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(LLVMContext *Context, 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::Int32Ty;
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() : 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 && SrcTy != TD->getIntPtrType()) ||
481 (Res == Instruction::PtrToInt && DstTy != TD->getIntPtrType()))
484 return Instruction::CastOps(Res);
487 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
488 /// in any code being generated. It does not require codegen if V is simple
489 /// enough or if the cast can be folded into other casts.
490 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
491 const Type *Ty, TargetData *TD) {
492 if (V->getType() == Ty || isa<Constant>(V)) return false;
494 // If this is another cast that can be eliminated, it isn't codegen either.
495 if (const CastInst *CI = dyn_cast<CastInst>(V))
496 if (isEliminableCastPair(CI, opcode, Ty, TD))
501 // SimplifyCommutative - This performs a few simplifications for commutative
504 // 1. Order operands such that they are listed from right (least complex) to
505 // left (most complex). This puts constants before unary operators before
508 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
509 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
511 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
512 bool Changed = false;
513 if (getComplexity(Context, I.getOperand(0)) <
514 getComplexity(Context, I.getOperand(1)))
515 Changed = !I.swapOperands();
517 if (!I.isAssociative()) return Changed;
518 Instruction::BinaryOps Opcode = I.getOpcode();
519 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
520 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
521 if (isa<Constant>(I.getOperand(1))) {
522 Constant *Folded = ConstantExpr::get(I.getOpcode(),
523 cast<Constant>(I.getOperand(1)),
524 cast<Constant>(Op->getOperand(1)));
525 I.setOperand(0, Op->getOperand(0));
526 I.setOperand(1, Folded);
528 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
529 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
530 isOnlyUse(Op) && isOnlyUse(Op1)) {
531 Constant *C1 = cast<Constant>(Op->getOperand(1));
532 Constant *C2 = cast<Constant>(Op1->getOperand(1));
534 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
535 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
536 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
540 I.setOperand(0, New);
541 I.setOperand(1, Folded);
548 /// SimplifyCompare - For a CmpInst this function just orders the operands
549 /// so that theyare listed from right (least complex) to left (most complex).
550 /// This puts constants before unary operators before binary operators.
551 bool InstCombiner::SimplifyCompare(CmpInst &I) {
552 if (getComplexity(Context, I.getOperand(0)) >=
553 getComplexity(Context, 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::Int32Ty));
1592 Elts.push_back(ConstantInt::get(Type::Int32Ty,
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::Int32Ty, 0U, false), "tmp"), *II);
1724 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1725 ConstantInt::get(Type::Int32Ty, 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::Int32Ty, 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 LLVMContext *Context = IC->getContext();
1873 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1874 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1877 // Figure out if the constant is the left or the right argument.
1878 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1879 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1881 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1883 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1884 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1887 Value *Op0 = SO, *Op1 = ConstOperand;
1889 std::swap(Op0, Op1);
1891 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1892 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1893 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1894 New = CmpInst::Create(*Context, CI->getOpcode(), CI->getPredicate(),
1895 Op0, Op1, SO->getName()+".cmp");
1897 llvm_unreachable("Unknown binary instruction type!");
1899 return IC->InsertNewInstBefore(New, I);
1902 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1903 // constant as the other operand, try to fold the binary operator into the
1904 // select arguments. This also works for Cast instructions, which obviously do
1905 // not have a second operand.
1906 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1908 // Don't modify shared select instructions
1909 if (!SI->hasOneUse()) return 0;
1910 Value *TV = SI->getOperand(1);
1911 Value *FV = SI->getOperand(2);
1913 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1914 // Bool selects with constant operands can be folded to logical ops.
1915 if (SI->getType() == Type::Int1Ty) return 0;
1917 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1918 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1920 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1927 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1928 /// node as operand #0, see if we can fold the instruction into the PHI (which
1929 /// is only possible if all operands to the PHI are constants).
1930 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1931 PHINode *PN = cast<PHINode>(I.getOperand(0));
1932 unsigned NumPHIValues = PN->getNumIncomingValues();
1933 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1935 // Check to see if all of the operands of the PHI are constants. If there is
1936 // one non-constant value, remember the BB it is. If there is more than one
1937 // or if *it* is a PHI, bail out.
1938 BasicBlock *NonConstBB = 0;
1939 for (unsigned i = 0; i != NumPHIValues; ++i)
1940 if (!isa<Constant>(PN->getIncomingValue(i))) {
1941 if (NonConstBB) return 0; // More than one non-const value.
1942 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1943 NonConstBB = PN->getIncomingBlock(i);
1945 // If the incoming non-constant value is in I's block, we have an infinite
1947 if (NonConstBB == I.getParent())
1951 // If there is exactly one non-constant value, we can insert a copy of the
1952 // operation in that block. However, if this is a critical edge, we would be
1953 // inserting the computation one some other paths (e.g. inside a loop). Only
1954 // do this if the pred block is unconditionally branching into the phi block.
1956 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1957 if (!BI || !BI->isUnconditional()) return 0;
1960 // Okay, we can do the transformation: create the new PHI node.
1961 PHINode *NewPN = PHINode::Create(I.getType(), "");
1962 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1963 InsertNewInstBefore(NewPN, *PN);
1964 NewPN->takeName(PN);
1966 // Next, add all of the operands to the PHI.
1967 if (I.getNumOperands() == 2) {
1968 Constant *C = cast<Constant>(I.getOperand(1));
1969 for (unsigned i = 0; i != NumPHIValues; ++i) {
1971 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1972 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1973 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1975 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1977 assert(PN->getIncomingBlock(i) == NonConstBB);
1978 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1979 InV = BinaryOperator::Create(BO->getOpcode(),
1980 PN->getIncomingValue(i), C, "phitmp",
1981 NonConstBB->getTerminator());
1982 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1983 InV = CmpInst::Create(*Context, CI->getOpcode(),
1985 PN->getIncomingValue(i), C, "phitmp",
1986 NonConstBB->getTerminator());
1988 llvm_unreachable("Unknown binop!");
1990 AddToWorkList(cast<Instruction>(InV));
1992 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1995 CastInst *CI = cast<CastInst>(&I);
1996 const Type *RetTy = CI->getType();
1997 for (unsigned i = 0; i != NumPHIValues; ++i) {
1999 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2000 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2002 assert(PN->getIncomingBlock(i) == NonConstBB);
2003 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2004 I.getType(), "phitmp",
2005 NonConstBB->getTerminator());
2006 AddToWorkList(cast<Instruction>(InV));
2008 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2011 return ReplaceInstUsesWith(I, NewPN);
2015 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2016 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2017 /// This basically requires proving that the add in the original type would not
2018 /// overflow to change the sign bit or have a carry out.
2019 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2020 // There are different heuristics we can use for this. Here are some simple
2023 // Add has the property that adding any two 2's complement numbers can only
2024 // have one carry bit which can change a sign. As such, if LHS and RHS each
2025 // have at least two sign bits, we know that the addition of the two values will
2026 // sign extend fine.
2027 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2031 // If one of the operands only has one non-zero bit, and if the other operand
2032 // has a known-zero bit in a more significant place than it (not including the
2033 // sign bit) the ripple may go up to and fill the zero, but won't change the
2034 // sign. For example, (X & ~4) + 1.
2042 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2043 bool Changed = SimplifyCommutative(I);
2044 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2046 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2047 // X + undef -> undef
2048 if (isa<UndefValue>(RHS))
2049 return ReplaceInstUsesWith(I, RHS);
2052 if (RHSC->isNullValue())
2053 return ReplaceInstUsesWith(I, LHS);
2055 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2056 // X + (signbit) --> X ^ signbit
2057 const APInt& Val = CI->getValue();
2058 uint32_t BitWidth = Val.getBitWidth();
2059 if (Val == APInt::getSignBit(BitWidth))
2060 return BinaryOperator::CreateXor(LHS, RHS);
2062 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2063 // (X & 254)+1 -> (X&254)|1
2064 if (SimplifyDemandedInstructionBits(I))
2067 // zext(bool) + C -> bool ? C + 1 : C
2068 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2069 if (ZI->getSrcTy() == Type::Int1Ty)
2070 return SelectInst::Create(ZI->getOperand(0), AddOne(CI), CI);
2073 if (isa<PHINode>(LHS))
2074 if (Instruction *NV = FoldOpIntoPhi(I))
2077 ConstantInt *XorRHS = 0;
2079 if (isa<ConstantInt>(RHSC) &&
2080 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2081 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2082 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2084 uint32_t Size = TySizeBits / 2;
2085 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2086 APInt CFF80Val(-C0080Val);
2088 if (TySizeBits > Size) {
2089 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2090 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2091 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2092 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2093 // This is a sign extend if the top bits are known zero.
2094 if (!MaskedValueIsZero(XorLHS,
2095 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2096 Size = 0; // Not a sign ext, but can't be any others either.
2101 C0080Val = APIntOps::lshr(C0080Val, Size);
2102 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2103 } while (Size >= 1);
2105 // FIXME: This shouldn't be necessary. When the backends can handle types
2106 // with funny bit widths then this switch statement should be removed. It
2107 // is just here to get the size of the "middle" type back up to something
2108 // that the back ends can handle.
2109 const Type *MiddleType = 0;
2112 case 32: MiddleType = Type::Int32Ty; break;
2113 case 16: MiddleType = Type::Int16Ty; break;
2114 case 8: MiddleType = Type::Int8Ty; break;
2117 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2118 InsertNewInstBefore(NewTrunc, I);
2119 return new SExtInst(NewTrunc, I.getType(), I.getName());
2124 if (I.getType() == Type::Int1Ty)
2125 return BinaryOperator::CreateXor(LHS, RHS);
2128 if (I.getType()->isInteger()) {
2129 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS)))
2132 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2133 if (RHSI->getOpcode() == Instruction::Sub)
2134 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2135 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2137 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2138 if (LHSI->getOpcode() == Instruction::Sub)
2139 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2140 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2145 // -A + -B --> -(A + B)
2146 if (Value *LHSV = dyn_castNegVal(LHS)) {
2147 if (LHS->getType()->isIntOrIntVector()) {
2148 if (Value *RHSV = dyn_castNegVal(RHS)) {
2149 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2150 InsertNewInstBefore(NewAdd, I);
2151 return BinaryOperator::CreateNeg(NewAdd);
2155 return BinaryOperator::CreateSub(RHS, LHSV);
2159 if (!isa<Constant>(RHS))
2160 if (Value *V = dyn_castNegVal(RHS))
2161 return BinaryOperator::CreateSub(LHS, V);
2165 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2166 if (X == RHS) // X*C + X --> X * (C+1)
2167 return BinaryOperator::CreateMul(RHS, AddOne(C2));
2169 // X*C1 + X*C2 --> X * (C1+C2)
2171 if (X == dyn_castFoldableMul(RHS, C1))
2172 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2175 // X + X*C --> X * (C+1)
2176 if (dyn_castFoldableMul(RHS, C2) == LHS)
2177 return BinaryOperator::CreateMul(LHS, AddOne(C2));
2179 // X + ~X --> -1 since ~X = -X-1
2180 if (dyn_castNotVal(LHS) == RHS ||
2181 dyn_castNotVal(RHS) == LHS)
2182 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2185 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2186 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2187 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2190 // A+B --> A|B iff A and B have no bits set in common.
2191 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2192 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2193 APInt LHSKnownOne(IT->getBitWidth(), 0);
2194 APInt LHSKnownZero(IT->getBitWidth(), 0);
2195 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2196 if (LHSKnownZero != 0) {
2197 APInt RHSKnownOne(IT->getBitWidth(), 0);
2198 APInt RHSKnownZero(IT->getBitWidth(), 0);
2199 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2201 // No bits in common -> bitwise or.
2202 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2203 return BinaryOperator::CreateOr(LHS, RHS);
2207 // W*X + Y*Z --> W * (X+Z) iff W == Y
2208 if (I.getType()->isIntOrIntVector()) {
2209 Value *W, *X, *Y, *Z;
2210 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2211 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2215 } else if (Y == X) {
2217 } else if (X == Z) {
2224 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2225 LHS->getName()), I);
2226 return BinaryOperator::CreateMul(W, NewAdd);
2231 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2233 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2234 return BinaryOperator::CreateSub(SubOne(CRHS), X);
2236 // (X & FF00) + xx00 -> (X+xx00) & FF00
2237 if (LHS->hasOneUse() &&
2238 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2239 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2240 if (Anded == CRHS) {
2241 // See if all bits from the first bit set in the Add RHS up are included
2242 // in the mask. First, get the rightmost bit.
2243 const APInt& AddRHSV = CRHS->getValue();
2245 // Form a mask of all bits from the lowest bit added through the top.
2246 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2248 // See if the and mask includes all of these bits.
2249 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2251 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2252 // Okay, the xform is safe. Insert the new add pronto.
2253 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2254 LHS->getName()), I);
2255 return BinaryOperator::CreateAnd(NewAdd, C2);
2260 // Try to fold constant add into select arguments.
2261 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2262 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2266 // add (select X 0 (sub n A)) A --> select X A n
2268 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2271 SI = dyn_cast<SelectInst>(RHS);
2274 if (SI && SI->hasOneUse()) {
2275 Value *TV = SI->getTrueValue();
2276 Value *FV = SI->getFalseValue();
2279 // Can we fold the add into the argument of the select?
2280 // We check both true and false select arguments for a matching subtract.
2281 if (match(FV, m_Zero()) &&
2282 match(TV, m_Sub(m_Value(N), m_Specific(A))))
2283 // Fold the add into the true select value.
2284 return SelectInst::Create(SI->getCondition(), N, A);
2285 if (match(TV, m_Zero()) &&
2286 match(FV, m_Sub(m_Value(N), m_Specific(A))))
2287 // Fold the add into the false select value.
2288 return SelectInst::Create(SI->getCondition(), A, N);
2292 // Check for (add (sext x), y), see if we can merge this into an
2293 // integer add followed by a sext.
2294 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2295 // (add (sext x), cst) --> (sext (add x, cst'))
2296 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2298 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2299 if (LHSConv->hasOneUse() &&
2300 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2301 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2302 // Insert the new, smaller add.
2303 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2305 InsertNewInstBefore(NewAdd, I);
2306 return new SExtInst(NewAdd, I.getType());
2310 // (add (sext x), (sext y)) --> (sext (add int x, y))
2311 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2312 // Only do this if x/y have the same type, if at last one of them has a
2313 // single use (so we don't increase the number of sexts), and if the
2314 // integer add will not overflow.
2315 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2316 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2317 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2318 RHSConv->getOperand(0))) {
2319 // Insert the new integer add.
2320 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2321 RHSConv->getOperand(0),
2323 InsertNewInstBefore(NewAdd, I);
2324 return new SExtInst(NewAdd, I.getType());
2329 return Changed ? &I : 0;
2332 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2333 bool Changed = SimplifyCommutative(I);
2334 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2336 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2338 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2339 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2340 (I.getType())->getValueAPF()))
2341 return ReplaceInstUsesWith(I, LHS);
2344 if (isa<PHINode>(LHS))
2345 if (Instruction *NV = FoldOpIntoPhi(I))
2350 // -A + -B --> -(A + B)
2351 if (Value *LHSV = dyn_castFNegVal(LHS))
2352 return BinaryOperator::CreateFSub(RHS, LHSV);
2355 if (!isa<Constant>(RHS))
2356 if (Value *V = dyn_castFNegVal(RHS))
2357 return BinaryOperator::CreateFSub(LHS, V);
2359 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2360 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2361 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2362 return ReplaceInstUsesWith(I, LHS);
2364 // Check for (add double (sitofp x), y), see if we can merge this into an
2365 // integer add followed by a promotion.
2366 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2367 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2368 // ... if the constant fits in the integer value. This is useful for things
2369 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2370 // requires a constant pool load, and generally allows the add to be better
2372 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2374 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2375 if (LHSConv->hasOneUse() &&
2376 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2377 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2378 // Insert the new integer add.
2379 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2381 InsertNewInstBefore(NewAdd, I);
2382 return new SIToFPInst(NewAdd, I.getType());
2386 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2387 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2388 // Only do this if x/y have the same type, if at last one of them has a
2389 // single use (so we don't increase the number of int->fp conversions),
2390 // and if the integer add will not overflow.
2391 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2392 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2393 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2394 RHSConv->getOperand(0))) {
2395 // Insert the new integer add.
2396 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2397 RHSConv->getOperand(0),
2399 InsertNewInstBefore(NewAdd, I);
2400 return new SIToFPInst(NewAdd, I.getType());
2405 return Changed ? &I : 0;
2408 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2409 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2411 if (Op0 == Op1) // sub X, X -> 0
2412 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2414 // If this is a 'B = x-(-A)', change to B = x+A...
2415 if (Value *V = dyn_castNegVal(Op1))
2416 return BinaryOperator::CreateAdd(Op0, V);
2418 if (isa<UndefValue>(Op0))
2419 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2420 if (isa<UndefValue>(Op1))
2421 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2423 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2424 // Replace (-1 - A) with (~A)...
2425 if (C->isAllOnesValue())
2426 return BinaryOperator::CreateNot(Op1);
2428 // C - ~X == X + (1+C)
2430 if (match(Op1, m_Not(m_Value(X))))
2431 return BinaryOperator::CreateAdd(X, AddOne(C));
2433 // -(X >>u 31) -> (X >>s 31)
2434 // -(X >>s 31) -> (X >>u 31)
2436 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2437 if (SI->getOpcode() == Instruction::LShr) {
2438 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2439 // Check to see if we are shifting out everything but the sign bit.
2440 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2441 SI->getType()->getPrimitiveSizeInBits()-1) {
2442 // Ok, the transformation is safe. Insert AShr.
2443 return BinaryOperator::Create(Instruction::AShr,
2444 SI->getOperand(0), CU, SI->getName());
2448 else if (SI->getOpcode() == Instruction::AShr) {
2449 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2450 // Check to see if we are shifting out everything but the sign bit.
2451 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2452 SI->getType()->getPrimitiveSizeInBits()-1) {
2453 // Ok, the transformation is safe. Insert LShr.
2454 return BinaryOperator::CreateLShr(
2455 SI->getOperand(0), CU, SI->getName());
2462 // Try to fold constant sub into select arguments.
2463 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2464 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2467 // C - zext(bool) -> bool ? C - 1 : C
2468 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2469 if (ZI->getSrcTy() == Type::Int1Ty)
2470 return SelectInst::Create(ZI->getOperand(0), SubOne(C), C);
2473 if (I.getType() == Type::Int1Ty)
2474 return BinaryOperator::CreateXor(Op0, Op1);
2476 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2477 if (Op1I->getOpcode() == Instruction::Add) {
2478 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2479 return BinaryOperator::CreateNeg(Op1I->getOperand(1),
2481 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2482 return BinaryOperator::CreateNeg(Op1I->getOperand(0),
2484 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2485 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2486 // C1-(X+C2) --> (C1-C2)-X
2487 return BinaryOperator::CreateSub(
2488 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2492 if (Op1I->hasOneUse()) {
2493 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2494 // is not used by anyone else...
2496 if (Op1I->getOpcode() == Instruction::Sub) {
2497 // Swap the two operands of the subexpr...
2498 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2499 Op1I->setOperand(0, IIOp1);
2500 Op1I->setOperand(1, IIOp0);
2502 // Create the new top level add instruction...
2503 return BinaryOperator::CreateAdd(Op0, Op1);
2506 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2508 if (Op1I->getOpcode() == Instruction::And &&
2509 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2510 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2513 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2514 return BinaryOperator::CreateAnd(Op0, NewNot);
2517 // 0 - (X sdiv C) -> (X sdiv -C)
2518 if (Op1I->getOpcode() == Instruction::SDiv)
2519 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2521 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2522 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2523 ConstantExpr::getNeg(DivRHS));
2525 // X - X*C --> X * (1-C)
2526 ConstantInt *C2 = 0;
2527 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2529 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2531 return BinaryOperator::CreateMul(Op0, CP1);
2536 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2537 if (Op0I->getOpcode() == Instruction::Add) {
2538 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2539 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2540 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2541 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2542 } else if (Op0I->getOpcode() == Instruction::Sub) {
2543 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2544 return BinaryOperator::CreateNeg(Op0I->getOperand(1),
2550 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2551 if (X == Op1) // X*C - X --> X * (C-1)
2552 return BinaryOperator::CreateMul(Op1, SubOne(C1));
2554 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2555 if (X == dyn_castFoldableMul(Op1, C2))
2556 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2561 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2562 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2564 // If this is a 'B = x-(-A)', change to B = x+A...
2565 if (Value *V = dyn_castFNegVal(Op1))
2566 return BinaryOperator::CreateFAdd(Op0, V);
2568 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2569 if (Op1I->getOpcode() == Instruction::FAdd) {
2570 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2571 return BinaryOperator::CreateFNeg(Op1I->getOperand(1),
2573 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2574 return BinaryOperator::CreateFNeg(Op1I->getOperand(0),
2582 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2583 /// comparison only checks the sign bit. If it only checks the sign bit, set
2584 /// TrueIfSigned if the result of the comparison is true when the input value is
2586 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2587 bool &TrueIfSigned) {
2589 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2590 TrueIfSigned = true;
2591 return RHS->isZero();
2592 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2593 TrueIfSigned = true;
2594 return RHS->isAllOnesValue();
2595 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2596 TrueIfSigned = false;
2597 return RHS->isAllOnesValue();
2598 case ICmpInst::ICMP_UGT:
2599 // True if LHS u> RHS and RHS == high-bit-mask - 1
2600 TrueIfSigned = true;
2601 return RHS->getValue() ==
2602 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2603 case ICmpInst::ICMP_UGE:
2604 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2605 TrueIfSigned = true;
2606 return RHS->getValue().isSignBit();
2612 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2613 bool Changed = SimplifyCommutative(I);
2614 Value *Op0 = I.getOperand(0);
2616 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2617 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2619 // Simplify mul instructions with a constant RHS...
2620 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2621 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2623 // ((X << C1)*C2) == (X * (C2 << C1))
2624 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2625 if (SI->getOpcode() == Instruction::Shl)
2626 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2627 return BinaryOperator::CreateMul(SI->getOperand(0),
2628 ConstantExpr::getShl(CI, ShOp));
2631 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2632 if (CI->equalsInt(1)) // X * 1 == X
2633 return ReplaceInstUsesWith(I, Op0);
2634 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2635 return BinaryOperator::CreateNeg(Op0, I.getName());
2637 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2638 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2639 return BinaryOperator::CreateShl(Op0,
2640 ConstantInt::get(Op0->getType(), Val.logBase2()));
2642 } else if (isa<VectorType>(Op1->getType())) {
2643 if (Op1->isNullValue())
2644 return ReplaceInstUsesWith(I, Op1);
2646 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2647 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2648 return BinaryOperator::CreateNeg(Op0, I.getName());
2650 // As above, vector X*splat(1.0) -> X in all defined cases.
2651 if (Constant *Splat = Op1V->getSplatValue()) {
2652 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2653 if (CI->equalsInt(1))
2654 return ReplaceInstUsesWith(I, Op0);
2659 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2660 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2661 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2662 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2663 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2665 InsertNewInstBefore(Add, I);
2666 Value *C1C2 = ConstantExpr::getMul(Op1,
2667 cast<Constant>(Op0I->getOperand(1)));
2668 return BinaryOperator::CreateAdd(Add, C1C2);
2672 // Try to fold constant mul into select arguments.
2673 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2674 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2677 if (isa<PHINode>(Op0))
2678 if (Instruction *NV = FoldOpIntoPhi(I))
2682 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2683 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2684 return BinaryOperator::CreateMul(Op0v, Op1v);
2686 // (X / Y) * Y = X - (X % Y)
2687 // (X / Y) * -Y = (X % Y) - X
2689 Value *Op1 = I.getOperand(1);
2690 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2692 (BO->getOpcode() != Instruction::UDiv &&
2693 BO->getOpcode() != Instruction::SDiv)) {
2695 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2697 Value *Neg = dyn_castNegVal(Op1);
2698 if (BO && BO->hasOneUse() &&
2699 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2700 (BO->getOpcode() == Instruction::UDiv ||
2701 BO->getOpcode() == Instruction::SDiv)) {
2702 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2705 if (BO->getOpcode() == Instruction::UDiv)
2706 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2708 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2710 InsertNewInstBefore(Rem, I);
2714 return BinaryOperator::CreateSub(Op0BO, Rem);
2716 return BinaryOperator::CreateSub(Rem, Op0BO);
2720 if (I.getType() == Type::Int1Ty)
2721 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2723 // If one of the operands of the multiply is a cast from a boolean value, then
2724 // we know the bool is either zero or one, so this is a 'masking' multiply.
2725 // See if we can simplify things based on how the boolean was originally
2727 CastInst *BoolCast = 0;
2728 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2729 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2732 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2733 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2736 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2737 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2738 const Type *SCOpTy = SCIOp0->getType();
2741 // If the icmp is true iff the sign bit of X is set, then convert this
2742 // multiply into a shift/and combination.
2743 if (isa<ConstantInt>(SCIOp1) &&
2744 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2746 // Shift the X value right to turn it into "all signbits".
2747 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2748 SCOpTy->getPrimitiveSizeInBits()-1);
2750 InsertNewInstBefore(
2751 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2752 BoolCast->getOperand(0)->getName()+
2755 // If the multiply type is not the same as the source type, sign extend
2756 // or truncate to the multiply type.
2757 if (I.getType() != V->getType()) {
2758 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2759 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2760 Instruction::CastOps opcode =
2761 (SrcBits == DstBits ? Instruction::BitCast :
2762 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2763 V = InsertCastBefore(opcode, V, I.getType(), I);
2766 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2767 return BinaryOperator::CreateAnd(V, OtherOp);
2772 return Changed ? &I : 0;
2775 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2776 bool Changed = SimplifyCommutative(I);
2777 Value *Op0 = I.getOperand(0);
2779 // Simplify mul instructions with a constant RHS...
2780 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2781 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2782 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2783 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2784 if (Op1F->isExactlyValue(1.0))
2785 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2786 } else if (isa<VectorType>(Op1->getType())) {
2787 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2788 // As above, vector X*splat(1.0) -> X in all defined cases.
2789 if (Constant *Splat = Op1V->getSplatValue()) {
2790 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2791 if (F->isExactlyValue(1.0))
2792 return ReplaceInstUsesWith(I, Op0);
2797 // Try to fold constant mul into select arguments.
2798 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2799 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2802 if (isa<PHINode>(Op0))
2803 if (Instruction *NV = FoldOpIntoPhi(I))
2807 if (Value *Op0v = dyn_castFNegVal(Op0)) // -X * -Y = X*Y
2808 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1)))
2809 return BinaryOperator::CreateFMul(Op0v, Op1v);
2811 return Changed ? &I : 0;
2814 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2816 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2817 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2819 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2820 int NonNullOperand = -1;
2821 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2822 if (ST->isNullValue())
2824 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2825 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2826 if (ST->isNullValue())
2829 if (NonNullOperand == -1)
2832 Value *SelectCond = SI->getOperand(0);
2834 // Change the div/rem to use 'Y' instead of the select.
2835 I.setOperand(1, SI->getOperand(NonNullOperand));
2837 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2838 // problem. However, the select, or the condition of the select may have
2839 // multiple uses. Based on our knowledge that the operand must be non-zero,
2840 // propagate the known value for the select into other uses of it, and
2841 // propagate a known value of the condition into its other users.
2843 // If the select and condition only have a single use, don't bother with this,
2845 if (SI->use_empty() && SelectCond->hasOneUse())
2848 // Scan the current block backward, looking for other uses of SI.
2849 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2851 while (BBI != BBFront) {
2853 // If we found a call to a function, we can't assume it will return, so
2854 // information from below it cannot be propagated above it.
2855 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2858 // Replace uses of the select or its condition with the known values.
2859 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2862 *I = SI->getOperand(NonNullOperand);
2864 } else if (*I == SelectCond) {
2865 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
2866 ConstantInt::getFalse(*Context);
2871 // If we past the instruction, quit looking for it.
2874 if (&*BBI == SelectCond)
2877 // If we ran out of things to eliminate, break out of the loop.
2878 if (SelectCond == 0 && SI == 0)
2886 /// This function implements the transforms on div instructions that work
2887 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2888 /// used by the visitors to those instructions.
2889 /// @brief Transforms common to all three div instructions
2890 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2891 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2893 // undef / X -> 0 for integer.
2894 // undef / X -> undef for FP (the undef could be a snan).
2895 if (isa<UndefValue>(Op0)) {
2896 if (Op0->getType()->isFPOrFPVector())
2897 return ReplaceInstUsesWith(I, Op0);
2898 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2901 // X / undef -> undef
2902 if (isa<UndefValue>(Op1))
2903 return ReplaceInstUsesWith(I, Op1);
2908 /// This function implements the transforms common to both integer division
2909 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2910 /// division instructions.
2911 /// @brief Common integer divide transforms
2912 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2913 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2915 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2917 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2918 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
2919 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2920 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2923 Constant *CI = ConstantInt::get(I.getType(), 1);
2924 return ReplaceInstUsesWith(I, CI);
2927 if (Instruction *Common = commonDivTransforms(I))
2930 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2931 // This does not apply for fdiv.
2932 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2935 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2937 if (RHS->equalsInt(1))
2938 return ReplaceInstUsesWith(I, Op0);
2940 // (X / C1) / C2 -> X / (C1*C2)
2941 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2942 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2943 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2944 if (MultiplyOverflows(RHS, LHSRHS,
2945 I.getOpcode()==Instruction::SDiv))
2946 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2948 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2949 ConstantExpr::getMul(RHS, LHSRHS));
2952 if (!RHS->isZero()) { // avoid X udiv 0
2953 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2954 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2956 if (isa<PHINode>(Op0))
2957 if (Instruction *NV = FoldOpIntoPhi(I))
2962 // 0 / X == 0, we don't need to preserve faults!
2963 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2964 if (LHS->equalsInt(0))
2965 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2967 // It can't be division by zero, hence it must be division by one.
2968 if (I.getType() == Type::Int1Ty)
2969 return ReplaceInstUsesWith(I, Op0);
2971 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2972 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2975 return ReplaceInstUsesWith(I, Op0);
2981 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2982 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2984 // Handle the integer div common cases
2985 if (Instruction *Common = commonIDivTransforms(I))
2988 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2989 // X udiv C^2 -> X >> C
2990 // Check to see if this is an unsigned division with an exact power of 2,
2991 // if so, convert to a right shift.
2992 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2993 return BinaryOperator::CreateLShr(Op0,
2994 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2996 // X udiv C, where C >= signbit
2997 if (C->getValue().isNegative()) {
2998 Value *IC = InsertNewInstBefore(new ICmpInst(*Context,
2999 ICmpInst::ICMP_ULT, Op0, C),
3001 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3002 ConstantInt::get(I.getType(), 1));
3006 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3007 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3008 if (RHSI->getOpcode() == Instruction::Shl &&
3009 isa<ConstantInt>(RHSI->getOperand(0))) {
3010 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3011 if (C1.isPowerOf2()) {
3012 Value *N = RHSI->getOperand(1);
3013 const Type *NTy = N->getType();
3014 if (uint32_t C2 = C1.logBase2()) {
3015 Constant *C2V = ConstantInt::get(NTy, C2);
3016 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3018 return BinaryOperator::CreateLShr(Op0, N);
3023 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3024 // where C1&C2 are powers of two.
3025 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3026 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3027 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3028 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3029 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3030 // Compute the shift amounts
3031 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3032 // Construct the "on true" case of the select
3033 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3034 Instruction *TSI = BinaryOperator::CreateLShr(
3035 Op0, TC, SI->getName()+".t");
3036 TSI = InsertNewInstBefore(TSI, I);
3038 // Construct the "on false" case of the select
3039 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3040 Instruction *FSI = BinaryOperator::CreateLShr(
3041 Op0, FC, SI->getName()+".f");
3042 FSI = InsertNewInstBefore(FSI, I);
3044 // construct the select instruction and return it.
3045 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3051 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3052 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3054 // Handle the integer div common cases
3055 if (Instruction *Common = commonIDivTransforms(I))
3058 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3060 if (RHS->isAllOnesValue())
3061 return BinaryOperator::CreateNeg(Op0);
3063 // sdiv X, C --> ashr X, log2(C)
3064 if (cast<SDivOperator>(&I)->isExact() &&
3065 RHS->getValue().isNonNegative() &&
3066 RHS->getValue().isPowerOf2()) {
3067 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3068 RHS->getValue().exactLogBase2());
3069 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3073 // If the sign bits of both operands are zero (i.e. we can prove they are
3074 // unsigned inputs), turn this into a udiv.
3075 if (I.getType()->isInteger()) {
3076 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3077 if (MaskedValueIsZero(Op0, Mask)) {
3078 if (MaskedValueIsZero(Op1, Mask)) {
3079 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3080 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3082 ConstantInt *ShiftedInt;
3083 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value())) &&
3084 ShiftedInt->getValue().isPowerOf2()) {
3085 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3086 // Safe because the only negative value (1 << Y) can take on is
3087 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3088 // the sign bit set.
3089 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3097 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3098 return commonDivTransforms(I);
3101 /// This function implements the transforms on rem instructions that work
3102 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3103 /// is used by the visitors to those instructions.
3104 /// @brief Transforms common to all three rem instructions
3105 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3106 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3108 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3109 if (I.getType()->isFPOrFPVector())
3110 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3111 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3113 if (isa<UndefValue>(Op1))
3114 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3116 // Handle cases involving: rem X, (select Cond, Y, Z)
3117 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3123 /// This function implements the transforms common to both integer remainder
3124 /// instructions (urem and srem). It is called by the visitors to those integer
3125 /// remainder instructions.
3126 /// @brief Common integer remainder transforms
3127 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3128 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3130 if (Instruction *common = commonRemTransforms(I))
3133 // 0 % X == 0 for integer, we don't need to preserve faults!
3134 if (Constant *LHS = dyn_cast<Constant>(Op0))
3135 if (LHS->isNullValue())
3136 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3138 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3139 // X % 0 == undef, we don't need to preserve faults!
3140 if (RHS->equalsInt(0))
3141 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3143 if (RHS->equalsInt(1)) // X % 1 == 0
3144 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3146 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3147 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3148 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3150 } else if (isa<PHINode>(Op0I)) {
3151 if (Instruction *NV = FoldOpIntoPhi(I))
3155 // See if we can fold away this rem instruction.
3156 if (SimplifyDemandedInstructionBits(I))
3164 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3165 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3167 if (Instruction *common = commonIRemTransforms(I))
3170 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3171 // X urem C^2 -> X and C
3172 // Check to see if this is an unsigned remainder with an exact power of 2,
3173 // if so, convert to a bitwise and.
3174 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3175 if (C->getValue().isPowerOf2())
3176 return BinaryOperator::CreateAnd(Op0, SubOne(C));
3179 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3180 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3181 if (RHSI->getOpcode() == Instruction::Shl &&
3182 isa<ConstantInt>(RHSI->getOperand(0))) {
3183 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3184 Constant *N1 = Constant::getAllOnesValue(I.getType());
3185 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3187 return BinaryOperator::CreateAnd(Op0, Add);
3192 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3193 // where C1&C2 are powers of two.
3194 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3195 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3196 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3197 // STO == 0 and SFO == 0 handled above.
3198 if ((STO->getValue().isPowerOf2()) &&
3199 (SFO->getValue().isPowerOf2())) {
3200 Value *TrueAnd = InsertNewInstBefore(
3201 BinaryOperator::CreateAnd(Op0, SubOne(STO),
3202 SI->getName()+".t"), I);
3203 Value *FalseAnd = InsertNewInstBefore(
3204 BinaryOperator::CreateAnd(Op0, SubOne(SFO),
3205 SI->getName()+".f"), I);
3206 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3214 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3215 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3217 // Handle the integer rem common cases
3218 if (Instruction *common = commonIRemTransforms(I))
3221 if (Value *RHSNeg = dyn_castNegVal(Op1))
3222 if (!isa<Constant>(RHSNeg) ||
3223 (isa<ConstantInt>(RHSNeg) &&
3224 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3226 AddUsesToWorkList(I);
3227 I.setOperand(1, RHSNeg);
3231 // If the sign bits of both operands are zero (i.e. we can prove they are
3232 // unsigned inputs), turn this into a urem.
3233 if (I.getType()->isInteger()) {
3234 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3235 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3236 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3237 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3241 // If it's a constant vector, flip any negative values positive.
3242 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3243 unsigned VWidth = RHSV->getNumOperands();
3245 bool hasNegative = false;
3246 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3247 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3248 if (RHS->getValue().isNegative())
3252 std::vector<Constant *> Elts(VWidth);
3253 for (unsigned i = 0; i != VWidth; ++i) {
3254 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3255 if (RHS->getValue().isNegative())
3256 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3262 Constant *NewRHSV = ConstantVector::get(Elts);
3263 if (NewRHSV != RHSV) {
3264 AddUsesToWorkList(I);
3265 I.setOperand(1, NewRHSV);
3274 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3275 return commonRemTransforms(I);
3278 // isOneBitSet - Return true if there is exactly one bit set in the specified
3280 static bool isOneBitSet(const ConstantInt *CI) {
3281 return CI->getValue().isPowerOf2();
3284 // isHighOnes - Return true if the constant is of the form 1+0+.
3285 // This is the same as lowones(~X).
3286 static bool isHighOnes(const ConstantInt *CI) {
3287 return (~CI->getValue() + 1).isPowerOf2();
3290 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3291 /// are carefully arranged to allow folding of expressions such as:
3293 /// (A < B) | (A > B) --> (A != B)
3295 /// Note that this is only valid if the first and second predicates have the
3296 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3298 /// Three bits are used to represent the condition, as follows:
3303 /// <=> Value Definition
3304 /// 000 0 Always false
3311 /// 111 7 Always true
3313 static unsigned getICmpCode(const ICmpInst *ICI) {
3314 switch (ICI->getPredicate()) {
3316 case ICmpInst::ICMP_UGT: return 1; // 001
3317 case ICmpInst::ICMP_SGT: return 1; // 001
3318 case ICmpInst::ICMP_EQ: return 2; // 010
3319 case ICmpInst::ICMP_UGE: return 3; // 011
3320 case ICmpInst::ICMP_SGE: return 3; // 011
3321 case ICmpInst::ICMP_ULT: return 4; // 100
3322 case ICmpInst::ICMP_SLT: return 4; // 100
3323 case ICmpInst::ICMP_NE: return 5; // 101
3324 case ICmpInst::ICMP_ULE: return 6; // 110
3325 case ICmpInst::ICMP_SLE: return 6; // 110
3328 llvm_unreachable("Invalid ICmp predicate!");
3333 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3334 /// predicate into a three bit mask. It also returns whether it is an ordered
3335 /// predicate by reference.
3336 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3339 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3340 case FCmpInst::FCMP_UNO: return 0; // 000
3341 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3342 case FCmpInst::FCMP_UGT: return 1; // 001
3343 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3344 case FCmpInst::FCMP_UEQ: return 2; // 010
3345 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3346 case FCmpInst::FCMP_UGE: return 3; // 011
3347 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3348 case FCmpInst::FCMP_ULT: return 4; // 100
3349 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3350 case FCmpInst::FCMP_UNE: return 5; // 101
3351 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3352 case FCmpInst::FCMP_ULE: return 6; // 110
3355 // Not expecting FCMP_FALSE and FCMP_TRUE;
3356 llvm_unreachable("Unexpected FCmp predicate!");
3361 /// getICmpValue - This is the complement of getICmpCode, which turns an
3362 /// opcode and two operands into either a constant true or false, or a brand
3363 /// new ICmp instruction. The sign is passed in to determine which kind
3364 /// of predicate to use in the new icmp instruction.
3365 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3366 LLVMContext *Context) {
3368 default: llvm_unreachable("Illegal ICmp code!");
3369 case 0: return ConstantInt::getFalse(*Context);
3372 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, LHS, RHS);
3374 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, LHS, RHS);
3375 case 2: return new ICmpInst(*Context, ICmpInst::ICMP_EQ, LHS, RHS);
3378 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHS, RHS);
3380 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHS, RHS);
3383 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHS, RHS);
3385 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHS, RHS);
3386 case 5: return new ICmpInst(*Context, ICmpInst::ICMP_NE, LHS, RHS);
3389 return new ICmpInst(*Context, ICmpInst::ICMP_SLE, LHS, RHS);
3391 return new ICmpInst(*Context, ICmpInst::ICMP_ULE, LHS, RHS);
3392 case 7: return ConstantInt::getTrue(*Context);
3396 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3397 /// opcode and two operands into either a FCmp instruction. isordered is passed
3398 /// in to determine which kind of predicate to use in the new fcmp instruction.
3399 static Value *getFCmpValue(bool isordered, unsigned code,
3400 Value *LHS, Value *RHS, LLVMContext *Context) {
3402 default: llvm_unreachable("Illegal FCmp code!");
3405 return new FCmpInst(*Context, FCmpInst::FCMP_ORD, LHS, RHS);
3407 return new FCmpInst(*Context, FCmpInst::FCMP_UNO, LHS, RHS);
3410 return new FCmpInst(*Context, FCmpInst::FCMP_OGT, LHS, RHS);
3412 return new FCmpInst(*Context, FCmpInst::FCMP_UGT, LHS, RHS);
3415 return new FCmpInst(*Context, FCmpInst::FCMP_OEQ, LHS, RHS);
3417 return new FCmpInst(*Context, FCmpInst::FCMP_UEQ, LHS, RHS);
3420 return new FCmpInst(*Context, FCmpInst::FCMP_OGE, LHS, RHS);
3422 return new FCmpInst(*Context, FCmpInst::FCMP_UGE, LHS, RHS);
3425 return new FCmpInst(*Context, FCmpInst::FCMP_OLT, LHS, RHS);
3427 return new FCmpInst(*Context, FCmpInst::FCMP_ULT, LHS, RHS);
3430 return new FCmpInst(*Context, FCmpInst::FCMP_ONE, LHS, RHS);
3432 return new FCmpInst(*Context, FCmpInst::FCMP_UNE, LHS, RHS);
3435 return new FCmpInst(*Context, FCmpInst::FCMP_OLE, LHS, RHS);
3437 return new FCmpInst(*Context, FCmpInst::FCMP_ULE, LHS, RHS);
3438 case 7: return ConstantInt::getTrue(*Context);
3442 /// PredicatesFoldable - Return true if both predicates match sign or if at
3443 /// least one of them is an equality comparison (which is signless).
3444 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3445 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3446 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3447 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3451 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3452 struct FoldICmpLogical {
3455 ICmpInst::Predicate pred;
3456 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3457 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3458 pred(ICI->getPredicate()) {}
3459 bool shouldApply(Value *V) const {
3460 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3461 if (PredicatesFoldable(pred, ICI->getPredicate()))
3462 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3463 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3466 Instruction *apply(Instruction &Log) const {
3467 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3468 if (ICI->getOperand(0) != LHS) {
3469 assert(ICI->getOperand(1) == LHS);
3470 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3473 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3474 unsigned LHSCode = getICmpCode(ICI);
3475 unsigned RHSCode = getICmpCode(RHSICI);
3477 switch (Log.getOpcode()) {
3478 case Instruction::And: Code = LHSCode & RHSCode; break;
3479 case Instruction::Or: Code = LHSCode | RHSCode; break;
3480 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3481 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3484 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3485 ICmpInst::isSignedPredicate(ICI->getPredicate());
3487 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3488 if (Instruction *I = dyn_cast<Instruction>(RV))
3490 // Otherwise, it's a constant boolean value...
3491 return IC.ReplaceInstUsesWith(Log, RV);
3494 } // end anonymous namespace
3496 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3497 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3498 // guaranteed to be a binary operator.
3499 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3501 ConstantInt *AndRHS,
3502 BinaryOperator &TheAnd) {
3503 Value *X = Op->getOperand(0);
3504 Constant *Together = 0;
3506 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3508 switch (Op->getOpcode()) {
3509 case Instruction::Xor:
3510 if (Op->hasOneUse()) {
3511 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3512 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3513 InsertNewInstBefore(And, TheAnd);
3515 return BinaryOperator::CreateXor(And, Together);
3518 case Instruction::Or:
3519 if (Together == AndRHS) // (X | C) & C --> C
3520 return ReplaceInstUsesWith(TheAnd, AndRHS);
3522 if (Op->hasOneUse() && Together != OpRHS) {
3523 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3524 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3525 InsertNewInstBefore(Or, TheAnd);
3527 return BinaryOperator::CreateAnd(Or, AndRHS);
3530 case Instruction::Add:
3531 if (Op->hasOneUse()) {
3532 // Adding a one to a single bit bit-field should be turned into an XOR
3533 // of the bit. First thing to check is to see if this AND is with a
3534 // single bit constant.
3535 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3537 // If there is only one bit set...
3538 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3539 // Ok, at this point, we know that we are masking the result of the
3540 // ADD down to exactly one bit. If the constant we are adding has
3541 // no bits set below this bit, then we can eliminate the ADD.
3542 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3544 // Check to see if any bits below the one bit set in AndRHSV are set.
3545 if ((AddRHS & (AndRHSV-1)) == 0) {
3546 // If not, the only thing that can effect the output of the AND is
3547 // the bit specified by AndRHSV. If that bit is set, the effect of
3548 // the XOR is to toggle the bit. If it is clear, then the ADD has
3550 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3551 TheAnd.setOperand(0, X);
3554 // Pull the XOR out of the AND.
3555 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3556 InsertNewInstBefore(NewAnd, TheAnd);
3557 NewAnd->takeName(Op);
3558 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3565 case Instruction::Shl: {
3566 // We know that the AND will not produce any of the bits shifted in, so if
3567 // the anded constant includes them, clear them now!
3569 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3570 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3571 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3572 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3574 if (CI->getValue() == ShlMask) {
3575 // Masking out bits that the shift already masks
3576 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3577 } else if (CI != AndRHS) { // Reducing bits set in and.
3578 TheAnd.setOperand(1, CI);
3583 case Instruction::LShr:
3585 // We know that the AND will not produce any of the bits shifted in, so if
3586 // the anded constant includes them, clear them now! This only applies to
3587 // unsigned shifts, because a signed shr may bring in set bits!
3589 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3590 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3591 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3592 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3594 if (CI->getValue() == ShrMask) {
3595 // Masking out bits that the shift already masks.
3596 return ReplaceInstUsesWith(TheAnd, Op);
3597 } else if (CI != AndRHS) {
3598 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3603 case Instruction::AShr:
3605 // See if this is shifting in some sign extension, then masking it out
3607 if (Op->hasOneUse()) {
3608 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3609 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3610 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3611 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3612 if (C == AndRHS) { // Masking out bits shifted in.
3613 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3614 // Make the argument unsigned.
3615 Value *ShVal = Op->getOperand(0);
3616 ShVal = InsertNewInstBefore(
3617 BinaryOperator::CreateLShr(ShVal, OpRHS,
3618 Op->getName()), TheAnd);
3619 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3628 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3629 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3630 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3631 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3632 /// insert new instructions.
3633 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3634 bool isSigned, bool Inside,
3636 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3637 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3638 "Lo is not <= Hi in range emission code!");
3641 if (Lo == Hi) // Trivially false.
3642 return new ICmpInst(*Context, ICmpInst::ICMP_NE, V, V);
3644 // V >= Min && V < Hi --> V < Hi
3645 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3646 ICmpInst::Predicate pred = (isSigned ?
3647 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3648 return new ICmpInst(*Context, pred, V, Hi);
3651 // Emit V-Lo <u Hi-Lo
3652 Constant *NegLo = ConstantExpr::getNeg(Lo);
3653 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3654 InsertNewInstBefore(Add, IB);
3655 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3656 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, UpperBound);
3659 if (Lo == Hi) // Trivially true.
3660 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, V, V);
3662 // V < Min || V >= Hi -> V > Hi-1
3663 Hi = SubOne(cast<ConstantInt>(Hi));
3664 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3665 ICmpInst::Predicate pred = (isSigned ?
3666 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3667 return new ICmpInst(*Context, pred, V, Hi);
3670 // Emit V-Lo >u Hi-1-Lo
3671 // Note that Hi has already had one subtracted from it, above.
3672 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3673 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3674 InsertNewInstBefore(Add, IB);
3675 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3676 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add, LowerBound);
3679 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3680 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3681 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3682 // not, since all 1s are not contiguous.
3683 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3684 const APInt& V = Val->getValue();
3685 uint32_t BitWidth = Val->getType()->getBitWidth();
3686 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3688 // look for the first zero bit after the run of ones
3689 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3690 // look for the first non-zero bit
3691 ME = V.getActiveBits();
3695 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3696 /// where isSub determines whether the operator is a sub. If we can fold one of
3697 /// the following xforms:
3699 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3700 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3701 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3703 /// return (A +/- B).
3705 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3706 ConstantInt *Mask, bool isSub,
3708 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3709 if (!LHSI || LHSI->getNumOperands() != 2 ||
3710 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3712 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3714 switch (LHSI->getOpcode()) {
3716 case Instruction::And:
3717 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3718 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3719 if ((Mask->getValue().countLeadingZeros() +
3720 Mask->getValue().countPopulation()) ==
3721 Mask->getValue().getBitWidth())
3724 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3725 // part, we don't need any explicit masks to take them out of A. If that
3726 // is all N is, ignore it.
3727 uint32_t MB = 0, ME = 0;
3728 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3729 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3730 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3731 if (MaskedValueIsZero(RHS, Mask))
3736 case Instruction::Or:
3737 case Instruction::Xor:
3738 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3739 if ((Mask->getValue().countLeadingZeros() +
3740 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3741 && ConstantExpr::getAnd(N, Mask)->isNullValue())
3748 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3750 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3751 return InsertNewInstBefore(New, I);
3754 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3755 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3756 ICmpInst *LHS, ICmpInst *RHS) {
3758 ConstantInt *LHSCst, *RHSCst;
3759 ICmpInst::Predicate LHSCC, RHSCC;
3761 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3762 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3763 m_ConstantInt(LHSCst))) ||
3764 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3765 m_ConstantInt(RHSCst))))
3768 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3769 // where C is a power of 2
3770 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3771 LHSCst->getValue().isPowerOf2()) {
3772 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3773 InsertNewInstBefore(NewOr, I);
3774 return new ICmpInst(*Context, LHSCC, NewOr, LHSCst);
3777 // From here on, we only handle:
3778 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3779 if (Val != Val2) return 0;
3781 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3782 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3783 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3784 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3785 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3788 // We can't fold (ugt x, C) & (sgt x, C2).
3789 if (!PredicatesFoldable(LHSCC, RHSCC))
3792 // Ensure that the larger constant is on the RHS.
3794 if (ICmpInst::isSignedPredicate(LHSCC) ||
3795 (ICmpInst::isEquality(LHSCC) &&
3796 ICmpInst::isSignedPredicate(RHSCC)))
3797 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3799 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3802 std::swap(LHS, RHS);
3803 std::swap(LHSCst, RHSCst);
3804 std::swap(LHSCC, RHSCC);
3807 // At this point, we know we have have two icmp instructions
3808 // comparing a value against two constants and and'ing the result
3809 // together. Because of the above check, we know that we only have
3810 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3811 // (from the FoldICmpLogical check above), that the two constants
3812 // are not equal and that the larger constant is on the RHS
3813 assert(LHSCst != RHSCst && "Compares not folded above?");
3816 default: llvm_unreachable("Unknown integer condition code!");
3817 case ICmpInst::ICMP_EQ:
3819 default: llvm_unreachable("Unknown integer condition code!");
3820 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3821 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3822 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3823 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3824 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3825 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3826 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3827 return ReplaceInstUsesWith(I, LHS);
3829 case ICmpInst::ICMP_NE:
3831 default: llvm_unreachable("Unknown integer condition code!");
3832 case ICmpInst::ICMP_ULT:
3833 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3834 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Val, LHSCst);
3835 break; // (X != 13 & X u< 15) -> no change
3836 case ICmpInst::ICMP_SLT:
3837 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3838 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Val, LHSCst);
3839 break; // (X != 13 & X s< 15) -> no change
3840 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3841 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3842 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3843 return ReplaceInstUsesWith(I, RHS);
3844 case ICmpInst::ICMP_NE:
3845 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3846 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3847 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3848 Val->getName()+".off");
3849 InsertNewInstBefore(Add, I);
3850 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add,
3851 ConstantInt::get(Add->getType(), 1));
3853 break; // (X != 13 & X != 15) -> no change
3856 case ICmpInst::ICMP_ULT:
3858 default: llvm_unreachable("Unknown integer condition code!");
3859 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3860 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3861 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3862 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3864 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3865 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3866 return ReplaceInstUsesWith(I, LHS);
3867 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3871 case ICmpInst::ICMP_SLT:
3873 default: llvm_unreachable("Unknown integer condition code!");
3874 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3875 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3876 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3877 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3879 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3880 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3881 return ReplaceInstUsesWith(I, LHS);
3882 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3886 case ICmpInst::ICMP_UGT:
3888 default: llvm_unreachable("Unknown integer condition code!");
3889 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3890 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3891 return ReplaceInstUsesWith(I, RHS);
3892 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3894 case ICmpInst::ICMP_NE:
3895 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3896 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3897 break; // (X u> 13 & X != 15) -> no change
3898 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3899 return InsertRangeTest(Val, AddOne(LHSCst),
3900 RHSCst, false, true, I);
3901 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3905 case ICmpInst::ICMP_SGT:
3907 default: llvm_unreachable("Unknown integer condition code!");
3908 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3909 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3910 return ReplaceInstUsesWith(I, RHS);
3911 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3913 case ICmpInst::ICMP_NE:
3914 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3915 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3916 break; // (X s> 13 & X != 15) -> no change
3917 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3918 return InsertRangeTest(Val, AddOne(LHSCst),
3919 RHSCst, true, true, I);
3920 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3929 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
3932 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3933 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3934 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3935 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3936 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3937 // If either of the constants are nans, then the whole thing returns
3939 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3940 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3941 return new FCmpInst(*Context, FCmpInst::FCMP_ORD,
3942 LHS->getOperand(0), RHS->getOperand(0));
3945 // Handle vector zeros. This occurs because the canonical form of
3946 // "fcmp ord x,x" is "fcmp ord x, 0".
3947 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
3948 isa<ConstantAggregateZero>(RHS->getOperand(1)))
3949 return new FCmpInst(*Context, FCmpInst::FCMP_ORD,
3950 LHS->getOperand(0), RHS->getOperand(0));
3954 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
3955 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
3956 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
3959 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
3960 // Swap RHS operands to match LHS.
3961 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
3962 std::swap(Op1LHS, Op1RHS);
3965 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
3966 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
3968 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
3970 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
3971 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3972 if (Op0CC == FCmpInst::FCMP_TRUE)
3973 return ReplaceInstUsesWith(I, RHS);
3974 if (Op1CC == FCmpInst::FCMP_TRUE)
3975 return ReplaceInstUsesWith(I, LHS);
3979 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
3980 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
3982 std::swap(LHS, RHS);
3983 std::swap(Op0Pred, Op1Pred);
3984 std::swap(Op0Ordered, Op1Ordered);
3987 // uno && ueq -> uno && (uno || eq) -> ueq
3988 // ord && olt -> ord && (ord && lt) -> olt
3989 if (Op0Ordered == Op1Ordered)
3990 return ReplaceInstUsesWith(I, RHS);
3992 // uno && oeq -> uno && (ord && eq) -> false
3993 // uno && ord -> false
3995 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3996 // ord && ueq -> ord && (uno || eq) -> oeq
3997 return cast<Instruction>(getFCmpValue(true, Op1Pred,
3998 Op0LHS, Op0RHS, Context));
4006 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4007 bool Changed = SimplifyCommutative(I);
4008 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4010 if (isa<UndefValue>(Op1)) // X & undef -> 0
4011 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4015 return ReplaceInstUsesWith(I, Op1);
4017 // See if we can simplify any instructions used by the instruction whose sole
4018 // purpose is to compute bits we don't care about.
4019 if (SimplifyDemandedInstructionBits(I))
4021 if (isa<VectorType>(I.getType())) {
4022 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4023 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4024 return ReplaceInstUsesWith(I, I.getOperand(0));
4025 } else if (isa<ConstantAggregateZero>(Op1)) {
4026 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4030 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4031 const APInt& AndRHSMask = AndRHS->getValue();
4032 APInt NotAndRHS(~AndRHSMask);
4034 // Optimize a variety of ((val OP C1) & C2) combinations...
4035 if (isa<BinaryOperator>(Op0)) {
4036 Instruction *Op0I = cast<Instruction>(Op0);
4037 Value *Op0LHS = Op0I->getOperand(0);
4038 Value *Op0RHS = Op0I->getOperand(1);
4039 switch (Op0I->getOpcode()) {
4040 case Instruction::Xor:
4041 case Instruction::Or:
4042 // If the mask is only needed on one incoming arm, push it up.
4043 if (Op0I->hasOneUse()) {
4044 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4045 // Not masking anything out for the LHS, move to RHS.
4046 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4047 Op0RHS->getName()+".masked");
4048 InsertNewInstBefore(NewRHS, I);
4049 return BinaryOperator::Create(
4050 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4052 if (!isa<Constant>(Op0RHS) &&
4053 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4054 // Not masking anything out for the RHS, move to LHS.
4055 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4056 Op0LHS->getName()+".masked");
4057 InsertNewInstBefore(NewLHS, I);
4058 return BinaryOperator::Create(
4059 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4064 case Instruction::Add:
4065 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4066 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4067 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4068 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4069 return BinaryOperator::CreateAnd(V, AndRHS);
4070 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4071 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4074 case Instruction::Sub:
4075 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4076 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4077 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4078 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4079 return BinaryOperator::CreateAnd(V, AndRHS);
4081 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4082 // has 1's for all bits that the subtraction with A might affect.
4083 if (Op0I->hasOneUse()) {
4084 uint32_t BitWidth = AndRHSMask.getBitWidth();
4085 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4086 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4088 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4089 if (!(A && A->isZero()) && // avoid infinite recursion.
4090 MaskedValueIsZero(Op0LHS, Mask)) {
4091 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
4092 InsertNewInstBefore(NewNeg, I);
4093 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4098 case Instruction::Shl:
4099 case Instruction::LShr:
4100 // (1 << x) & 1 --> zext(x == 0)
4101 // (1 >> x) & 1 --> zext(x == 0)
4102 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4103 Instruction *NewICmp = new ICmpInst(*Context, ICmpInst::ICMP_EQ,
4104 Op0RHS, Constant::getNullValue(I.getType()));
4105 InsertNewInstBefore(NewICmp, I);
4106 return new ZExtInst(NewICmp, I.getType());
4111 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4112 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4114 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4115 // If this is an integer truncation or change from signed-to-unsigned, and
4116 // if the source is an and/or with immediate, transform it. This
4117 // frequently occurs for bitfield accesses.
4118 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4119 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4120 CastOp->getNumOperands() == 2)
4121 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4122 if (CastOp->getOpcode() == Instruction::And) {
4123 // Change: and (cast (and X, C1) to T), C2
4124 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4125 // This will fold the two constants together, which may allow
4126 // other simplifications.
4127 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4128 CastOp->getOperand(0), I.getType(),
4129 CastOp->getName()+".shrunk");
4130 NewCast = InsertNewInstBefore(NewCast, I);
4131 // trunc_or_bitcast(C1)&C2
4133 ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4134 C3 = ConstantExpr::getAnd(C3, AndRHS);
4135 return BinaryOperator::CreateAnd(NewCast, C3);
4136 } else if (CastOp->getOpcode() == Instruction::Or) {
4137 // Change: and (cast (or X, C1) to T), C2
4138 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4140 ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4141 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4143 return ReplaceInstUsesWith(I, AndRHS);
4149 // Try to fold constant and into select arguments.
4150 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4151 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4153 if (isa<PHINode>(Op0))
4154 if (Instruction *NV = FoldOpIntoPhi(I))
4158 Value *Op0NotVal = dyn_castNotVal(Op0);
4159 Value *Op1NotVal = dyn_castNotVal(Op1);
4161 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4162 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4164 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4165 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4166 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4167 I.getName()+".demorgan");
4168 InsertNewInstBefore(Or, I);
4169 return BinaryOperator::CreateNot(Or);
4173 Value *A = 0, *B = 0, *C = 0, *D = 0;
4174 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
4175 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4176 return ReplaceInstUsesWith(I, Op1);
4178 // (A|B) & ~(A&B) -> A^B
4179 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
4180 if ((A == C && B == D) || (A == D && B == C))
4181 return BinaryOperator::CreateXor(A, B);
4185 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
4186 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4187 return ReplaceInstUsesWith(I, Op0);
4189 // ~(A&B) & (A|B) -> A^B
4190 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
4191 if ((A == C && B == D) || (A == D && B == C))
4192 return BinaryOperator::CreateXor(A, B);
4196 if (Op0->hasOneUse() &&
4197 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
4198 if (A == Op1) { // (A^B)&A -> A&(A^B)
4199 I.swapOperands(); // Simplify below
4200 std::swap(Op0, Op1);
4201 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4202 cast<BinaryOperator>(Op0)->swapOperands();
4203 I.swapOperands(); // Simplify below
4204 std::swap(Op0, Op1);
4208 if (Op1->hasOneUse() &&
4209 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
4210 if (B == Op0) { // B&(A^B) -> B&(B^A)
4211 cast<BinaryOperator>(Op1)->swapOperands();
4214 if (A == Op0) { // A&(A^B) -> A & ~B
4215 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4216 InsertNewInstBefore(NotB, I);
4217 return BinaryOperator::CreateAnd(A, NotB);
4221 // (A&((~A)|B)) -> A&B
4222 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A))) ||
4223 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1)))))
4224 return BinaryOperator::CreateAnd(A, Op1);
4225 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A))) ||
4226 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0)))))
4227 return BinaryOperator::CreateAnd(A, Op0);
4230 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4231 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4232 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4235 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4236 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4240 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4241 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4242 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4243 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4244 const Type *SrcTy = Op0C->getOperand(0)->getType();
4245 if (SrcTy == Op1C->getOperand(0)->getType() &&
4246 SrcTy->isIntOrIntVector() &&
4247 // Only do this if the casts both really cause code to be generated.
4248 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4250 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4252 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4253 Op1C->getOperand(0),
4255 InsertNewInstBefore(NewOp, I);
4256 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4260 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4261 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4262 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4263 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4264 SI0->getOperand(1) == SI1->getOperand(1) &&
4265 (SI0->hasOneUse() || SI1->hasOneUse())) {
4266 Instruction *NewOp =
4267 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4269 SI0->getName()), I);
4270 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4271 SI1->getOperand(1));
4275 // If and'ing two fcmp, try combine them into one.
4276 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4277 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4278 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4282 return Changed ? &I : 0;
4285 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4286 /// capable of providing pieces of a bswap. The subexpression provides pieces
4287 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4288 /// the expression came from the corresponding "byte swapped" byte in some other
4289 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4290 /// we know that the expression deposits the low byte of %X into the high byte
4291 /// of the bswap result and that all other bytes are zero. This expression is
4292 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4295 /// This function returns true if the match was unsuccessful and false if so.
4296 /// On entry to the function the "OverallLeftShift" is a signed integer value
4297 /// indicating the number of bytes that the subexpression is later shifted. For
4298 /// example, if the expression is later right shifted by 16 bits, the
4299 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4300 /// byte of ByteValues is actually being set.
4302 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4303 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4304 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4305 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4306 /// always in the local (OverallLeftShift) coordinate space.
4308 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4309 SmallVector<Value*, 8> &ByteValues) {
4310 if (Instruction *I = dyn_cast<Instruction>(V)) {
4311 // If this is an or instruction, it may be an inner node of the bswap.
4312 if (I->getOpcode() == Instruction::Or) {
4313 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4315 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4319 // If this is a logical shift by a constant multiple of 8, recurse with
4320 // OverallLeftShift and ByteMask adjusted.
4321 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4323 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4324 // Ensure the shift amount is defined and of a byte value.
4325 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4328 unsigned ByteShift = ShAmt >> 3;
4329 if (I->getOpcode() == Instruction::Shl) {
4330 // X << 2 -> collect(X, +2)
4331 OverallLeftShift += ByteShift;
4332 ByteMask >>= ByteShift;
4334 // X >>u 2 -> collect(X, -2)
4335 OverallLeftShift -= ByteShift;
4336 ByteMask <<= ByteShift;
4337 ByteMask &= (~0U >> (32-ByteValues.size()));
4340 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4341 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4343 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4347 // If this is a logical 'and' with a mask that clears bytes, clear the
4348 // corresponding bytes in ByteMask.
4349 if (I->getOpcode() == Instruction::And &&
4350 isa<ConstantInt>(I->getOperand(1))) {
4351 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4352 unsigned NumBytes = ByteValues.size();
4353 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4354 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4356 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4357 // If this byte is masked out by a later operation, we don't care what
4359 if ((ByteMask & (1 << i)) == 0)
4362 // If the AndMask is all zeros for this byte, clear the bit.
4363 APInt MaskB = AndMask & Byte;
4365 ByteMask &= ~(1U << i);
4369 // If the AndMask is not all ones for this byte, it's not a bytezap.
4373 // Otherwise, this byte is kept.
4376 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4381 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4382 // the input value to the bswap. Some observations: 1) if more than one byte
4383 // is demanded from this input, then it could not be successfully assembled
4384 // into a byteswap. At least one of the two bytes would not be aligned with
4385 // their ultimate destination.
4386 if (!isPowerOf2_32(ByteMask)) return true;
4387 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4389 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4390 // is demanded, it needs to go into byte 0 of the result. This means that the
4391 // byte needs to be shifted until it lands in the right byte bucket. The
4392 // shift amount depends on the position: if the byte is coming from the high
4393 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4394 // low part, it must be shifted left.
4395 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4396 if (InputByteNo < ByteValues.size()/2) {
4397 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4400 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4404 // If the destination byte value is already defined, the values are or'd
4405 // together, which isn't a bswap (unless it's an or of the same bits).
4406 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4408 ByteValues[DestByteNo] = V;
4412 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4413 /// If so, insert the new bswap intrinsic and return it.
4414 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4415 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4416 if (!ITy || ITy->getBitWidth() % 16 ||
4417 // ByteMask only allows up to 32-byte values.
4418 ITy->getBitWidth() > 32*8)
4419 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4421 /// ByteValues - For each byte of the result, we keep track of which value
4422 /// defines each byte.
4423 SmallVector<Value*, 8> ByteValues;
4424 ByteValues.resize(ITy->getBitWidth()/8);
4426 // Try to find all the pieces corresponding to the bswap.
4427 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4428 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4431 // Check to see if all of the bytes come from the same value.
4432 Value *V = ByteValues[0];
4433 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4435 // Check to make sure that all of the bytes come from the same value.
4436 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4437 if (ByteValues[i] != V)
4439 const Type *Tys[] = { ITy };
4440 Module *M = I.getParent()->getParent()->getParent();
4441 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4442 return CallInst::Create(F, V);
4445 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4446 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4447 /// we can simplify this expression to "cond ? C : D or B".
4448 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4450 LLVMContext *Context) {
4451 // If A is not a select of -1/0, this cannot match.
4453 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond))))
4456 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4457 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond))))
4458 return SelectInst::Create(Cond, C, B);
4459 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4460 return SelectInst::Create(Cond, C, B);
4461 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4462 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond))))
4463 return SelectInst::Create(Cond, C, D);
4464 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond)))))
4465 return SelectInst::Create(Cond, C, D);
4469 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4470 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4471 ICmpInst *LHS, ICmpInst *RHS) {
4473 ConstantInt *LHSCst, *RHSCst;
4474 ICmpInst::Predicate LHSCC, RHSCC;
4476 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4477 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4478 m_ConstantInt(LHSCst))) ||
4479 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4480 m_ConstantInt(RHSCst))))
4483 // From here on, we only handle:
4484 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4485 if (Val != Val2) return 0;
4487 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4488 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4489 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4490 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4491 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4494 // We can't fold (ugt x, C) | (sgt x, C2).
4495 if (!PredicatesFoldable(LHSCC, RHSCC))
4498 // Ensure that the larger constant is on the RHS.
4500 if (ICmpInst::isSignedPredicate(LHSCC) ||
4501 (ICmpInst::isEquality(LHSCC) &&
4502 ICmpInst::isSignedPredicate(RHSCC)))
4503 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4505 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4508 std::swap(LHS, RHS);
4509 std::swap(LHSCst, RHSCst);
4510 std::swap(LHSCC, RHSCC);
4513 // At this point, we know we have have two icmp instructions
4514 // comparing a value against two constants and or'ing the result
4515 // together. Because of the above check, we know that we only have
4516 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4517 // FoldICmpLogical check above), that the two constants are not
4519 assert(LHSCst != RHSCst && "Compares not folded above?");
4522 default: llvm_unreachable("Unknown integer condition code!");
4523 case ICmpInst::ICMP_EQ:
4525 default: llvm_unreachable("Unknown integer condition code!");
4526 case ICmpInst::ICMP_EQ:
4527 if (LHSCst == SubOne(RHSCst)) {
4528 // (X == 13 | X == 14) -> X-13 <u 2
4529 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4530 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4531 Val->getName()+".off");
4532 InsertNewInstBefore(Add, I);
4533 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
4534 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, AddCST);
4536 break; // (X == 13 | X == 15) -> no change
4537 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4538 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4540 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4541 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4542 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4543 return ReplaceInstUsesWith(I, RHS);
4546 case ICmpInst::ICMP_NE:
4548 default: llvm_unreachable("Unknown integer condition code!");
4549 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4550 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4551 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4552 return ReplaceInstUsesWith(I, LHS);
4553 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4554 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4555 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4556 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4559 case ICmpInst::ICMP_ULT:
4561 default: llvm_unreachable("Unknown integer condition code!");
4562 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4564 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4565 // If RHSCst is [us]MAXINT, it is always false. Not handling
4566 // this can cause overflow.
4567 if (RHSCst->isMaxValue(false))
4568 return ReplaceInstUsesWith(I, LHS);
4569 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4571 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4573 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4574 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4575 return ReplaceInstUsesWith(I, RHS);
4576 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4580 case ICmpInst::ICMP_SLT:
4582 default: llvm_unreachable("Unknown integer condition code!");
4583 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4585 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4586 // If RHSCst is [us]MAXINT, it is always false. Not handling
4587 // this can cause overflow.
4588 if (RHSCst->isMaxValue(true))
4589 return ReplaceInstUsesWith(I, LHS);
4590 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst),
4592 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4594 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4595 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4596 return ReplaceInstUsesWith(I, RHS);
4597 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4601 case ICmpInst::ICMP_UGT:
4603 default: llvm_unreachable("Unknown integer condition code!");
4604 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4605 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4606 return ReplaceInstUsesWith(I, LHS);
4607 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4609 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4610 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4611 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4612 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4616 case ICmpInst::ICMP_SGT:
4618 default: llvm_unreachable("Unknown integer condition code!");
4619 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4620 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4621 return ReplaceInstUsesWith(I, LHS);
4622 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4624 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4625 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4626 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4627 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4635 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
4637 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4638 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4639 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4640 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4641 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4642 // If either of the constants are nans, then the whole thing returns
4644 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4645 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4647 // Otherwise, no need to compare the two constants, compare the
4649 return new FCmpInst(*Context, FCmpInst::FCMP_UNO,
4650 LHS->getOperand(0), RHS->getOperand(0));
4653 // Handle vector zeros. This occurs because the canonical form of
4654 // "fcmp uno x,x" is "fcmp uno x, 0".
4655 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4656 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4657 return new FCmpInst(*Context, FCmpInst::FCMP_UNO,
4658 LHS->getOperand(0), RHS->getOperand(0));
4663 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4664 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4665 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4667 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4668 // Swap RHS operands to match LHS.
4669 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4670 std::swap(Op1LHS, Op1RHS);
4672 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4673 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4675 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4677 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
4678 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4679 if (Op0CC == FCmpInst::FCMP_FALSE)
4680 return ReplaceInstUsesWith(I, RHS);
4681 if (Op1CC == FCmpInst::FCMP_FALSE)
4682 return ReplaceInstUsesWith(I, LHS);
4685 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4686 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4687 if (Op0Ordered == Op1Ordered) {
4688 // If both are ordered or unordered, return a new fcmp with
4689 // or'ed predicates.
4690 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4691 Op0LHS, Op0RHS, Context);
4692 if (Instruction *I = dyn_cast<Instruction>(RV))
4694 // Otherwise, it's a constant boolean value...
4695 return ReplaceInstUsesWith(I, RV);
4701 /// FoldOrWithConstants - This helper function folds:
4703 /// ((A | B) & C1) | (B & C2)
4709 /// when the XOR of the two constants is "all ones" (-1).
4710 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4711 Value *A, Value *B, Value *C) {
4712 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4716 ConstantInt *CI2 = 0;
4717 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)))) return 0;
4719 APInt Xor = CI1->getValue() ^ CI2->getValue();
4720 if (!Xor.isAllOnesValue()) return 0;
4722 if (V1 == A || V1 == B) {
4723 Instruction *NewOp =
4724 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4725 return BinaryOperator::CreateOr(NewOp, V1);
4731 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4732 bool Changed = SimplifyCommutative(I);
4733 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4735 if (isa<UndefValue>(Op1)) // X | undef -> -1
4736 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4740 return ReplaceInstUsesWith(I, Op0);
4742 // See if we can simplify any instructions used by the instruction whose sole
4743 // purpose is to compute bits we don't care about.
4744 if (SimplifyDemandedInstructionBits(I))
4746 if (isa<VectorType>(I.getType())) {
4747 if (isa<ConstantAggregateZero>(Op1)) {
4748 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4749 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4750 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4751 return ReplaceInstUsesWith(I, I.getOperand(1));
4756 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4757 ConstantInt *C1 = 0; Value *X = 0;
4758 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4759 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) &&
4761 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4762 InsertNewInstBefore(Or, I);
4764 return BinaryOperator::CreateAnd(Or,
4765 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
4768 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4769 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) &&
4771 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4772 InsertNewInstBefore(Or, I);
4774 return BinaryOperator::CreateXor(Or,
4775 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
4778 // Try to fold constant and into select arguments.
4779 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4780 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4782 if (isa<PHINode>(Op0))
4783 if (Instruction *NV = FoldOpIntoPhi(I))
4787 Value *A = 0, *B = 0;
4788 ConstantInt *C1 = 0, *C2 = 0;
4790 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4791 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4792 return ReplaceInstUsesWith(I, Op1);
4793 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4794 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4795 return ReplaceInstUsesWith(I, Op0);
4797 // (A | B) | C and A | (B | C) -> bswap if possible.
4798 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4799 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4800 match(Op1, m_Or(m_Value(), m_Value())) ||
4801 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4802 match(Op1, m_Shift(m_Value(), m_Value())))) {
4803 if (Instruction *BSwap = MatchBSwap(I))
4807 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4808 if (Op0->hasOneUse() &&
4809 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4810 MaskedValueIsZero(Op1, C1->getValue())) {
4811 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4812 InsertNewInstBefore(NOr, I);
4814 return BinaryOperator::CreateXor(NOr, C1);
4817 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4818 if (Op1->hasOneUse() &&
4819 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4820 MaskedValueIsZero(Op0, C1->getValue())) {
4821 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4822 InsertNewInstBefore(NOr, I);
4824 return BinaryOperator::CreateXor(NOr, C1);
4828 Value *C = 0, *D = 0;
4829 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4830 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4831 Value *V1 = 0, *V2 = 0, *V3 = 0;
4832 C1 = dyn_cast<ConstantInt>(C);
4833 C2 = dyn_cast<ConstantInt>(D);
4834 if (C1 && C2) { // (A & C1)|(B & C2)
4835 // If we have: ((V + N) & C1) | (V & C2)
4836 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4837 // replace with V+N.
4838 if (C1->getValue() == ~C2->getValue()) {
4839 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4840 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4841 // Add commutes, try both ways.
4842 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4843 return ReplaceInstUsesWith(I, A);
4844 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4845 return ReplaceInstUsesWith(I, A);
4847 // Or commutes, try both ways.
4848 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4849 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4850 // Add commutes, try both ways.
4851 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4852 return ReplaceInstUsesWith(I, B);
4853 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4854 return ReplaceInstUsesWith(I, B);
4857 V1 = 0; V2 = 0; V3 = 0;
4860 // Check to see if we have any common things being and'ed. If so, find the
4861 // terms for V1 & (V2|V3).
4862 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4863 if (A == B) // (A & C)|(A & D) == A & (C|D)
4864 V1 = A, V2 = C, V3 = D;
4865 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4866 V1 = A, V2 = B, V3 = C;
4867 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4868 V1 = C, V2 = A, V3 = D;
4869 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4870 V1 = C, V2 = A, V3 = B;
4874 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4875 return BinaryOperator::CreateAnd(V1, Or);
4879 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4880 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4882 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4884 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4886 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4889 // ((A&~B)|(~A&B)) -> A^B
4890 if ((match(C, m_Not(m_Specific(D))) &&
4891 match(B, m_Not(m_Specific(A)))))
4892 return BinaryOperator::CreateXor(A, D);
4893 // ((~B&A)|(~A&B)) -> A^B
4894 if ((match(A, m_Not(m_Specific(D))) &&
4895 match(B, m_Not(m_Specific(C)))))
4896 return BinaryOperator::CreateXor(C, D);
4897 // ((A&~B)|(B&~A)) -> A^B
4898 if ((match(C, m_Not(m_Specific(B))) &&
4899 match(D, m_Not(m_Specific(A)))))
4900 return BinaryOperator::CreateXor(A, B);
4901 // ((~B&A)|(B&~A)) -> A^B
4902 if ((match(A, m_Not(m_Specific(B))) &&
4903 match(D, m_Not(m_Specific(C)))))
4904 return BinaryOperator::CreateXor(C, B);
4907 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4908 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4909 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4910 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4911 SI0->getOperand(1) == SI1->getOperand(1) &&
4912 (SI0->hasOneUse() || SI1->hasOneUse())) {
4913 Instruction *NewOp =
4914 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4916 SI0->getName()), I);
4917 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4918 SI1->getOperand(1));
4922 // ((A|B)&1)|(B&-2) -> (A&1) | B
4923 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4924 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4925 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4926 if (Ret) return Ret;
4928 // (B&-2)|((A|B)&1) -> (A&1) | B
4929 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C))) ||
4930 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))))) {
4931 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4932 if (Ret) return Ret;
4935 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4936 if (A == Op1) // ~A | A == -1
4937 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4941 // Note, A is still live here!
4942 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4944 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4946 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4947 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4948 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4949 I.getName()+".demorgan"), I);
4950 return BinaryOperator::CreateNot(And);
4954 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4955 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4956 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4959 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4960 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4964 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4965 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4966 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4967 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4968 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4969 !isa<ICmpInst>(Op1C->getOperand(0))) {
4970 const Type *SrcTy = Op0C->getOperand(0)->getType();
4971 if (SrcTy == Op1C->getOperand(0)->getType() &&
4972 SrcTy->isIntOrIntVector() &&
4973 // Only do this if the casts both really cause code to be
4975 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4977 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4979 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4980 Op1C->getOperand(0),
4982 InsertNewInstBefore(NewOp, I);
4983 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4990 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4991 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4992 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4993 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
4997 return Changed ? &I : 0;
5002 // XorSelf - Implements: X ^ X --> 0
5005 XorSelf(Value *rhs) : RHS(rhs) {}
5006 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5007 Instruction *apply(BinaryOperator &Xor) const {
5014 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5015 bool Changed = SimplifyCommutative(I);
5016 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5018 if (isa<UndefValue>(Op1)) {
5019 if (isa<UndefValue>(Op0))
5020 // Handle undef ^ undef -> 0 special case. This is a common
5022 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5023 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5026 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5027 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
5028 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5029 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5032 // See if we can simplify any instructions used by the instruction whose sole
5033 // purpose is to compute bits we don't care about.
5034 if (SimplifyDemandedInstructionBits(I))
5036 if (isa<VectorType>(I.getType()))
5037 if (isa<ConstantAggregateZero>(Op1))
5038 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5040 // Is this a ~ operation?
5041 if (Value *NotOp = dyn_castNotVal(&I)) {
5042 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5043 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5044 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5045 if (Op0I->getOpcode() == Instruction::And ||
5046 Op0I->getOpcode() == Instruction::Or) {
5047 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
5048 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
5050 BinaryOperator::CreateNot(Op0I->getOperand(1),
5051 Op0I->getOperand(1)->getName()+".not");
5052 InsertNewInstBefore(NotY, I);
5053 if (Op0I->getOpcode() == Instruction::And)
5054 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5056 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5063 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5064 if (RHS == ConstantInt::getTrue(*Context) && Op0->hasOneUse()) {
5065 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5066 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5067 return new ICmpInst(*Context, ICI->getInversePredicate(),
5068 ICI->getOperand(0), ICI->getOperand(1));
5070 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5071 return new FCmpInst(*Context, FCI->getInversePredicate(),
5072 FCI->getOperand(0), FCI->getOperand(1));
5075 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5076 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5077 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5078 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5079 Instruction::CastOps Opcode = Op0C->getOpcode();
5080 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
5081 if (RHS == ConstantExpr::getCast(Opcode,
5082 ConstantInt::getTrue(*Context),
5083 Op0C->getDestTy())) {
5084 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
5086 CI->getOpcode(), CI->getInversePredicate(),
5087 CI->getOperand(0), CI->getOperand(1)), I);
5088 NewCI->takeName(CI);
5089 return CastInst::Create(Opcode, NewCI, Op0C->getType());
5096 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5097 // ~(c-X) == X-c-1 == X+(-c-1)
5098 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5099 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5100 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5101 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5102 ConstantInt::get(I.getType(), 1));
5103 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5106 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5107 if (Op0I->getOpcode() == Instruction::Add) {
5108 // ~(X-c) --> (-c-1)-X
5109 if (RHS->isAllOnesValue()) {
5110 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5111 return BinaryOperator::CreateSub(
5112 ConstantExpr::getSub(NegOp0CI,
5113 ConstantInt::get(I.getType(), 1)),
5114 Op0I->getOperand(0));
5115 } else if (RHS->getValue().isSignBit()) {
5116 // (X + C) ^ signbit -> (X + C + signbit)
5117 Constant *C = ConstantInt::get(*Context,
5118 RHS->getValue() + Op0CI->getValue());
5119 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5122 } else if (Op0I->getOpcode() == Instruction::Or) {
5123 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5124 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5125 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5126 // Anything in both C1 and C2 is known to be zero, remove it from
5128 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5129 NewRHS = ConstantExpr::getAnd(NewRHS,
5130 ConstantExpr::getNot(CommonBits));
5131 AddToWorkList(Op0I);
5132 I.setOperand(0, Op0I->getOperand(0));
5133 I.setOperand(1, NewRHS);
5140 // Try to fold constant and into select arguments.
5141 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5142 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5144 if (isa<PHINode>(Op0))
5145 if (Instruction *NV = FoldOpIntoPhi(I))
5149 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
5151 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5153 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
5155 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5158 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5161 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
5162 if (A == Op0) { // B^(B|A) == (A|B)^B
5163 Op1I->swapOperands();
5165 std::swap(Op0, Op1);
5166 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5167 I.swapOperands(); // Simplified below.
5168 std::swap(Op0, Op1);
5170 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)))) {
5171 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5172 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)))) {
5173 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5174 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) &&
5176 if (A == Op0) { // A^(A&B) -> A^(B&A)
5177 Op1I->swapOperands();
5180 if (B == Op0) { // A^(B&A) -> (B&A)^A
5181 I.swapOperands(); // Simplified below.
5182 std::swap(Op0, Op1);
5187 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5190 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5191 Op0I->hasOneUse()) {
5192 if (A == Op1) // (B|A)^B == (A|B)^B
5194 if (B == Op1) { // (A|B)^B == A & ~B
5196 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5197 return BinaryOperator::CreateAnd(A, NotB);
5199 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)))) {
5200 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5201 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)))) {
5202 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5203 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5205 if (A == Op1) // (A&B)^A -> (B&A)^A
5207 if (B == Op1 && // (B&A)^A == ~B & A
5208 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5210 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5211 return BinaryOperator::CreateAnd(N, Op1);
5216 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5217 if (Op0I && Op1I && Op0I->isShift() &&
5218 Op0I->getOpcode() == Op1I->getOpcode() &&
5219 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5220 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5221 Instruction *NewOp =
5222 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5223 Op1I->getOperand(0),
5224 Op0I->getName()), I);
5225 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5226 Op1I->getOperand(1));
5230 Value *A, *B, *C, *D;
5231 // (A & B)^(A | B) -> A ^ B
5232 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5233 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
5234 if ((A == C && B == D) || (A == D && B == C))
5235 return BinaryOperator::CreateXor(A, B);
5237 // (A | B)^(A & B) -> A ^ B
5238 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
5239 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5240 if ((A == C && B == D) || (A == D && B == C))
5241 return BinaryOperator::CreateXor(A, B);
5245 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5246 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
5247 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
5248 // (X & Y)^(X & Y) -> (Y^Z) & X
5249 Value *X = 0, *Y = 0, *Z = 0;
5251 X = A, Y = B, Z = D;
5253 X = A, Y = B, Z = C;
5255 X = B, Y = A, Z = D;
5257 X = B, Y = A, Z = C;
5260 Instruction *NewOp =
5261 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5262 return BinaryOperator::CreateAnd(NewOp, X);
5267 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5268 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5269 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
5272 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5273 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5274 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5275 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5276 const Type *SrcTy = Op0C->getOperand(0)->getType();
5277 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5278 // Only do this if the casts both really cause code to be generated.
5279 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5281 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5283 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5284 Op1C->getOperand(0),
5286 InsertNewInstBefore(NewOp, I);
5287 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5292 return Changed ? &I : 0;
5295 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5296 LLVMContext *Context) {
5297 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5300 static bool HasAddOverflow(ConstantInt *Result,
5301 ConstantInt *In1, ConstantInt *In2,
5304 if (In2->getValue().isNegative())
5305 return Result->getValue().sgt(In1->getValue());
5307 return Result->getValue().slt(In1->getValue());
5309 return Result->getValue().ult(In1->getValue());
5312 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5313 /// overflowed for this type.
5314 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5315 Constant *In2, LLVMContext *Context,
5316 bool IsSigned = false) {
5317 Result = ConstantExpr::getAdd(In1, In2);
5319 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5320 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5321 Constant *Idx = ConstantInt::get(Type::Int32Ty, i);
5322 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5323 ExtractElement(In1, Idx, Context),
5324 ExtractElement(In2, Idx, Context),
5331 return HasAddOverflow(cast<ConstantInt>(Result),
5332 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5336 static bool HasSubOverflow(ConstantInt *Result,
5337 ConstantInt *In1, ConstantInt *In2,
5340 if (In2->getValue().isNegative())
5341 return Result->getValue().slt(In1->getValue());
5343 return Result->getValue().sgt(In1->getValue());
5345 return Result->getValue().ugt(In1->getValue());
5348 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5349 /// overflowed for this type.
5350 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5351 Constant *In2, LLVMContext *Context,
5352 bool IsSigned = false) {
5353 Result = ConstantExpr::getSub(In1, In2);
5355 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5356 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5357 Constant *Idx = ConstantInt::get(Type::Int32Ty, i);
5358 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5359 ExtractElement(In1, Idx, Context),
5360 ExtractElement(In2, Idx, Context),
5367 return HasSubOverflow(cast<ConstantInt>(Result),
5368 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5372 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5373 /// code necessary to compute the offset from the base pointer (without adding
5374 /// in the base pointer). Return the result as a signed integer of intptr size.
5375 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5376 TargetData &TD = *IC.getTargetData();
5377 gep_type_iterator GTI = gep_type_begin(GEP);
5378 const Type *IntPtrTy = TD.getIntPtrType();
5379 LLVMContext *Context = IC.getContext();
5380 Value *Result = Constant::getNullValue(IntPtrTy);
5382 // Build a mask for high order bits.
5383 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5384 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5386 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5389 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5390 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5391 if (OpC->isZero()) continue;
5393 // Handle a struct index, which adds its field offset to the pointer.
5394 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5395 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5397 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5399 ConstantInt::get(*Context,
5400 RC->getValue() + APInt(IntPtrWidth, Size));
5402 Result = IC.InsertNewInstBefore(
5403 BinaryOperator::CreateAdd(Result,
5404 ConstantInt::get(IntPtrTy, Size),
5405 GEP->getName()+".offs"), I);
5409 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5411 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5412 Scale = ConstantExpr::getMul(OC, Scale);
5413 if (Constant *RC = dyn_cast<Constant>(Result))
5414 Result = ConstantExpr::getAdd(RC, Scale);
5416 // Emit an add instruction.
5417 Result = IC.InsertNewInstBefore(
5418 BinaryOperator::CreateAdd(Result, Scale,
5419 GEP->getName()+".offs"), I);
5423 // Convert to correct type.
5424 if (Op->getType() != IntPtrTy) {
5425 if (Constant *OpC = dyn_cast<Constant>(Op))
5426 Op = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true);
5428 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5430 Op->getName()+".c"), I);
5433 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5434 if (Constant *OpC = dyn_cast<Constant>(Op))
5435 Op = ConstantExpr::getMul(OpC, Scale);
5436 else // We'll let instcombine(mul) convert this to a shl if possible.
5437 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5438 GEP->getName()+".idx"), I);
5441 // Emit an add instruction.
5442 if (isa<Constant>(Op) && isa<Constant>(Result))
5443 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5444 cast<Constant>(Result));
5446 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5447 GEP->getName()+".offs"), I);
5453 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5454 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5455 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5456 /// be complex, and scales are involved. The above expression would also be
5457 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5458 /// This later form is less amenable to optimization though, and we are allowed
5459 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5461 /// If we can't emit an optimized form for this expression, this returns null.
5463 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5465 TargetData &TD = *IC.getTargetData();
5466 gep_type_iterator GTI = gep_type_begin(GEP);
5468 // Check to see if this gep only has a single variable index. If so, and if
5469 // any constant indices are a multiple of its scale, then we can compute this
5470 // in terms of the scale of the variable index. For example, if the GEP
5471 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5472 // because the expression will cross zero at the same point.
5473 unsigned i, e = GEP->getNumOperands();
5475 for (i = 1; i != e; ++i, ++GTI) {
5476 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5477 // Compute the aggregate offset of constant indices.
5478 if (CI->isZero()) continue;
5480 // Handle a struct index, which adds its field offset to the pointer.
5481 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5482 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5484 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5485 Offset += Size*CI->getSExtValue();
5488 // Found our variable index.
5493 // If there are no variable indices, we must have a constant offset, just
5494 // evaluate it the general way.
5495 if (i == e) return 0;
5497 Value *VariableIdx = GEP->getOperand(i);
5498 // Determine the scale factor of the variable element. For example, this is
5499 // 4 if the variable index is into an array of i32.
5500 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5502 // Verify that there are no other variable indices. If so, emit the hard way.
5503 for (++i, ++GTI; i != e; ++i, ++GTI) {
5504 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5507 // Compute the aggregate offset of constant indices.
5508 if (CI->isZero()) continue;
5510 // Handle a struct index, which adds its field offset to the pointer.
5511 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5512 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5514 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5515 Offset += Size*CI->getSExtValue();
5519 // Okay, we know we have a single variable index, which must be a
5520 // pointer/array/vector index. If there is no offset, life is simple, return
5522 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5524 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5525 // we don't need to bother extending: the extension won't affect where the
5526 // computation crosses zero.
5527 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5528 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5529 VariableIdx->getName(), &I);
5533 // Otherwise, there is an index. The computation we will do will be modulo
5534 // the pointer size, so get it.
5535 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5537 Offset &= PtrSizeMask;
5538 VariableScale &= PtrSizeMask;
5540 // To do this transformation, any constant index must be a multiple of the
5541 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5542 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5543 // multiple of the variable scale.
5544 int64_t NewOffs = Offset / (int64_t)VariableScale;
5545 if (Offset != NewOffs*(int64_t)VariableScale)
5548 // Okay, we can do this evaluation. Start by converting the index to intptr.
5549 const Type *IntPtrTy = TD.getIntPtrType();
5550 if (VariableIdx->getType() != IntPtrTy)
5551 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5553 VariableIdx->getName(), &I);
5554 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5555 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5559 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5560 /// else. At this point we know that the GEP is on the LHS of the comparison.
5561 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5562 ICmpInst::Predicate Cond,
5564 // Look through bitcasts.
5565 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5566 RHS = BCI->getOperand(0);
5568 Value *PtrBase = GEPLHS->getOperand(0);
5569 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5570 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5571 // This transformation (ignoring the base and scales) is valid because we
5572 // know pointers can't overflow since the gep is inbounds. See if we can
5573 // output an optimized form.
5574 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5576 // If not, synthesize the offset the hard way.
5578 Offset = EmitGEPOffset(GEPLHS, I, *this);
5579 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), Offset,
5580 Constant::getNullValue(Offset->getType()));
5581 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5582 // If the base pointers are different, but the indices are the same, just
5583 // compare the base pointer.
5584 if (PtrBase != GEPRHS->getOperand(0)) {
5585 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5586 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5587 GEPRHS->getOperand(0)->getType();
5589 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5590 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5591 IndicesTheSame = false;
5595 // If all indices are the same, just compare the base pointers.
5597 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond),
5598 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5600 // Otherwise, the base pointers are different and the indices are
5601 // different, bail out.
5605 // If one of the GEPs has all zero indices, recurse.
5606 bool AllZeros = true;
5607 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5608 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5609 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5614 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5615 ICmpInst::getSwappedPredicate(Cond), I);
5617 // If the other GEP has all zero indices, recurse.
5619 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5620 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5621 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5626 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5628 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5629 // If the GEPs only differ by one index, compare it.
5630 unsigned NumDifferences = 0; // Keep track of # differences.
5631 unsigned DiffOperand = 0; // The operand that differs.
5632 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5633 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5634 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5635 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5636 // Irreconcilable differences.
5640 if (NumDifferences++) break;
5645 if (NumDifferences == 0) // SAME GEP?
5646 return ReplaceInstUsesWith(I, // No comparison is needed here.
5647 ConstantInt::get(Type::Int1Ty,
5648 ICmpInst::isTrueWhenEqual(Cond)));
5650 else if (NumDifferences == 1) {
5651 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5652 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5653 // Make sure we do a signed comparison here.
5654 return new ICmpInst(*Context,
5655 ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5659 // Only lower this if the icmp is the only user of the GEP or if we expect
5660 // the result to fold to a constant!
5662 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5663 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5664 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5665 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5666 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5667 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), L, R);
5673 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5675 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5678 if (!isa<ConstantFP>(RHSC)) return 0;
5679 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5681 // Get the width of the mantissa. We don't want to hack on conversions that
5682 // might lose information from the integer, e.g. "i64 -> float"
5683 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5684 if (MantissaWidth == -1) return 0; // Unknown.
5686 // Check to see that the input is converted from an integer type that is small
5687 // enough that preserves all bits. TODO: check here for "known" sign bits.
5688 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5689 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5691 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5692 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5696 // If the conversion would lose info, don't hack on this.
5697 if ((int)InputSize > MantissaWidth)
5700 // Otherwise, we can potentially simplify the comparison. We know that it
5701 // will always come through as an integer value and we know the constant is
5702 // not a NAN (it would have been previously simplified).
5703 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5705 ICmpInst::Predicate Pred;
5706 switch (I.getPredicate()) {
5707 default: llvm_unreachable("Unexpected predicate!");
5708 case FCmpInst::FCMP_UEQ:
5709 case FCmpInst::FCMP_OEQ:
5710 Pred = ICmpInst::ICMP_EQ;
5712 case FCmpInst::FCMP_UGT:
5713 case FCmpInst::FCMP_OGT:
5714 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5716 case FCmpInst::FCMP_UGE:
5717 case FCmpInst::FCMP_OGE:
5718 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5720 case FCmpInst::FCMP_ULT:
5721 case FCmpInst::FCMP_OLT:
5722 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5724 case FCmpInst::FCMP_ULE:
5725 case FCmpInst::FCMP_OLE:
5726 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5728 case FCmpInst::FCMP_UNE:
5729 case FCmpInst::FCMP_ONE:
5730 Pred = ICmpInst::ICMP_NE;
5732 case FCmpInst::FCMP_ORD:
5733 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5734 case FCmpInst::FCMP_UNO:
5735 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5738 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5740 // Now we know that the APFloat is a normal number, zero or inf.
5742 // See if the FP constant is too large for the integer. For example,
5743 // comparing an i8 to 300.0.
5744 unsigned IntWidth = IntTy->getScalarSizeInBits();
5747 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5748 // and large values.
5749 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5750 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5751 APFloat::rmNearestTiesToEven);
5752 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5753 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5754 Pred == ICmpInst::ICMP_SLE)
5755 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5756 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5759 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5760 // +INF and large values.
5761 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5762 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5763 APFloat::rmNearestTiesToEven);
5764 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5765 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5766 Pred == ICmpInst::ICMP_ULE)
5767 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5768 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5773 // See if the RHS value is < SignedMin.
5774 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5775 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5776 APFloat::rmNearestTiesToEven);
5777 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5778 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5779 Pred == ICmpInst::ICMP_SGE)
5780 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5781 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5785 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5786 // [0, UMAX], but it may still be fractional. See if it is fractional by
5787 // casting the FP value to the integer value and back, checking for equality.
5788 // Don't do this for zero, because -0.0 is not fractional.
5789 Constant *RHSInt = LHSUnsigned
5790 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5791 : ConstantExpr::getFPToSI(RHSC, IntTy);
5792 if (!RHS.isZero()) {
5793 bool Equal = LHSUnsigned
5794 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5795 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5797 // If we had a comparison against a fractional value, we have to adjust
5798 // the compare predicate and sometimes the value. RHSC is rounded towards
5799 // zero at this point.
5801 default: llvm_unreachable("Unexpected integer comparison!");
5802 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5803 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5804 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5805 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5806 case ICmpInst::ICMP_ULE:
5807 // (float)int <= 4.4 --> int <= 4
5808 // (float)int <= -4.4 --> false
5809 if (RHS.isNegative())
5810 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5812 case ICmpInst::ICMP_SLE:
5813 // (float)int <= 4.4 --> int <= 4
5814 // (float)int <= -4.4 --> int < -4
5815 if (RHS.isNegative())
5816 Pred = ICmpInst::ICMP_SLT;
5818 case ICmpInst::ICMP_ULT:
5819 // (float)int < -4.4 --> false
5820 // (float)int < 4.4 --> int <= 4
5821 if (RHS.isNegative())
5822 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5823 Pred = ICmpInst::ICMP_ULE;
5825 case ICmpInst::ICMP_SLT:
5826 // (float)int < -4.4 --> int < -4
5827 // (float)int < 4.4 --> int <= 4
5828 if (!RHS.isNegative())
5829 Pred = ICmpInst::ICMP_SLE;
5831 case ICmpInst::ICMP_UGT:
5832 // (float)int > 4.4 --> int > 4
5833 // (float)int > -4.4 --> true
5834 if (RHS.isNegative())
5835 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5837 case ICmpInst::ICMP_SGT:
5838 // (float)int > 4.4 --> int > 4
5839 // (float)int > -4.4 --> int >= -4
5840 if (RHS.isNegative())
5841 Pred = ICmpInst::ICMP_SGE;
5843 case ICmpInst::ICMP_UGE:
5844 // (float)int >= -4.4 --> true
5845 // (float)int >= 4.4 --> int > 4
5846 if (!RHS.isNegative())
5847 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5848 Pred = ICmpInst::ICMP_UGT;
5850 case ICmpInst::ICMP_SGE:
5851 // (float)int >= -4.4 --> int >= -4
5852 // (float)int >= 4.4 --> int > 4
5853 if (!RHS.isNegative())
5854 Pred = ICmpInst::ICMP_SGT;
5860 // Lower this FP comparison into an appropriate integer version of the
5862 return new ICmpInst(*Context, Pred, LHSI->getOperand(0), RHSInt);
5865 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5866 bool Changed = SimplifyCompare(I);
5867 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5869 // Fold trivial predicates.
5870 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5871 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5872 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5873 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5875 // Simplify 'fcmp pred X, X'
5877 switch (I.getPredicate()) {
5878 default: llvm_unreachable("Unknown predicate!");
5879 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5880 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5881 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5882 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5883 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5884 case FCmpInst::FCMP_OLT: // True if ordered and less than
5885 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5886 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5888 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5889 case FCmpInst::FCMP_ULT: // True if unordered or less than
5890 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5891 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5892 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5893 I.setPredicate(FCmpInst::FCMP_UNO);
5894 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5897 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5898 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5899 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5900 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5901 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5902 I.setPredicate(FCmpInst::FCMP_ORD);
5903 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5908 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5909 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5911 // Handle fcmp with constant RHS
5912 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5913 // If the constant is a nan, see if we can fold the comparison based on it.
5914 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5915 if (CFP->getValueAPF().isNaN()) {
5916 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5917 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5918 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5919 "Comparison must be either ordered or unordered!");
5920 // True if unordered.
5921 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5925 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5926 switch (LHSI->getOpcode()) {
5927 case Instruction::PHI:
5928 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5929 // block. If in the same block, we're encouraging jump threading. If
5930 // not, we are just pessimizing the code by making an i1 phi.
5931 if (LHSI->getParent() == I.getParent())
5932 if (Instruction *NV = FoldOpIntoPhi(I))
5935 case Instruction::SIToFP:
5936 case Instruction::UIToFP:
5937 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5940 case Instruction::Select:
5941 // If either operand of the select is a constant, we can fold the
5942 // comparison into the select arms, which will cause one to be
5943 // constant folded and the select turned into a bitwise or.
5944 Value *Op1 = 0, *Op2 = 0;
5945 if (LHSI->hasOneUse()) {
5946 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5947 // Fold the known value into the constant operand.
5948 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5949 // Insert a new FCmp of the other select operand.
5950 Op2 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5951 LHSI->getOperand(2), RHSC,
5953 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5954 // Fold the known value into the constant operand.
5955 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5956 // Insert a new FCmp of the other select operand.
5957 Op1 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5958 LHSI->getOperand(1), RHSC,
5964 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5969 return Changed ? &I : 0;
5972 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5973 bool Changed = SimplifyCompare(I);
5974 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5975 const Type *Ty = Op0->getType();
5979 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5980 I.isTrueWhenEqual()));
5982 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5983 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5985 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5986 // addresses never equal each other! We already know that Op0 != Op1.
5987 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5988 isa<ConstantPointerNull>(Op0)) &&
5989 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5990 isa<ConstantPointerNull>(Op1)))
5991 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5992 !I.isTrueWhenEqual()));
5994 // icmp's with boolean values can always be turned into bitwise operations
5995 if (Ty == Type::Int1Ty) {
5996 switch (I.getPredicate()) {
5997 default: llvm_unreachable("Invalid icmp instruction!");
5998 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5999 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
6000 InsertNewInstBefore(Xor, I);
6001 return BinaryOperator::CreateNot(Xor);
6003 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6004 return BinaryOperator::CreateXor(Op0, Op1);
6006 case ICmpInst::ICMP_UGT:
6007 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6009 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6010 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
6011 InsertNewInstBefore(Not, I);
6012 return BinaryOperator::CreateAnd(Not, Op1);
6014 case ICmpInst::ICMP_SGT:
6015 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6017 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6018 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
6019 InsertNewInstBefore(Not, I);
6020 return BinaryOperator::CreateAnd(Not, Op0);
6022 case ICmpInst::ICMP_UGE:
6023 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6025 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6026 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
6027 InsertNewInstBefore(Not, I);
6028 return BinaryOperator::CreateOr(Not, Op1);
6030 case ICmpInst::ICMP_SGE:
6031 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6033 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6034 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
6035 InsertNewInstBefore(Not, I);
6036 return BinaryOperator::CreateOr(Not, Op0);
6041 unsigned BitWidth = 0;
6043 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6044 else if (Ty->isIntOrIntVector())
6045 BitWidth = Ty->getScalarSizeInBits();
6047 bool isSignBit = false;
6049 // See if we are doing a comparison with a constant.
6050 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6051 Value *A = 0, *B = 0;
6053 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6054 if (I.isEquality() && CI->isNullValue() &&
6055 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
6056 // (icmp cond A B) if cond is equality
6057 return new ICmpInst(*Context, I.getPredicate(), A, B);
6060 // If we have an icmp le or icmp ge instruction, turn it into the
6061 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6062 // them being folded in the code below.
6063 switch (I.getPredicate()) {
6065 case ICmpInst::ICMP_ULE:
6066 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6067 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6068 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Op0,
6070 case ICmpInst::ICMP_SLE:
6071 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6072 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6073 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6075 case ICmpInst::ICMP_UGE:
6076 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6077 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6078 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Op0,
6080 case ICmpInst::ICMP_SGE:
6081 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6082 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6083 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6087 // If this comparison is a normal comparison, it demands all
6088 // bits, if it is a sign bit comparison, it only demands the sign bit.
6090 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6093 // See if we can fold the comparison based on range information we can get
6094 // by checking whether bits are known to be zero or one in the input.
6095 if (BitWidth != 0) {
6096 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6097 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6099 if (SimplifyDemandedBits(I.getOperandUse(0),
6100 isSignBit ? APInt::getSignBit(BitWidth)
6101 : APInt::getAllOnesValue(BitWidth),
6102 Op0KnownZero, Op0KnownOne, 0))
6104 if (SimplifyDemandedBits(I.getOperandUse(1),
6105 APInt::getAllOnesValue(BitWidth),
6106 Op1KnownZero, Op1KnownOne, 0))
6109 // Given the known and unknown bits, compute a range that the LHS could be
6110 // in. Compute the Min, Max and RHS values based on the known bits. For the
6111 // EQ and NE we use unsigned values.
6112 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6113 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6114 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6115 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6117 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6120 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6122 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6126 // If Min and Max are known to be the same, then SimplifyDemandedBits
6127 // figured out that the LHS is a constant. Just constant fold this now so
6128 // that code below can assume that Min != Max.
6129 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6130 return new ICmpInst(*Context, I.getPredicate(),
6131 ConstantInt::get(*Context, Op0Min), Op1);
6132 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6133 return new ICmpInst(*Context, I.getPredicate(), Op0,
6134 ConstantInt::get(*Context, Op1Min));
6136 // Based on the range information we know about the LHS, see if we can
6137 // simplify this comparison. For example, (x&4) < 8 is always true.
6138 switch (I.getPredicate()) {
6139 default: llvm_unreachable("Unknown icmp opcode!");
6140 case ICmpInst::ICMP_EQ:
6141 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6142 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6144 case ICmpInst::ICMP_NE:
6145 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6146 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6148 case ICmpInst::ICMP_ULT:
6149 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6150 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6151 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6152 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6153 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6154 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6155 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6156 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6157 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6160 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6161 if (CI->isMinValue(true))
6162 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6163 Constant::getAllOnesValue(Op0->getType()));
6166 case ICmpInst::ICMP_UGT:
6167 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6168 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6169 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6170 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6172 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6173 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6174 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6175 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6176 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6179 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6180 if (CI->isMaxValue(true))
6181 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6182 Constant::getNullValue(Op0->getType()));
6185 case ICmpInst::ICMP_SLT:
6186 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6187 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6188 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6189 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6190 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6191 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6192 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6193 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6194 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6198 case ICmpInst::ICMP_SGT:
6199 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6200 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6201 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6202 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6204 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6205 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6206 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6207 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6208 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6212 case ICmpInst::ICMP_SGE:
6213 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6214 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6215 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6216 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6217 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6219 case ICmpInst::ICMP_SLE:
6220 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6221 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6222 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6223 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6224 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6226 case ICmpInst::ICMP_UGE:
6227 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6228 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6229 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6230 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6231 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6233 case ICmpInst::ICMP_ULE:
6234 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6235 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6236 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6237 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6238 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6242 // Turn a signed comparison into an unsigned one if both operands
6243 // are known to have the same sign.
6244 if (I.isSignedPredicate() &&
6245 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6246 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6247 return new ICmpInst(*Context, I.getUnsignedPredicate(), Op0, Op1);
6250 // Test if the ICmpInst instruction is used exclusively by a select as
6251 // part of a minimum or maximum operation. If so, refrain from doing
6252 // any other folding. This helps out other analyses which understand
6253 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6254 // and CodeGen. And in this case, at least one of the comparison
6255 // operands has at least one user besides the compare (the select),
6256 // which would often largely negate the benefit of folding anyway.
6258 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6259 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6260 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6263 // See if we are doing a comparison between a constant and an instruction that
6264 // can be folded into the comparison.
6265 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6266 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6267 // instruction, see if that instruction also has constants so that the
6268 // instruction can be folded into the icmp
6269 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6270 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6274 // Handle icmp with constant (but not simple integer constant) RHS
6275 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6276 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6277 switch (LHSI->getOpcode()) {
6278 case Instruction::GetElementPtr:
6279 if (RHSC->isNullValue()) {
6280 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6281 bool isAllZeros = true;
6282 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6283 if (!isa<Constant>(LHSI->getOperand(i)) ||
6284 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6289 return new ICmpInst(*Context, I.getPredicate(), LHSI->getOperand(0),
6290 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6294 case Instruction::PHI:
6295 // Only fold icmp into the PHI if the phi and fcmp are in the same
6296 // block. If in the same block, we're encouraging jump threading. If
6297 // not, we are just pessimizing the code by making an i1 phi.
6298 if (LHSI->getParent() == I.getParent())
6299 if (Instruction *NV = FoldOpIntoPhi(I))
6302 case Instruction::Select: {
6303 // If either operand of the select is a constant, we can fold the
6304 // comparison into the select arms, which will cause one to be
6305 // constant folded and the select turned into a bitwise or.
6306 Value *Op1 = 0, *Op2 = 0;
6307 if (LHSI->hasOneUse()) {
6308 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6309 // Fold the known value into the constant operand.
6310 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6311 // Insert a new ICmp of the other select operand.
6312 Op2 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6313 LHSI->getOperand(2), RHSC,
6315 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6316 // Fold the known value into the constant operand.
6317 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6318 // Insert a new ICmp of the other select operand.
6319 Op1 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6320 LHSI->getOperand(1), RHSC,
6326 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6329 case Instruction::Malloc:
6330 // If we have (malloc != null), and if the malloc has a single use, we
6331 // can assume it is successful and remove the malloc.
6332 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6333 AddToWorkList(LHSI);
6334 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6335 !I.isTrueWhenEqual()));
6341 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6342 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6343 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6345 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6346 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6347 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6350 // Test to see if the operands of the icmp are casted versions of other
6351 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6353 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6354 if (isa<PointerType>(Op0->getType()) &&
6355 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6356 // We keep moving the cast from the left operand over to the right
6357 // operand, where it can often be eliminated completely.
6358 Op0 = CI->getOperand(0);
6360 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6361 // so eliminate it as well.
6362 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6363 Op1 = CI2->getOperand(0);
6365 // If Op1 is a constant, we can fold the cast into the constant.
6366 if (Op0->getType() != Op1->getType()) {
6367 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6368 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6370 // Otherwise, cast the RHS right before the icmp
6371 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6374 return new ICmpInst(*Context, I.getPredicate(), Op0, Op1);
6378 if (isa<CastInst>(Op0)) {
6379 // Handle the special case of: icmp (cast bool to X), <cst>
6380 // This comes up when you have code like
6383 // For generality, we handle any zero-extension of any operand comparison
6384 // with a constant or another cast from the same type.
6385 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6386 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6390 // See if it's the same type of instruction on the left and right.
6391 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6392 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6393 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6394 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6395 switch (Op0I->getOpcode()) {
6397 case Instruction::Add:
6398 case Instruction::Sub:
6399 case Instruction::Xor:
6400 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6401 return new ICmpInst(*Context, I.getPredicate(), Op0I->getOperand(0),
6402 Op1I->getOperand(0));
6403 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6404 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6405 if (CI->getValue().isSignBit()) {
6406 ICmpInst::Predicate Pred = I.isSignedPredicate()
6407 ? I.getUnsignedPredicate()
6408 : I.getSignedPredicate();
6409 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6410 Op1I->getOperand(0));
6413 if (CI->getValue().isMaxSignedValue()) {
6414 ICmpInst::Predicate Pred = I.isSignedPredicate()
6415 ? I.getUnsignedPredicate()
6416 : I.getSignedPredicate();
6417 Pred = I.getSwappedPredicate(Pred);
6418 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6419 Op1I->getOperand(0));
6423 case Instruction::Mul:
6424 if (!I.isEquality())
6427 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6428 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6429 // Mask = -1 >> count-trailing-zeros(Cst).
6430 if (!CI->isZero() && !CI->isOne()) {
6431 const APInt &AP = CI->getValue();
6432 ConstantInt *Mask = ConstantInt::get(*Context,
6433 APInt::getLowBitsSet(AP.getBitWidth(),
6435 AP.countTrailingZeros()));
6436 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6438 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6440 InsertNewInstBefore(And1, I);
6441 InsertNewInstBefore(And2, I);
6442 return new ICmpInst(*Context, I.getPredicate(), And1, And2);
6451 // ~x < ~y --> y < x
6453 if (match(Op0, m_Not(m_Value(A))) &&
6454 match(Op1, m_Not(m_Value(B))))
6455 return new ICmpInst(*Context, I.getPredicate(), B, A);
6458 if (I.isEquality()) {
6459 Value *A, *B, *C, *D;
6461 // -x == -y --> x == y
6462 if (match(Op0, m_Neg(m_Value(A))) &&
6463 match(Op1, m_Neg(m_Value(B))))
6464 return new ICmpInst(*Context, I.getPredicate(), A, B);
6466 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
6467 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6468 Value *OtherVal = A == Op1 ? B : A;
6469 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6470 Constant::getNullValue(A->getType()));
6473 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
6474 // A^c1 == C^c2 --> A == C^(c1^c2)
6475 ConstantInt *C1, *C2;
6476 if (match(B, m_ConstantInt(C1)) &&
6477 match(D, m_ConstantInt(C2)) && Op1->hasOneUse()) {
6479 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6480 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6481 return new ICmpInst(*Context, I.getPredicate(), A,
6482 InsertNewInstBefore(Xor, I));
6485 // A^B == A^D -> B == D
6486 if (A == C) return new ICmpInst(*Context, I.getPredicate(), B, D);
6487 if (A == D) return new ICmpInst(*Context, I.getPredicate(), B, C);
6488 if (B == C) return new ICmpInst(*Context, I.getPredicate(), A, D);
6489 if (B == D) return new ICmpInst(*Context, I.getPredicate(), A, C);
6493 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
6494 (A == Op0 || B == Op0)) {
6495 // A == (A^B) -> B == 0
6496 Value *OtherVal = A == Op0 ? B : A;
6497 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6498 Constant::getNullValue(A->getType()));
6501 // (A-B) == A -> B == 0
6502 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B))))
6503 return new ICmpInst(*Context, I.getPredicate(), B,
6504 Constant::getNullValue(B->getType()));
6506 // A == (A-B) -> B == 0
6507 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B))))
6508 return new ICmpInst(*Context, I.getPredicate(), B,
6509 Constant::getNullValue(B->getType()));
6511 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6512 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6513 match(Op0, m_And(m_Value(A), m_Value(B))) &&
6514 match(Op1, m_And(m_Value(C), m_Value(D)))) {
6515 Value *X = 0, *Y = 0, *Z = 0;
6518 X = B; Y = D; Z = A;
6519 } else if (A == D) {
6520 X = B; Y = C; Z = A;
6521 } else if (B == C) {
6522 X = A; Y = D; Z = B;
6523 } else if (B == D) {
6524 X = A; Y = C; Z = B;
6527 if (X) { // Build (X^Y) & Z
6528 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6529 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6530 I.setOperand(0, Op1);
6531 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6536 return Changed ? &I : 0;
6540 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6541 /// and CmpRHS are both known to be integer constants.
6542 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6543 ConstantInt *DivRHS) {
6544 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6545 const APInt &CmpRHSV = CmpRHS->getValue();
6547 // FIXME: If the operand types don't match the type of the divide
6548 // then don't attempt this transform. The code below doesn't have the
6549 // logic to deal with a signed divide and an unsigned compare (and
6550 // vice versa). This is because (x /s C1) <s C2 produces different
6551 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6552 // (x /u C1) <u C2. Simply casting the operands and result won't
6553 // work. :( The if statement below tests that condition and bails
6555 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6556 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6558 if (DivRHS->isZero())
6559 return 0; // The ProdOV computation fails on divide by zero.
6560 if (DivIsSigned && DivRHS->isAllOnesValue())
6561 return 0; // The overflow computation also screws up here
6562 if (DivRHS->isOne())
6563 return 0; // Not worth bothering, and eliminates some funny cases
6566 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6567 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6568 // C2 (CI). By solving for X we can turn this into a range check
6569 // instead of computing a divide.
6570 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6572 // Determine if the product overflows by seeing if the product is
6573 // not equal to the divide. Make sure we do the same kind of divide
6574 // as in the LHS instruction that we're folding.
6575 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6576 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6578 // Get the ICmp opcode
6579 ICmpInst::Predicate Pred = ICI.getPredicate();
6581 // Figure out the interval that is being checked. For example, a comparison
6582 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6583 // Compute this interval based on the constants involved and the signedness of
6584 // the compare/divide. This computes a half-open interval, keeping track of
6585 // whether either value in the interval overflows. After analysis each
6586 // overflow variable is set to 0 if it's corresponding bound variable is valid
6587 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6588 int LoOverflow = 0, HiOverflow = 0;
6589 Constant *LoBound = 0, *HiBound = 0;
6591 if (!DivIsSigned) { // udiv
6592 // e.g. X/5 op 3 --> [15, 20)
6594 HiOverflow = LoOverflow = ProdOV;
6596 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6597 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6598 if (CmpRHSV == 0) { // (X / pos) op 0
6599 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6600 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
6602 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6603 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6604 HiOverflow = LoOverflow = ProdOV;
6606 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6607 } else { // (X / pos) op neg
6608 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6609 HiBound = AddOne(Prod);
6610 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6612 ConstantInt* DivNeg =
6613 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6614 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6618 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6619 if (CmpRHSV == 0) { // (X / neg) op 0
6620 // e.g. X/-5 op 0 --> [-4, 5)
6621 LoBound = AddOne(DivRHS);
6622 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6623 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6624 HiOverflow = 1; // [INTMIN+1, overflow)
6625 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6627 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6628 // e.g. X/-5 op 3 --> [-19, -14)
6629 HiBound = AddOne(Prod);
6630 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6632 LoOverflow = AddWithOverflow(LoBound, HiBound,
6633 DivRHS, Context, true) ? -1 : 0;
6634 } else { // (X / neg) op neg
6635 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6636 LoOverflow = HiOverflow = ProdOV;
6638 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6641 // Dividing by a negative swaps the condition. LT <-> GT
6642 Pred = ICmpInst::getSwappedPredicate(Pred);
6645 Value *X = DivI->getOperand(0);
6647 default: llvm_unreachable("Unhandled icmp opcode!");
6648 case ICmpInst::ICMP_EQ:
6649 if (LoOverflow && HiOverflow)
6650 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6651 else if (HiOverflow)
6652 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6653 ICmpInst::ICMP_UGE, X, LoBound);
6654 else if (LoOverflow)
6655 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6656 ICmpInst::ICMP_ULT, X, HiBound);
6658 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6659 case ICmpInst::ICMP_NE:
6660 if (LoOverflow && HiOverflow)
6661 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6662 else if (HiOverflow)
6663 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6664 ICmpInst::ICMP_ULT, X, LoBound);
6665 else if (LoOverflow)
6666 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6667 ICmpInst::ICMP_UGE, X, HiBound);
6669 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6670 case ICmpInst::ICMP_ULT:
6671 case ICmpInst::ICMP_SLT:
6672 if (LoOverflow == +1) // Low bound is greater than input range.
6673 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6674 if (LoOverflow == -1) // Low bound is less than input range.
6675 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6676 return new ICmpInst(*Context, Pred, X, LoBound);
6677 case ICmpInst::ICMP_UGT:
6678 case ICmpInst::ICMP_SGT:
6679 if (HiOverflow == +1) // High bound greater than input range.
6680 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6681 else if (HiOverflow == -1) // High bound less than input range.
6682 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6683 if (Pred == ICmpInst::ICMP_UGT)
6684 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, X, HiBound);
6686 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, X, HiBound);
6691 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6693 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6696 const APInt &RHSV = RHS->getValue();
6698 switch (LHSI->getOpcode()) {
6699 case Instruction::Trunc:
6700 if (ICI.isEquality() && LHSI->hasOneUse()) {
6701 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6702 // of the high bits truncated out of x are known.
6703 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6704 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6705 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6706 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6707 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6709 // If all the high bits are known, we can do this xform.
6710 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6711 // Pull in the high bits from known-ones set.
6712 APInt NewRHS(RHS->getValue());
6713 NewRHS.zext(SrcBits);
6715 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6716 ConstantInt::get(*Context, NewRHS));
6721 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6722 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6723 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6725 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6726 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6727 Value *CompareVal = LHSI->getOperand(0);
6729 // If the sign bit of the XorCST is not set, there is no change to
6730 // the operation, just stop using the Xor.
6731 if (!XorCST->getValue().isNegative()) {
6732 ICI.setOperand(0, CompareVal);
6733 AddToWorkList(LHSI);
6737 // Was the old condition true if the operand is positive?
6738 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6740 // If so, the new one isn't.
6741 isTrueIfPositive ^= true;
6743 if (isTrueIfPositive)
6744 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, CompareVal,
6747 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, CompareVal,
6751 if (LHSI->hasOneUse()) {
6752 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6753 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6754 const APInt &SignBit = XorCST->getValue();
6755 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6756 ? ICI.getUnsignedPredicate()
6757 : ICI.getSignedPredicate();
6758 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6759 ConstantInt::get(*Context, RHSV ^ SignBit));
6762 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6763 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6764 const APInt &NotSignBit = XorCST->getValue();
6765 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6766 ? ICI.getUnsignedPredicate()
6767 : ICI.getSignedPredicate();
6768 Pred = ICI.getSwappedPredicate(Pred);
6769 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6770 ConstantInt::get(*Context, RHSV ^ NotSignBit));
6775 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6776 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6777 LHSI->getOperand(0)->hasOneUse()) {
6778 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6780 // If the LHS is an AND of a truncating cast, we can widen the
6781 // and/compare to be the input width without changing the value
6782 // produced, eliminating a cast.
6783 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6784 // We can do this transformation if either the AND constant does not
6785 // have its sign bit set or if it is an equality comparison.
6786 // Extending a relational comparison when we're checking the sign
6787 // bit would not work.
6788 if (Cast->hasOneUse() &&
6789 (ICI.isEquality() ||
6790 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6792 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6793 APInt NewCST = AndCST->getValue();
6794 NewCST.zext(BitWidth);
6796 NewCI.zext(BitWidth);
6797 Instruction *NewAnd =
6798 BinaryOperator::CreateAnd(Cast->getOperand(0),
6799 ConstantInt::get(*Context, NewCST), LHSI->getName());
6800 InsertNewInstBefore(NewAnd, ICI);
6801 return new ICmpInst(*Context, ICI.getPredicate(), NewAnd,
6802 ConstantInt::get(*Context, NewCI));
6806 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6807 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6808 // happens a LOT in code produced by the C front-end, for bitfield
6810 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6811 if (Shift && !Shift->isShift())
6815 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6816 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6817 const Type *AndTy = AndCST->getType(); // Type of the and.
6819 // We can fold this as long as we can't shift unknown bits
6820 // into the mask. This can only happen with signed shift
6821 // rights, as they sign-extend.
6823 bool CanFold = Shift->isLogicalShift();
6825 // To test for the bad case of the signed shr, see if any
6826 // of the bits shifted in could be tested after the mask.
6827 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6828 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6830 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6831 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6832 AndCST->getValue()) == 0)
6838 if (Shift->getOpcode() == Instruction::Shl)
6839 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6841 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6843 // Check to see if we are shifting out any of the bits being
6845 if (ConstantExpr::get(Shift->getOpcode(),
6846 NewCst, ShAmt) != RHS) {
6847 // If we shifted bits out, the fold is not going to work out.
6848 // As a special case, check to see if this means that the
6849 // result is always true or false now.
6850 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6851 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6852 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6853 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6855 ICI.setOperand(1, NewCst);
6856 Constant *NewAndCST;
6857 if (Shift->getOpcode() == Instruction::Shl)
6858 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6860 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6861 LHSI->setOperand(1, NewAndCST);
6862 LHSI->setOperand(0, Shift->getOperand(0));
6863 AddToWorkList(Shift); // Shift is dead.
6864 AddUsesToWorkList(ICI);
6870 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6871 // preferable because it allows the C<<Y expression to be hoisted out
6872 // of a loop if Y is invariant and X is not.
6873 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6874 ICI.isEquality() && !Shift->isArithmeticShift() &&
6875 !isa<Constant>(Shift->getOperand(0))) {
6878 if (Shift->getOpcode() == Instruction::LShr) {
6879 NS = BinaryOperator::CreateShl(AndCST,
6880 Shift->getOperand(1), "tmp");
6882 // Insert a logical shift.
6883 NS = BinaryOperator::CreateLShr(AndCST,
6884 Shift->getOperand(1), "tmp");
6886 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6888 // Compute X & (C << Y).
6889 Instruction *NewAnd =
6890 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6891 InsertNewInstBefore(NewAnd, ICI);
6893 ICI.setOperand(0, NewAnd);
6899 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6900 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6903 uint32_t TypeBits = RHSV.getBitWidth();
6905 // Check that the shift amount is in range. If not, don't perform
6906 // undefined shifts. When the shift is visited it will be
6908 if (ShAmt->uge(TypeBits))
6911 if (ICI.isEquality()) {
6912 // If we are comparing against bits always shifted out, the
6913 // comparison cannot succeed.
6915 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
6917 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6918 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6919 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6920 return ReplaceInstUsesWith(ICI, Cst);
6923 if (LHSI->hasOneUse()) {
6924 // Otherwise strength reduce the shift into an and.
6925 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6927 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
6928 TypeBits-ShAmtVal));
6931 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6932 Mask, LHSI->getName()+".mask");
6933 Value *And = InsertNewInstBefore(AndI, ICI);
6934 return new ICmpInst(*Context, ICI.getPredicate(), And,
6935 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
6939 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6940 bool TrueIfSigned = false;
6941 if (LHSI->hasOneUse() &&
6942 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6943 // (X << 31) <s 0 --> (X&1) != 0
6944 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
6945 (TypeBits-ShAmt->getZExtValue()-1));
6947 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6948 Mask, LHSI->getName()+".mask");
6949 Value *And = InsertNewInstBefore(AndI, ICI);
6951 return new ICmpInst(*Context,
6952 TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6953 And, Constant::getNullValue(And->getType()));
6958 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6959 case Instruction::AShr: {
6960 // Only handle equality comparisons of shift-by-constant.
6961 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6962 if (!ShAmt || !ICI.isEquality()) break;
6964 // Check that the shift amount is in range. If not, don't perform
6965 // undefined shifts. When the shift is visited it will be
6967 uint32_t TypeBits = RHSV.getBitWidth();
6968 if (ShAmt->uge(TypeBits))
6971 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6973 // If we are comparing against bits always shifted out, the
6974 // comparison cannot succeed.
6975 APInt Comp = RHSV << ShAmtVal;
6976 if (LHSI->getOpcode() == Instruction::LShr)
6977 Comp = Comp.lshr(ShAmtVal);
6979 Comp = Comp.ashr(ShAmtVal);
6981 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6982 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6983 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6984 return ReplaceInstUsesWith(ICI, Cst);
6987 // Otherwise, check to see if the bits shifted out are known to be zero.
6988 // If so, we can compare against the unshifted value:
6989 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6990 if (LHSI->hasOneUse() &&
6991 MaskedValueIsZero(LHSI->getOperand(0),
6992 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6993 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6994 ConstantExpr::getShl(RHS, ShAmt));
6997 if (LHSI->hasOneUse()) {
6998 // Otherwise strength reduce the shift into an and.
6999 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7000 Constant *Mask = ConstantInt::get(*Context, Val);
7003 BinaryOperator::CreateAnd(LHSI->getOperand(0),
7004 Mask, LHSI->getName()+".mask");
7005 Value *And = InsertNewInstBefore(AndI, ICI);
7006 return new ICmpInst(*Context, ICI.getPredicate(), And,
7007 ConstantExpr::getShl(RHS, ShAmt));
7012 case Instruction::SDiv:
7013 case Instruction::UDiv:
7014 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7015 // Fold this div into the comparison, producing a range check.
7016 // Determine, based on the divide type, what the range is being
7017 // checked. If there is an overflow on the low or high side, remember
7018 // it, otherwise compute the range [low, hi) bounding the new value.
7019 // See: InsertRangeTest above for the kinds of replacements possible.
7020 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7021 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7026 case Instruction::Add:
7027 // Fold: icmp pred (add, X, C1), C2
7029 if (!ICI.isEquality()) {
7030 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7032 const APInt &LHSV = LHSC->getValue();
7034 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7037 if (ICI.isSignedPredicate()) {
7038 if (CR.getLower().isSignBit()) {
7039 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7040 ConstantInt::get(*Context, CR.getUpper()));
7041 } else if (CR.getUpper().isSignBit()) {
7042 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7043 ConstantInt::get(*Context, CR.getLower()));
7046 if (CR.getLower().isMinValue()) {
7047 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7048 ConstantInt::get(*Context, CR.getUpper()));
7049 } else if (CR.getUpper().isMinValue()) {
7050 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7051 ConstantInt::get(*Context, CR.getLower()));
7058 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7059 if (ICI.isEquality()) {
7060 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7062 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7063 // the second operand is a constant, simplify a bit.
7064 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7065 switch (BO->getOpcode()) {
7066 case Instruction::SRem:
7067 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7068 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7069 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7070 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7071 Instruction *NewRem =
7072 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
7074 InsertNewInstBefore(NewRem, ICI);
7075 return new ICmpInst(*Context, ICI.getPredicate(), NewRem,
7076 Constant::getNullValue(BO->getType()));
7080 case Instruction::Add:
7081 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7082 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7083 if (BO->hasOneUse())
7084 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7085 ConstantExpr::getSub(RHS, BOp1C));
7086 } else if (RHSV == 0) {
7087 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7088 // efficiently invertible, or if the add has just this one use.
7089 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7091 if (Value *NegVal = dyn_castNegVal(BOp1))
7092 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, NegVal);
7093 else if (Value *NegVal = dyn_castNegVal(BOp0))
7094 return new ICmpInst(*Context, ICI.getPredicate(), NegVal, BOp1);
7095 else if (BO->hasOneUse()) {
7096 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
7097 InsertNewInstBefore(Neg, ICI);
7099 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, Neg);
7103 case Instruction::Xor:
7104 // For the xor case, we can xor two constants together, eliminating
7105 // the explicit xor.
7106 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7107 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7108 ConstantExpr::getXor(RHS, BOC));
7111 case Instruction::Sub:
7112 // Replace (([sub|xor] A, B) != 0) with (A != B)
7114 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7118 case Instruction::Or:
7119 // If bits are being or'd in that are not present in the constant we
7120 // are comparing against, then the comparison could never succeed!
7121 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7122 Constant *NotCI = ConstantExpr::getNot(RHS);
7123 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7124 return ReplaceInstUsesWith(ICI,
7125 ConstantInt::get(Type::Int1Ty,
7130 case Instruction::And:
7131 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7132 // If bits are being compared against that are and'd out, then the
7133 // comparison can never succeed!
7134 if ((RHSV & ~BOC->getValue()) != 0)
7135 return ReplaceInstUsesWith(ICI,
7136 ConstantInt::get(Type::Int1Ty,
7139 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7140 if (RHS == BOC && RHSV.isPowerOf2())
7141 return new ICmpInst(*Context, isICMP_NE ? ICmpInst::ICMP_EQ :
7142 ICmpInst::ICMP_NE, LHSI,
7143 Constant::getNullValue(RHS->getType()));
7145 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7146 if (BOC->getValue().isSignBit()) {
7147 Value *X = BO->getOperand(0);
7148 Constant *Zero = Constant::getNullValue(X->getType());
7149 ICmpInst::Predicate pred = isICMP_NE ?
7150 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7151 return new ICmpInst(*Context, pred, X, Zero);
7154 // ((X & ~7) == 0) --> X < 8
7155 if (RHSV == 0 && isHighOnes(BOC)) {
7156 Value *X = BO->getOperand(0);
7157 Constant *NegX = ConstantExpr::getNeg(BOC);
7158 ICmpInst::Predicate pred = isICMP_NE ?
7159 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7160 return new ICmpInst(*Context, pred, X, NegX);
7165 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7166 // Handle icmp {eq|ne} <intrinsic>, intcst.
7167 if (II->getIntrinsicID() == Intrinsic::bswap) {
7169 ICI.setOperand(0, II->getOperand(1));
7170 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7178 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7179 /// We only handle extending casts so far.
7181 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7182 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7183 Value *LHSCIOp = LHSCI->getOperand(0);
7184 const Type *SrcTy = LHSCIOp->getType();
7185 const Type *DestTy = LHSCI->getType();
7188 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7189 // integer type is the same size as the pointer type.
7190 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7191 TD->getPointerSizeInBits() ==
7192 cast<IntegerType>(DestTy)->getBitWidth()) {
7194 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7195 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7196 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7197 RHSOp = RHSC->getOperand(0);
7198 // If the pointer types don't match, insert a bitcast.
7199 if (LHSCIOp->getType() != RHSOp->getType())
7200 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
7204 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSOp);
7207 // The code below only handles extension cast instructions, so far.
7209 if (LHSCI->getOpcode() != Instruction::ZExt &&
7210 LHSCI->getOpcode() != Instruction::SExt)
7213 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7214 bool isSignedCmp = ICI.isSignedPredicate();
7216 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7217 // Not an extension from the same type?
7218 RHSCIOp = CI->getOperand(0);
7219 if (RHSCIOp->getType() != LHSCIOp->getType())
7222 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7223 // and the other is a zext), then we can't handle this.
7224 if (CI->getOpcode() != LHSCI->getOpcode())
7227 // Deal with equality cases early.
7228 if (ICI.isEquality())
7229 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7231 // A signed comparison of sign extended values simplifies into a
7232 // signed comparison.
7233 if (isSignedCmp && isSignedExt)
7234 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7236 // The other three cases all fold into an unsigned comparison.
7237 return new ICmpInst(*Context, ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7240 // If we aren't dealing with a constant on the RHS, exit early
7241 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7245 // Compute the constant that would happen if we truncated to SrcTy then
7246 // reextended to DestTy.
7247 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7248 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7251 // If the re-extended constant didn't change...
7253 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7254 // For example, we might have:
7255 // %A = sext i16 %X to i32
7256 // %B = icmp ugt i32 %A, 1330
7257 // It is incorrect to transform this into
7258 // %B = icmp ugt i16 %X, 1330
7259 // because %A may have negative value.
7261 // However, we allow this when the compare is EQ/NE, because they are
7263 if (isSignedExt == isSignedCmp || ICI.isEquality())
7264 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, Res1);
7268 // The re-extended constant changed so the constant cannot be represented
7269 // in the shorter type. Consequently, we cannot emit a simple comparison.
7271 // First, handle some easy cases. We know the result cannot be equal at this
7272 // point so handle the ICI.isEquality() cases
7273 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7274 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7275 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7276 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7278 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7279 // should have been folded away previously and not enter in here.
7282 // We're performing a signed comparison.
7283 if (cast<ConstantInt>(CI)->getValue().isNegative())
7284 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7286 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7288 // We're performing an unsigned comparison.
7290 // We're performing an unsigned comp with a sign extended value.
7291 // This is true if the input is >= 0. [aka >s -1]
7292 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7293 Result = InsertNewInstBefore(new ICmpInst(*Context, ICmpInst::ICMP_SGT,
7294 LHSCIOp, NegOne, ICI.getName()), ICI);
7296 // Unsigned extend & unsigned compare -> always true.
7297 Result = ConstantInt::getTrue(*Context);
7301 // Finally, return the value computed.
7302 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7303 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7304 return ReplaceInstUsesWith(ICI, Result);
7306 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7307 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7308 "ICmp should be folded!");
7309 if (Constant *CI = dyn_cast<Constant>(Result))
7310 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7311 return BinaryOperator::CreateNot(Result);
7314 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7315 return commonShiftTransforms(I);
7318 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7319 return commonShiftTransforms(I);
7322 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7323 if (Instruction *R = commonShiftTransforms(I))
7326 Value *Op0 = I.getOperand(0);
7328 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7329 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7330 if (CSI->isAllOnesValue())
7331 return ReplaceInstUsesWith(I, CSI);
7333 // See if we can turn a signed shr into an unsigned shr.
7334 if (MaskedValueIsZero(Op0,
7335 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7336 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7338 // Arithmetic shifting an all-sign-bit value is a no-op.
7339 unsigned NumSignBits = ComputeNumSignBits(Op0);
7340 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7341 return ReplaceInstUsesWith(I, Op0);
7346 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7347 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7348 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7350 // shl X, 0 == X and shr X, 0 == X
7351 // shl 0, X == 0 and shr 0, X == 0
7352 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7353 Op0 == Constant::getNullValue(Op0->getType()))
7354 return ReplaceInstUsesWith(I, Op0);
7356 if (isa<UndefValue>(Op0)) {
7357 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7358 return ReplaceInstUsesWith(I, Op0);
7359 else // undef << X -> 0, undef >>u X -> 0
7360 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7362 if (isa<UndefValue>(Op1)) {
7363 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7364 return ReplaceInstUsesWith(I, Op0);
7365 else // X << undef, X >>u undef -> 0
7366 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7369 // See if we can fold away this shift.
7370 if (SimplifyDemandedInstructionBits(I))
7373 // Try to fold constant and into select arguments.
7374 if (isa<Constant>(Op0))
7375 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7376 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7379 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7380 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7385 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7386 BinaryOperator &I) {
7387 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7389 // See if we can simplify any instructions used by the instruction whose sole
7390 // purpose is to compute bits we don't care about.
7391 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7393 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7396 if (Op1->uge(TypeBits)) {
7397 if (I.getOpcode() != Instruction::AShr)
7398 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7400 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7405 // ((X*C1) << C2) == (X * (C1 << C2))
7406 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7407 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7408 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7409 return BinaryOperator::CreateMul(BO->getOperand(0),
7410 ConstantExpr::getShl(BOOp, Op1));
7412 // Try to fold constant and into select arguments.
7413 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7414 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7416 if (isa<PHINode>(Op0))
7417 if (Instruction *NV = FoldOpIntoPhi(I))
7420 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7421 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7422 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7423 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7424 // place. Don't try to do this transformation in this case. Also, we
7425 // require that the input operand is a shift-by-constant so that we have
7426 // confidence that the shifts will get folded together. We could do this
7427 // xform in more cases, but it is unlikely to be profitable.
7428 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7429 isa<ConstantInt>(TrOp->getOperand(1))) {
7430 // Okay, we'll do this xform. Make the shift of shift.
7431 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7432 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7434 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7436 // For logical shifts, the truncation has the effect of making the high
7437 // part of the register be zeros. Emulate this by inserting an AND to
7438 // clear the top bits as needed. This 'and' will usually be zapped by
7439 // other xforms later if dead.
7440 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7441 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7442 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7444 // The mask we constructed says what the trunc would do if occurring
7445 // between the shifts. We want to know the effect *after* the second
7446 // shift. We know that it is a logical shift by a constant, so adjust the
7447 // mask as appropriate.
7448 if (I.getOpcode() == Instruction::Shl)
7449 MaskV <<= Op1->getZExtValue();
7451 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7452 MaskV = MaskV.lshr(Op1->getZExtValue());
7456 BinaryOperator::CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7458 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7460 // Return the value truncated to the interesting size.
7461 return new TruncInst(And, I.getType());
7465 if (Op0->hasOneUse()) {
7466 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7467 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7470 switch (Op0BO->getOpcode()) {
7472 case Instruction::Add:
7473 case Instruction::And:
7474 case Instruction::Or:
7475 case Instruction::Xor: {
7476 // These operators commute.
7477 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7478 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7479 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7481 Instruction *YS = BinaryOperator::CreateShl(
7482 Op0BO->getOperand(0), Op1,
7484 InsertNewInstBefore(YS, I); // (Y << C)
7486 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7487 Op0BO->getOperand(1)->getName());
7488 InsertNewInstBefore(X, I); // (X + (Y << C))
7489 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7490 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7491 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7494 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7495 Value *Op0BOOp1 = Op0BO->getOperand(1);
7496 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7498 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7499 m_ConstantInt(CC))) &&
7500 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7501 Instruction *YS = BinaryOperator::CreateShl(
7502 Op0BO->getOperand(0), Op1,
7504 InsertNewInstBefore(YS, I); // (Y << C)
7506 BinaryOperator::CreateAnd(V1,
7507 ConstantExpr::getShl(CC, Op1),
7508 V1->getName()+".mask");
7509 InsertNewInstBefore(XM, I); // X & (CC << C)
7511 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7516 case Instruction::Sub: {
7517 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7518 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7519 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7520 m_Specific(Op1)))) {
7521 Instruction *YS = BinaryOperator::CreateShl(
7522 Op0BO->getOperand(1), Op1,
7524 InsertNewInstBefore(YS, I); // (Y << C)
7526 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7527 Op0BO->getOperand(0)->getName());
7528 InsertNewInstBefore(X, I); // (X + (Y << C))
7529 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7530 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7531 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7534 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7535 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7536 match(Op0BO->getOperand(0),
7537 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7538 m_ConstantInt(CC))) && V2 == Op1 &&
7539 cast<BinaryOperator>(Op0BO->getOperand(0))
7540 ->getOperand(0)->hasOneUse()) {
7541 Instruction *YS = BinaryOperator::CreateShl(
7542 Op0BO->getOperand(1), Op1,
7544 InsertNewInstBefore(YS, I); // (Y << C)
7546 BinaryOperator::CreateAnd(V1,
7547 ConstantExpr::getShl(CC, Op1),
7548 V1->getName()+".mask");
7549 InsertNewInstBefore(XM, I); // X & (CC << C)
7551 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7559 // If the operand is an bitwise operator with a constant RHS, and the
7560 // shift is the only use, we can pull it out of the shift.
7561 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7562 bool isValid = true; // Valid only for And, Or, Xor
7563 bool highBitSet = false; // Transform if high bit of constant set?
7565 switch (Op0BO->getOpcode()) {
7566 default: isValid = false; break; // Do not perform transform!
7567 case Instruction::Add:
7568 isValid = isLeftShift;
7570 case Instruction::Or:
7571 case Instruction::Xor:
7574 case Instruction::And:
7579 // If this is a signed shift right, and the high bit is modified
7580 // by the logical operation, do not perform the transformation.
7581 // The highBitSet boolean indicates the value of the high bit of
7582 // the constant which would cause it to be modified for this
7585 if (isValid && I.getOpcode() == Instruction::AShr)
7586 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7589 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7591 Instruction *NewShift =
7592 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7593 InsertNewInstBefore(NewShift, I);
7594 NewShift->takeName(Op0BO);
7596 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7603 // Find out if this is a shift of a shift by a constant.
7604 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7605 if (ShiftOp && !ShiftOp->isShift())
7608 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7609 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7610 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7611 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7612 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7613 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7614 Value *X = ShiftOp->getOperand(0);
7616 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7618 const IntegerType *Ty = cast<IntegerType>(I.getType());
7620 // Check for (X << c1) << c2 and (X >> c1) >> c2
7621 if (I.getOpcode() == ShiftOp->getOpcode()) {
7622 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7624 if (AmtSum >= TypeBits) {
7625 if (I.getOpcode() != Instruction::AShr)
7626 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7627 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7630 return BinaryOperator::Create(I.getOpcode(), X,
7631 ConstantInt::get(Ty, AmtSum));
7632 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7633 I.getOpcode() == Instruction::AShr) {
7634 if (AmtSum >= TypeBits)
7635 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7637 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7638 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7639 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7640 I.getOpcode() == Instruction::LShr) {
7641 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7642 if (AmtSum >= TypeBits)
7643 AmtSum = TypeBits-1;
7645 Instruction *Shift =
7646 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7647 InsertNewInstBefore(Shift, I);
7649 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7650 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7653 // Okay, if we get here, one shift must be left, and the other shift must be
7654 // right. See if the amounts are equal.
7655 if (ShiftAmt1 == ShiftAmt2) {
7656 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7657 if (I.getOpcode() == Instruction::Shl) {
7658 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7659 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7661 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7662 if (I.getOpcode() == Instruction::LShr) {
7663 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7664 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7666 // We can simplify ((X << C) >>s C) into a trunc + sext.
7667 // NOTE: we could do this for any C, but that would make 'unusual' integer
7668 // types. For now, just stick to ones well-supported by the code
7670 const Type *SExtType = 0;
7671 switch (Ty->getBitWidth() - ShiftAmt1) {
7678 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7683 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7684 InsertNewInstBefore(NewTrunc, I);
7685 return new SExtInst(NewTrunc, Ty);
7687 // Otherwise, we can't handle it yet.
7688 } else if (ShiftAmt1 < ShiftAmt2) {
7689 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7691 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7692 if (I.getOpcode() == Instruction::Shl) {
7693 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7694 ShiftOp->getOpcode() == Instruction::AShr);
7695 Instruction *Shift =
7696 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7697 InsertNewInstBefore(Shift, I);
7699 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7700 return BinaryOperator::CreateAnd(Shift,
7701 ConstantInt::get(*Context, Mask));
7704 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7705 if (I.getOpcode() == Instruction::LShr) {
7706 assert(ShiftOp->getOpcode() == Instruction::Shl);
7707 Instruction *Shift =
7708 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7709 InsertNewInstBefore(Shift, I);
7711 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7712 return BinaryOperator::CreateAnd(Shift,
7713 ConstantInt::get(*Context, Mask));
7716 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7718 assert(ShiftAmt2 < ShiftAmt1);
7719 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7721 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7722 if (I.getOpcode() == Instruction::Shl) {
7723 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7724 ShiftOp->getOpcode() == Instruction::AShr);
7725 Instruction *Shift =
7726 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7727 ConstantInt::get(Ty, ShiftDiff));
7728 InsertNewInstBefore(Shift, I);
7730 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7731 return BinaryOperator::CreateAnd(Shift,
7732 ConstantInt::get(*Context, Mask));
7735 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7736 if (I.getOpcode() == Instruction::LShr) {
7737 assert(ShiftOp->getOpcode() == Instruction::Shl);
7738 Instruction *Shift =
7739 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7740 InsertNewInstBefore(Shift, I);
7742 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7743 return BinaryOperator::CreateAnd(Shift,
7744 ConstantInt::get(*Context, Mask));
7747 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7754 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7755 /// expression. If so, decompose it, returning some value X, such that Val is
7758 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7759 int &Offset, LLVMContext *Context) {
7760 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7761 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7762 Offset = CI->getZExtValue();
7764 return ConstantInt::get(Type::Int32Ty, 0);
7765 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7766 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7767 if (I->getOpcode() == Instruction::Shl) {
7768 // This is a value scaled by '1 << the shift amt'.
7769 Scale = 1U << RHS->getZExtValue();
7771 return I->getOperand(0);
7772 } else if (I->getOpcode() == Instruction::Mul) {
7773 // This value is scaled by 'RHS'.
7774 Scale = RHS->getZExtValue();
7776 return I->getOperand(0);
7777 } else if (I->getOpcode() == Instruction::Add) {
7778 // We have X+C. Check to see if we really have (X*C2)+C1,
7779 // where C1 is divisible by C2.
7782 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7784 Offset += RHS->getZExtValue();
7791 // Otherwise, we can't look past this.
7798 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7799 /// try to eliminate the cast by moving the type information into the alloc.
7800 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7801 AllocationInst &AI) {
7802 const PointerType *PTy = cast<PointerType>(CI.getType());
7804 // Remove any uses of AI that are dead.
7805 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7807 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7808 Instruction *User = cast<Instruction>(*UI++);
7809 if (isInstructionTriviallyDead(User)) {
7810 while (UI != E && *UI == User)
7811 ++UI; // If this instruction uses AI more than once, don't break UI.
7814 DOUT << "IC: DCE: " << *User << '\n';
7815 EraseInstFromFunction(*User);
7819 // This requires TargetData to get the alloca alignment and size information.
7822 // Get the type really allocated and the type casted to.
7823 const Type *AllocElTy = AI.getAllocatedType();
7824 const Type *CastElTy = PTy->getElementType();
7825 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7827 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7828 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7829 if (CastElTyAlign < AllocElTyAlign) return 0;
7831 // If the allocation has multiple uses, only promote it if we are strictly
7832 // increasing the alignment of the resultant allocation. If we keep it the
7833 // same, we open the door to infinite loops of various kinds. (A reference
7834 // from a dbg.declare doesn't count as a use for this purpose.)
7835 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7836 CastElTyAlign == AllocElTyAlign) return 0;
7838 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7839 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7840 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7842 // See if we can satisfy the modulus by pulling a scale out of the array
7844 unsigned ArraySizeScale;
7846 Value *NumElements = // See if the array size is a decomposable linear expr.
7847 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7848 ArrayOffset, Context);
7850 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7852 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7853 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7855 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7860 // If the allocation size is constant, form a constant mul expression
7861 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7862 if (isa<ConstantInt>(NumElements))
7863 Amt = ConstantExpr::getMul(cast<ConstantInt>(NumElements),
7864 cast<ConstantInt>(Amt));
7865 // otherwise multiply the amount and the number of elements
7867 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7868 Amt = InsertNewInstBefore(Tmp, AI);
7872 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7873 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7874 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7875 Amt = InsertNewInstBefore(Tmp, AI);
7878 AllocationInst *New;
7879 if (isa<MallocInst>(AI))
7880 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7882 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7883 InsertNewInstBefore(New, AI);
7886 // If the allocation has one real use plus a dbg.declare, just remove the
7888 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7889 EraseInstFromFunction(*DI);
7891 // If the allocation has multiple real uses, insert a cast and change all
7892 // things that used it to use the new cast. This will also hack on CI, but it
7894 else if (!AI.hasOneUse()) {
7895 AddUsesToWorkList(AI);
7896 // New is the allocation instruction, pointer typed. AI is the original
7897 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7898 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7899 InsertNewInstBefore(NewCast, AI);
7900 AI.replaceAllUsesWith(NewCast);
7902 return ReplaceInstUsesWith(CI, New);
7905 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7906 /// and return it as type Ty without inserting any new casts and without
7907 /// changing the computed value. This is used by code that tries to decide
7908 /// whether promoting or shrinking integer operations to wider or smaller types
7909 /// will allow us to eliminate a truncate or extend.
7911 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7912 /// extension operation if Ty is larger.
7914 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7915 /// should return true if trunc(V) can be computed by computing V in the smaller
7916 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7917 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7918 /// efficiently truncated.
7920 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7921 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7922 /// the final result.
7923 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7925 int &NumCastsRemoved){
7926 // We can always evaluate constants in another type.
7927 if (isa<Constant>(V))
7930 Instruction *I = dyn_cast<Instruction>(V);
7931 if (!I) return false;
7933 const Type *OrigTy = V->getType();
7935 // If this is an extension or truncate, we can often eliminate it.
7936 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7937 // If this is a cast from the destination type, we can trivially eliminate
7938 // it, and this will remove a cast overall.
7939 if (I->getOperand(0)->getType() == Ty) {
7940 // If the first operand is itself a cast, and is eliminable, do not count
7941 // this as an eliminable cast. We would prefer to eliminate those two
7943 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7949 // We can't extend or shrink something that has multiple uses: doing so would
7950 // require duplicating the instruction in general, which isn't profitable.
7951 if (!I->hasOneUse()) return false;
7953 unsigned Opc = I->getOpcode();
7955 case Instruction::Add:
7956 case Instruction::Sub:
7957 case Instruction::Mul:
7958 case Instruction::And:
7959 case Instruction::Or:
7960 case Instruction::Xor:
7961 // These operators can all arbitrarily be extended or truncated.
7962 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7964 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7967 case Instruction::UDiv:
7968 case Instruction::URem: {
7969 // UDiv and URem can be truncated if all the truncated bits are zero.
7970 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7971 uint32_t BitWidth = Ty->getScalarSizeInBits();
7972 if (BitWidth < OrigBitWidth) {
7973 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7974 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7975 MaskedValueIsZero(I->getOperand(1), Mask)) {
7976 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7978 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7984 case Instruction::Shl:
7985 // If we are truncating the result of this SHL, and if it's a shift of a
7986 // constant amount, we can always perform a SHL in a smaller type.
7987 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7988 uint32_t BitWidth = Ty->getScalarSizeInBits();
7989 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7990 CI->getLimitedValue(BitWidth) < BitWidth)
7991 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7995 case Instruction::LShr:
7996 // If this is a truncate of a logical shr, we can truncate it to a smaller
7997 // lshr iff we know that the bits we would otherwise be shifting in are
7999 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8000 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8001 uint32_t BitWidth = Ty->getScalarSizeInBits();
8002 if (BitWidth < OrigBitWidth &&
8003 MaskedValueIsZero(I->getOperand(0),
8004 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
8005 CI->getLimitedValue(BitWidth) < BitWidth) {
8006 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8011 case Instruction::ZExt:
8012 case Instruction::SExt:
8013 case Instruction::Trunc:
8014 // If this is the same kind of case as our original (e.g. zext+zext), we
8015 // can safely replace it. Note that replacing it does not reduce the number
8016 // of casts in the input.
8020 // sext (zext ty1), ty2 -> zext ty2
8021 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8024 case Instruction::Select: {
8025 SelectInst *SI = cast<SelectInst>(I);
8026 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8028 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8031 case Instruction::PHI: {
8032 // We can change a phi if we can change all operands.
8033 PHINode *PN = cast<PHINode>(I);
8034 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8035 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8041 // TODO: Can handle more cases here.
8048 /// EvaluateInDifferentType - Given an expression that
8049 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8050 /// evaluate the expression.
8051 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8053 if (Constant *C = dyn_cast<Constant>(V))
8054 return ConstantExpr::getIntegerCast(C, Ty,
8055 isSigned /*Sext or ZExt*/);
8057 // Otherwise, it must be an instruction.
8058 Instruction *I = cast<Instruction>(V);
8059 Instruction *Res = 0;
8060 unsigned Opc = I->getOpcode();
8062 case Instruction::Add:
8063 case Instruction::Sub:
8064 case Instruction::Mul:
8065 case Instruction::And:
8066 case Instruction::Or:
8067 case Instruction::Xor:
8068 case Instruction::AShr:
8069 case Instruction::LShr:
8070 case Instruction::Shl:
8071 case Instruction::UDiv:
8072 case Instruction::URem: {
8073 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8074 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8075 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8078 case Instruction::Trunc:
8079 case Instruction::ZExt:
8080 case Instruction::SExt:
8081 // If the source type of the cast is the type we're trying for then we can
8082 // just return the source. There's no need to insert it because it is not
8084 if (I->getOperand(0)->getType() == Ty)
8085 return I->getOperand(0);
8087 // Otherwise, must be the same type of cast, so just reinsert a new one.
8088 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8091 case Instruction::Select: {
8092 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8093 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8094 Res = SelectInst::Create(I->getOperand(0), True, False);
8097 case Instruction::PHI: {
8098 PHINode *OPN = cast<PHINode>(I);
8099 PHINode *NPN = PHINode::Create(Ty);
8100 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8101 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8102 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8108 // TODO: Can handle more cases here.
8109 llvm_unreachable("Unreachable!");
8114 return InsertNewInstBefore(Res, *I);
8117 /// @brief Implement the transforms common to all CastInst visitors.
8118 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8119 Value *Src = CI.getOperand(0);
8121 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8122 // eliminate it now.
8123 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8124 if (Instruction::CastOps opc =
8125 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8126 // The first cast (CSrc) is eliminable so we need to fix up or replace
8127 // the second cast (CI). CSrc will then have a good chance of being dead.
8128 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8132 // If we are casting a select then fold the cast into the select
8133 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8134 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8137 // If we are casting a PHI then fold the cast into the PHI
8138 if (isa<PHINode>(Src))
8139 if (Instruction *NV = FoldOpIntoPhi(CI))
8145 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8146 /// or not there is a sequence of GEP indices into the type that will land us at
8147 /// the specified offset. If so, fill them into NewIndices and return the
8148 /// resultant element type, otherwise return null.
8149 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8150 SmallVectorImpl<Value*> &NewIndices,
8151 const TargetData *TD,
8152 LLVMContext *Context) {
8154 if (!Ty->isSized()) return 0;
8156 // Start with the index over the outer type. Note that the type size
8157 // might be zero (even if the offset isn't zero) if the indexed type
8158 // is something like [0 x {int, int}]
8159 const Type *IntPtrTy = TD->getIntPtrType();
8160 int64_t FirstIdx = 0;
8161 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8162 FirstIdx = Offset/TySize;
8163 Offset -= FirstIdx*TySize;
8165 // Handle hosts where % returns negative instead of values [0..TySize).
8169 assert(Offset >= 0);
8171 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8174 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8176 // Index into the types. If we fail, set OrigBase to null.
8178 // Indexing into tail padding between struct/array elements.
8179 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8182 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8183 const StructLayout *SL = TD->getStructLayout(STy);
8184 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8185 "Offset must stay within the indexed type");
8187 unsigned Elt = SL->getElementContainingOffset(Offset);
8188 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
8190 Offset -= SL->getElementOffset(Elt);
8191 Ty = STy->getElementType(Elt);
8192 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8193 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8194 assert(EltSize && "Cannot index into a zero-sized array");
8195 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8197 Ty = AT->getElementType();
8199 // Otherwise, we can't index into the middle of this atomic type, bail.
8207 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8208 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8209 Value *Src = CI.getOperand(0);
8211 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8212 // If casting the result of a getelementptr instruction with no offset, turn
8213 // this into a cast of the original pointer!
8214 if (GEP->hasAllZeroIndices()) {
8215 // Changing the cast operand is usually not a good idea but it is safe
8216 // here because the pointer operand is being replaced with another
8217 // pointer operand so the opcode doesn't need to change.
8219 CI.setOperand(0, GEP->getOperand(0));
8223 // If the GEP has a single use, and the base pointer is a bitcast, and the
8224 // GEP computes a constant offset, see if we can convert these three
8225 // instructions into fewer. This typically happens with unions and other
8226 // non-type-safe code.
8227 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8228 if (GEP->hasAllConstantIndices()) {
8229 // We are guaranteed to get a constant from EmitGEPOffset.
8230 ConstantInt *OffsetV =
8231 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8232 int64_t Offset = OffsetV->getSExtValue();
8234 // Get the base pointer input of the bitcast, and the type it points to.
8235 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8236 const Type *GEPIdxTy =
8237 cast<PointerType>(OrigBase->getType())->getElementType();
8238 SmallVector<Value*, 8> NewIndices;
8239 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8240 // If we were able to index down into an element, create the GEP
8241 // and bitcast the result. This eliminates one bitcast, potentially
8243 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
8245 NewIndices.end(), "");
8246 InsertNewInstBefore(NGEP, CI);
8247 NGEP->takeName(GEP);
8248 if (cast<GEPOperator>(GEP)->isInBounds())
8249 cast<GEPOperator>(NGEP)->setIsInBounds(true);
8251 if (isa<BitCastInst>(CI))
8252 return new BitCastInst(NGEP, CI.getType());
8253 assert(isa<PtrToIntInst>(CI));
8254 return new PtrToIntInst(NGEP, CI.getType());
8260 return commonCastTransforms(CI);
8263 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8264 /// type like i42. We don't want to introduce operations on random non-legal
8265 /// integer types where they don't already exist in the code. In the future,
8266 /// we should consider making this based off target-data, so that 32-bit targets
8267 /// won't get i64 operations etc.
8268 static bool isSafeIntegerType(const Type *Ty) {
8269 switch (Ty->getPrimitiveSizeInBits()) {
8280 /// commonIntCastTransforms - This function implements the common transforms
8281 /// for trunc, zext, and sext.
8282 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8283 if (Instruction *Result = commonCastTransforms(CI))
8286 Value *Src = CI.getOperand(0);
8287 const Type *SrcTy = Src->getType();
8288 const Type *DestTy = CI.getType();
8289 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8290 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8292 // See if we can simplify any instructions used by the LHS whose sole
8293 // purpose is to compute bits we don't care about.
8294 if (SimplifyDemandedInstructionBits(CI))
8297 // If the source isn't an instruction or has more than one use then we
8298 // can't do anything more.
8299 Instruction *SrcI = dyn_cast<Instruction>(Src);
8300 if (!SrcI || !Src->hasOneUse())
8303 // Attempt to propagate the cast into the instruction for int->int casts.
8304 int NumCastsRemoved = 0;
8305 // Only do this if the dest type is a simple type, don't convert the
8306 // expression tree to something weird like i93 unless the source is also
8308 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8309 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8310 CanEvaluateInDifferentType(SrcI, DestTy,
8311 CI.getOpcode(), NumCastsRemoved)) {
8312 // If this cast is a truncate, evaluting in a different type always
8313 // eliminates the cast, so it is always a win. If this is a zero-extension,
8314 // we need to do an AND to maintain the clear top-part of the computation,
8315 // so we require that the input have eliminated at least one cast. If this
8316 // is a sign extension, we insert two new casts (to do the extension) so we
8317 // require that two casts have been eliminated.
8318 bool DoXForm = false;
8319 bool JustReplace = false;
8320 switch (CI.getOpcode()) {
8322 // All the others use floating point so we shouldn't actually
8323 // get here because of the check above.
8324 llvm_unreachable("Unknown cast type");
8325 case Instruction::Trunc:
8328 case Instruction::ZExt: {
8329 DoXForm = NumCastsRemoved >= 1;
8330 if (!DoXForm && 0) {
8331 // If it's unnecessary to issue an AND to clear the high bits, it's
8332 // always profitable to do this xform.
8333 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8334 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8335 if (MaskedValueIsZero(TryRes, Mask))
8336 return ReplaceInstUsesWith(CI, TryRes);
8338 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8339 if (TryI->use_empty())
8340 EraseInstFromFunction(*TryI);
8344 case Instruction::SExt: {
8345 DoXForm = NumCastsRemoved >= 2;
8346 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8347 // If we do not have to emit the truncate + sext pair, then it's always
8348 // profitable to do this xform.
8350 // It's not safe to eliminate the trunc + sext pair if one of the
8351 // eliminated cast is a truncate. e.g.
8352 // t2 = trunc i32 t1 to i16
8353 // t3 = sext i16 t2 to i32
8356 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8357 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8358 if (NumSignBits > (DestBitSize - SrcBitSize))
8359 return ReplaceInstUsesWith(CI, TryRes);
8361 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8362 if (TryI->use_empty())
8363 EraseInstFromFunction(*TryI);
8370 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
8372 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8373 CI.getOpcode() == Instruction::SExt);
8375 // Just replace this cast with the result.
8376 return ReplaceInstUsesWith(CI, Res);
8378 assert(Res->getType() == DestTy);
8379 switch (CI.getOpcode()) {
8380 default: llvm_unreachable("Unknown cast type!");
8381 case Instruction::Trunc:
8382 // Just replace this cast with the result.
8383 return ReplaceInstUsesWith(CI, Res);
8384 case Instruction::ZExt: {
8385 assert(SrcBitSize < DestBitSize && "Not a zext?");
8387 // If the high bits are already zero, just replace this cast with the
8389 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8390 if (MaskedValueIsZero(Res, Mask))
8391 return ReplaceInstUsesWith(CI, Res);
8393 // We need to emit an AND to clear the high bits.
8394 Constant *C = ConstantInt::get(*Context,
8395 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8396 return BinaryOperator::CreateAnd(Res, C);
8398 case Instruction::SExt: {
8399 // If the high bits are already filled with sign bit, just replace this
8400 // cast with the result.
8401 unsigned NumSignBits = ComputeNumSignBits(Res);
8402 if (NumSignBits > (DestBitSize - SrcBitSize))
8403 return ReplaceInstUsesWith(CI, Res);
8405 // We need to emit a cast to truncate, then a cast to sext.
8406 return CastInst::Create(Instruction::SExt,
8407 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8414 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8415 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8417 switch (SrcI->getOpcode()) {
8418 case Instruction::Add:
8419 case Instruction::Mul:
8420 case Instruction::And:
8421 case Instruction::Or:
8422 case Instruction::Xor:
8423 // If we are discarding information, rewrite.
8424 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8425 // Don't insert two casts unless at least one can be eliminated.
8426 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8427 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8428 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8429 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8430 return BinaryOperator::Create(
8431 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8435 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8436 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8437 SrcI->getOpcode() == Instruction::Xor &&
8438 Op1 == ConstantInt::getTrue(*Context) &&
8439 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8440 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8441 return BinaryOperator::CreateXor(New,
8442 ConstantInt::get(CI.getType(), 1));
8446 case Instruction::Shl: {
8447 // Canonicalize trunc inside shl, if we can.
8448 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8449 if (CI && DestBitSize < SrcBitSize &&
8450 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8451 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8452 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8453 return BinaryOperator::CreateShl(Op0c, Op1c);
8461 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8462 if (Instruction *Result = commonIntCastTransforms(CI))
8465 Value *Src = CI.getOperand(0);
8466 const Type *Ty = CI.getType();
8467 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8468 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8470 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8471 if (DestBitWidth == 1) {
8472 Constant *One = ConstantInt::get(Src->getType(), 1);
8473 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8474 Value *Zero = Constant::getNullValue(Src->getType());
8475 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Src, Zero);
8478 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8479 ConstantInt *ShAmtV = 0;
8481 if (Src->hasOneUse() &&
8482 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)))) {
8483 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8485 // Get a mask for the bits shifting in.
8486 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8487 if (MaskedValueIsZero(ShiftOp, Mask)) {
8488 if (ShAmt >= DestBitWidth) // All zeros.
8489 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8491 // Okay, we can shrink this. Truncate the input, then return a new
8493 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8494 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8495 return BinaryOperator::CreateLShr(V1, V2);
8502 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8503 /// in order to eliminate the icmp.
8504 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8506 // If we are just checking for a icmp eq of a single bit and zext'ing it
8507 // to an integer, then shift the bit to the appropriate place and then
8508 // cast to integer to avoid the comparison.
8509 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8510 const APInt &Op1CV = Op1C->getValue();
8512 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8513 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8514 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8515 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8516 if (!DoXform) return ICI;
8518 Value *In = ICI->getOperand(0);
8519 Value *Sh = ConstantInt::get(In->getType(),
8520 In->getType()->getScalarSizeInBits()-1);
8521 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8522 In->getName()+".lobit"),
8524 if (In->getType() != CI.getType())
8525 In = CastInst::CreateIntegerCast(In, CI.getType(),
8526 false/*ZExt*/, "tmp", &CI);
8528 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8529 Constant *One = ConstantInt::get(In->getType(), 1);
8530 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8531 In->getName()+".not"),
8535 return ReplaceInstUsesWith(CI, In);
8540 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8541 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8542 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8543 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8544 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8545 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8546 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8547 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8548 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8549 // This only works for EQ and NE
8550 ICI->isEquality()) {
8551 // If Op1C some other power of two, convert:
8552 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8553 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8554 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8555 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8557 APInt KnownZeroMask(~KnownZero);
8558 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8559 if (!DoXform) return ICI;
8561 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8562 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8563 // (X&4) == 2 --> false
8564 // (X&4) != 2 --> true
8565 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8566 Res = ConstantExpr::getZExt(Res, CI.getType());
8567 return ReplaceInstUsesWith(CI, Res);
8570 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8571 Value *In = ICI->getOperand(0);
8573 // Perform a logical shr by shiftamt.
8574 // Insert the shift to put the result in the low bit.
8575 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8576 ConstantInt::get(In->getType(), ShiftAmt),
8577 In->getName()+".lobit"), CI);
8580 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8581 Constant *One = ConstantInt::get(In->getType(), 1);
8582 In = BinaryOperator::CreateXor(In, One, "tmp");
8583 InsertNewInstBefore(cast<Instruction>(In), CI);
8586 if (CI.getType() == In->getType())
8587 return ReplaceInstUsesWith(CI, In);
8589 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8597 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8598 // If one of the common conversion will work ..
8599 if (Instruction *Result = commonIntCastTransforms(CI))
8602 Value *Src = CI.getOperand(0);
8604 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8605 // types and if the sizes are just right we can convert this into a logical
8606 // 'and' which will be much cheaper than the pair of casts.
8607 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8608 // Get the sizes of the types involved. We know that the intermediate type
8609 // will be smaller than A or C, but don't know the relation between A and C.
8610 Value *A = CSrc->getOperand(0);
8611 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8612 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8613 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8614 // If we're actually extending zero bits, then if
8615 // SrcSize < DstSize: zext(a & mask)
8616 // SrcSize == DstSize: a & mask
8617 // SrcSize > DstSize: trunc(a) & mask
8618 if (SrcSize < DstSize) {
8619 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8620 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8622 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8623 InsertNewInstBefore(And, CI);
8624 return new ZExtInst(And, CI.getType());
8625 } else if (SrcSize == DstSize) {
8626 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8627 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8629 } else if (SrcSize > DstSize) {
8630 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8631 InsertNewInstBefore(Trunc, CI);
8632 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8633 return BinaryOperator::CreateAnd(Trunc,
8634 ConstantInt::get(Trunc->getType(),
8639 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8640 return transformZExtICmp(ICI, CI);
8642 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8643 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8644 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8645 // of the (zext icmp) will be transformed.
8646 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8647 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8648 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8649 (transformZExtICmp(LHS, CI, false) ||
8650 transformZExtICmp(RHS, CI, false))) {
8651 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8652 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8653 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8657 // zext(trunc(t) & C) -> (t & zext(C)).
8658 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8659 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8660 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8661 Value *TI0 = TI->getOperand(0);
8662 if (TI0->getType() == CI.getType())
8664 BinaryOperator::CreateAnd(TI0,
8665 ConstantExpr::getZExt(C, CI.getType()));
8668 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8669 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8670 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8671 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8672 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8673 And->getOperand(1) == C)
8674 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8675 Value *TI0 = TI->getOperand(0);
8676 if (TI0->getType() == CI.getType()) {
8677 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8678 Instruction *NewAnd = BinaryOperator::CreateAnd(TI0, ZC, "tmp");
8679 InsertNewInstBefore(NewAnd, *And);
8680 return BinaryOperator::CreateXor(NewAnd, ZC);
8687 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8688 if (Instruction *I = commonIntCastTransforms(CI))
8691 Value *Src = CI.getOperand(0);
8693 // Canonicalize sign-extend from i1 to a select.
8694 if (Src->getType() == Type::Int1Ty)
8695 return SelectInst::Create(Src,
8696 Constant::getAllOnesValue(CI.getType()),
8697 Constant::getNullValue(CI.getType()));
8699 // See if the value being truncated is already sign extended. If so, just
8700 // eliminate the trunc/sext pair.
8701 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8702 Value *Op = cast<User>(Src)->getOperand(0);
8703 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8704 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8705 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8706 unsigned NumSignBits = ComputeNumSignBits(Op);
8708 if (OpBits == DestBits) {
8709 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8710 // bits, it is already ready.
8711 if (NumSignBits > DestBits-MidBits)
8712 return ReplaceInstUsesWith(CI, Op);
8713 } else if (OpBits < DestBits) {
8714 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8715 // bits, just sext from i32.
8716 if (NumSignBits > OpBits-MidBits)
8717 return new SExtInst(Op, CI.getType(), "tmp");
8719 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8720 // bits, just truncate to i32.
8721 if (NumSignBits > OpBits-MidBits)
8722 return new TruncInst(Op, CI.getType(), "tmp");
8726 // If the input is a shl/ashr pair of a same constant, then this is a sign
8727 // extension from a smaller value. If we could trust arbitrary bitwidth
8728 // integers, we could turn this into a truncate to the smaller bit and then
8729 // use a sext for the whole extension. Since we don't, look deeper and check
8730 // for a truncate. If the source and dest are the same type, eliminate the
8731 // trunc and extend and just do shifts. For example, turn:
8732 // %a = trunc i32 %i to i8
8733 // %b = shl i8 %a, 6
8734 // %c = ashr i8 %b, 6
8735 // %d = sext i8 %c to i32
8737 // %a = shl i32 %i, 30
8738 // %d = ashr i32 %a, 30
8740 ConstantInt *BA = 0, *CA = 0;
8741 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8742 m_ConstantInt(CA))) &&
8743 BA == CA && isa<TruncInst>(A)) {
8744 Value *I = cast<TruncInst>(A)->getOperand(0);
8745 if (I->getType() == CI.getType()) {
8746 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8747 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8748 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8749 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8750 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8752 return BinaryOperator::CreateAShr(I, ShAmtV);
8759 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8760 /// in the specified FP type without changing its value.
8761 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8762 LLVMContext *Context) {
8764 APFloat F = CFP->getValueAPF();
8765 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8767 return ConstantFP::get(*Context, F);
8771 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8772 /// through it until we get the source value.
8773 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8774 if (Instruction *I = dyn_cast<Instruction>(V))
8775 if (I->getOpcode() == Instruction::FPExt)
8776 return LookThroughFPExtensions(I->getOperand(0), Context);
8778 // If this value is a constant, return the constant in the smallest FP type
8779 // that can accurately represent it. This allows us to turn
8780 // (float)((double)X+2.0) into x+2.0f.
8781 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8782 if (CFP->getType() == Type::PPC_FP128Ty)
8783 return V; // No constant folding of this.
8784 // See if the value can be truncated to float and then reextended.
8785 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8787 if (CFP->getType() == Type::DoubleTy)
8788 return V; // Won't shrink.
8789 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8791 // Don't try to shrink to various long double types.
8797 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8798 if (Instruction *I = commonCastTransforms(CI))
8801 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8802 // smaller than the destination type, we can eliminate the truncate by doing
8803 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8804 // many builtins (sqrt, etc).
8805 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8806 if (OpI && OpI->hasOneUse()) {
8807 switch (OpI->getOpcode()) {
8809 case Instruction::FAdd:
8810 case Instruction::FSub:
8811 case Instruction::FMul:
8812 case Instruction::FDiv:
8813 case Instruction::FRem:
8814 const Type *SrcTy = OpI->getType();
8815 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8816 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8817 if (LHSTrunc->getType() != SrcTy &&
8818 RHSTrunc->getType() != SrcTy) {
8819 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8820 // If the source types were both smaller than the destination type of
8821 // the cast, do this xform.
8822 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8823 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8824 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8826 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8828 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8837 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8838 return commonCastTransforms(CI);
8841 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8842 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8844 return commonCastTransforms(FI);
8846 // fptoui(uitofp(X)) --> X
8847 // fptoui(sitofp(X)) --> X
8848 // This is safe if the intermediate type has enough bits in its mantissa to
8849 // accurately represent all values of X. For example, do not do this with
8850 // i64->float->i64. This is also safe for sitofp case, because any negative
8851 // 'X' value would cause an undefined result for the fptoui.
8852 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8853 OpI->getOperand(0)->getType() == FI.getType() &&
8854 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8855 OpI->getType()->getFPMantissaWidth())
8856 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8858 return commonCastTransforms(FI);
8861 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8862 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8864 return commonCastTransforms(FI);
8866 // fptosi(sitofp(X)) --> X
8867 // fptosi(uitofp(X)) --> X
8868 // This is safe if the intermediate type has enough bits in its mantissa to
8869 // accurately represent all values of X. For example, do not do this with
8870 // i64->float->i64. This is also safe for sitofp case, because any negative
8871 // 'X' value would cause an undefined result for the fptoui.
8872 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8873 OpI->getOperand(0)->getType() == FI.getType() &&
8874 (int)FI.getType()->getScalarSizeInBits() <=
8875 OpI->getType()->getFPMantissaWidth())
8876 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8878 return commonCastTransforms(FI);
8881 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8882 return commonCastTransforms(CI);
8885 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8886 return commonCastTransforms(CI);
8889 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8890 // If the destination integer type is smaller than the intptr_t type for
8891 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8892 // trunc to be exposed to other transforms. Don't do this for extending
8893 // ptrtoint's, because we don't know if the target sign or zero extends its
8896 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8897 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8898 TD->getIntPtrType(),
8900 return new TruncInst(P, CI.getType());
8903 return commonPointerCastTransforms(CI);
8906 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8907 // If the source integer type is larger than the intptr_t type for
8908 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8909 // allows the trunc to be exposed to other transforms. Don't do this for
8910 // extending inttoptr's, because we don't know if the target sign or zero
8911 // extends to pointers.
8913 CI.getOperand(0)->getType()->getScalarSizeInBits() >
8914 TD->getPointerSizeInBits()) {
8915 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8916 TD->getIntPtrType(),
8918 return new IntToPtrInst(P, CI.getType());
8921 if (Instruction *I = commonCastTransforms(CI))
8927 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8928 // If the operands are integer typed then apply the integer transforms,
8929 // otherwise just apply the common ones.
8930 Value *Src = CI.getOperand(0);
8931 const Type *SrcTy = Src->getType();
8932 const Type *DestTy = CI.getType();
8934 if (isa<PointerType>(SrcTy)) {
8935 if (Instruction *I = commonPointerCastTransforms(CI))
8938 if (Instruction *Result = commonCastTransforms(CI))
8943 // Get rid of casts from one type to the same type. These are useless and can
8944 // be replaced by the operand.
8945 if (DestTy == Src->getType())
8946 return ReplaceInstUsesWith(CI, Src);
8948 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8949 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8950 const Type *DstElTy = DstPTy->getElementType();
8951 const Type *SrcElTy = SrcPTy->getElementType();
8953 // If the address spaces don't match, don't eliminate the bitcast, which is
8954 // required for changing types.
8955 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8958 // If we are casting a malloc or alloca to a pointer to a type of the same
8959 // size, rewrite the allocation instruction to allocate the "right" type.
8960 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8961 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8964 // If the source and destination are pointers, and this cast is equivalent
8965 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8966 // This can enhance SROA and other transforms that want type-safe pointers.
8967 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8968 unsigned NumZeros = 0;
8969 while (SrcElTy != DstElTy &&
8970 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8971 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8972 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8976 // If we found a path from the src to dest, create the getelementptr now.
8977 if (SrcElTy == DstElTy) {
8978 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8979 Instruction *GEP = GetElementPtrInst::Create(Src,
8980 Idxs.begin(), Idxs.end(), "",
8981 ((Instruction*) NULL));
8982 cast<GEPOperator>(GEP)->setIsInBounds(true);
8987 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
8988 if (DestVTy->getNumElements() == 1) {
8989 if (!isa<VectorType>(SrcTy)) {
8990 Value *Elem = InsertCastBefore(Instruction::BitCast, Src,
8991 DestVTy->getElementType(), CI);
8992 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
8993 Constant::getNullValue(Type::Int32Ty));
8995 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
8999 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
9000 if (SrcVTy->getNumElements() == 1) {
9001 if (!isa<VectorType>(DestTy)) {
9003 ExtractElementInst::Create(Src, Constant::getNullValue(Type::Int32Ty));
9004 InsertNewInstBefore(Elem, CI);
9005 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
9010 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9011 if (SVI->hasOneUse()) {
9012 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9013 // a bitconvert to a vector with the same # elts.
9014 if (isa<VectorType>(DestTy) &&
9015 cast<VectorType>(DestTy)->getNumElements() ==
9016 SVI->getType()->getNumElements() &&
9017 SVI->getType()->getNumElements() ==
9018 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9020 // If either of the operands is a cast from CI.getType(), then
9021 // evaluating the shuffle in the casted destination's type will allow
9022 // us to eliminate at least one cast.
9023 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9024 Tmp->getOperand(0)->getType() == DestTy) ||
9025 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9026 Tmp->getOperand(0)->getType() == DestTy)) {
9027 Value *LHS = InsertCastBefore(Instruction::BitCast,
9028 SVI->getOperand(0), DestTy, CI);
9029 Value *RHS = InsertCastBefore(Instruction::BitCast,
9030 SVI->getOperand(1), DestTy, CI);
9031 // Return a new shuffle vector. Use the same element ID's, as we
9032 // know the vector types match #elts.
9033 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9041 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9043 /// %D = select %cond, %C, %A
9045 /// %C = select %cond, %B, 0
9048 /// Assuming that the specified instruction is an operand to the select, return
9049 /// a bitmask indicating which operands of this instruction are foldable if they
9050 /// equal the other incoming value of the select.
9052 static unsigned GetSelectFoldableOperands(Instruction *I) {
9053 switch (I->getOpcode()) {
9054 case Instruction::Add:
9055 case Instruction::Mul:
9056 case Instruction::And:
9057 case Instruction::Or:
9058 case Instruction::Xor:
9059 return 3; // Can fold through either operand.
9060 case Instruction::Sub: // Can only fold on the amount subtracted.
9061 case Instruction::Shl: // Can only fold on the shift amount.
9062 case Instruction::LShr:
9063 case Instruction::AShr:
9066 return 0; // Cannot fold
9070 /// GetSelectFoldableConstant - For the same transformation as the previous
9071 /// function, return the identity constant that goes into the select.
9072 static Constant *GetSelectFoldableConstant(Instruction *I,
9073 LLVMContext *Context) {
9074 switch (I->getOpcode()) {
9075 default: llvm_unreachable("This cannot happen!");
9076 case Instruction::Add:
9077 case Instruction::Sub:
9078 case Instruction::Or:
9079 case Instruction::Xor:
9080 case Instruction::Shl:
9081 case Instruction::LShr:
9082 case Instruction::AShr:
9083 return Constant::getNullValue(I->getType());
9084 case Instruction::And:
9085 return Constant::getAllOnesValue(I->getType());
9086 case Instruction::Mul:
9087 return ConstantInt::get(I->getType(), 1);
9091 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9092 /// have the same opcode and only one use each. Try to simplify this.
9093 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9095 if (TI->getNumOperands() == 1) {
9096 // If this is a non-volatile load or a cast from the same type,
9099 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9102 return 0; // unknown unary op.
9105 // Fold this by inserting a select from the input values.
9106 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9107 FI->getOperand(0), SI.getName()+".v");
9108 InsertNewInstBefore(NewSI, SI);
9109 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9113 // Only handle binary operators here.
9114 if (!isa<BinaryOperator>(TI))
9117 // Figure out if the operations have any operands in common.
9118 Value *MatchOp, *OtherOpT, *OtherOpF;
9120 if (TI->getOperand(0) == FI->getOperand(0)) {
9121 MatchOp = TI->getOperand(0);
9122 OtherOpT = TI->getOperand(1);
9123 OtherOpF = FI->getOperand(1);
9124 MatchIsOpZero = true;
9125 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9126 MatchOp = TI->getOperand(1);
9127 OtherOpT = TI->getOperand(0);
9128 OtherOpF = FI->getOperand(0);
9129 MatchIsOpZero = false;
9130 } else if (!TI->isCommutative()) {
9132 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9133 MatchOp = TI->getOperand(0);
9134 OtherOpT = TI->getOperand(1);
9135 OtherOpF = FI->getOperand(0);
9136 MatchIsOpZero = true;
9137 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9138 MatchOp = TI->getOperand(1);
9139 OtherOpT = TI->getOperand(0);
9140 OtherOpF = FI->getOperand(1);
9141 MatchIsOpZero = true;
9146 // If we reach here, they do have operations in common.
9147 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9148 OtherOpF, SI.getName()+".v");
9149 InsertNewInstBefore(NewSI, SI);
9151 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9153 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9155 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9157 llvm_unreachable("Shouldn't get here");
9161 static bool isSelect01(Constant *C1, Constant *C2) {
9162 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9165 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9168 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9171 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9172 /// facilitate further optimization.
9173 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9175 // See the comment above GetSelectFoldableOperands for a description of the
9176 // transformation we are doing here.
9177 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9178 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9179 !isa<Constant>(FalseVal)) {
9180 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9181 unsigned OpToFold = 0;
9182 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9184 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9189 Constant *C = GetSelectFoldableConstant(TVI, Context);
9190 Value *OOp = TVI->getOperand(2-OpToFold);
9191 // Avoid creating select between 2 constants unless it's selecting
9193 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9194 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9195 InsertNewInstBefore(NewSel, SI);
9196 NewSel->takeName(TVI);
9197 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9198 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9199 llvm_unreachable("Unknown instruction!!");
9206 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9207 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9208 !isa<Constant>(TrueVal)) {
9209 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9210 unsigned OpToFold = 0;
9211 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9213 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9218 Constant *C = GetSelectFoldableConstant(FVI, Context);
9219 Value *OOp = FVI->getOperand(2-OpToFold);
9220 // Avoid creating select between 2 constants unless it's selecting
9222 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9223 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9224 InsertNewInstBefore(NewSel, SI);
9225 NewSel->takeName(FVI);
9226 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9227 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9228 llvm_unreachable("Unknown instruction!!");
9238 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9239 /// ICmpInst as its first operand.
9241 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9243 bool Changed = false;
9244 ICmpInst::Predicate Pred = ICI->getPredicate();
9245 Value *CmpLHS = ICI->getOperand(0);
9246 Value *CmpRHS = ICI->getOperand(1);
9247 Value *TrueVal = SI.getTrueValue();
9248 Value *FalseVal = SI.getFalseValue();
9250 // Check cases where the comparison is with a constant that
9251 // can be adjusted to fit the min/max idiom. We may edit ICI in
9252 // place here, so make sure the select is the only user.
9253 if (ICI->hasOneUse())
9254 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9257 case ICmpInst::ICMP_ULT:
9258 case ICmpInst::ICMP_SLT: {
9259 // X < MIN ? T : F --> F
9260 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9261 return ReplaceInstUsesWith(SI, FalseVal);
9262 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9263 Constant *AdjustedRHS = SubOne(CI);
9264 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9265 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9266 Pred = ICmpInst::getSwappedPredicate(Pred);
9267 CmpRHS = AdjustedRHS;
9268 std::swap(FalseVal, TrueVal);
9269 ICI->setPredicate(Pred);
9270 ICI->setOperand(1, CmpRHS);
9271 SI.setOperand(1, TrueVal);
9272 SI.setOperand(2, FalseVal);
9277 case ICmpInst::ICMP_UGT:
9278 case ICmpInst::ICMP_SGT: {
9279 // X > MAX ? T : F --> F
9280 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9281 return ReplaceInstUsesWith(SI, FalseVal);
9282 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9283 Constant *AdjustedRHS = AddOne(CI);
9284 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9285 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9286 Pred = ICmpInst::getSwappedPredicate(Pred);
9287 CmpRHS = AdjustedRHS;
9288 std::swap(FalseVal, TrueVal);
9289 ICI->setPredicate(Pred);
9290 ICI->setOperand(1, CmpRHS);
9291 SI.setOperand(1, TrueVal);
9292 SI.setOperand(2, FalseVal);
9299 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9300 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9301 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9302 if (match(TrueVal, m_ConstantInt<-1>()) &&
9303 match(FalseVal, m_ConstantInt<0>()))
9304 Pred = ICI->getPredicate();
9305 else if (match(TrueVal, m_ConstantInt<0>()) &&
9306 match(FalseVal, m_ConstantInt<-1>()))
9307 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9309 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9310 // If we are just checking for a icmp eq of a single bit and zext'ing it
9311 // to an integer, then shift the bit to the appropriate place and then
9312 // cast to integer to avoid the comparison.
9313 const APInt &Op1CV = CI->getValue();
9315 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9316 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9317 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9318 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9319 Value *In = ICI->getOperand(0);
9320 Value *Sh = ConstantInt::get(In->getType(),
9321 In->getType()->getScalarSizeInBits()-1);
9322 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9323 In->getName()+".lobit"),
9325 if (In->getType() != SI.getType())
9326 In = CastInst::CreateIntegerCast(In, SI.getType(),
9327 true/*SExt*/, "tmp", ICI);
9329 if (Pred == ICmpInst::ICMP_SGT)
9330 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9331 In->getName()+".not"), *ICI);
9333 return ReplaceInstUsesWith(SI, In);
9338 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9339 // Transform (X == Y) ? X : Y -> Y
9340 if (Pred == ICmpInst::ICMP_EQ)
9341 return ReplaceInstUsesWith(SI, FalseVal);
9342 // Transform (X != Y) ? X : Y -> X
9343 if (Pred == ICmpInst::ICMP_NE)
9344 return ReplaceInstUsesWith(SI, TrueVal);
9345 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9347 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9348 // Transform (X == Y) ? Y : X -> X
9349 if (Pred == ICmpInst::ICMP_EQ)
9350 return ReplaceInstUsesWith(SI, FalseVal);
9351 // Transform (X != Y) ? Y : X -> Y
9352 if (Pred == ICmpInst::ICMP_NE)
9353 return ReplaceInstUsesWith(SI, TrueVal);
9354 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9357 /// NOTE: if we wanted to, this is where to detect integer ABS
9359 return Changed ? &SI : 0;
9362 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9363 Value *CondVal = SI.getCondition();
9364 Value *TrueVal = SI.getTrueValue();
9365 Value *FalseVal = SI.getFalseValue();
9367 // select true, X, Y -> X
9368 // select false, X, Y -> Y
9369 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9370 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9372 // select C, X, X -> X
9373 if (TrueVal == FalseVal)
9374 return ReplaceInstUsesWith(SI, TrueVal);
9376 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9377 return ReplaceInstUsesWith(SI, FalseVal);
9378 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9379 return ReplaceInstUsesWith(SI, TrueVal);
9380 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9381 if (isa<Constant>(TrueVal))
9382 return ReplaceInstUsesWith(SI, TrueVal);
9384 return ReplaceInstUsesWith(SI, FalseVal);
9387 if (SI.getType() == Type::Int1Ty) {
9388 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9389 if (C->getZExtValue()) {
9390 // Change: A = select B, true, C --> A = or B, C
9391 return BinaryOperator::CreateOr(CondVal, FalseVal);
9393 // Change: A = select B, false, C --> A = and !B, C
9395 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9396 "not."+CondVal->getName()), SI);
9397 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9399 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9400 if (C->getZExtValue() == false) {
9401 // Change: A = select B, C, false --> A = and B, C
9402 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9404 // Change: A = select B, C, true --> A = or !B, C
9406 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9407 "not."+CondVal->getName()), SI);
9408 return BinaryOperator::CreateOr(NotCond, TrueVal);
9412 // select a, b, a -> a&b
9413 // select a, a, b -> a|b
9414 if (CondVal == TrueVal)
9415 return BinaryOperator::CreateOr(CondVal, FalseVal);
9416 else if (CondVal == FalseVal)
9417 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9420 // Selecting between two integer constants?
9421 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9422 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9423 // select C, 1, 0 -> zext C to int
9424 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9425 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9426 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9427 // select C, 0, 1 -> zext !C to int
9429 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9430 "not."+CondVal->getName()), SI);
9431 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9434 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9435 // If one of the constants is zero (we know they can't both be) and we
9436 // have an icmp instruction with zero, and we have an 'and' with the
9437 // non-constant value, eliminate this whole mess. This corresponds to
9438 // cases like this: ((X & 27) ? 27 : 0)
9439 if (TrueValC->isZero() || FalseValC->isZero())
9440 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9441 cast<Constant>(IC->getOperand(1))->isNullValue())
9442 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9443 if (ICA->getOpcode() == Instruction::And &&
9444 isa<ConstantInt>(ICA->getOperand(1)) &&
9445 (ICA->getOperand(1) == TrueValC ||
9446 ICA->getOperand(1) == FalseValC) &&
9447 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9448 // Okay, now we know that everything is set up, we just don't
9449 // know whether we have a icmp_ne or icmp_eq and whether the
9450 // true or false val is the zero.
9451 bool ShouldNotVal = !TrueValC->isZero();
9452 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9455 V = InsertNewInstBefore(BinaryOperator::Create(
9456 Instruction::Xor, V, ICA->getOperand(1)), SI);
9457 return ReplaceInstUsesWith(SI, V);
9462 // See if we are selecting two values based on a comparison of the two values.
9463 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9464 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9465 // Transform (X == Y) ? X : Y -> Y
9466 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9467 // This is not safe in general for floating point:
9468 // consider X== -0, Y== +0.
9469 // It becomes safe if either operand is a nonzero constant.
9470 ConstantFP *CFPt, *CFPf;
9471 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9472 !CFPt->getValueAPF().isZero()) ||
9473 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9474 !CFPf->getValueAPF().isZero()))
9475 return ReplaceInstUsesWith(SI, FalseVal);
9477 // Transform (X != Y) ? X : Y -> X
9478 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9479 return ReplaceInstUsesWith(SI, TrueVal);
9480 // NOTE: if we wanted to, this is where to detect MIN/MAX
9482 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9483 // Transform (X == Y) ? Y : X -> X
9484 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9485 // This is not safe in general for floating point:
9486 // consider X== -0, Y== +0.
9487 // It becomes safe if either operand is a nonzero constant.
9488 ConstantFP *CFPt, *CFPf;
9489 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9490 !CFPt->getValueAPF().isZero()) ||
9491 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9492 !CFPf->getValueAPF().isZero()))
9493 return ReplaceInstUsesWith(SI, FalseVal);
9495 // Transform (X != Y) ? Y : X -> Y
9496 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9497 return ReplaceInstUsesWith(SI, TrueVal);
9498 // NOTE: if we wanted to, this is where to detect MIN/MAX
9500 // NOTE: if we wanted to, this is where to detect ABS
9503 // See if we are selecting two values based on a comparison of the two values.
9504 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9505 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9508 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9509 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9510 if (TI->hasOneUse() && FI->hasOneUse()) {
9511 Instruction *AddOp = 0, *SubOp = 0;
9513 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9514 if (TI->getOpcode() == FI->getOpcode())
9515 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9518 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9519 // even legal for FP.
9520 if ((TI->getOpcode() == Instruction::Sub &&
9521 FI->getOpcode() == Instruction::Add) ||
9522 (TI->getOpcode() == Instruction::FSub &&
9523 FI->getOpcode() == Instruction::FAdd)) {
9524 AddOp = FI; SubOp = TI;
9525 } else if ((FI->getOpcode() == Instruction::Sub &&
9526 TI->getOpcode() == Instruction::Add) ||
9527 (FI->getOpcode() == Instruction::FSub &&
9528 TI->getOpcode() == Instruction::FAdd)) {
9529 AddOp = TI; SubOp = FI;
9533 Value *OtherAddOp = 0;
9534 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9535 OtherAddOp = AddOp->getOperand(1);
9536 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9537 OtherAddOp = AddOp->getOperand(0);
9541 // So at this point we know we have (Y -> OtherAddOp):
9542 // select C, (add X, Y), (sub X, Z)
9543 Value *NegVal; // Compute -Z
9544 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9545 NegVal = ConstantExpr::getNeg(C);
9547 NegVal = InsertNewInstBefore(
9548 BinaryOperator::CreateNeg(SubOp->getOperand(1),
9552 Value *NewTrueOp = OtherAddOp;
9553 Value *NewFalseOp = NegVal;
9555 std::swap(NewTrueOp, NewFalseOp);
9556 Instruction *NewSel =
9557 SelectInst::Create(CondVal, NewTrueOp,
9558 NewFalseOp, SI.getName() + ".p");
9560 NewSel = InsertNewInstBefore(NewSel, SI);
9561 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9566 // See if we can fold the select into one of our operands.
9567 if (SI.getType()->isInteger()) {
9568 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9573 if (BinaryOperator::isNot(CondVal)) {
9574 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9575 SI.setOperand(1, FalseVal);
9576 SI.setOperand(2, TrueVal);
9583 /// EnforceKnownAlignment - If the specified pointer points to an object that
9584 /// we control, modify the object's alignment to PrefAlign. This isn't
9585 /// often possible though. If alignment is important, a more reliable approach
9586 /// is to simply align all global variables and allocation instructions to
9587 /// their preferred alignment from the beginning.
9589 static unsigned EnforceKnownAlignment(Value *V,
9590 unsigned Align, unsigned PrefAlign) {
9592 User *U = dyn_cast<User>(V);
9593 if (!U) return Align;
9595 switch (Operator::getOpcode(U)) {
9597 case Instruction::BitCast:
9598 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9599 case Instruction::GetElementPtr: {
9600 // If all indexes are zero, it is just the alignment of the base pointer.
9601 bool AllZeroOperands = true;
9602 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9603 if (!isa<Constant>(*i) ||
9604 !cast<Constant>(*i)->isNullValue()) {
9605 AllZeroOperands = false;
9609 if (AllZeroOperands) {
9610 // Treat this like a bitcast.
9611 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9617 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9618 // If there is a large requested alignment and we can, bump up the alignment
9620 if (!GV->isDeclaration()) {
9621 if (GV->getAlignment() >= PrefAlign)
9622 Align = GV->getAlignment();
9624 GV->setAlignment(PrefAlign);
9628 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9629 // If there is a requested alignment and if this is an alloca, round up. We
9630 // don't do this for malloc, because some systems can't respect the request.
9631 if (isa<AllocaInst>(AI)) {
9632 if (AI->getAlignment() >= PrefAlign)
9633 Align = AI->getAlignment();
9635 AI->setAlignment(PrefAlign);
9644 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9645 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9646 /// and it is more than the alignment of the ultimate object, see if we can
9647 /// increase the alignment of the ultimate object, making this check succeed.
9648 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9649 unsigned PrefAlign) {
9650 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9651 sizeof(PrefAlign) * CHAR_BIT;
9652 APInt Mask = APInt::getAllOnesValue(BitWidth);
9653 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9654 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9655 unsigned TrailZ = KnownZero.countTrailingOnes();
9656 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9658 if (PrefAlign > Align)
9659 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9661 // We don't need to make any adjustment.
9665 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9666 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9667 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9668 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9669 unsigned CopyAlign = MI->getAlignment();
9671 if (CopyAlign < MinAlign) {
9672 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9677 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9679 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9680 if (MemOpLength == 0) return 0;
9682 // Source and destination pointer types are always "i8*" for intrinsic. See
9683 // if the size is something we can handle with a single primitive load/store.
9684 // A single load+store correctly handles overlapping memory in the memmove
9686 unsigned Size = MemOpLength->getZExtValue();
9687 if (Size == 0) return MI; // Delete this mem transfer.
9689 if (Size > 8 || (Size&(Size-1)))
9690 return 0; // If not 1/2/4/8 bytes, exit.
9692 // Use an integer load+store unless we can find something better.
9694 PointerType::getUnqual(IntegerType::get(Size<<3));
9696 // Memcpy forces the use of i8* for the source and destination. That means
9697 // that if you're using memcpy to move one double around, you'll get a cast
9698 // from double* to i8*. We'd much rather use a double load+store rather than
9699 // an i64 load+store, here because this improves the odds that the source or
9700 // dest address will be promotable. See if we can find a better type than the
9701 // integer datatype.
9702 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9703 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9704 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9705 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9706 // down through these levels if so.
9707 while (!SrcETy->isSingleValueType()) {
9708 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9709 if (STy->getNumElements() == 1)
9710 SrcETy = STy->getElementType(0);
9713 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9714 if (ATy->getNumElements() == 1)
9715 SrcETy = ATy->getElementType();
9722 if (SrcETy->isSingleValueType())
9723 NewPtrTy = PointerType::getUnqual(SrcETy);
9728 // If the memcpy/memmove provides better alignment info than we can
9730 SrcAlign = std::max(SrcAlign, CopyAlign);
9731 DstAlign = std::max(DstAlign, CopyAlign);
9733 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9734 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9735 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9736 InsertNewInstBefore(L, *MI);
9737 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9739 // Set the size of the copy to 0, it will be deleted on the next iteration.
9740 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9744 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9745 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9746 if (MI->getAlignment() < Alignment) {
9747 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9752 // Extract the length and alignment and fill if they are constant.
9753 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9754 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9755 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9757 uint64_t Len = LenC->getZExtValue();
9758 Alignment = MI->getAlignment();
9760 // If the length is zero, this is a no-op
9761 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9763 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9764 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9765 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9767 Value *Dest = MI->getDest();
9768 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9770 // Alignment 0 is identity for alignment 1 for memset, but not store.
9771 if (Alignment == 0) Alignment = 1;
9773 // Extract the fill value and store.
9774 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9775 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
9776 Dest, false, Alignment), *MI);
9778 // Set the size of the copy to 0, it will be deleted on the next iteration.
9779 MI->setLength(Constant::getNullValue(LenC->getType()));
9787 /// visitCallInst - CallInst simplification. This mostly only handles folding
9788 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9789 /// the heavy lifting.
9791 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9792 // If the caller function is nounwind, mark the call as nounwind, even if the
9794 if (CI.getParent()->getParent()->doesNotThrow() &&
9795 !CI.doesNotThrow()) {
9796 CI.setDoesNotThrow();
9802 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9803 if (!II) return visitCallSite(&CI);
9805 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9807 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9808 bool Changed = false;
9810 // memmove/cpy/set of zero bytes is a noop.
9811 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9812 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9814 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9815 if (CI->getZExtValue() == 1) {
9816 // Replace the instruction with just byte operations. We would
9817 // transform other cases to loads/stores, but we don't know if
9818 // alignment is sufficient.
9822 // If we have a memmove and the source operation is a constant global,
9823 // then the source and dest pointers can't alias, so we can change this
9824 // into a call to memcpy.
9825 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9826 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9827 if (GVSrc->isConstant()) {
9828 Module *M = CI.getParent()->getParent()->getParent();
9829 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9831 Tys[0] = CI.getOperand(3)->getType();
9833 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9837 // memmove(x,x,size) -> noop.
9838 if (MMI->getSource() == MMI->getDest())
9839 return EraseInstFromFunction(CI);
9842 // If we can determine a pointer alignment that is bigger than currently
9843 // set, update the alignment.
9844 if (isa<MemTransferInst>(MI)) {
9845 if (Instruction *I = SimplifyMemTransfer(MI))
9847 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9848 if (Instruction *I = SimplifyMemSet(MSI))
9852 if (Changed) return II;
9855 switch (II->getIntrinsicID()) {
9857 case Intrinsic::bswap:
9858 // bswap(bswap(x)) -> x
9859 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9860 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9861 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9863 case Intrinsic::ppc_altivec_lvx:
9864 case Intrinsic::ppc_altivec_lvxl:
9865 case Intrinsic::x86_sse_loadu_ps:
9866 case Intrinsic::x86_sse2_loadu_pd:
9867 case Intrinsic::x86_sse2_loadu_dq:
9868 // Turn PPC lvx -> load if the pointer is known aligned.
9869 // Turn X86 loadups -> load if the pointer is known aligned.
9870 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9871 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9872 PointerType::getUnqual(II->getType()),
9874 return new LoadInst(Ptr);
9877 case Intrinsic::ppc_altivec_stvx:
9878 case Intrinsic::ppc_altivec_stvxl:
9879 // Turn stvx -> store if the pointer is known aligned.
9880 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9881 const Type *OpPtrTy =
9882 PointerType::getUnqual(II->getOperand(1)->getType());
9883 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9884 return new StoreInst(II->getOperand(1), Ptr);
9887 case Intrinsic::x86_sse_storeu_ps:
9888 case Intrinsic::x86_sse2_storeu_pd:
9889 case Intrinsic::x86_sse2_storeu_dq:
9890 // Turn X86 storeu -> store if the pointer is known aligned.
9891 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9892 const Type *OpPtrTy =
9893 PointerType::getUnqual(II->getOperand(2)->getType());
9894 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9895 return new StoreInst(II->getOperand(2), Ptr);
9899 case Intrinsic::x86_sse_cvttss2si: {
9900 // These intrinsics only demands the 0th element of its input vector. If
9901 // we can simplify the input based on that, do so now.
9903 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9904 APInt DemandedElts(VWidth, 1);
9905 APInt UndefElts(VWidth, 0);
9906 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9908 II->setOperand(1, V);
9914 case Intrinsic::ppc_altivec_vperm:
9915 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9916 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9917 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9919 // Check that all of the elements are integer constants or undefs.
9920 bool AllEltsOk = true;
9921 for (unsigned i = 0; i != 16; ++i) {
9922 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9923 !isa<UndefValue>(Mask->getOperand(i))) {
9930 // Cast the input vectors to byte vectors.
9931 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9932 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9933 Value *Result = UndefValue::get(Op0->getType());
9935 // Only extract each element once.
9936 Value *ExtractedElts[32];
9937 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9939 for (unsigned i = 0; i != 16; ++i) {
9940 if (isa<UndefValue>(Mask->getOperand(i)))
9942 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9943 Idx &= 31; // Match the hardware behavior.
9945 if (ExtractedElts[Idx] == 0) {
9947 ExtractElementInst::Create(Idx < 16 ? Op0 : Op1,
9948 ConstantInt::get(Type::Int32Ty, Idx&15, false), "tmp");
9949 InsertNewInstBefore(Elt, CI);
9950 ExtractedElts[Idx] = Elt;
9953 // Insert this value into the result vector.
9954 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9955 ConstantInt::get(Type::Int32Ty, i, false),
9957 InsertNewInstBefore(cast<Instruction>(Result), CI);
9959 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9964 case Intrinsic::stackrestore: {
9965 // If the save is right next to the restore, remove the restore. This can
9966 // happen when variable allocas are DCE'd.
9967 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9968 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9969 BasicBlock::iterator BI = SS;
9971 return EraseInstFromFunction(CI);
9975 // Scan down this block to see if there is another stack restore in the
9976 // same block without an intervening call/alloca.
9977 BasicBlock::iterator BI = II;
9978 TerminatorInst *TI = II->getParent()->getTerminator();
9979 bool CannotRemove = false;
9980 for (++BI; &*BI != TI; ++BI) {
9981 if (isa<AllocaInst>(BI)) {
9982 CannotRemove = true;
9985 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9986 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9987 // If there is a stackrestore below this one, remove this one.
9988 if (II->getIntrinsicID() == Intrinsic::stackrestore)
9989 return EraseInstFromFunction(CI);
9990 // Otherwise, ignore the intrinsic.
9992 // If we found a non-intrinsic call, we can't remove the stack
9994 CannotRemove = true;
10000 // If the stack restore is in a return/unwind block and if there are no
10001 // allocas or calls between the restore and the return, nuke the restore.
10002 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10003 return EraseInstFromFunction(CI);
10008 return visitCallSite(II);
10011 // InvokeInst simplification
10013 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10014 return visitCallSite(&II);
10017 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10018 /// passed through the varargs area, we can eliminate the use of the cast.
10019 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10020 const CastInst * const CI,
10021 const TargetData * const TD,
10023 if (!CI->isLosslessCast())
10026 // The size of ByVal arguments is derived from the type, so we
10027 // can't change to a type with a different size. If the size were
10028 // passed explicitly we could avoid this check.
10029 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10032 const Type* SrcTy =
10033 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10034 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10035 if (!SrcTy->isSized() || !DstTy->isSized())
10037 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10042 // visitCallSite - Improvements for call and invoke instructions.
10044 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10045 bool Changed = false;
10047 // If the callee is a constexpr cast of a function, attempt to move the cast
10048 // to the arguments of the call/invoke.
10049 if (transformConstExprCastCall(CS)) return 0;
10051 Value *Callee = CS.getCalledValue();
10053 if (Function *CalleeF = dyn_cast<Function>(Callee))
10054 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10055 Instruction *OldCall = CS.getInstruction();
10056 // If the call and callee calling conventions don't match, this call must
10057 // be unreachable, as the call is undefined.
10058 new StoreInst(ConstantInt::getTrue(*Context),
10059 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
10061 if (!OldCall->use_empty())
10062 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
10063 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10064 return EraseInstFromFunction(*OldCall);
10068 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10069 // This instruction is not reachable, just remove it. We insert a store to
10070 // undef so that we know that this code is not reachable, despite the fact
10071 // that we can't modify the CFG here.
10072 new StoreInst(ConstantInt::getTrue(*Context),
10073 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
10074 CS.getInstruction());
10076 if (!CS.getInstruction()->use_empty())
10077 CS.getInstruction()->
10078 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
10080 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10081 // Don't break the CFG, insert a dummy cond branch.
10082 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10083 ConstantInt::getTrue(*Context), II);
10085 return EraseInstFromFunction(*CS.getInstruction());
10088 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10089 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10090 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10091 return transformCallThroughTrampoline(CS);
10093 const PointerType *PTy = cast<PointerType>(Callee->getType());
10094 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10095 if (FTy->isVarArg()) {
10096 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10097 // See if we can optimize any arguments passed through the varargs area of
10099 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10100 E = CS.arg_end(); I != E; ++I, ++ix) {
10101 CastInst *CI = dyn_cast<CastInst>(*I);
10102 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10103 *I = CI->getOperand(0);
10109 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10110 // Inline asm calls cannot throw - mark them 'nounwind'.
10111 CS.setDoesNotThrow();
10115 return Changed ? CS.getInstruction() : 0;
10118 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10119 // attempt to move the cast to the arguments of the call/invoke.
10121 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10122 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10123 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10124 if (CE->getOpcode() != Instruction::BitCast ||
10125 !isa<Function>(CE->getOperand(0)))
10127 Function *Callee = cast<Function>(CE->getOperand(0));
10128 Instruction *Caller = CS.getInstruction();
10129 const AttrListPtr &CallerPAL = CS.getAttributes();
10131 // Okay, this is a cast from a function to a different type. Unless doing so
10132 // would cause a type conversion of one of our arguments, change this call to
10133 // be a direct call with arguments casted to the appropriate types.
10135 const FunctionType *FT = Callee->getFunctionType();
10136 const Type *OldRetTy = Caller->getType();
10137 const Type *NewRetTy = FT->getReturnType();
10139 if (isa<StructType>(NewRetTy))
10140 return false; // TODO: Handle multiple return values.
10142 // Check to see if we are changing the return type...
10143 if (OldRetTy != NewRetTy) {
10144 if (Callee->isDeclaration() &&
10145 // Conversion is ok if changing from one pointer type to another or from
10146 // a pointer to an integer of the same size.
10147 !((isa<PointerType>(OldRetTy) || !TD ||
10148 OldRetTy == TD->getIntPtrType()) &&
10149 (isa<PointerType>(NewRetTy) || !TD ||
10150 NewRetTy == TD->getIntPtrType())))
10151 return false; // Cannot transform this return value.
10153 if (!Caller->use_empty() &&
10154 // void -> non-void is handled specially
10155 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
10156 return false; // Cannot transform this return value.
10158 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10159 Attributes RAttrs = CallerPAL.getRetAttributes();
10160 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10161 return false; // Attribute not compatible with transformed value.
10164 // If the callsite is an invoke instruction, and the return value is used by
10165 // a PHI node in a successor, we cannot change the return type of the call
10166 // because there is no place to put the cast instruction (without breaking
10167 // the critical edge). Bail out in this case.
10168 if (!Caller->use_empty())
10169 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10170 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10172 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10173 if (PN->getParent() == II->getNormalDest() ||
10174 PN->getParent() == II->getUnwindDest())
10178 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10179 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10181 CallSite::arg_iterator AI = CS.arg_begin();
10182 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10183 const Type *ParamTy = FT->getParamType(i);
10184 const Type *ActTy = (*AI)->getType();
10186 if (!CastInst::isCastable(ActTy, ParamTy))
10187 return false; // Cannot transform this parameter value.
10189 if (CallerPAL.getParamAttributes(i + 1)
10190 & Attribute::typeIncompatible(ParamTy))
10191 return false; // Attribute not compatible with transformed value.
10193 // Converting from one pointer type to another or between a pointer and an
10194 // integer of the same size is safe even if we do not have a body.
10195 bool isConvertible = ActTy == ParamTy ||
10196 (TD && ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
10197 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType())));
10198 if (Callee->isDeclaration() && !isConvertible) return false;
10201 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10202 Callee->isDeclaration())
10203 return false; // Do not delete arguments unless we have a function body.
10205 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10206 !CallerPAL.isEmpty())
10207 // In this case we have more arguments than the new function type, but we
10208 // won't be dropping them. Check that these extra arguments have attributes
10209 // that are compatible with being a vararg call argument.
10210 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10211 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10213 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10214 if (PAttrs & Attribute::VarArgsIncompatible)
10218 // Okay, we decided that this is a safe thing to do: go ahead and start
10219 // inserting cast instructions as necessary...
10220 std::vector<Value*> Args;
10221 Args.reserve(NumActualArgs);
10222 SmallVector<AttributeWithIndex, 8> attrVec;
10223 attrVec.reserve(NumCommonArgs);
10225 // Get any return attributes.
10226 Attributes RAttrs = CallerPAL.getRetAttributes();
10228 // If the return value is not being used, the type may not be compatible
10229 // with the existing attributes. Wipe out any problematic attributes.
10230 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10232 // Add the new return attributes.
10234 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10236 AI = CS.arg_begin();
10237 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10238 const Type *ParamTy = FT->getParamType(i);
10239 if ((*AI)->getType() == ParamTy) {
10240 Args.push_back(*AI);
10242 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10243 false, ParamTy, false);
10244 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
10245 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
10248 // Add any parameter attributes.
10249 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10250 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10253 // If the function takes more arguments than the call was taking, add them
10255 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10256 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10258 // If we are removing arguments to the function, emit an obnoxious warning...
10259 if (FT->getNumParams() < NumActualArgs) {
10260 if (!FT->isVarArg()) {
10261 errs() << "WARNING: While resolving call to function '"
10262 << Callee->getName() << "' arguments were dropped!\n";
10264 // Add all of the arguments in their promoted form to the arg list...
10265 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10266 const Type *PTy = getPromotedType((*AI)->getType());
10267 if (PTy != (*AI)->getType()) {
10268 // Must promote to pass through va_arg area!
10269 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
10271 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
10272 InsertNewInstBefore(Cast, *Caller);
10273 Args.push_back(Cast);
10275 Args.push_back(*AI);
10278 // Add any parameter attributes.
10279 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10280 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10285 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10286 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10288 if (NewRetTy == Type::VoidTy)
10289 Caller->setName(""); // Void type should not have a name.
10291 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10295 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10296 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10297 Args.begin(), Args.end(),
10298 Caller->getName(), Caller);
10299 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10300 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10302 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10303 Caller->getName(), Caller);
10304 CallInst *CI = cast<CallInst>(Caller);
10305 if (CI->isTailCall())
10306 cast<CallInst>(NC)->setTailCall();
10307 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10308 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10311 // Insert a cast of the return type as necessary.
10313 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10314 if (NV->getType() != Type::VoidTy) {
10315 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10317 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10319 // If this is an invoke instruction, we should insert it after the first
10320 // non-phi, instruction in the normal successor block.
10321 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10322 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10323 InsertNewInstBefore(NC, *I);
10325 // Otherwise, it's a call, just insert cast right after the call instr
10326 InsertNewInstBefore(NC, *Caller);
10328 AddUsersToWorkList(*Caller);
10330 NV = UndefValue::get(Caller->getType());
10334 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10335 Caller->replaceAllUsesWith(NV);
10336 Caller->eraseFromParent();
10337 RemoveFromWorkList(Caller);
10341 // transformCallThroughTrampoline - Turn a call to a function created by the
10342 // init_trampoline intrinsic into a direct call to the underlying function.
10344 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10345 Value *Callee = CS.getCalledValue();
10346 const PointerType *PTy = cast<PointerType>(Callee->getType());
10347 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10348 const AttrListPtr &Attrs = CS.getAttributes();
10350 // If the call already has the 'nest' attribute somewhere then give up -
10351 // otherwise 'nest' would occur twice after splicing in the chain.
10352 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10355 IntrinsicInst *Tramp =
10356 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10358 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10359 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10360 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10362 const AttrListPtr &NestAttrs = NestF->getAttributes();
10363 if (!NestAttrs.isEmpty()) {
10364 unsigned NestIdx = 1;
10365 const Type *NestTy = 0;
10366 Attributes NestAttr = Attribute::None;
10368 // Look for a parameter marked with the 'nest' attribute.
10369 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10370 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10371 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10372 // Record the parameter type and any other attributes.
10374 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10379 Instruction *Caller = CS.getInstruction();
10380 std::vector<Value*> NewArgs;
10381 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10383 SmallVector<AttributeWithIndex, 8> NewAttrs;
10384 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10386 // Insert the nest argument into the call argument list, which may
10387 // mean appending it. Likewise for attributes.
10389 // Add any result attributes.
10390 if (Attributes Attr = Attrs.getRetAttributes())
10391 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10395 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10397 if (Idx == NestIdx) {
10398 // Add the chain argument and attributes.
10399 Value *NestVal = Tramp->getOperand(3);
10400 if (NestVal->getType() != NestTy)
10401 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10402 NewArgs.push_back(NestVal);
10403 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10409 // Add the original argument and attributes.
10410 NewArgs.push_back(*I);
10411 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10413 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10419 // Add any function attributes.
10420 if (Attributes Attr = Attrs.getFnAttributes())
10421 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10423 // The trampoline may have been bitcast to a bogus type (FTy).
10424 // Handle this by synthesizing a new function type, equal to FTy
10425 // with the chain parameter inserted.
10427 std::vector<const Type*> NewTypes;
10428 NewTypes.reserve(FTy->getNumParams()+1);
10430 // Insert the chain's type into the list of parameter types, which may
10431 // mean appending it.
10434 FunctionType::param_iterator I = FTy->param_begin(),
10435 E = FTy->param_end();
10438 if (Idx == NestIdx)
10439 // Add the chain's type.
10440 NewTypes.push_back(NestTy);
10445 // Add the original type.
10446 NewTypes.push_back(*I);
10452 // Replace the trampoline call with a direct call. Let the generic
10453 // code sort out any function type mismatches.
10454 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10456 Constant *NewCallee =
10457 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10458 NestF : ConstantExpr::getBitCast(NestF,
10459 PointerType::getUnqual(NewFTy));
10460 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10463 Instruction *NewCaller;
10464 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10465 NewCaller = InvokeInst::Create(NewCallee,
10466 II->getNormalDest(), II->getUnwindDest(),
10467 NewArgs.begin(), NewArgs.end(),
10468 Caller->getName(), Caller);
10469 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10470 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10472 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10473 Caller->getName(), Caller);
10474 if (cast<CallInst>(Caller)->isTailCall())
10475 cast<CallInst>(NewCaller)->setTailCall();
10476 cast<CallInst>(NewCaller)->
10477 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10478 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10480 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10481 Caller->replaceAllUsesWith(NewCaller);
10482 Caller->eraseFromParent();
10483 RemoveFromWorkList(Caller);
10488 // Replace the trampoline call with a direct call. Since there is no 'nest'
10489 // parameter, there is no need to adjust the argument list. Let the generic
10490 // code sort out any function type mismatches.
10491 Constant *NewCallee =
10492 NestF->getType() == PTy ? NestF :
10493 ConstantExpr::getBitCast(NestF, PTy);
10494 CS.setCalledFunction(NewCallee);
10495 return CS.getInstruction();
10498 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10499 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10500 /// and a single binop.
10501 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10502 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10503 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10504 unsigned Opc = FirstInst->getOpcode();
10505 Value *LHSVal = FirstInst->getOperand(0);
10506 Value *RHSVal = FirstInst->getOperand(1);
10508 const Type *LHSType = LHSVal->getType();
10509 const Type *RHSType = RHSVal->getType();
10511 // Scan to see if all operands are the same opcode, all have one use, and all
10512 // kill their operands (i.e. the operands have one use).
10513 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10514 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10515 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10516 // Verify type of the LHS matches so we don't fold cmp's of different
10517 // types or GEP's with different index types.
10518 I->getOperand(0)->getType() != LHSType ||
10519 I->getOperand(1)->getType() != RHSType)
10522 // If they are CmpInst instructions, check their predicates
10523 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10524 if (cast<CmpInst>(I)->getPredicate() !=
10525 cast<CmpInst>(FirstInst)->getPredicate())
10528 // Keep track of which operand needs a phi node.
10529 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10530 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10533 // Otherwise, this is safe to transform!
10535 Value *InLHS = FirstInst->getOperand(0);
10536 Value *InRHS = FirstInst->getOperand(1);
10537 PHINode *NewLHS = 0, *NewRHS = 0;
10539 NewLHS = PHINode::Create(LHSType,
10540 FirstInst->getOperand(0)->getName() + ".pn");
10541 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10542 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10543 InsertNewInstBefore(NewLHS, PN);
10548 NewRHS = PHINode::Create(RHSType,
10549 FirstInst->getOperand(1)->getName() + ".pn");
10550 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10551 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10552 InsertNewInstBefore(NewRHS, PN);
10556 // Add all operands to the new PHIs.
10557 if (NewLHS || NewRHS) {
10558 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10559 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10561 Value *NewInLHS = InInst->getOperand(0);
10562 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10565 Value *NewInRHS = InInst->getOperand(1);
10566 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10571 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10572 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10573 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10574 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10578 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10579 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10581 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10582 FirstInst->op_end());
10583 // This is true if all GEP bases are allocas and if all indices into them are
10585 bool AllBasePointersAreAllocas = true;
10587 // Scan to see if all operands are the same opcode, all have one use, and all
10588 // kill their operands (i.e. the operands have one use).
10589 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10590 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10591 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10592 GEP->getNumOperands() != FirstInst->getNumOperands())
10595 // Keep track of whether or not all GEPs are of alloca pointers.
10596 if (AllBasePointersAreAllocas &&
10597 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10598 !GEP->hasAllConstantIndices()))
10599 AllBasePointersAreAllocas = false;
10601 // Compare the operand lists.
10602 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10603 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10606 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10607 // if one of the PHIs has a constant for the index. The index may be
10608 // substantially cheaper to compute for the constants, so making it a
10609 // variable index could pessimize the path. This also handles the case
10610 // for struct indices, which must always be constant.
10611 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10612 isa<ConstantInt>(GEP->getOperand(op)))
10615 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10617 FixedOperands[op] = 0; // Needs a PHI.
10621 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10622 // bother doing this transformation. At best, this will just save a bit of
10623 // offset calculation, but all the predecessors will have to materialize the
10624 // stack address into a register anyway. We'd actually rather *clone* the
10625 // load up into the predecessors so that we have a load of a gep of an alloca,
10626 // which can usually all be folded into the load.
10627 if (AllBasePointersAreAllocas)
10630 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10631 // that is variable.
10632 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10634 bool HasAnyPHIs = false;
10635 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10636 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10637 Value *FirstOp = FirstInst->getOperand(i);
10638 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10639 FirstOp->getName()+".pn");
10640 InsertNewInstBefore(NewPN, PN);
10642 NewPN->reserveOperandSpace(e);
10643 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10644 OperandPhis[i] = NewPN;
10645 FixedOperands[i] = NewPN;
10650 // Add all operands to the new PHIs.
10652 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10653 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10654 BasicBlock *InBB = PN.getIncomingBlock(i);
10656 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10657 if (PHINode *OpPhi = OperandPhis[op])
10658 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10662 Value *Base = FixedOperands[0];
10663 GetElementPtrInst *GEP =
10664 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10665 FixedOperands.end());
10666 if (cast<GEPOperator>(FirstInst)->isInBounds())
10667 cast<GEPOperator>(GEP)->setIsInBounds(true);
10672 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10673 /// sink the load out of the block that defines it. This means that it must be
10674 /// obvious the value of the load is not changed from the point of the load to
10675 /// the end of the block it is in.
10677 /// Finally, it is safe, but not profitable, to sink a load targetting a
10678 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10680 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10681 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10683 for (++BBI; BBI != E; ++BBI)
10684 if (BBI->mayWriteToMemory())
10687 // Check for non-address taken alloca. If not address-taken already, it isn't
10688 // profitable to do this xform.
10689 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10690 bool isAddressTaken = false;
10691 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10693 if (isa<LoadInst>(UI)) continue;
10694 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10695 // If storing TO the alloca, then the address isn't taken.
10696 if (SI->getOperand(1) == AI) continue;
10698 isAddressTaken = true;
10702 if (!isAddressTaken && AI->isStaticAlloca())
10706 // If this load is a load from a GEP with a constant offset from an alloca,
10707 // then we don't want to sink it. In its present form, it will be
10708 // load [constant stack offset]. Sinking it will cause us to have to
10709 // materialize the stack addresses in each predecessor in a register only to
10710 // do a shared load from register in the successor.
10711 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10712 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10713 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10720 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10721 // operator and they all are only used by the PHI, PHI together their
10722 // inputs, and do the operation once, to the result of the PHI.
10723 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10724 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10726 // Scan the instruction, looking for input operations that can be folded away.
10727 // If all input operands to the phi are the same instruction (e.g. a cast from
10728 // the same type or "+42") we can pull the operation through the PHI, reducing
10729 // code size and simplifying code.
10730 Constant *ConstantOp = 0;
10731 const Type *CastSrcTy = 0;
10732 bool isVolatile = false;
10733 if (isa<CastInst>(FirstInst)) {
10734 CastSrcTy = FirstInst->getOperand(0)->getType();
10735 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10736 // Can fold binop, compare or shift here if the RHS is a constant,
10737 // otherwise call FoldPHIArgBinOpIntoPHI.
10738 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10739 if (ConstantOp == 0)
10740 return FoldPHIArgBinOpIntoPHI(PN);
10741 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10742 isVolatile = LI->isVolatile();
10743 // We can't sink the load if the loaded value could be modified between the
10744 // load and the PHI.
10745 if (LI->getParent() != PN.getIncomingBlock(0) ||
10746 !isSafeAndProfitableToSinkLoad(LI))
10749 // If the PHI is of volatile loads and the load block has multiple
10750 // successors, sinking it would remove a load of the volatile value from
10751 // the path through the other successor.
10753 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10756 } else if (isa<GetElementPtrInst>(FirstInst)) {
10757 return FoldPHIArgGEPIntoPHI(PN);
10759 return 0; // Cannot fold this operation.
10762 // Check to see if all arguments are the same operation.
10763 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10764 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10765 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10766 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10769 if (I->getOperand(0)->getType() != CastSrcTy)
10770 return 0; // Cast operation must match.
10771 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10772 // We can't sink the load if the loaded value could be modified between
10773 // the load and the PHI.
10774 if (LI->isVolatile() != isVolatile ||
10775 LI->getParent() != PN.getIncomingBlock(i) ||
10776 !isSafeAndProfitableToSinkLoad(LI))
10779 // If the PHI is of volatile loads and the load block has multiple
10780 // successors, sinking it would remove a load of the volatile value from
10781 // the path through the other successor.
10783 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10786 } else if (I->getOperand(1) != ConstantOp) {
10791 // Okay, they are all the same operation. Create a new PHI node of the
10792 // correct type, and PHI together all of the LHS's of the instructions.
10793 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10794 PN.getName()+".in");
10795 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10797 Value *InVal = FirstInst->getOperand(0);
10798 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10800 // Add all operands to the new PHI.
10801 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10802 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10803 if (NewInVal != InVal)
10805 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10810 // The new PHI unions all of the same values together. This is really
10811 // common, so we handle it intelligently here for compile-time speed.
10815 InsertNewInstBefore(NewPN, PN);
10819 // Insert and return the new operation.
10820 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10821 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10822 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10823 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10824 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10825 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10826 PhiVal, ConstantOp);
10827 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10829 // If this was a volatile load that we are merging, make sure to loop through
10830 // and mark all the input loads as non-volatile. If we don't do this, we will
10831 // insert a new volatile load and the old ones will not be deletable.
10833 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10834 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10836 return new LoadInst(PhiVal, "", isVolatile);
10839 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10841 static bool DeadPHICycle(PHINode *PN,
10842 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10843 if (PN->use_empty()) return true;
10844 if (!PN->hasOneUse()) return false;
10846 // Remember this node, and if we find the cycle, return.
10847 if (!PotentiallyDeadPHIs.insert(PN))
10850 // Don't scan crazily complex things.
10851 if (PotentiallyDeadPHIs.size() == 16)
10854 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10855 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10860 /// PHIsEqualValue - Return true if this phi node is always equal to
10861 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10862 /// z = some value; x = phi (y, z); y = phi (x, z)
10863 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10864 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10865 // See if we already saw this PHI node.
10866 if (!ValueEqualPHIs.insert(PN))
10869 // Don't scan crazily complex things.
10870 if (ValueEqualPHIs.size() == 16)
10873 // Scan the operands to see if they are either phi nodes or are equal to
10875 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10876 Value *Op = PN->getIncomingValue(i);
10877 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10878 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10880 } else if (Op != NonPhiInVal)
10888 // PHINode simplification
10890 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10891 // If LCSSA is around, don't mess with Phi nodes
10892 if (MustPreserveLCSSA) return 0;
10894 if (Value *V = PN.hasConstantValue())
10895 return ReplaceInstUsesWith(PN, V);
10897 // If all PHI operands are the same operation, pull them through the PHI,
10898 // reducing code size.
10899 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10900 isa<Instruction>(PN.getIncomingValue(1)) &&
10901 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10902 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10903 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10904 // than themselves more than once.
10905 PN.getIncomingValue(0)->hasOneUse())
10906 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10909 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10910 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10911 // PHI)... break the cycle.
10912 if (PN.hasOneUse()) {
10913 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10914 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10915 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10916 PotentiallyDeadPHIs.insert(&PN);
10917 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10918 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10921 // If this phi has a single use, and if that use just computes a value for
10922 // the next iteration of a loop, delete the phi. This occurs with unused
10923 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10924 // common case here is good because the only other things that catch this
10925 // are induction variable analysis (sometimes) and ADCE, which is only run
10927 if (PHIUser->hasOneUse() &&
10928 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10929 PHIUser->use_back() == &PN) {
10930 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10934 // We sometimes end up with phi cycles that non-obviously end up being the
10935 // same value, for example:
10936 // z = some value; x = phi (y, z); y = phi (x, z)
10937 // where the phi nodes don't necessarily need to be in the same block. Do a
10938 // quick check to see if the PHI node only contains a single non-phi value, if
10939 // so, scan to see if the phi cycle is actually equal to that value.
10941 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10942 // Scan for the first non-phi operand.
10943 while (InValNo != NumOperandVals &&
10944 isa<PHINode>(PN.getIncomingValue(InValNo)))
10947 if (InValNo != NumOperandVals) {
10948 Value *NonPhiInVal = PN.getOperand(InValNo);
10950 // Scan the rest of the operands to see if there are any conflicts, if so
10951 // there is no need to recursively scan other phis.
10952 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10953 Value *OpVal = PN.getIncomingValue(InValNo);
10954 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10958 // If we scanned over all operands, then we have one unique value plus
10959 // phi values. Scan PHI nodes to see if they all merge in each other or
10961 if (InValNo == NumOperandVals) {
10962 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10963 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10964 return ReplaceInstUsesWith(PN, NonPhiInVal);
10971 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10972 Instruction *InsertPoint,
10973 InstCombiner *IC) {
10974 unsigned PtrSize = DTy->getScalarSizeInBits();
10975 unsigned VTySize = V->getType()->getScalarSizeInBits();
10976 // We must cast correctly to the pointer type. Ensure that we
10977 // sign extend the integer value if it is smaller as this is
10978 // used for address computation.
10979 Instruction::CastOps opcode =
10980 (VTySize < PtrSize ? Instruction::SExt :
10981 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10982 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10986 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10987 Value *PtrOp = GEP.getOperand(0);
10988 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
10989 // If so, eliminate the noop.
10990 if (GEP.getNumOperands() == 1)
10991 return ReplaceInstUsesWith(GEP, PtrOp);
10993 if (isa<UndefValue>(GEP.getOperand(0)))
10994 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
10996 bool HasZeroPointerIndex = false;
10997 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
10998 HasZeroPointerIndex = C->isNullValue();
11000 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11001 return ReplaceInstUsesWith(GEP, PtrOp);
11003 // Eliminate unneeded casts for indices.
11004 bool MadeChange = false;
11006 gep_type_iterator GTI = gep_type_begin(GEP);
11007 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
11008 i != e; ++i, ++GTI) {
11009 if (TD && isa<SequentialType>(*GTI)) {
11010 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
11011 if (CI->getOpcode() == Instruction::ZExt ||
11012 CI->getOpcode() == Instruction::SExt) {
11013 const Type *SrcTy = CI->getOperand(0)->getType();
11014 // We can eliminate a cast from i32 to i64 iff the target
11015 // is a 32-bit pointer target.
11016 if (SrcTy->getScalarSizeInBits() >= TD->getPointerSizeInBits()) {
11018 *i = CI->getOperand(0);
11022 // If we are using a wider index than needed for this platform, shrink it
11023 // to what we need. If narrower, sign-extend it to what we need.
11024 // If the incoming value needs a cast instruction,
11025 // insert it. This explicit cast can make subsequent optimizations more
11028 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
11029 if (Constant *C = dyn_cast<Constant>(Op)) {
11030 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
11033 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
11038 } else if (TD->getTypeSizeInBits(Op->getType())
11039 < TD->getPointerSizeInBits()) {
11040 if (Constant *C = dyn_cast<Constant>(Op)) {
11041 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
11044 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
11052 if (MadeChange) return &GEP;
11054 // Combine Indices - If the source pointer to this getelementptr instruction
11055 // is a getelementptr instruction, combine the indices of the two
11056 // getelementptr instructions into a single instruction.
11058 SmallVector<Value*, 8> SrcGEPOperands;
11059 bool BothInBounds = cast<GEPOperator>(&GEP)->isInBounds();
11060 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
11061 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
11062 if (!Src->isInBounds())
11063 BothInBounds = false;
11066 if (!SrcGEPOperands.empty()) {
11067 // Note that if our source is a gep chain itself that we wait for that
11068 // chain to be resolved before we perform this transformation. This
11069 // avoids us creating a TON of code in some cases.
11071 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
11072 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
11073 return 0; // Wait until our source is folded to completion.
11075 SmallVector<Value*, 8> Indices;
11077 // Find out whether the last index in the source GEP is a sequential idx.
11078 bool EndsWithSequential = false;
11079 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
11080 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
11081 EndsWithSequential = !isa<StructType>(*I);
11083 // Can we combine the two pointer arithmetics offsets?
11084 if (EndsWithSequential) {
11085 // Replace: gep (gep %P, long B), long A, ...
11086 // With: T = long A+B; gep %P, T, ...
11088 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
11089 if (SO1 == Constant::getNullValue(SO1->getType())) {
11091 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
11094 // If they aren't the same type, convert both to an integer of the
11095 // target's pointer size.
11096 if (SO1->getType() != GO1->getType()) {
11097 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
11099 ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
11100 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
11102 ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
11104 unsigned PS = TD->getPointerSizeInBits();
11105 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
11106 // Convert GO1 to SO1's type.
11107 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
11109 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
11110 // Convert SO1 to GO1's type.
11111 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
11113 const Type *PT = TD->getIntPtrType();
11114 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
11115 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
11119 if (isa<Constant>(SO1) && isa<Constant>(GO1))
11120 Sum = ConstantExpr::getAdd(cast<Constant>(SO1),
11121 cast<Constant>(GO1));
11123 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11124 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
11128 // Recycle the GEP we already have if possible.
11129 if (SrcGEPOperands.size() == 2) {
11130 GEP.setOperand(0, SrcGEPOperands[0]);
11131 GEP.setOperand(1, Sum);
11134 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11135 SrcGEPOperands.end()-1);
11136 Indices.push_back(Sum);
11137 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
11139 } else if (isa<Constant>(*GEP.idx_begin()) &&
11140 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11141 SrcGEPOperands.size() != 1) {
11142 // Otherwise we can do the fold if the first index of the GEP is a zero
11143 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11144 SrcGEPOperands.end());
11145 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
11148 if (!Indices.empty()) {
11149 GetElementPtrInst *NewGEP = GetElementPtrInst::Create(SrcGEPOperands[0],
11154 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11158 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
11159 // GEP of global variable. If all of the indices for this GEP are
11160 // constants, we can promote this to a constexpr instead of an instruction.
11162 // Scan for nonconstants...
11163 SmallVector<Constant*, 8> Indices;
11164 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
11165 for (; I != E && isa<Constant>(*I); ++I)
11166 Indices.push_back(cast<Constant>(*I));
11168 if (I == E) { // If they are all constants...
11169 Constant *CE = ConstantExpr::getGetElementPtr(GV,
11170 &Indices[0],Indices.size());
11172 // Replace all uses of the GEP with the new constexpr...
11173 return ReplaceInstUsesWith(GEP, CE);
11175 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
11176 if (!isa<PointerType>(X->getType())) {
11177 // Not interesting. Source pointer must be a cast from pointer.
11178 } else if (HasZeroPointerIndex) {
11179 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11180 // into : GEP [10 x i8]* X, i32 0, ...
11182 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11183 // into : GEP i8* X, ...
11185 // This occurs when the program declares an array extern like "int X[];"
11186 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11187 const PointerType *XTy = cast<PointerType>(X->getType());
11188 if (const ArrayType *CATy =
11189 dyn_cast<ArrayType>(CPTy->getElementType())) {
11190 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11191 if (CATy->getElementType() == XTy->getElementType()) {
11192 // -> GEP i8* X, ...
11193 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11194 GetElementPtrInst *NewGEP =
11195 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11197 if (cast<GEPOperator>(&GEP)->isInBounds())
11198 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11200 } else if (const ArrayType *XATy =
11201 dyn_cast<ArrayType>(XTy->getElementType())) {
11202 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11203 if (CATy->getElementType() == XATy->getElementType()) {
11204 // -> GEP [10 x i8]* X, i32 0, ...
11205 // At this point, we know that the cast source type is a pointer
11206 // to an array of the same type as the destination pointer
11207 // array. Because the array type is never stepped over (there
11208 // is a leading zero) we can fold the cast into this GEP.
11209 GEP.setOperand(0, X);
11214 } else if (GEP.getNumOperands() == 2) {
11215 // Transform things like:
11216 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11217 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11218 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11219 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11220 if (TD && isa<ArrayType>(SrcElTy) &&
11221 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11222 TD->getTypeAllocSize(ResElTy)) {
11224 Idx[0] = Constant::getNullValue(Type::Int32Ty);
11225 Idx[1] = GEP.getOperand(1);
11226 GetElementPtrInst *NewGEP =
11227 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11228 if (cast<GEPOperator>(&GEP)->isInBounds())
11229 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11230 Value *V = InsertNewInstBefore(NewGEP, GEP);
11231 // V and GEP are both pointer types --> BitCast
11232 return new BitCastInst(V, GEP.getType());
11235 // Transform things like:
11236 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11237 // (where tmp = 8*tmp2) into:
11238 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11240 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
11241 uint64_t ArrayEltSize =
11242 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11244 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11245 // allow either a mul, shift, or constant here.
11247 ConstantInt *Scale = 0;
11248 if (ArrayEltSize == 1) {
11249 NewIdx = GEP.getOperand(1);
11251 ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11252 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11253 NewIdx = ConstantInt::get(CI->getType(), 1);
11255 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11256 if (Inst->getOpcode() == Instruction::Shl &&
11257 isa<ConstantInt>(Inst->getOperand(1))) {
11258 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11259 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11260 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11262 NewIdx = Inst->getOperand(0);
11263 } else if (Inst->getOpcode() == Instruction::Mul &&
11264 isa<ConstantInt>(Inst->getOperand(1))) {
11265 Scale = cast<ConstantInt>(Inst->getOperand(1));
11266 NewIdx = Inst->getOperand(0);
11270 // If the index will be to exactly the right offset with the scale taken
11271 // out, perform the transformation. Note, we don't know whether Scale is
11272 // signed or not. We'll use unsigned version of division/modulo
11273 // operation after making sure Scale doesn't have the sign bit set.
11274 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11275 Scale->getZExtValue() % ArrayEltSize == 0) {
11276 Scale = ConstantInt::get(Scale->getType(),
11277 Scale->getZExtValue() / ArrayEltSize);
11278 if (Scale->getZExtValue() != 1) {
11280 ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11282 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
11283 NewIdx = InsertNewInstBefore(Sc, GEP);
11286 // Insert the new GEP instruction.
11288 Idx[0] = Constant::getNullValue(Type::Int32Ty);
11290 Instruction *NewGEP =
11291 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11292 if (cast<GEPOperator>(&GEP)->isInBounds())
11293 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11294 NewGEP = InsertNewInstBefore(NewGEP, GEP);
11295 // The NewGEP must be pointer typed, so must the old one -> BitCast
11296 return new BitCastInst(NewGEP, GEP.getType());
11302 /// See if we can simplify:
11303 /// X = bitcast A to B*
11304 /// Y = gep X, <...constant indices...>
11305 /// into a gep of the original struct. This is important for SROA and alias
11306 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11307 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11309 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11310 // Determine how much the GEP moves the pointer. We are guaranteed to get
11311 // a constant back from EmitGEPOffset.
11312 ConstantInt *OffsetV =
11313 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11314 int64_t Offset = OffsetV->getSExtValue();
11316 // If this GEP instruction doesn't move the pointer, just replace the GEP
11317 // with a bitcast of the real input to the dest type.
11319 // If the bitcast is of an allocation, and the allocation will be
11320 // converted to match the type of the cast, don't touch this.
11321 if (isa<AllocationInst>(BCI->getOperand(0))) {
11322 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11323 if (Instruction *I = visitBitCast(*BCI)) {
11326 BCI->getParent()->getInstList().insert(BCI, I);
11327 ReplaceInstUsesWith(*BCI, I);
11332 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11335 // Otherwise, if the offset is non-zero, we need to find out if there is a
11336 // field at Offset in 'A's type. If so, we can pull the cast through the
11338 SmallVector<Value*, 8> NewIndices;
11340 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11341 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11342 Instruction *NGEP =
11343 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11345 if (NGEP->getType() == GEP.getType()) return NGEP;
11346 if (cast<GEPOperator>(&GEP)->isInBounds())
11347 cast<GEPOperator>(NGEP)->setIsInBounds(true);
11348 InsertNewInstBefore(NGEP, GEP);
11349 NGEP->takeName(&GEP);
11350 return new BitCastInst(NGEP, GEP.getType());
11358 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11359 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11360 if (AI.isArrayAllocation()) { // Check C != 1
11361 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11362 const Type *NewTy =
11363 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11364 AllocationInst *New = 0;
11366 // Create and insert the replacement instruction...
11367 if (isa<MallocInst>(AI))
11368 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11370 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11371 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11374 InsertNewInstBefore(New, AI);
11376 // Scan to the end of the allocation instructions, to skip over a block of
11377 // allocas if possible...also skip interleaved debug info
11379 BasicBlock::iterator It = New;
11380 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11382 // Now that I is pointing to the first non-allocation-inst in the block,
11383 // insert our getelementptr instruction...
11385 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
11389 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11390 New->getName()+".sub", It);
11391 cast<GEPOperator>(V)->setIsInBounds(true);
11393 // Now make everything use the getelementptr instead of the original
11395 return ReplaceInstUsesWith(AI, V);
11396 } else if (isa<UndefValue>(AI.getArraySize())) {
11397 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11401 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11402 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11403 // Note that we only do this for alloca's, because malloc should allocate
11404 // and return a unique pointer, even for a zero byte allocation.
11405 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11406 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11408 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11409 if (AI.getAlignment() == 0)
11410 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11416 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11417 Value *Op = FI.getOperand(0);
11419 // free undef -> unreachable.
11420 if (isa<UndefValue>(Op)) {
11421 // Insert a new store to null because we cannot modify the CFG here.
11422 new StoreInst(ConstantInt::getTrue(*Context),
11423 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
11424 return EraseInstFromFunction(FI);
11427 // If we have 'free null' delete the instruction. This can happen in stl code
11428 // when lots of inlining happens.
11429 if (isa<ConstantPointerNull>(Op))
11430 return EraseInstFromFunction(FI);
11432 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11433 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11434 FI.setOperand(0, CI->getOperand(0));
11438 // Change free (gep X, 0,0,0,0) into free(X)
11439 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11440 if (GEPI->hasAllZeroIndices()) {
11441 AddToWorkList(GEPI);
11442 FI.setOperand(0, GEPI->getOperand(0));
11447 // Change free(malloc) into nothing, if the malloc has a single use.
11448 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11449 if (MI->hasOneUse()) {
11450 EraseInstFromFunction(FI);
11451 return EraseInstFromFunction(*MI);
11458 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11459 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11460 const TargetData *TD) {
11461 User *CI = cast<User>(LI.getOperand(0));
11462 Value *CastOp = CI->getOperand(0);
11463 LLVMContext *Context = IC.getContext();
11466 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11467 // Instead of loading constant c string, use corresponding integer value
11468 // directly if string length is small enough.
11470 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11471 unsigned len = Str.length();
11472 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11473 unsigned numBits = Ty->getPrimitiveSizeInBits();
11474 // Replace LI with immediate integer store.
11475 if ((numBits >> 3) == len + 1) {
11476 APInt StrVal(numBits, 0);
11477 APInt SingleChar(numBits, 0);
11478 if (TD->isLittleEndian()) {
11479 for (signed i = len-1; i >= 0; i--) {
11480 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11481 StrVal = (StrVal << 8) | SingleChar;
11484 for (unsigned i = 0; i < len; i++) {
11485 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11486 StrVal = (StrVal << 8) | SingleChar;
11488 // Append NULL at the end.
11490 StrVal = (StrVal << 8) | SingleChar;
11492 Value *NL = ConstantInt::get(*Context, StrVal);
11493 return IC.ReplaceInstUsesWith(LI, NL);
11499 const PointerType *DestTy = cast<PointerType>(CI->getType());
11500 const Type *DestPTy = DestTy->getElementType();
11501 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11503 // If the address spaces don't match, don't eliminate the cast.
11504 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11507 const Type *SrcPTy = SrcTy->getElementType();
11509 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11510 isa<VectorType>(DestPTy)) {
11511 // If the source is an array, the code below will not succeed. Check to
11512 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11514 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11515 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11516 if (ASrcTy->getNumElements() != 0) {
11518 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
11519 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11520 SrcTy = cast<PointerType>(CastOp->getType());
11521 SrcPTy = SrcTy->getElementType();
11524 if (IC.getTargetData() &&
11525 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11526 isa<VectorType>(SrcPTy)) &&
11527 // Do not allow turning this into a load of an integer, which is then
11528 // casted to a pointer, this pessimizes pointer analysis a lot.
11529 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11530 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
11531 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
11533 // Okay, we are casting from one integer or pointer type to another of
11534 // the same size. Instead of casting the pointer before the load, cast
11535 // the result of the loaded value.
11536 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11538 LI.isVolatile()),LI);
11539 // Now cast the result of the load.
11540 return new BitCastInst(NewLoad, LI.getType());
11547 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11548 Value *Op = LI.getOperand(0);
11550 // Attempt to improve the alignment.
11552 unsigned KnownAlign =
11553 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11555 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11556 LI.getAlignment()))
11557 LI.setAlignment(KnownAlign);
11560 // load (cast X) --> cast (load X) iff safe
11561 if (isa<CastInst>(Op))
11562 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11565 // None of the following transforms are legal for volatile loads.
11566 if (LI.isVolatile()) return 0;
11568 // Do really simple store-to-load forwarding and load CSE, to catch cases
11569 // where there are several consequtive memory accesses to the same location,
11570 // separated by a few arithmetic operations.
11571 BasicBlock::iterator BBI = &LI;
11572 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11573 return ReplaceInstUsesWith(LI, AvailableVal);
11575 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11576 const Value *GEPI0 = GEPI->getOperand(0);
11577 // TODO: Consider a target hook for valid address spaces for this xform.
11578 if (isa<ConstantPointerNull>(GEPI0) &&
11579 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11580 // Insert a new store to null instruction before the load to indicate
11581 // that this code is not reachable. We do this instead of inserting
11582 // an unreachable instruction directly because we cannot modify the
11584 new StoreInst(UndefValue::get(LI.getType()),
11585 Constant::getNullValue(Op->getType()), &LI);
11586 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11590 if (Constant *C = dyn_cast<Constant>(Op)) {
11591 // load null/undef -> undef
11592 // TODO: Consider a target hook for valid address spaces for this xform.
11593 if (isa<UndefValue>(C) || (C->isNullValue() &&
11594 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11595 // Insert a new store to null instruction before the load to indicate that
11596 // this code is not reachable. We do this instead of inserting an
11597 // unreachable instruction directly because we cannot modify the CFG.
11598 new StoreInst(UndefValue::get(LI.getType()),
11599 Constant::getNullValue(Op->getType()), &LI);
11600 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11603 // Instcombine load (constant global) into the value loaded.
11604 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11605 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11606 return ReplaceInstUsesWith(LI, GV->getInitializer());
11608 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11609 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11610 if (CE->getOpcode() == Instruction::GetElementPtr) {
11611 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11612 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11614 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11616 return ReplaceInstUsesWith(LI, V);
11617 if (CE->getOperand(0)->isNullValue()) {
11618 // Insert a new store to null instruction before the load to indicate
11619 // that this code is not reachable. We do this instead of inserting
11620 // an unreachable instruction directly because we cannot modify the
11622 new StoreInst(UndefValue::get(LI.getType()),
11623 Constant::getNullValue(Op->getType()), &LI);
11624 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11627 } else if (CE->isCast()) {
11628 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11634 // If this load comes from anywhere in a constant global, and if the global
11635 // is all undef or zero, we know what it loads.
11636 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11637 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11638 if (GV->getInitializer()->isNullValue())
11639 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11640 else if (isa<UndefValue>(GV->getInitializer()))
11641 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11645 if (Op->hasOneUse()) {
11646 // Change select and PHI nodes to select values instead of addresses: this
11647 // helps alias analysis out a lot, allows many others simplifications, and
11648 // exposes redundancy in the code.
11650 // Note that we cannot do the transformation unless we know that the
11651 // introduced loads cannot trap! Something like this is valid as long as
11652 // the condition is always false: load (select bool %C, int* null, int* %G),
11653 // but it would not be valid if we transformed it to load from null
11654 // unconditionally.
11656 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11657 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11658 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11659 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11660 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11661 SI->getOperand(1)->getName()+".val"), LI);
11662 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11663 SI->getOperand(2)->getName()+".val"), LI);
11664 return SelectInst::Create(SI->getCondition(), V1, V2);
11667 // load (select (cond, null, P)) -> load P
11668 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11669 if (C->isNullValue()) {
11670 LI.setOperand(0, SI->getOperand(2));
11674 // load (select (cond, P, null)) -> load P
11675 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11676 if (C->isNullValue()) {
11677 LI.setOperand(0, SI->getOperand(1));
11685 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11686 /// when possible. This makes it generally easy to do alias analysis and/or
11687 /// SROA/mem2reg of the memory object.
11688 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11689 User *CI = cast<User>(SI.getOperand(1));
11690 Value *CastOp = CI->getOperand(0);
11692 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11693 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11694 if (SrcTy == 0) return 0;
11696 const Type *SrcPTy = SrcTy->getElementType();
11698 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11701 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11702 /// to its first element. This allows us to handle things like:
11703 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11704 /// on 32-bit hosts.
11705 SmallVector<Value*, 4> NewGEPIndices;
11707 // If the source is an array, the code below will not succeed. Check to
11708 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11710 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11711 // Index through pointer.
11712 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
11713 NewGEPIndices.push_back(Zero);
11716 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11717 if (!STy->getNumElements()) /* Struct can be empty {} */
11719 NewGEPIndices.push_back(Zero);
11720 SrcPTy = STy->getElementType(0);
11721 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11722 NewGEPIndices.push_back(Zero);
11723 SrcPTy = ATy->getElementType();
11729 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11732 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11735 // If the pointers point into different address spaces or if they point to
11736 // values with different sizes, we can't do the transformation.
11737 if (!IC.getTargetData() ||
11738 SrcTy->getAddressSpace() !=
11739 cast<PointerType>(CI->getType())->getAddressSpace() ||
11740 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
11741 IC.getTargetData()->getTypeSizeInBits(DestPTy))
11744 // Okay, we are casting from one integer or pointer type to another of
11745 // the same size. Instead of casting the pointer before
11746 // the store, cast the value to be stored.
11748 Value *SIOp0 = SI.getOperand(0);
11749 Instruction::CastOps opcode = Instruction::BitCast;
11750 const Type* CastSrcTy = SIOp0->getType();
11751 const Type* CastDstTy = SrcPTy;
11752 if (isa<PointerType>(CastDstTy)) {
11753 if (CastSrcTy->isInteger())
11754 opcode = Instruction::IntToPtr;
11755 } else if (isa<IntegerType>(CastDstTy)) {
11756 if (isa<PointerType>(SIOp0->getType()))
11757 opcode = Instruction::PtrToInt;
11760 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11761 // emit a GEP to index into its first field.
11762 if (!NewGEPIndices.empty()) {
11763 if (Constant *C = dyn_cast<Constant>(CastOp))
11764 CastOp = ConstantExpr::getGetElementPtr(C, &NewGEPIndices[0],
11765 NewGEPIndices.size());
11767 CastOp = IC.InsertNewInstBefore(
11768 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11769 NewGEPIndices.end()), SI);
11770 cast<GEPOperator>(CastOp)->setIsInBounds(true);
11773 if (Constant *C = dyn_cast<Constant>(SIOp0))
11774 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11776 NewCast = IC.InsertNewInstBefore(
11777 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11779 return new StoreInst(NewCast, CastOp);
11782 /// equivalentAddressValues - Test if A and B will obviously have the same
11783 /// value. This includes recognizing that %t0 and %t1 will have the same
11784 /// value in code like this:
11785 /// %t0 = getelementptr \@a, 0, 3
11786 /// store i32 0, i32* %t0
11787 /// %t1 = getelementptr \@a, 0, 3
11788 /// %t2 = load i32* %t1
11790 static bool equivalentAddressValues(Value *A, Value *B) {
11791 // Test if the values are trivially equivalent.
11792 if (A == B) return true;
11794 // Test if the values come form identical arithmetic instructions.
11795 if (isa<BinaryOperator>(A) ||
11796 isa<CastInst>(A) ||
11798 isa<GetElementPtrInst>(A))
11799 if (Instruction *BI = dyn_cast<Instruction>(B))
11800 if (cast<Instruction>(A)->isIdenticalTo(BI))
11803 // Otherwise they may not be equivalent.
11807 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11808 // return the llvm.dbg.declare.
11809 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11810 if (!V->hasNUses(2))
11812 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11814 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11816 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11817 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11824 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11825 Value *Val = SI.getOperand(0);
11826 Value *Ptr = SI.getOperand(1);
11828 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11829 EraseInstFromFunction(SI);
11834 // If the RHS is an alloca with a single use, zapify the store, making the
11836 // If the RHS is an alloca with a two uses, the other one being a
11837 // llvm.dbg.declare, zapify the store and the declare, making the
11838 // alloca dead. We must do this to prevent declare's from affecting
11840 if (!SI.isVolatile()) {
11841 if (Ptr->hasOneUse()) {
11842 if (isa<AllocaInst>(Ptr)) {
11843 EraseInstFromFunction(SI);
11847 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11848 if (isa<AllocaInst>(GEP->getOperand(0))) {
11849 if (GEP->getOperand(0)->hasOneUse()) {
11850 EraseInstFromFunction(SI);
11854 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11855 EraseInstFromFunction(*DI);
11856 EraseInstFromFunction(SI);
11863 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11864 EraseInstFromFunction(*DI);
11865 EraseInstFromFunction(SI);
11871 // Attempt to improve the alignment.
11873 unsigned KnownAlign =
11874 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11876 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11877 SI.getAlignment()))
11878 SI.setAlignment(KnownAlign);
11881 // Do really simple DSE, to catch cases where there are several consecutive
11882 // stores to the same location, separated by a few arithmetic operations. This
11883 // situation often occurs with bitfield accesses.
11884 BasicBlock::iterator BBI = &SI;
11885 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11888 // Don't count debug info directives, lest they affect codegen,
11889 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11890 // It is necessary for correctness to skip those that feed into a
11891 // llvm.dbg.declare, as these are not present when debugging is off.
11892 if (isa<DbgInfoIntrinsic>(BBI) ||
11893 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11898 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11899 // Prev store isn't volatile, and stores to the same location?
11900 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11901 SI.getOperand(1))) {
11904 EraseInstFromFunction(*PrevSI);
11910 // If this is a load, we have to stop. However, if the loaded value is from
11911 // the pointer we're loading and is producing the pointer we're storing,
11912 // then *this* store is dead (X = load P; store X -> P).
11913 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11914 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11915 !SI.isVolatile()) {
11916 EraseInstFromFunction(SI);
11920 // Otherwise, this is a load from some other location. Stores before it
11921 // may not be dead.
11925 // Don't skip over loads or things that can modify memory.
11926 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11931 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11933 // store X, null -> turns into 'unreachable' in SimplifyCFG
11934 if (isa<ConstantPointerNull>(Ptr) &&
11935 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11936 if (!isa<UndefValue>(Val)) {
11937 SI.setOperand(0, UndefValue::get(Val->getType()));
11938 if (Instruction *U = dyn_cast<Instruction>(Val))
11939 AddToWorkList(U); // Dropped a use.
11942 return 0; // Do not modify these!
11945 // store undef, Ptr -> noop
11946 if (isa<UndefValue>(Val)) {
11947 EraseInstFromFunction(SI);
11952 // If the pointer destination is a cast, see if we can fold the cast into the
11954 if (isa<CastInst>(Ptr))
11955 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11957 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11959 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11963 // If this store is the last instruction in the basic block (possibly
11964 // excepting debug info instructions and the pointer bitcasts that feed
11965 // into them), and if the block ends with an unconditional branch, try
11966 // to move it to the successor block.
11970 } while (isa<DbgInfoIntrinsic>(BBI) ||
11971 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11972 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11973 if (BI->isUnconditional())
11974 if (SimplifyStoreAtEndOfBlock(SI))
11975 return 0; // xform done!
11980 /// SimplifyStoreAtEndOfBlock - Turn things like:
11981 /// if () { *P = v1; } else { *P = v2 }
11982 /// into a phi node with a store in the successor.
11984 /// Simplify things like:
11985 /// *P = v1; if () { *P = v2; }
11986 /// into a phi node with a store in the successor.
11988 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11989 BasicBlock *StoreBB = SI.getParent();
11991 // Check to see if the successor block has exactly two incoming edges. If
11992 // so, see if the other predecessor contains a store to the same location.
11993 // if so, insert a PHI node (if needed) and move the stores down.
11994 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11996 // Determine whether Dest has exactly two predecessors and, if so, compute
11997 // the other predecessor.
11998 pred_iterator PI = pred_begin(DestBB);
11999 BasicBlock *OtherBB = 0;
12000 if (*PI != StoreBB)
12003 if (PI == pred_end(DestBB))
12006 if (*PI != StoreBB) {
12011 if (++PI != pred_end(DestBB))
12014 // Bail out if all the relevant blocks aren't distinct (this can happen,
12015 // for example, if SI is in an infinite loop)
12016 if (StoreBB == DestBB || OtherBB == DestBB)
12019 // Verify that the other block ends in a branch and is not otherwise empty.
12020 BasicBlock::iterator BBI = OtherBB->getTerminator();
12021 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
12022 if (!OtherBr || BBI == OtherBB->begin())
12025 // If the other block ends in an unconditional branch, check for the 'if then
12026 // else' case. there is an instruction before the branch.
12027 StoreInst *OtherStore = 0;
12028 if (OtherBr->isUnconditional()) {
12030 // Skip over debugging info.
12031 while (isa<DbgInfoIntrinsic>(BBI) ||
12032 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12033 if (BBI==OtherBB->begin())
12037 // If this isn't a store, or isn't a store to the same location, bail out.
12038 OtherStore = dyn_cast<StoreInst>(BBI);
12039 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
12042 // Otherwise, the other block ended with a conditional branch. If one of the
12043 // destinations is StoreBB, then we have the if/then case.
12044 if (OtherBr->getSuccessor(0) != StoreBB &&
12045 OtherBr->getSuccessor(1) != StoreBB)
12048 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12049 // if/then triangle. See if there is a store to the same ptr as SI that
12050 // lives in OtherBB.
12052 // Check to see if we find the matching store.
12053 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12054 if (OtherStore->getOperand(1) != SI.getOperand(1))
12058 // If we find something that may be using or overwriting the stored
12059 // value, or if we run out of instructions, we can't do the xform.
12060 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12061 BBI == OtherBB->begin())
12065 // In order to eliminate the store in OtherBr, we have to
12066 // make sure nothing reads or overwrites the stored value in
12068 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12069 // FIXME: This should really be AA driven.
12070 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12075 // Insert a PHI node now if we need it.
12076 Value *MergedVal = OtherStore->getOperand(0);
12077 if (MergedVal != SI.getOperand(0)) {
12078 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12079 PN->reserveOperandSpace(2);
12080 PN->addIncoming(SI.getOperand(0), SI.getParent());
12081 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12082 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12085 // Advance to a place where it is safe to insert the new store and
12087 BBI = DestBB->getFirstNonPHI();
12088 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12089 OtherStore->isVolatile()), *BBI);
12091 // Nuke the old stores.
12092 EraseInstFromFunction(SI);
12093 EraseInstFromFunction(*OtherStore);
12099 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12100 // Change br (not X), label True, label False to: br X, label False, True
12102 BasicBlock *TrueDest;
12103 BasicBlock *FalseDest;
12104 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
12105 !isa<Constant>(X)) {
12106 // Swap Destinations and condition...
12107 BI.setCondition(X);
12108 BI.setSuccessor(0, FalseDest);
12109 BI.setSuccessor(1, TrueDest);
12113 // Cannonicalize fcmp_one -> fcmp_oeq
12114 FCmpInst::Predicate FPred; Value *Y;
12115 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12116 TrueDest, FalseDest)))
12117 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12118 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
12119 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
12120 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
12121 Instruction *NewSCC = new FCmpInst(I, NewPred, X, Y, "");
12122 NewSCC->takeName(I);
12123 // Swap Destinations and condition...
12124 BI.setCondition(NewSCC);
12125 BI.setSuccessor(0, FalseDest);
12126 BI.setSuccessor(1, TrueDest);
12127 RemoveFromWorkList(I);
12128 I->eraseFromParent();
12129 AddToWorkList(NewSCC);
12133 // Cannonicalize icmp_ne -> icmp_eq
12134 ICmpInst::Predicate IPred;
12135 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12136 TrueDest, FalseDest)))
12137 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12138 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12139 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
12140 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
12141 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
12142 Instruction *NewSCC = new ICmpInst(I, NewPred, X, Y, "");
12143 NewSCC->takeName(I);
12144 // Swap Destinations and condition...
12145 BI.setCondition(NewSCC);
12146 BI.setSuccessor(0, FalseDest);
12147 BI.setSuccessor(1, TrueDest);
12148 RemoveFromWorkList(I);
12149 I->eraseFromParent();;
12150 AddToWorkList(NewSCC);
12157 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12158 Value *Cond = SI.getCondition();
12159 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12160 if (I->getOpcode() == Instruction::Add)
12161 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12162 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12163 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12165 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
12167 SI.setOperand(0, I->getOperand(0));
12175 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12176 Value *Agg = EV.getAggregateOperand();
12178 if (!EV.hasIndices())
12179 return ReplaceInstUsesWith(EV, Agg);
12181 if (Constant *C = dyn_cast<Constant>(Agg)) {
12182 if (isa<UndefValue>(C))
12183 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12185 if (isa<ConstantAggregateZero>(C))
12186 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12188 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12189 // Extract the element indexed by the first index out of the constant
12190 Value *V = C->getOperand(*EV.idx_begin());
12191 if (EV.getNumIndices() > 1)
12192 // Extract the remaining indices out of the constant indexed by the
12194 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12196 return ReplaceInstUsesWith(EV, V);
12198 return 0; // Can't handle other constants
12200 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12201 // We're extracting from an insertvalue instruction, compare the indices
12202 const unsigned *exti, *exte, *insi, *inse;
12203 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12204 exte = EV.idx_end(), inse = IV->idx_end();
12205 exti != exte && insi != inse;
12207 if (*insi != *exti)
12208 // The insert and extract both reference distinctly different elements.
12209 // This means the extract is not influenced by the insert, and we can
12210 // replace the aggregate operand of the extract with the aggregate
12211 // operand of the insert. i.e., replace
12212 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12213 // %E = extractvalue { i32, { i32 } } %I, 0
12215 // %E = extractvalue { i32, { i32 } } %A, 0
12216 return ExtractValueInst::Create(IV->getAggregateOperand(),
12217 EV.idx_begin(), EV.idx_end());
12219 if (exti == exte && insi == inse)
12220 // Both iterators are at the end: Index lists are identical. Replace
12221 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12222 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12224 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12225 if (exti == exte) {
12226 // The extract list is a prefix of the insert list. i.e. replace
12227 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12228 // %E = extractvalue { i32, { i32 } } %I, 1
12230 // %X = extractvalue { i32, { i32 } } %A, 1
12231 // %E = insertvalue { i32 } %X, i32 42, 0
12232 // by switching the order of the insert and extract (though the
12233 // insertvalue should be left in, since it may have other uses).
12234 Value *NewEV = InsertNewInstBefore(
12235 ExtractValueInst::Create(IV->getAggregateOperand(),
12236 EV.idx_begin(), EV.idx_end()),
12238 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12242 // The insert list is a prefix of the extract list
12243 // We can simply remove the common indices from the extract and make it
12244 // operate on the inserted value instead of the insertvalue result.
12246 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12247 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12249 // %E extractvalue { i32 } { i32 42 }, 0
12250 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12253 // Can't simplify extracts from other values. Note that nested extracts are
12254 // already simplified implicitely by the above (extract ( extract (insert) )
12255 // will be translated into extract ( insert ( extract ) ) first and then just
12256 // the value inserted, if appropriate).
12260 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12261 /// is to leave as a vector operation.
12262 static bool CheapToScalarize(Value *V, bool isConstant) {
12263 if (isa<ConstantAggregateZero>(V))
12265 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12266 if (isConstant) return true;
12267 // If all elts are the same, we can extract.
12268 Constant *Op0 = C->getOperand(0);
12269 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12270 if (C->getOperand(i) != Op0)
12274 Instruction *I = dyn_cast<Instruction>(V);
12275 if (!I) return false;
12277 // Insert element gets simplified to the inserted element or is deleted if
12278 // this is constant idx extract element and its a constant idx insertelt.
12279 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12280 isa<ConstantInt>(I->getOperand(2)))
12282 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12284 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12285 if (BO->hasOneUse() &&
12286 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12287 CheapToScalarize(BO->getOperand(1), isConstant)))
12289 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12290 if (CI->hasOneUse() &&
12291 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12292 CheapToScalarize(CI->getOperand(1), isConstant)))
12298 /// Read and decode a shufflevector mask.
12300 /// It turns undef elements into values that are larger than the number of
12301 /// elements in the input.
12302 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12303 unsigned NElts = SVI->getType()->getNumElements();
12304 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12305 return std::vector<unsigned>(NElts, 0);
12306 if (isa<UndefValue>(SVI->getOperand(2)))
12307 return std::vector<unsigned>(NElts, 2*NElts);
12309 std::vector<unsigned> Result;
12310 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12311 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12312 if (isa<UndefValue>(*i))
12313 Result.push_back(NElts*2); // undef -> 8
12315 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12319 /// FindScalarElement - Given a vector and an element number, see if the scalar
12320 /// value is already around as a register, for example if it were inserted then
12321 /// extracted from the vector.
12322 static Value *FindScalarElement(Value *V, unsigned EltNo,
12323 LLVMContext *Context) {
12324 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12325 const VectorType *PTy = cast<VectorType>(V->getType());
12326 unsigned Width = PTy->getNumElements();
12327 if (EltNo >= Width) // Out of range access.
12328 return UndefValue::get(PTy->getElementType());
12330 if (isa<UndefValue>(V))
12331 return UndefValue::get(PTy->getElementType());
12332 else if (isa<ConstantAggregateZero>(V))
12333 return Constant::getNullValue(PTy->getElementType());
12334 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12335 return CP->getOperand(EltNo);
12336 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12337 // If this is an insert to a variable element, we don't know what it is.
12338 if (!isa<ConstantInt>(III->getOperand(2)))
12340 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12342 // If this is an insert to the element we are looking for, return the
12344 if (EltNo == IIElt)
12345 return III->getOperand(1);
12347 // Otherwise, the insertelement doesn't modify the value, recurse on its
12349 return FindScalarElement(III->getOperand(0), EltNo, Context);
12350 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12351 unsigned LHSWidth =
12352 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12353 unsigned InEl = getShuffleMask(SVI)[EltNo];
12354 if (InEl < LHSWidth)
12355 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12356 else if (InEl < LHSWidth*2)
12357 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12359 return UndefValue::get(PTy->getElementType());
12362 // Otherwise, we don't know.
12366 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12367 // If vector val is undef, replace extract with scalar undef.
12368 if (isa<UndefValue>(EI.getOperand(0)))
12369 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12371 // If vector val is constant 0, replace extract with scalar 0.
12372 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12373 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12375 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12376 // If vector val is constant with all elements the same, replace EI with
12377 // that element. When the elements are not identical, we cannot replace yet
12378 // (we do that below, but only when the index is constant).
12379 Constant *op0 = C->getOperand(0);
12380 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12381 if (C->getOperand(i) != op0) {
12386 return ReplaceInstUsesWith(EI, op0);
12389 // If extracting a specified index from the vector, see if we can recursively
12390 // find a previously computed scalar that was inserted into the vector.
12391 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12392 unsigned IndexVal = IdxC->getZExtValue();
12393 unsigned VectorWidth =
12394 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12396 // If this is extracting an invalid index, turn this into undef, to avoid
12397 // crashing the code below.
12398 if (IndexVal >= VectorWidth)
12399 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12401 // This instruction only demands the single element from the input vector.
12402 // If the input vector has a single use, simplify it based on this use
12404 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12405 APInt UndefElts(VectorWidth, 0);
12406 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12407 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12408 DemandedMask, UndefElts)) {
12409 EI.setOperand(0, V);
12414 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12415 return ReplaceInstUsesWith(EI, Elt);
12417 // If the this extractelement is directly using a bitcast from a vector of
12418 // the same number of elements, see if we can find the source element from
12419 // it. In this case, we will end up needing to bitcast the scalars.
12420 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12421 if (const VectorType *VT =
12422 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12423 if (VT->getNumElements() == VectorWidth)
12424 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12425 IndexVal, Context))
12426 return new BitCastInst(Elt, EI.getType());
12430 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12431 if (I->hasOneUse()) {
12432 // Push extractelement into predecessor operation if legal and
12433 // profitable to do so
12434 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12435 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12436 if (CheapToScalarize(BO, isConstantElt)) {
12437 ExtractElementInst *newEI0 =
12438 ExtractElementInst::Create(BO->getOperand(0), EI.getOperand(1),
12439 EI.getName()+".lhs");
12440 ExtractElementInst *newEI1 =
12441 ExtractElementInst::Create(BO->getOperand(1), EI.getOperand(1),
12442 EI.getName()+".rhs");
12443 InsertNewInstBefore(newEI0, EI);
12444 InsertNewInstBefore(newEI1, EI);
12445 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12447 } else if (isa<LoadInst>(I)) {
12449 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12450 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12451 PointerType::get(EI.getType(), AS),EI);
12452 GetElementPtrInst *GEP =
12453 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12454 cast<GEPOperator>(GEP)->setIsInBounds(true);
12455 InsertNewInstBefore(GEP, EI);
12456 return new LoadInst(GEP);
12459 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12460 // Extracting the inserted element?
12461 if (IE->getOperand(2) == EI.getOperand(1))
12462 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12463 // If the inserted and extracted elements are constants, they must not
12464 // be the same value, extract from the pre-inserted value instead.
12465 if (isa<Constant>(IE->getOperand(2)) &&
12466 isa<Constant>(EI.getOperand(1))) {
12467 AddUsesToWorkList(EI);
12468 EI.setOperand(0, IE->getOperand(0));
12471 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12472 // If this is extracting an element from a shufflevector, figure out where
12473 // it came from and extract from the appropriate input element instead.
12474 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12475 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12477 unsigned LHSWidth =
12478 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12480 if (SrcIdx < LHSWidth)
12481 Src = SVI->getOperand(0);
12482 else if (SrcIdx < LHSWidth*2) {
12483 SrcIdx -= LHSWidth;
12484 Src = SVI->getOperand(1);
12486 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12488 return ExtractElementInst::Create(Src,
12489 ConstantInt::get(Type::Int32Ty, SrcIdx, false));
12492 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12497 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12498 /// elements from either LHS or RHS, return the shuffle mask and true.
12499 /// Otherwise, return false.
12500 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12501 std::vector<Constant*> &Mask,
12502 LLVMContext *Context) {
12503 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12504 "Invalid CollectSingleShuffleElements");
12505 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12507 if (isa<UndefValue>(V)) {
12508 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12510 } else if (V == LHS) {
12511 for (unsigned i = 0; i != NumElts; ++i)
12512 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12514 } else if (V == RHS) {
12515 for (unsigned i = 0; i != NumElts; ++i)
12516 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
12518 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12519 // If this is an insert of an extract from some other vector, include it.
12520 Value *VecOp = IEI->getOperand(0);
12521 Value *ScalarOp = IEI->getOperand(1);
12522 Value *IdxOp = IEI->getOperand(2);
12524 if (!isa<ConstantInt>(IdxOp))
12526 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12528 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12529 // Okay, we can handle this if the vector we are insertinting into is
12530 // transitively ok.
12531 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12532 // If so, update the mask to reflect the inserted undef.
12533 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
12536 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12537 if (isa<ConstantInt>(EI->getOperand(1)) &&
12538 EI->getOperand(0)->getType() == V->getType()) {
12539 unsigned ExtractedIdx =
12540 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12542 // This must be extracting from either LHS or RHS.
12543 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12544 // Okay, we can handle this if the vector we are insertinting into is
12545 // transitively ok.
12546 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12547 // If so, update the mask to reflect the inserted value.
12548 if (EI->getOperand(0) == LHS) {
12549 Mask[InsertedIdx % NumElts] =
12550 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12552 assert(EI->getOperand(0) == RHS);
12553 Mask[InsertedIdx % NumElts] =
12554 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
12563 // TODO: Handle shufflevector here!
12568 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12569 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12570 /// that computes V and the LHS value of the shuffle.
12571 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12572 Value *&RHS, LLVMContext *Context) {
12573 assert(isa<VectorType>(V->getType()) &&
12574 (RHS == 0 || V->getType() == RHS->getType()) &&
12575 "Invalid shuffle!");
12576 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12578 if (isa<UndefValue>(V)) {
12579 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12581 } else if (isa<ConstantAggregateZero>(V)) {
12582 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
12584 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12585 // If this is an insert of an extract from some other vector, include it.
12586 Value *VecOp = IEI->getOperand(0);
12587 Value *ScalarOp = IEI->getOperand(1);
12588 Value *IdxOp = IEI->getOperand(2);
12590 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12591 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12592 EI->getOperand(0)->getType() == V->getType()) {
12593 unsigned ExtractedIdx =
12594 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12595 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12597 // Either the extracted from or inserted into vector must be RHSVec,
12598 // otherwise we'd end up with a shuffle of three inputs.
12599 if (EI->getOperand(0) == RHS || RHS == 0) {
12600 RHS = EI->getOperand(0);
12601 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12602 Mask[InsertedIdx % NumElts] =
12603 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
12607 if (VecOp == RHS) {
12608 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12610 // Everything but the extracted element is replaced with the RHS.
12611 for (unsigned i = 0; i != NumElts; ++i) {
12612 if (i != InsertedIdx)
12613 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
12618 // If this insertelement is a chain that comes from exactly these two
12619 // vectors, return the vector and the effective shuffle.
12620 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12622 return EI->getOperand(0);
12627 // TODO: Handle shufflevector here!
12629 // Otherwise, can't do anything fancy. Return an identity vector.
12630 for (unsigned i = 0; i != NumElts; ++i)
12631 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12635 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12636 Value *VecOp = IE.getOperand(0);
12637 Value *ScalarOp = IE.getOperand(1);
12638 Value *IdxOp = IE.getOperand(2);
12640 // Inserting an undef or into an undefined place, remove this.
12641 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12642 ReplaceInstUsesWith(IE, VecOp);
12644 // If the inserted element was extracted from some other vector, and if the
12645 // indexes are constant, try to turn this into a shufflevector operation.
12646 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12647 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12648 EI->getOperand(0)->getType() == IE.getType()) {
12649 unsigned NumVectorElts = IE.getType()->getNumElements();
12650 unsigned ExtractedIdx =
12651 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12652 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12654 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12655 return ReplaceInstUsesWith(IE, VecOp);
12657 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12658 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12660 // If we are extracting a value from a vector, then inserting it right
12661 // back into the same place, just use the input vector.
12662 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12663 return ReplaceInstUsesWith(IE, VecOp);
12665 // We could theoretically do this for ANY input. However, doing so could
12666 // turn chains of insertelement instructions into a chain of shufflevector
12667 // instructions, and right now we do not merge shufflevectors. As such,
12668 // only do this in a situation where it is clear that there is benefit.
12669 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12670 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12671 // the values of VecOp, except then one read from EIOp0.
12672 // Build a new shuffle mask.
12673 std::vector<Constant*> Mask;
12674 if (isa<UndefValue>(VecOp))
12675 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
12677 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12678 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
12681 Mask[InsertedIdx] =
12682 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12683 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12684 ConstantVector::get(Mask));
12687 // If this insertelement isn't used by some other insertelement, turn it
12688 // (and any insertelements it points to), into one big shuffle.
12689 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12690 std::vector<Constant*> Mask;
12692 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12693 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12694 // We now have a shuffle of LHS, RHS, Mask.
12695 return new ShuffleVectorInst(LHS, RHS,
12696 ConstantVector::get(Mask));
12701 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12702 APInt UndefElts(VWidth, 0);
12703 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12704 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12711 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12712 Value *LHS = SVI.getOperand(0);
12713 Value *RHS = SVI.getOperand(1);
12714 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12716 bool MadeChange = false;
12718 // Undefined shuffle mask -> undefined value.
12719 if (isa<UndefValue>(SVI.getOperand(2)))
12720 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12722 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12724 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12727 APInt UndefElts(VWidth, 0);
12728 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12729 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12730 LHS = SVI.getOperand(0);
12731 RHS = SVI.getOperand(1);
12735 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12736 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12737 if (LHS == RHS || isa<UndefValue>(LHS)) {
12738 if (isa<UndefValue>(LHS) && LHS == RHS) {
12739 // shuffle(undef,undef,mask) -> undef.
12740 return ReplaceInstUsesWith(SVI, LHS);
12743 // Remap any references to RHS to use LHS.
12744 std::vector<Constant*> Elts;
12745 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12746 if (Mask[i] >= 2*e)
12747 Elts.push_back(UndefValue::get(Type::Int32Ty));
12749 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12750 (Mask[i] < e && isa<UndefValue>(LHS))) {
12751 Mask[i] = 2*e; // Turn into undef.
12752 Elts.push_back(UndefValue::get(Type::Int32Ty));
12754 Mask[i] = Mask[i] % e; // Force to LHS.
12755 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
12759 SVI.setOperand(0, SVI.getOperand(1));
12760 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12761 SVI.setOperand(2, ConstantVector::get(Elts));
12762 LHS = SVI.getOperand(0);
12763 RHS = SVI.getOperand(1);
12767 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12768 bool isLHSID = true, isRHSID = true;
12770 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12771 if (Mask[i] >= e*2) continue; // Ignore undef values.
12772 // Is this an identity shuffle of the LHS value?
12773 isLHSID &= (Mask[i] == i);
12775 // Is this an identity shuffle of the RHS value?
12776 isRHSID &= (Mask[i]-e == i);
12779 // Eliminate identity shuffles.
12780 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12781 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12783 // If the LHS is a shufflevector itself, see if we can combine it with this
12784 // one without producing an unusual shuffle. Here we are really conservative:
12785 // we are absolutely afraid of producing a shuffle mask not in the input
12786 // program, because the code gen may not be smart enough to turn a merged
12787 // shuffle into two specific shuffles: it may produce worse code. As such,
12788 // we only merge two shuffles if the result is one of the two input shuffle
12789 // masks. In this case, merging the shuffles just removes one instruction,
12790 // which we know is safe. This is good for things like turning:
12791 // (splat(splat)) -> splat.
12792 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12793 if (isa<UndefValue>(RHS)) {
12794 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12796 std::vector<unsigned> NewMask;
12797 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12798 if (Mask[i] >= 2*e)
12799 NewMask.push_back(2*e);
12801 NewMask.push_back(LHSMask[Mask[i]]);
12803 // If the result mask is equal to the src shuffle or this shuffle mask, do
12804 // the replacement.
12805 if (NewMask == LHSMask || NewMask == Mask) {
12806 unsigned LHSInNElts =
12807 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12808 std::vector<Constant*> Elts;
12809 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12810 if (NewMask[i] >= LHSInNElts*2) {
12811 Elts.push_back(UndefValue::get(Type::Int32Ty));
12813 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
12816 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12817 LHSSVI->getOperand(1),
12818 ConstantVector::get(Elts));
12823 return MadeChange ? &SVI : 0;
12829 /// TryToSinkInstruction - Try to move the specified instruction from its
12830 /// current block into the beginning of DestBlock, which can only happen if it's
12831 /// safe to move the instruction past all of the instructions between it and the
12832 /// end of its block.
12833 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12834 assert(I->hasOneUse() && "Invariants didn't hold!");
12836 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12837 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12840 // Do not sink alloca instructions out of the entry block.
12841 if (isa<AllocaInst>(I) && I->getParent() ==
12842 &DestBlock->getParent()->getEntryBlock())
12845 // We can only sink load instructions if there is nothing between the load and
12846 // the end of block that could change the value.
12847 if (I->mayReadFromMemory()) {
12848 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12850 if (Scan->mayWriteToMemory())
12854 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12856 CopyPrecedingStopPoint(I, InsertPos);
12857 I->moveBefore(InsertPos);
12863 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12864 /// all reachable code to the worklist.
12866 /// This has a couple of tricks to make the code faster and more powerful. In
12867 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12868 /// them to the worklist (this significantly speeds up instcombine on code where
12869 /// many instructions are dead or constant). Additionally, if we find a branch
12870 /// whose condition is a known constant, we only visit the reachable successors.
12872 static void AddReachableCodeToWorklist(BasicBlock *BB,
12873 SmallPtrSet<BasicBlock*, 64> &Visited,
12875 const TargetData *TD) {
12876 SmallVector<BasicBlock*, 256> Worklist;
12877 Worklist.push_back(BB);
12879 while (!Worklist.empty()) {
12880 BB = Worklist.back();
12881 Worklist.pop_back();
12883 // We have now visited this block! If we've already been here, ignore it.
12884 if (!Visited.insert(BB)) continue;
12886 DbgInfoIntrinsic *DBI_Prev = NULL;
12887 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12888 Instruction *Inst = BBI++;
12890 // DCE instruction if trivially dead.
12891 if (isInstructionTriviallyDead(Inst)) {
12893 DOUT << "IC: DCE: " << *Inst << '\n';
12894 Inst->eraseFromParent();
12898 // ConstantProp instruction if trivially constant.
12899 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12900 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst << '\n';
12901 Inst->replaceAllUsesWith(C);
12903 Inst->eraseFromParent();
12907 // If there are two consecutive llvm.dbg.stoppoint calls then
12908 // it is likely that the optimizer deleted code in between these
12910 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12913 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12914 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12915 IC.RemoveFromWorkList(DBI_Prev);
12916 DBI_Prev->eraseFromParent();
12918 DBI_Prev = DBI_Next;
12923 IC.AddToWorkList(Inst);
12926 // Recursively visit successors. If this is a branch or switch on a
12927 // constant, only visit the reachable successor.
12928 TerminatorInst *TI = BB->getTerminator();
12929 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12930 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12931 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12932 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12933 Worklist.push_back(ReachableBB);
12936 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12937 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12938 // See if this is an explicit destination.
12939 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12940 if (SI->getCaseValue(i) == Cond) {
12941 BasicBlock *ReachableBB = SI->getSuccessor(i);
12942 Worklist.push_back(ReachableBB);
12946 // Otherwise it is the default destination.
12947 Worklist.push_back(SI->getSuccessor(0));
12952 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12953 Worklist.push_back(TI->getSuccessor(i));
12957 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12958 bool Changed = false;
12959 TD = getAnalysisIfAvailable<TargetData>();
12961 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12962 << F.getNameStr() << "\n");
12965 // Do a depth-first traversal of the function, populate the worklist with
12966 // the reachable instructions. Ignore blocks that are not reachable. Keep
12967 // track of which blocks we visit.
12968 SmallPtrSet<BasicBlock*, 64> Visited;
12969 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12971 // Do a quick scan over the function. If we find any blocks that are
12972 // unreachable, remove any instructions inside of them. This prevents
12973 // the instcombine code from having to deal with some bad special cases.
12974 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12975 if (!Visited.count(BB)) {
12976 Instruction *Term = BB->getTerminator();
12977 while (Term != BB->begin()) { // Remove instrs bottom-up
12978 BasicBlock::iterator I = Term; --I;
12980 DOUT << "IC: DCE: " << *I << '\n';
12981 // A debug intrinsic shouldn't force another iteration if we weren't
12982 // going to do one without it.
12983 if (!isa<DbgInfoIntrinsic>(I)) {
12987 if (!I->use_empty())
12988 I->replaceAllUsesWith(UndefValue::get(I->getType()));
12989 I->eraseFromParent();
12994 while (!Worklist.empty()) {
12995 Instruction *I = RemoveOneFromWorkList();
12996 if (I == 0) continue; // skip null values.
12998 // Check to see if we can DCE the instruction.
12999 if (isInstructionTriviallyDead(I)) {
13000 // Add operands to the worklist.
13001 if (I->getNumOperands() < 4)
13002 AddUsesToWorkList(*I);
13005 DOUT << "IC: DCE: " << *I << '\n';
13007 I->eraseFromParent();
13008 RemoveFromWorkList(I);
13013 // Instruction isn't dead, see if we can constant propagate it.
13014 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
13015 DOUT << "IC: ConstFold to: " << *C << " from: " << *I << '\n';
13017 // Add operands to the worklist.
13018 AddUsesToWorkList(*I);
13019 ReplaceInstUsesWith(*I, C);
13022 I->eraseFromParent();
13023 RemoveFromWorkList(I);
13029 // See if we can constant fold its operands.
13030 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
13031 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
13032 if (Constant *NewC = ConstantFoldConstantExpression(CE,
13033 F.getContext(), TD))
13040 // See if we can trivially sink this instruction to a successor basic block.
13041 if (I->hasOneUse()) {
13042 BasicBlock *BB = I->getParent();
13043 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
13044 if (UserParent != BB) {
13045 bool UserIsSuccessor = false;
13046 // See if the user is one of our successors.
13047 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13048 if (*SI == UserParent) {
13049 UserIsSuccessor = true;
13053 // If the user is one of our immediate successors, and if that successor
13054 // only has us as a predecessors (we'd have to split the critical edge
13055 // otherwise), we can keep going.
13056 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
13057 next(pred_begin(UserParent)) == pred_end(UserParent))
13058 // Okay, the CFG is simple enough, try to sink this instruction.
13059 Changed |= TryToSinkInstruction(I, UserParent);
13063 // Now that we have an instruction, try combining it to simplify it...
13067 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
13068 if (Instruction *Result = visit(*I)) {
13070 // Should we replace the old instruction with a new one?
13072 DOUT << "IC: Old = " << *I << '\n'
13073 << " New = " << *Result << '\n';
13075 // Everything uses the new instruction now.
13076 I->replaceAllUsesWith(Result);
13078 // Push the new instruction and any users onto the worklist.
13079 AddToWorkList(Result);
13080 AddUsersToWorkList(*Result);
13082 // Move the name to the new instruction first.
13083 Result->takeName(I);
13085 // Insert the new instruction into the basic block...
13086 BasicBlock *InstParent = I->getParent();
13087 BasicBlock::iterator InsertPos = I;
13089 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13090 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13093 InstParent->getInstList().insert(InsertPos, Result);
13095 // Make sure that we reprocess all operands now that we reduced their
13097 AddUsesToWorkList(*I);
13099 // Instructions can end up on the worklist more than once. Make sure
13100 // we do not process an instruction that has been deleted.
13101 RemoveFromWorkList(I);
13103 // Erase the old instruction.
13104 InstParent->getInstList().erase(I);
13107 DOUT << "IC: Mod = " << OrigI << '\n'
13108 << " New = " << *I << '\n';
13111 // If the instruction was modified, it's possible that it is now dead.
13112 // if so, remove it.
13113 if (isInstructionTriviallyDead(I)) {
13114 // Make sure we process all operands now that we are reducing their
13116 AddUsesToWorkList(*I);
13118 // Instructions may end up in the worklist more than once. Erase all
13119 // occurrences of this instruction.
13120 RemoveFromWorkList(I);
13121 I->eraseFromParent();
13124 AddUsersToWorkList(*I);
13131 assert(WorklistMap.empty() && "Worklist empty, but map not?");
13133 // Do an explicit clear, this shrinks the map if needed.
13134 WorklistMap.clear();
13139 bool InstCombiner::runOnFunction(Function &F) {
13140 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13141 Context = &F.getContext();
13143 bool EverMadeChange = false;
13145 // Iterate while there is work to do.
13146 unsigned Iteration = 0;
13147 while (DoOneIteration(F, Iteration++))
13148 EverMadeChange = true;
13149 return EverMadeChange;
13152 FunctionPass *llvm::createInstructionCombiningPass() {
13153 return new InstCombiner();