) {
12568 unsigned ExtractedIdx =
12569 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12570 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12572 // Either the extracted from or inserted into vector must be RHSVec,
12573 // otherwise we'd end up with a shuffle of three inputs.
12574 if (EI->getOperand(0) == RHS || RHS == 0) {
12575 RHS = EI->getOperand(0);
12576 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12577 Mask[InsertedIdx % NumElts] =
12578 ConstantInt::get(Type::getInt32Ty(*Context), NumElts+ExtractedIdx);
12582 if (VecOp == RHS) {
12583 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12585 // Everything but the extracted element is replaced with the RHS.
12586 for (unsigned i = 0; i != NumElts; ++i) {
12587 if (i != InsertedIdx)
12588 Mask[i] = ConstantInt::get(Type::getInt32Ty(*Context), NumElts+i);
12593 // If this insertelement is a chain that comes from exactly these two
12594 // vectors, return the vector and the effective shuffle.
12595 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12597 return EI->getOperand(0);
12602 // TODO: Handle shufflevector here!
12604 // Otherwise, can't do anything fancy. Return an identity vector.
12605 for (unsigned i = 0; i != NumElts; ++i)
12606 Mask.push_back(ConstantInt::get(Type::getInt32Ty(*Context), i));
12610 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12611 Value *VecOp = IE.getOperand(0);
12612 Value *ScalarOp = IE.getOperand(1);
12613 Value *IdxOp = IE.getOperand(2);
12615 // Inserting an undef or into an undefined place, remove this.
12616 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12617 ReplaceInstUsesWith(IE, VecOp);
12619 // If the inserted element was extracted from some other vector, and if the
12620 // indexes are constant, try to turn this into a shufflevector operation.
12621 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12622 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12623 EI->getOperand(0)->getType() == IE.getType()) {
12624 unsigned NumVectorElts = IE.getType()->getNumElements();
12625 unsigned ExtractedIdx =
12626 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12627 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12629 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12630 return ReplaceInstUsesWith(IE, VecOp);
12632 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12633 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12635 // If we are extracting a value from a vector, then inserting it right
12636 // back into the same place, just use the input vector.
12637 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12638 return ReplaceInstUsesWith(IE, VecOp);
12640 // We could theoretically do this for ANY input. However, doing so could
12641 // turn chains of insertelement instructions into a chain of shufflevector
12642 // instructions, and right now we do not merge shufflevectors. As such,
12643 // only do this in a situation where it is clear that there is benefit.
12644 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12645 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12646 // the values of VecOp, except then one read from EIOp0.
12647 // Build a new shuffle mask.
12648 std::vector<Constant*> Mask;
12649 if (isa<UndefValue>(VecOp))
12650 Mask.assign(NumVectorElts, UndefValue::get(Type::getInt32Ty(*Context)));
12652 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12653 Mask.assign(NumVectorElts, ConstantInt::get(Type::getInt32Ty(*Context),
12656 Mask[InsertedIdx] =
12657 ConstantInt::get(Type::getInt32Ty(*Context), ExtractedIdx);
12658 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12659 ConstantVector::get(Mask));
12662 // If this insertelement isn't used by some other insertelement, turn it
12663 // (and any insertelements it points to), into one big shuffle.
12664 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12665 std::vector<Constant*> Mask;
12667 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12668 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12669 // We now have a shuffle of LHS, RHS, Mask.
12670 return new ShuffleVectorInst(LHS, RHS,
12671 ConstantVector::get(Mask));
12676 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12677 APInt UndefElts(VWidth, 0);
12678 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12679 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12686 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12687 Value *LHS = SVI.getOperand(0);
12688 Value *RHS = SVI.getOperand(1);
12689 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12691 bool MadeChange = false;
12693 // Undefined shuffle mask -> undefined value.
12694 if (isa<UndefValue>(SVI.getOperand(2)))
12695 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12697 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12699 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12702 APInt UndefElts(VWidth, 0);
12703 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12704 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12705 LHS = SVI.getOperand(0);
12706 RHS = SVI.getOperand(1);
12710 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12711 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12712 if (LHS == RHS || isa<UndefValue>(LHS)) {
12713 if (isa<UndefValue>(LHS) && LHS == RHS) {
12714 // shuffle(undef,undef,mask) -> undef.
12715 return ReplaceInstUsesWith(SVI, LHS);
12718 // Remap any references to RHS to use LHS.
12719 std::vector<Constant*> Elts;
12720 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12721 if (Mask[i] >= 2*e)
12722 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12724 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12725 (Mask[i] < e && isa<UndefValue>(LHS))) {
12726 Mask[i] = 2*e; // Turn into undef.
12727 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12729 Mask[i] = Mask[i] % e; // Force to LHS.
12730 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), Mask[i]));
12734 SVI.setOperand(0, SVI.getOperand(1));
12735 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12736 SVI.setOperand(2, ConstantVector::get(Elts));
12737 LHS = SVI.getOperand(0);
12738 RHS = SVI.getOperand(1);
12742 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12743 bool isLHSID = true, isRHSID = true;
12745 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12746 if (Mask[i] >= e*2) continue; // Ignore undef values.
12747 // Is this an identity shuffle of the LHS value?
12748 isLHSID &= (Mask[i] == i);
12750 // Is this an identity shuffle of the RHS value?
12751 isRHSID &= (Mask[i]-e == i);
12754 // Eliminate identity shuffles.
12755 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12756 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12758 // If the LHS is a shufflevector itself, see if we can combine it with this
12759 // one without producing an unusual shuffle. Here we are really conservative:
12760 // we are absolutely afraid of producing a shuffle mask not in the input
12761 // program, because the code gen may not be smart enough to turn a merged
12762 // shuffle into two specific shuffles: it may produce worse code. As such,
12763 // we only merge two shuffles if the result is one of the two input shuffle
12764 // masks. In this case, merging the shuffles just removes one instruction,
12765 // which we know is safe. This is good for things like turning:
12766 // (splat(splat)) -> splat.
12767 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12768 if (isa<UndefValue>(RHS)) {
12769 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12771 std::vector<unsigned> NewMask;
12772 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12773 if (Mask[i] >= 2*e)
12774 NewMask.push_back(2*e);
12776 NewMask.push_back(LHSMask[Mask[i]]);
12778 // If the result mask is equal to the src shuffle or this shuffle mask, do
12779 // the replacement.
12780 if (NewMask == LHSMask || NewMask == Mask) {
12781 unsigned LHSInNElts =
12782 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12783 std::vector<Constant*> Elts;
12784 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12785 if (NewMask[i] >= LHSInNElts*2) {
12786 Elts.push_back(UndefValue::get(Type::getInt32Ty(*Context)));
12788 Elts.push_back(ConstantInt::get(Type::getInt32Ty(*Context), NewMask[i]));
12791 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12792 LHSSVI->getOperand(1),
12793 ConstantVector::get(Elts));
12798 return MadeChange ? &SVI : 0;
12804 /// TryToSinkInstruction - Try to move the specified instruction from its
12805 /// current block into the beginning of DestBlock, which can only happen if it's
12806 /// safe to move the instruction past all of the instructions between it and the
12807 /// end of its block.
12808 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12809 assert(I->hasOneUse() && "Invariants didn't hold!");
12811 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12812 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12815 // Do not sink alloca instructions out of the entry block.
12816 if (isa<AllocaInst>(I) && I->getParent() ==
12817 &DestBlock->getParent()->getEntryBlock())
12820 // We can only sink load instructions if there is nothing between the load and
12821 // the end of block that could change the value.
12822 if (I->mayReadFromMemory()) {
12823 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12825 if (Scan->mayWriteToMemory())
12829 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12831 CopyPrecedingStopPoint(I, InsertPos);
12832 I->moveBefore(InsertPos);
12838 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12839 /// all reachable code to the worklist.
12841 /// This has a couple of tricks to make the code faster and more powerful. In
12842 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12843 /// them to the worklist (this significantly speeds up instcombine on code where
12844 /// many instructions are dead or constant). Additionally, if we find a branch
12845 /// whose condition is a known constant, we only visit the reachable successors.
12847 static void AddReachableCodeToWorklist(BasicBlock *BB,
12848 SmallPtrSet<BasicBlock*, 64> &Visited,
12850 const TargetData *TD) {
12851 SmallVector<BasicBlock*, 256> Worklist;
12852 Worklist.push_back(BB);
12854 while (!Worklist.empty()) {
12855 BB = Worklist.back();
12856 Worklist.pop_back();
12858 // We have now visited this block! If we've already been here, ignore it.
12859 if (!Visited.insert(BB)) continue;
12861 DbgInfoIntrinsic *DBI_Prev = NULL;
12862 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12863 Instruction *Inst = BBI++;
12865 // DCE instruction if trivially dead.
12866 if (isInstructionTriviallyDead(Inst)) {
12868 DEBUG(errs() << "IC: DCE: " << *Inst << '\n');
12869 Inst->eraseFromParent();
12873 // ConstantProp instruction if trivially constant.
12874 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12875 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: "
12877 Inst->replaceAllUsesWith(C);
12879 Inst->eraseFromParent();
12883 // If there are two consecutive llvm.dbg.stoppoint calls then
12884 // it is likely that the optimizer deleted code in between these
12886 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12889 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12890 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12891 IC.RemoveFromWorkList(DBI_Prev);
12892 DBI_Prev->eraseFromParent();
12894 DBI_Prev = DBI_Next;
12899 IC.AddToWorkList(Inst);
12902 // Recursively visit successors. If this is a branch or switch on a
12903 // constant, only visit the reachable successor.
12904 TerminatorInst *TI = BB->getTerminator();
12905 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12906 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12907 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12908 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12909 Worklist.push_back(ReachableBB);
12912 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12913 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12914 // See if this is an explicit destination.
12915 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12916 if (SI->getCaseValue(i) == Cond) {
12917 BasicBlock *ReachableBB = SI->getSuccessor(i);
12918 Worklist.push_back(ReachableBB);
12922 // Otherwise it is the default destination.
12923 Worklist.push_back(SI->getSuccessor(0));
12928 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12929 Worklist.push_back(TI->getSuccessor(i));
12933 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12934 bool Changed = false;
12935 TD = getAnalysisIfAvailable<TargetData>();
12937 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12938 << F.getNameStr() << "\n");
12941 // Do a depth-first traversal of the function, populate the worklist with
12942 // the reachable instructions. Ignore blocks that are not reachable. Keep
12943 // track of which blocks we visit.
12944 SmallPtrSet<BasicBlock*, 64> Visited;
12945 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12947 // Do a quick scan over the function. If we find any blocks that are
12948 // unreachable, remove any instructions inside of them. This prevents
12949 // the instcombine code from having to deal with some bad special cases.
12950 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12951 if (!Visited.count(BB)) {
12952 Instruction *Term = BB->getTerminator();
12953 while (Term != BB->begin()) { // Remove instrs bottom-up
12954 BasicBlock::iterator I = Term; --I;
12956 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12957 // A debug intrinsic shouldn't force another iteration if we weren't
12958 // going to do one without it.
12959 if (!isa<DbgInfoIntrinsic>(I)) {
12963 if (!I->use_empty())
12964 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12965 I->eraseFromParent();
12970 while (!Worklist.empty()) {
12971 Instruction *I = RemoveOneFromWorkList();
12972 if (I == 0) continue; // skip null values.
12974 // Check to see if we can DCE the instruction.
12975 if (isInstructionTriviallyDead(I)) {
12976 // Add operands to the worklist.
12977 if (I->getNumOperands() < 4)
12978 AddUsesToWorkList(*I);
12981 DEBUG(errs() << "IC: DCE: " << *I << '\n');
12983 I->eraseFromParent();
12984 RemoveFromWorkList(I);
12989 // Instruction isn't dead, see if we can constant propagate it.
12990 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
12991 DEBUG(errs() << "IC: ConstFold to: " << *C << " from: " << *I << '\n');
12993 // Add operands to the worklist.
12994 AddUsesToWorkList(*I);
12995 ReplaceInstUsesWith(*I, C);
12998 I->eraseFromParent();
12999 RemoveFromWorkList(I);
13005 // See if we can constant fold its operands.
13006 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
13007 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
13008 if (Constant *NewC = ConstantFoldConstantExpression(CE,
13009 F.getContext(), TD))
13016 // See if we can trivially sink this instruction to a successor basic block.
13017 if (I->hasOneUse()) {
13018 BasicBlock *BB = I->getParent();
13019 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
13020 if (UserParent != BB) {
13021 bool UserIsSuccessor = false;
13022 // See if the user is one of our successors.
13023 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13024 if (*SI == UserParent) {
13025 UserIsSuccessor = true;
13029 // If the user is one of our immediate successors, and if that successor
13030 // only has us as a predecessors (we'd have to split the critical edge
13031 // otherwise), we can keep going.
13032 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
13033 next(pred_begin(UserParent)) == pred_end(UserParent))
13034 // Okay, the CFG is simple enough, try to sink this instruction.
13035 Changed |= TryToSinkInstruction(I, UserParent);
13039 // Now that we have an instruction, try combining it to simplify it...
13043 DEBUG(raw_string_ostream SS(OrigI); I->print(SS); OrigI = SS.str(););
13044 if (Instruction *Result = visit(*I)) {
13046 // Should we replace the old instruction with a new one?
13048 DEBUG(errs() << "IC: Old = " << *I << '\n'
13049 << " New = " << *Result << '\n');
13051 // Everything uses the new instruction now.
13052 I->replaceAllUsesWith(Result);
13054 // Push the new instruction and any users onto the worklist.
13055 AddToWorkList(Result);
13056 AddUsersToWorkList(*Result);
13058 // Move the name to the new instruction first.
13059 Result->takeName(I);
13061 // Insert the new instruction into the basic block...
13062 BasicBlock *InstParent = I->getParent();
13063 BasicBlock::iterator InsertPos = I;
13065 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13066 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13069 InstParent->getInstList().insert(InsertPos, Result);
13071 // Make sure that we reprocess all operands now that we reduced their
13073 AddUsesToWorkList(*I);
13075 // Instructions can end up on the worklist more than once. Make sure
13076 // we do not process an instruction that has been deleted.
13077 RemoveFromWorkList(I);
13079 // Erase the old instruction.
13080 InstParent->getInstList().erase(I);
13083 DEBUG(errs() << "IC: Mod = " << OrigI << '\n'
13084 << " New = " << *I << '\n');
13087 // If the instruction was modified, it's possible that it is now dead.
13088 // if so, remove it.
13089 if (isInstructionTriviallyDead(I)) {
13090 // Make sure we process all operands now that we are reducing their
13092 AddUsesToWorkList(*I);
13094 // Instructions may end up in the worklist more than once. Erase all
13095 // occurrences of this instruction.
13096 RemoveFromWorkList(I);
13097 I->eraseFromParent();
13100 AddUsersToWorkList(*I);
13107 assert(WorklistMap.empty() && "Worklist empty, but map not?");
13109 // Do an explicit clear, this shrinks the map if needed.
13110 WorklistMap.clear();
13115 bool InstCombiner::runOnFunction(Function &F) {
13116 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13117 Context = &F.getContext();
13119 bool EverMadeChange = false;
13121 // Iterate while there is work to do.
13122 unsigned Iteration = 0;
13123 while (DoOneIteration(F, Iteration++))
13124 EverMadeChange = true;
13125 return EverMadeChange;
13128 FunctionPass *llvm::createInstructionCombiningPass() {
13129 return new InstCombiner();