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, LLVMContext *Context) {
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, LLVMContext *Context) {
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, LLVMContext *Context) {
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(*Context, ~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 LLVMContext *Context) {
614 if (V->hasOneUse() && V->getType()->isInteger())
615 if (Instruction *I = dyn_cast<Instruction>(V)) {
616 if (I->getOpcode() == Instruction::Mul)
617 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
618 return I->getOperand(0);
619 if (I->getOpcode() == Instruction::Shl)
620 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
621 // The multiplier is really 1 << CST.
622 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
623 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
624 CST = ConstantInt::get(*Context, APInt(BitWidth, 1).shl(CSTVal));
625 return I->getOperand(0);
631 /// AddOne - Add one to a ConstantInt
632 static Constant *AddOne(Constant *C, LLVMContext *Context) {
633 return ConstantExpr::getAdd(C,
634 ConstantInt::get(C->getType(), 1));
636 /// SubOne - Subtract one from a ConstantInt
637 static Constant *SubOne(ConstantInt *C, LLVMContext *Context) {
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 LLVMContext *Context) {
645 uint32_t W = C1->getBitWidth();
646 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
655 APInt MulExt = LHSExt * RHSExt;
658 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
659 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
660 return MulExt.slt(Min) || MulExt.sgt(Max);
662 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
666 /// ShrinkDemandedConstant - Check to see if the specified operand of the
667 /// specified instruction is a constant integer. If so, check to see if there
668 /// are any bits set in the constant that are not demanded. If so, shrink the
669 /// constant and return true.
670 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
671 APInt Demanded, LLVMContext *Context) {
672 assert(I && "No instruction?");
673 assert(OpNo < I->getNumOperands() && "Operand index too large");
675 // If the operand is not a constant integer, nothing to do.
676 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
677 if (!OpC) return false;
679 // If there are no bits set that aren't demanded, nothing to do.
680 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
681 if ((~Demanded & OpC->getValue()) == 0)
684 // This instruction is producing bits that are not demanded. Shrink the RHS.
685 Demanded &= OpC->getValue();
686 I->setOperand(OpNo, ConstantInt::get(*Context, Demanded));
690 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
691 // set of known zero and one bits, compute the maximum and minimum values that
692 // could have the specified known zero and known one bits, returning them in
694 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
695 const APInt& KnownOne,
696 APInt& Min, APInt& Max) {
697 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
698 KnownZero.getBitWidth() == Min.getBitWidth() &&
699 KnownZero.getBitWidth() == Max.getBitWidth() &&
700 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
701 APInt UnknownBits = ~(KnownZero|KnownOne);
703 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
704 // bit if it is unknown.
706 Max = KnownOne|UnknownBits;
708 if (UnknownBits.isNegative()) { // Sign bit is unknown
709 Min.set(Min.getBitWidth()-1);
710 Max.clear(Max.getBitWidth()-1);
714 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
715 // a set of known zero and one bits, compute the maximum and minimum values that
716 // could have the specified known zero and known one bits, returning them in
718 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
719 const APInt &KnownOne,
720 APInt &Min, APInt &Max) {
721 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
722 KnownZero.getBitWidth() == Min.getBitWidth() &&
723 KnownZero.getBitWidth() == Max.getBitWidth() &&
724 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
725 APInt UnknownBits = ~(KnownZero|KnownOne);
727 // The minimum value is when the unknown bits are all zeros.
729 // The maximum value is when the unknown bits are all ones.
730 Max = KnownOne|UnknownBits;
733 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
734 /// SimplifyDemandedBits knows about. See if the instruction has any
735 /// properties that allow us to simplify its operands.
736 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
737 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
738 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
739 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
741 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
742 KnownZero, KnownOne, 0);
743 if (V == 0) return false;
744 if (V == &Inst) return true;
745 ReplaceInstUsesWith(Inst, V);
749 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
750 /// specified instruction operand if possible, updating it in place. It returns
751 /// true if it made any change and false otherwise.
752 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
753 APInt &KnownZero, APInt &KnownOne,
755 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
756 KnownZero, KnownOne, Depth);
757 if (NewVal == 0) return false;
763 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
764 /// value based on the demanded bits. When this function is called, it is known
765 /// that only the bits set in DemandedMask of the result of V are ever used
766 /// downstream. Consequently, depending on the mask and V, it may be possible
767 /// to replace V with a constant or one of its operands. In such cases, this
768 /// function does the replacement and returns true. In all other cases, it
769 /// returns false after analyzing the expression and setting KnownOne and known
770 /// to be one in the expression. KnownZero contains all the bits that are known
771 /// to be zero in the expression. These are provided to potentially allow the
772 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
773 /// the expression. KnownOne and KnownZero always follow the invariant that
774 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
775 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
776 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
777 /// and KnownOne must all be the same.
779 /// This returns null if it did not change anything and it permits no
780 /// simplification. This returns V itself if it did some simplification of V's
781 /// operands based on the information about what bits are demanded. This returns
782 /// some other non-null value if it found out that V is equal to another value
783 /// in the context where the specified bits are demanded, but not for all users.
784 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
785 APInt &KnownZero, APInt &KnownOne,
787 assert(V != 0 && "Null pointer of Value???");
788 assert(Depth <= 6 && "Limit Search Depth");
789 uint32_t BitWidth = DemandedMask.getBitWidth();
790 const Type *VTy = V->getType();
791 assert((TD || !isa<PointerType>(VTy)) &&
792 "SimplifyDemandedBits needs to know bit widths!");
793 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
794 (!VTy->isIntOrIntVector() ||
795 VTy->getScalarSizeInBits() == BitWidth) &&
796 KnownZero.getBitWidth() == BitWidth &&
797 KnownOne.getBitWidth() == BitWidth &&
798 "Value *V, DemandedMask, KnownZero and KnownOne "
799 "must have same BitWidth");
800 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
801 // We know all of the bits for a constant!
802 KnownOne = CI->getValue() & DemandedMask;
803 KnownZero = ~KnownOne & DemandedMask;
806 if (isa<ConstantPointerNull>(V)) {
807 // We know all of the bits for a constant!
809 KnownZero = DemandedMask;
815 if (DemandedMask == 0) { // Not demanding any bits from V.
816 if (isa<UndefValue>(V))
818 return UndefValue::get(VTy);
821 if (Depth == 6) // Limit search depth.
824 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
825 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
827 Instruction *I = dyn_cast<Instruction>(V);
829 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
830 return 0; // Only analyze instructions.
833 // If there are multiple uses of this value and we aren't at the root, then
834 // we can't do any simplifications of the operands, because DemandedMask
835 // only reflects the bits demanded by *one* of the users.
836 if (Depth != 0 && !I->hasOneUse()) {
837 // Despite the fact that we can't simplify this instruction in all User's
838 // context, we can at least compute the knownzero/knownone bits, and we can
839 // do simplifications that apply to *just* the one user if we know that
840 // this instruction has a simpler value in that context.
841 if (I->getOpcode() == Instruction::And) {
842 // If either the LHS or the RHS are Zero, the result is zero.
843 ComputeMaskedBits(I->getOperand(1), DemandedMask,
844 RHSKnownZero, RHSKnownOne, Depth+1);
845 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
846 LHSKnownZero, LHSKnownOne, Depth+1);
848 // If all of the demanded bits are known 1 on one side, return the other.
849 // These bits cannot contribute to the result of the 'and' in this
851 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
852 (DemandedMask & ~LHSKnownZero))
853 return I->getOperand(0);
854 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
855 (DemandedMask & ~RHSKnownZero))
856 return I->getOperand(1);
858 // If all of the demanded bits in the inputs are known zeros, return zero.
859 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
860 return Constant::getNullValue(VTy);
862 } else if (I->getOpcode() == Instruction::Or) {
863 // We can simplify (X|Y) -> X or Y in the user's context if we know that
864 // only bits from X or Y are demanded.
866 // If either the LHS or the RHS are One, the result is One.
867 ComputeMaskedBits(I->getOperand(1), DemandedMask,
868 RHSKnownZero, RHSKnownOne, Depth+1);
869 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
870 LHSKnownZero, LHSKnownOne, Depth+1);
872 // If all of the demanded bits are known zero on one side, return the
873 // other. These bits cannot contribute to the result of the 'or' in this
875 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
876 (DemandedMask & ~LHSKnownOne))
877 return I->getOperand(0);
878 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
879 (DemandedMask & ~RHSKnownOne))
880 return I->getOperand(1);
882 // If all of the potentially set bits on one side are known to be set on
883 // the other side, just use the 'other' side.
884 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
885 (DemandedMask & (~RHSKnownZero)))
886 return I->getOperand(0);
887 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
888 (DemandedMask & (~LHSKnownZero)))
889 return I->getOperand(1);
892 // Compute the KnownZero/KnownOne bits to simplify things downstream.
893 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
897 // If this is the root being simplified, allow it to have multiple uses,
898 // just set the DemandedMask to all bits so that we can try to simplify the
899 // operands. This allows visitTruncInst (for example) to simplify the
900 // operand of a trunc without duplicating all the logic below.
901 if (Depth == 0 && !V->hasOneUse())
902 DemandedMask = APInt::getAllOnesValue(BitWidth);
904 switch (I->getOpcode()) {
906 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
908 case Instruction::And:
909 // If either the LHS or the RHS are Zero, the result is zero.
910 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
911 RHSKnownZero, RHSKnownOne, Depth+1) ||
912 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
913 LHSKnownZero, LHSKnownOne, Depth+1))
915 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
916 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
918 // If all of the demanded bits are known 1 on one side, return the other.
919 // These bits cannot contribute to the result of the 'and'.
920 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
921 (DemandedMask & ~LHSKnownZero))
922 return I->getOperand(0);
923 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
924 (DemandedMask & ~RHSKnownZero))
925 return I->getOperand(1);
927 // If all of the demanded bits in the inputs are known zeros, return zero.
928 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
929 return Constant::getNullValue(VTy);
931 // If the RHS is a constant, see if we can simplify it.
932 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero, Context))
935 // Output known-1 bits are only known if set in both the LHS & RHS.
936 RHSKnownOne &= LHSKnownOne;
937 // Output known-0 are known to be clear if zero in either the LHS | RHS.
938 RHSKnownZero |= LHSKnownZero;
940 case Instruction::Or:
941 // If either the LHS or the RHS are One, the result is One.
942 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
943 RHSKnownZero, RHSKnownOne, Depth+1) ||
944 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
945 LHSKnownZero, LHSKnownOne, Depth+1))
947 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
948 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
950 // If all of the demanded bits are known zero on one side, return the other.
951 // These bits cannot contribute to the result of the 'or'.
952 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
953 (DemandedMask & ~LHSKnownOne))
954 return I->getOperand(0);
955 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
956 (DemandedMask & ~RHSKnownOne))
957 return I->getOperand(1);
959 // If all of the potentially set bits on one side are known to be set on
960 // the other side, just use the 'other' side.
961 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
962 (DemandedMask & (~RHSKnownZero)))
963 return I->getOperand(0);
964 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
965 (DemandedMask & (~LHSKnownZero)))
966 return I->getOperand(1);
968 // If the RHS is a constant, see if we can simplify it.
969 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context))
972 // Output known-0 bits are only known if clear in both the LHS & RHS.
973 RHSKnownZero &= LHSKnownZero;
974 // Output known-1 are known to be set if set in either the LHS | RHS.
975 RHSKnownOne |= LHSKnownOne;
977 case Instruction::Xor: {
978 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
979 RHSKnownZero, RHSKnownOne, Depth+1) ||
980 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
981 LHSKnownZero, LHSKnownOne, Depth+1))
983 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
984 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
986 // If all of the demanded bits are known zero on one side, return the other.
987 // These bits cannot contribute to the result of the 'xor'.
988 if ((DemandedMask & RHSKnownZero) == DemandedMask)
989 return I->getOperand(0);
990 if ((DemandedMask & LHSKnownZero) == DemandedMask)
991 return I->getOperand(1);
993 // Output known-0 bits are known if clear or set in both the LHS & RHS.
994 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
995 (RHSKnownOne & LHSKnownOne);
996 // Output known-1 are known to be set if set in only one of the LHS, RHS.
997 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
998 (RHSKnownOne & LHSKnownZero);
1000 // If all of the demanded bits are known to be zero on one side or the
1001 // other, turn this into an *inclusive* or.
1002 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1003 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1005 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1007 return InsertNewInstBefore(Or, *I);
1010 // If all of the demanded bits on one side are known, and all of the set
1011 // bits on that side are also known to be set on the other side, turn this
1012 // into an AND, as we know the bits will be cleared.
1013 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1014 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1016 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1017 Constant *AndC = Constant::getIntegerValue(VTy,
1018 ~RHSKnownOne & DemandedMask);
1020 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1021 return InsertNewInstBefore(And, *I);
1025 // If the RHS is a constant, see if we can simplify it.
1026 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1027 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context))
1030 RHSKnownZero = KnownZeroOut;
1031 RHSKnownOne = KnownOneOut;
1034 case Instruction::Select:
1035 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1036 RHSKnownZero, RHSKnownOne, Depth+1) ||
1037 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1038 LHSKnownZero, LHSKnownOne, Depth+1))
1040 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1041 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1043 // If the operands are constants, see if we can simplify them.
1044 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context) ||
1045 ShrinkDemandedConstant(I, 2, DemandedMask, Context))
1048 // Only known if known in both the LHS and RHS.
1049 RHSKnownOne &= LHSKnownOne;
1050 RHSKnownZero &= LHSKnownZero;
1052 case Instruction::Trunc: {
1053 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1054 DemandedMask.zext(truncBf);
1055 RHSKnownZero.zext(truncBf);
1056 RHSKnownOne.zext(truncBf);
1057 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1058 RHSKnownZero, RHSKnownOne, Depth+1))
1060 DemandedMask.trunc(BitWidth);
1061 RHSKnownZero.trunc(BitWidth);
1062 RHSKnownOne.trunc(BitWidth);
1063 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1066 case Instruction::BitCast:
1067 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1068 return false; // vector->int or fp->int?
1070 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1071 if (const VectorType *SrcVTy =
1072 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1073 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1074 // Don't touch a bitcast between vectors of different element counts.
1077 // Don't touch a scalar-to-vector bitcast.
1079 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1080 // Don't touch a vector-to-scalar bitcast.
1083 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1084 RHSKnownZero, RHSKnownOne, Depth+1))
1086 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1088 case Instruction::ZExt: {
1089 // Compute the bits in the result that are not present in the input.
1090 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1092 DemandedMask.trunc(SrcBitWidth);
1093 RHSKnownZero.trunc(SrcBitWidth);
1094 RHSKnownOne.trunc(SrcBitWidth);
1095 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1096 RHSKnownZero, RHSKnownOne, Depth+1))
1098 DemandedMask.zext(BitWidth);
1099 RHSKnownZero.zext(BitWidth);
1100 RHSKnownOne.zext(BitWidth);
1101 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1102 // The top bits are known to be zero.
1103 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1106 case Instruction::SExt: {
1107 // Compute the bits in the result that are not present in the input.
1108 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1110 APInt InputDemandedBits = DemandedMask &
1111 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1113 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1114 // If any of the sign extended bits are demanded, we know that the sign
1116 if ((NewBits & DemandedMask) != 0)
1117 InputDemandedBits.set(SrcBitWidth-1);
1119 InputDemandedBits.trunc(SrcBitWidth);
1120 RHSKnownZero.trunc(SrcBitWidth);
1121 RHSKnownOne.trunc(SrcBitWidth);
1122 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1123 RHSKnownZero, RHSKnownOne, Depth+1))
1125 InputDemandedBits.zext(BitWidth);
1126 RHSKnownZero.zext(BitWidth);
1127 RHSKnownOne.zext(BitWidth);
1128 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1130 // If the sign bit of the input is known set or clear, then we know the
1131 // top bits of the result.
1133 // If the input sign bit is known zero, or if the NewBits are not demanded
1134 // convert this into a zero extension.
1135 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1136 // Convert to ZExt cast
1137 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1138 return InsertNewInstBefore(NewCast, *I);
1139 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1140 RHSKnownOne |= NewBits;
1144 case Instruction::Add: {
1145 // Figure out what the input bits are. If the top bits of the and result
1146 // are not demanded, then the add doesn't demand them from its input
1148 unsigned NLZ = DemandedMask.countLeadingZeros();
1150 // If there is a constant on the RHS, there are a variety of xformations
1152 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1153 // If null, this should be simplified elsewhere. Some of the xforms here
1154 // won't work if the RHS is zero.
1158 // If the top bit of the output is demanded, demand everything from the
1159 // input. Otherwise, we demand all the input bits except NLZ top bits.
1160 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1162 // Find information about known zero/one bits in the input.
1163 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1164 LHSKnownZero, LHSKnownOne, Depth+1))
1167 // If the RHS of the add has bits set that can't affect the input, reduce
1169 if (ShrinkDemandedConstant(I, 1, InDemandedBits, Context))
1172 // Avoid excess work.
1173 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1176 // Turn it into OR if input bits are zero.
1177 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1179 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1181 return InsertNewInstBefore(Or, *I);
1184 // We can say something about the output known-zero and known-one bits,
1185 // depending on potential carries from the input constant and the
1186 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1187 // bits set and the RHS constant is 0x01001, then we know we have a known
1188 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1190 // To compute this, we first compute the potential carry bits. These are
1191 // the bits which may be modified. I'm not aware of a better way to do
1193 const APInt &RHSVal = RHS->getValue();
1194 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1196 // Now that we know which bits have carries, compute the known-1/0 sets.
1198 // Bits are known one if they are known zero in one operand and one in the
1199 // other, and there is no input carry.
1200 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1201 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1203 // Bits are known zero if they are known zero in both operands and there
1204 // is no input carry.
1205 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1207 // If the high-bits of this ADD are not demanded, then it does not demand
1208 // the high bits of its LHS or RHS.
1209 if (DemandedMask[BitWidth-1] == 0) {
1210 // Right fill the mask of bits for this ADD to demand the most
1211 // significant bit and all those below it.
1212 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1213 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1214 LHSKnownZero, LHSKnownOne, Depth+1) ||
1215 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1216 LHSKnownZero, LHSKnownOne, Depth+1))
1222 case Instruction::Sub:
1223 // If the high-bits of this SUB are not demanded, then it does not demand
1224 // the high bits of its LHS or RHS.
1225 if (DemandedMask[BitWidth-1] == 0) {
1226 // Right fill the mask of bits for this SUB to demand the most
1227 // significant bit and all those below it.
1228 uint32_t NLZ = DemandedMask.countLeadingZeros();
1229 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1230 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1231 LHSKnownZero, LHSKnownOne, Depth+1) ||
1232 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1233 LHSKnownZero, LHSKnownOne, Depth+1))
1236 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1237 // the known zeros and ones.
1238 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1240 case Instruction::Shl:
1241 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1242 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1243 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1244 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1245 RHSKnownZero, RHSKnownOne, Depth+1))
1247 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1248 RHSKnownZero <<= ShiftAmt;
1249 RHSKnownOne <<= ShiftAmt;
1250 // low bits known zero.
1252 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1255 case Instruction::LShr:
1256 // For a logical shift right
1257 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1258 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1260 // Unsigned shift right.
1261 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1262 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1263 RHSKnownZero, RHSKnownOne, Depth+1))
1265 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1266 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1267 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1269 // Compute the new bits that are at the top now.
1270 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1271 RHSKnownZero |= HighBits; // high bits known zero.
1275 case Instruction::AShr:
1276 // If this is an arithmetic shift right and only the low-bit is set, we can
1277 // always convert this into a logical shr, even if the shift amount is
1278 // variable. The low bit of the shift cannot be an input sign bit unless
1279 // the shift amount is >= the size of the datatype, which is undefined.
1280 if (DemandedMask == 1) {
1281 // Perform the logical shift right.
1282 Instruction *NewVal = BinaryOperator::CreateLShr(
1283 I->getOperand(0), I->getOperand(1), I->getName());
1284 return InsertNewInstBefore(NewVal, *I);
1287 // If the sign bit is the only bit demanded by this ashr, then there is no
1288 // need to do it, the shift doesn't change the high bit.
1289 if (DemandedMask.isSignBit())
1290 return I->getOperand(0);
1292 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1293 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1295 // Signed shift right.
1296 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1297 // If any of the "high bits" are demanded, we should set the sign bit as
1299 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1300 DemandedMaskIn.set(BitWidth-1);
1301 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1302 RHSKnownZero, RHSKnownOne, Depth+1))
1304 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1305 // Compute the new bits that are at the top now.
1306 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1307 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1308 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1310 // Handle the sign bits.
1311 APInt SignBit(APInt::getSignBit(BitWidth));
1312 // Adjust to where it is now in the mask.
1313 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1315 // If the input sign bit is known to be zero, or if none of the top bits
1316 // are demanded, turn this into an unsigned shift right.
1317 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1318 (HighBits & ~DemandedMask) == HighBits) {
1319 // Perform the logical shift right.
1320 Instruction *NewVal = BinaryOperator::CreateLShr(
1321 I->getOperand(0), SA, I->getName());
1322 return InsertNewInstBefore(NewVal, *I);
1323 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1324 RHSKnownOne |= HighBits;
1328 case Instruction::SRem:
1329 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1330 APInt RA = Rem->getValue().abs();
1331 if (RA.isPowerOf2()) {
1332 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1333 return I->getOperand(0);
1335 APInt LowBits = RA - 1;
1336 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1337 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1338 LHSKnownZero, LHSKnownOne, Depth+1))
1341 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1342 LHSKnownZero |= ~LowBits;
1344 KnownZero |= LHSKnownZero & DemandedMask;
1346 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1350 case Instruction::URem: {
1351 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1352 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1353 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1354 KnownZero2, KnownOne2, Depth+1) ||
1355 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1356 KnownZero2, KnownOne2, Depth+1))
1359 unsigned Leaders = KnownZero2.countLeadingOnes();
1360 Leaders = std::max(Leaders,
1361 KnownZero2.countLeadingOnes());
1362 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1365 case Instruction::Call:
1366 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1367 switch (II->getIntrinsicID()) {
1369 case Intrinsic::bswap: {
1370 // If the only bits demanded come from one byte of the bswap result,
1371 // just shift the input byte into position to eliminate the bswap.
1372 unsigned NLZ = DemandedMask.countLeadingZeros();
1373 unsigned NTZ = DemandedMask.countTrailingZeros();
1375 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1376 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1377 // have 14 leading zeros, round to 8.
1380 // If we need exactly one byte, we can do this transformation.
1381 if (BitWidth-NLZ-NTZ == 8) {
1382 unsigned ResultBit = NTZ;
1383 unsigned InputBit = BitWidth-NTZ-8;
1385 // Replace this with either a left or right shift to get the byte into
1387 Instruction *NewVal;
1388 if (InputBit > ResultBit)
1389 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1390 ConstantInt::get(I->getType(), InputBit-ResultBit));
1392 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1393 ConstantInt::get(I->getType(), ResultBit-InputBit));
1394 NewVal->takeName(I);
1395 return InsertNewInstBefore(NewVal, *I);
1398 // TODO: Could compute known zero/one bits based on the input.
1403 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1407 // If the client is only demanding bits that we know, return the known
1409 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1410 return Constant::getIntegerValue(VTy, RHSKnownOne);
1415 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1416 /// any number of elements. DemandedElts contains the set of elements that are
1417 /// actually used by the caller. This method analyzes which elements of the
1418 /// operand are undef and returns that information in UndefElts.
1420 /// If the information about demanded elements can be used to simplify the
1421 /// operation, the operation is simplified, then the resultant value is
1422 /// returned. This returns null if no change was made.
1423 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1426 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1427 APInt EltMask(APInt::getAllOnesValue(VWidth));
1428 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1430 if (isa<UndefValue>(V)) {
1431 // If the entire vector is undefined, just return this info.
1432 UndefElts = EltMask;
1434 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1435 UndefElts = EltMask;
1436 return UndefValue::get(V->getType());
1440 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1441 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1442 Constant *Undef = UndefValue::get(EltTy);
1444 std::vector<Constant*> Elts;
1445 for (unsigned i = 0; i != VWidth; ++i)
1446 if (!DemandedElts[i]) { // If not demanded, set to undef.
1447 Elts.push_back(Undef);
1449 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1450 Elts.push_back(Undef);
1452 } else { // Otherwise, defined.
1453 Elts.push_back(CP->getOperand(i));
1456 // If we changed the constant, return it.
1457 Constant *NewCP = ConstantVector::get(Elts);
1458 return NewCP != CP ? NewCP : 0;
1459 } else if (isa<ConstantAggregateZero>(V)) {
1460 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1463 // Check if this is identity. If so, return 0 since we are not simplifying
1465 if (DemandedElts == ((1ULL << VWidth) -1))
1468 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1469 Constant *Zero = Constant::getNullValue(EltTy);
1470 Constant *Undef = UndefValue::get(EltTy);
1471 std::vector<Constant*> Elts;
1472 for (unsigned i = 0; i != VWidth; ++i) {
1473 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1474 Elts.push_back(Elt);
1476 UndefElts = DemandedElts ^ EltMask;
1477 return ConstantVector::get(Elts);
1480 // Limit search depth.
1484 // If multiple users are using the root value, procede with
1485 // simplification conservatively assuming that all elements
1487 if (!V->hasOneUse()) {
1488 // Quit if we find multiple users of a non-root value though.
1489 // They'll be handled when it's their turn to be visited by
1490 // the main instcombine process.
1492 // TODO: Just compute the UndefElts information recursively.
1495 // Conservatively assume that all elements are needed.
1496 DemandedElts = EltMask;
1499 Instruction *I = dyn_cast<Instruction>(V);
1500 if (!I) return 0; // Only analyze instructions.
1502 bool MadeChange = false;
1503 APInt UndefElts2(VWidth, 0);
1505 switch (I->getOpcode()) {
1508 case Instruction::InsertElement: {
1509 // If this is a variable index, we don't know which element it overwrites.
1510 // demand exactly the same input as we produce.
1511 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1513 // Note that we can't propagate undef elt info, because we don't know
1514 // which elt is getting updated.
1515 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1516 UndefElts2, Depth+1);
1517 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1521 // If this is inserting an element that isn't demanded, remove this
1523 unsigned IdxNo = Idx->getZExtValue();
1524 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1525 return AddSoonDeadInstToWorklist(*I, 0);
1527 // Otherwise, the element inserted overwrites whatever was there, so the
1528 // input demanded set is simpler than the output set.
1529 APInt DemandedElts2 = DemandedElts;
1530 DemandedElts2.clear(IdxNo);
1531 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1532 UndefElts, Depth+1);
1533 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1535 // The inserted element is defined.
1536 UndefElts.clear(IdxNo);
1539 case Instruction::ShuffleVector: {
1540 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1541 uint64_t LHSVWidth =
1542 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1543 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1544 for (unsigned i = 0; i < VWidth; i++) {
1545 if (DemandedElts[i]) {
1546 unsigned MaskVal = Shuffle->getMaskValue(i);
1547 if (MaskVal != -1u) {
1548 assert(MaskVal < LHSVWidth * 2 &&
1549 "shufflevector mask index out of range!");
1550 if (MaskVal < LHSVWidth)
1551 LeftDemanded.set(MaskVal);
1553 RightDemanded.set(MaskVal - LHSVWidth);
1558 APInt UndefElts4(LHSVWidth, 0);
1559 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1560 UndefElts4, Depth+1);
1561 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1563 APInt UndefElts3(LHSVWidth, 0);
1564 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1565 UndefElts3, Depth+1);
1566 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1568 bool NewUndefElts = false;
1569 for (unsigned i = 0; i < VWidth; i++) {
1570 unsigned MaskVal = Shuffle->getMaskValue(i);
1571 if (MaskVal == -1u) {
1573 } else if (MaskVal < LHSVWidth) {
1574 if (UndefElts4[MaskVal]) {
1575 NewUndefElts = true;
1579 if (UndefElts3[MaskVal - LHSVWidth]) {
1580 NewUndefElts = true;
1587 // Add additional discovered undefs.
1588 std::vector<Constant*> Elts;
1589 for (unsigned i = 0; i < VWidth; ++i) {
1591 Elts.push_back(UndefValue::get(Type::Int32Ty));
1593 Elts.push_back(ConstantInt::get(Type::Int32Ty,
1594 Shuffle->getMaskValue(i)));
1596 I->setOperand(2, ConstantVector::get(Elts));
1601 case Instruction::BitCast: {
1602 // Vector->vector casts only.
1603 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1605 unsigned InVWidth = VTy->getNumElements();
1606 APInt InputDemandedElts(InVWidth, 0);
1609 if (VWidth == InVWidth) {
1610 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1611 // elements as are demanded of us.
1613 InputDemandedElts = DemandedElts;
1614 } else if (VWidth > InVWidth) {
1618 // If there are more elements in the result than there are in the source,
1619 // then an input element is live if any of the corresponding output
1620 // elements are live.
1621 Ratio = VWidth/InVWidth;
1622 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1623 if (DemandedElts[OutIdx])
1624 InputDemandedElts.set(OutIdx/Ratio);
1630 // If there are more elements in the source than there are in the result,
1631 // then an input element is live if the corresponding output element is
1633 Ratio = InVWidth/VWidth;
1634 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1635 if (DemandedElts[InIdx/Ratio])
1636 InputDemandedElts.set(InIdx);
1639 // div/rem demand all inputs, because they don't want divide by zero.
1640 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1641 UndefElts2, Depth+1);
1643 I->setOperand(0, TmpV);
1647 UndefElts = UndefElts2;
1648 if (VWidth > InVWidth) {
1649 llvm_unreachable("Unimp");
1650 // If there are more elements in the result than there are in the source,
1651 // then an output element is undef if the corresponding input element is
1653 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1654 if (UndefElts2[OutIdx/Ratio])
1655 UndefElts.set(OutIdx);
1656 } else if (VWidth < InVWidth) {
1657 llvm_unreachable("Unimp");
1658 // If there are more elements in the source than there are in the result,
1659 // then a result element is undef if all of the corresponding input
1660 // elements are undef.
1661 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1662 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1663 if (!UndefElts2[InIdx]) // Not undef?
1664 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1668 case Instruction::And:
1669 case Instruction::Or:
1670 case Instruction::Xor:
1671 case Instruction::Add:
1672 case Instruction::Sub:
1673 case Instruction::Mul:
1674 // div/rem demand all inputs, because they don't want divide by zero.
1675 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1676 UndefElts, Depth+1);
1677 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1678 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1679 UndefElts2, Depth+1);
1680 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1682 // Output elements are undefined if both are undefined. Consider things
1683 // like undef&0. The result is known zero, not undef.
1684 UndefElts &= UndefElts2;
1687 case Instruction::Call: {
1688 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1690 switch (II->getIntrinsicID()) {
1693 // Binary vector operations that work column-wise. A dest element is a
1694 // function of the corresponding input elements from the two inputs.
1695 case Intrinsic::x86_sse_sub_ss:
1696 case Intrinsic::x86_sse_mul_ss:
1697 case Intrinsic::x86_sse_min_ss:
1698 case Intrinsic::x86_sse_max_ss:
1699 case Intrinsic::x86_sse2_sub_sd:
1700 case Intrinsic::x86_sse2_mul_sd:
1701 case Intrinsic::x86_sse2_min_sd:
1702 case Intrinsic::x86_sse2_max_sd:
1703 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1704 UndefElts, Depth+1);
1705 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1706 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1707 UndefElts2, Depth+1);
1708 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1710 // If only the low elt is demanded and this is a scalarizable intrinsic,
1711 // scalarize it now.
1712 if (DemandedElts == 1) {
1713 switch (II->getIntrinsicID()) {
1715 case Intrinsic::x86_sse_sub_ss:
1716 case Intrinsic::x86_sse_mul_ss:
1717 case Intrinsic::x86_sse2_sub_sd:
1718 case Intrinsic::x86_sse2_mul_sd:
1719 // TODO: Lower MIN/MAX/ABS/etc
1720 Value *LHS = II->getOperand(1);
1721 Value *RHS = II->getOperand(2);
1722 // Extract the element as scalars.
1723 LHS = InsertNewInstBefore(ExtractElementInst::Create(LHS,
1724 ConstantInt::get(Type::Int32Ty, 0U, false), "tmp"), *II);
1725 RHS = InsertNewInstBefore(ExtractElementInst::Create(RHS,
1726 ConstantInt::get(Type::Int32Ty, 0U, false), "tmp"), *II);
1728 switch (II->getIntrinsicID()) {
1729 default: llvm_unreachable("Case stmts out of sync!");
1730 case Intrinsic::x86_sse_sub_ss:
1731 case Intrinsic::x86_sse2_sub_sd:
1732 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1733 II->getName()), *II);
1735 case Intrinsic::x86_sse_mul_ss:
1736 case Intrinsic::x86_sse2_mul_sd:
1737 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1738 II->getName()), *II);
1743 InsertElementInst::Create(
1744 UndefValue::get(II->getType()), TmpV,
1745 ConstantInt::get(Type::Int32Ty, 0U, false), II->getName());
1746 InsertNewInstBefore(New, *II);
1747 AddSoonDeadInstToWorklist(*II, 0);
1752 // Output elements are undefined if both are undefined. Consider things
1753 // like undef&0. The result is known zero, not undef.
1754 UndefElts &= UndefElts2;
1760 return MadeChange ? I : 0;
1764 /// AssociativeOpt - Perform an optimization on an associative operator. This
1765 /// function is designed to check a chain of associative operators for a
1766 /// potential to apply a certain optimization. Since the optimization may be
1767 /// applicable if the expression was reassociated, this checks the chain, then
1768 /// reassociates the expression as necessary to expose the optimization
1769 /// opportunity. This makes use of a special Functor, which must define
1770 /// 'shouldApply' and 'apply' methods.
1772 template<typename Functor>
1773 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F,
1774 LLVMContext *Context) {
1775 unsigned Opcode = Root.getOpcode();
1776 Value *LHS = Root.getOperand(0);
1778 // Quick check, see if the immediate LHS matches...
1779 if (F.shouldApply(LHS))
1780 return F.apply(Root);
1782 // Otherwise, if the LHS is not of the same opcode as the root, return.
1783 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1784 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1785 // Should we apply this transform to the RHS?
1786 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1788 // If not to the RHS, check to see if we should apply to the LHS...
1789 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1790 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1794 // If the functor wants to apply the optimization to the RHS of LHSI,
1795 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1797 // Now all of the instructions are in the current basic block, go ahead
1798 // and perform the reassociation.
1799 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1801 // First move the selected RHS to the LHS of the root...
1802 Root.setOperand(0, LHSI->getOperand(1));
1804 // Make what used to be the LHS of the root be the user of the root...
1805 Value *ExtraOperand = TmpLHSI->getOperand(1);
1806 if (&Root == TmpLHSI) {
1807 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1810 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1811 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1812 BasicBlock::iterator ARI = &Root; ++ARI;
1813 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1816 // Now propagate the ExtraOperand down the chain of instructions until we
1818 while (TmpLHSI != LHSI) {
1819 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1820 // Move the instruction to immediately before the chain we are
1821 // constructing to avoid breaking dominance properties.
1822 NextLHSI->moveBefore(ARI);
1825 Value *NextOp = NextLHSI->getOperand(1);
1826 NextLHSI->setOperand(1, ExtraOperand);
1828 ExtraOperand = NextOp;
1831 // Now that the instructions are reassociated, have the functor perform
1832 // the transformation...
1833 return F.apply(Root);
1836 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1843 // AddRHS - Implements: X + X --> X << 1
1846 LLVMContext *Context;
1847 AddRHS(Value *rhs, LLVMContext *C) : RHS(rhs), Context(C) {}
1848 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1849 Instruction *apply(BinaryOperator &Add) const {
1850 return BinaryOperator::CreateShl(Add.getOperand(0),
1851 ConstantInt::get(Add.getType(), 1));
1855 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1857 struct AddMaskingAnd {
1859 LLVMContext *Context;
1860 AddMaskingAnd(Constant *c, LLVMContext *C) : C2(c), Context(C) {}
1861 bool shouldApply(Value *LHS) const {
1863 return match(LHS, m_And(m_Value(), m_ConstantInt(C1)), *Context) &&
1864 ConstantExpr::getAnd(C1, C2)->isNullValue();
1866 Instruction *apply(BinaryOperator &Add) const {
1867 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1873 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1875 LLVMContext *Context = IC->getContext();
1877 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1878 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1881 // Figure out if the constant is the left or the right argument.
1882 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1883 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1885 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1887 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1888 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1891 Value *Op0 = SO, *Op1 = ConstOperand;
1893 std::swap(Op0, Op1);
1895 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1896 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1897 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1898 New = CmpInst::Create(*Context, CI->getOpcode(), CI->getPredicate(),
1899 Op0, Op1, SO->getName()+".cmp");
1901 llvm_unreachable("Unknown binary instruction type!");
1903 return IC->InsertNewInstBefore(New, I);
1906 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1907 // constant as the other operand, try to fold the binary operator into the
1908 // select arguments. This also works for Cast instructions, which obviously do
1909 // not have a second operand.
1910 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1912 // Don't modify shared select instructions
1913 if (!SI->hasOneUse()) return 0;
1914 Value *TV = SI->getOperand(1);
1915 Value *FV = SI->getOperand(2);
1917 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1918 // Bool selects with constant operands can be folded to logical ops.
1919 if (SI->getType() == Type::Int1Ty) return 0;
1921 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1922 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1924 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1931 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1932 /// node as operand #0, see if we can fold the instruction into the PHI (which
1933 /// is only possible if all operands to the PHI are constants).
1934 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1935 PHINode *PN = cast<PHINode>(I.getOperand(0));
1936 unsigned NumPHIValues = PN->getNumIncomingValues();
1937 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1939 // Check to see if all of the operands of the PHI are constants. If there is
1940 // one non-constant value, remember the BB it is. If there is more than one
1941 // or if *it* is a PHI, bail out.
1942 BasicBlock *NonConstBB = 0;
1943 for (unsigned i = 0; i != NumPHIValues; ++i)
1944 if (!isa<Constant>(PN->getIncomingValue(i))) {
1945 if (NonConstBB) return 0; // More than one non-const value.
1946 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1947 NonConstBB = PN->getIncomingBlock(i);
1949 // If the incoming non-constant value is in I's block, we have an infinite
1951 if (NonConstBB == I.getParent())
1955 // If there is exactly one non-constant value, we can insert a copy of the
1956 // operation in that block. However, if this is a critical edge, we would be
1957 // inserting the computation one some other paths (e.g. inside a loop). Only
1958 // do this if the pred block is unconditionally branching into the phi block.
1960 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1961 if (!BI || !BI->isUnconditional()) return 0;
1964 // Okay, we can do the transformation: create the new PHI node.
1965 PHINode *NewPN = PHINode::Create(I.getType(), "");
1966 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1967 InsertNewInstBefore(NewPN, *PN);
1968 NewPN->takeName(PN);
1970 // Next, add all of the operands to the PHI.
1971 if (I.getNumOperands() == 2) {
1972 Constant *C = cast<Constant>(I.getOperand(1));
1973 for (unsigned i = 0; i != NumPHIValues; ++i) {
1975 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1976 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1977 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1979 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1981 assert(PN->getIncomingBlock(i) == NonConstBB);
1982 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1983 InV = BinaryOperator::Create(BO->getOpcode(),
1984 PN->getIncomingValue(i), C, "phitmp",
1985 NonConstBB->getTerminator());
1986 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1987 InV = CmpInst::Create(*Context, CI->getOpcode(),
1989 PN->getIncomingValue(i), C, "phitmp",
1990 NonConstBB->getTerminator());
1992 llvm_unreachable("Unknown binop!");
1994 AddToWorkList(cast<Instruction>(InV));
1996 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1999 CastInst *CI = cast<CastInst>(&I);
2000 const Type *RetTy = CI->getType();
2001 for (unsigned i = 0; i != NumPHIValues; ++i) {
2003 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2004 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2006 assert(PN->getIncomingBlock(i) == NonConstBB);
2007 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2008 I.getType(), "phitmp",
2009 NonConstBB->getTerminator());
2010 AddToWorkList(cast<Instruction>(InV));
2012 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2015 return ReplaceInstUsesWith(I, NewPN);
2019 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2020 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2021 /// This basically requires proving that the add in the original type would not
2022 /// overflow to change the sign bit or have a carry out.
2023 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2024 // There are different heuristics we can use for this. Here are some simple
2027 // Add has the property that adding any two 2's complement numbers can only
2028 // have one carry bit which can change a sign. As such, if LHS and RHS each
2029 // have at least two sign bits, we know that the addition of the two values will
2030 // sign extend fine.
2031 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2035 // If one of the operands only has one non-zero bit, and if the other operand
2036 // has a known-zero bit in a more significant place than it (not including the
2037 // sign bit) the ripple may go up to and fill the zero, but won't change the
2038 // sign. For example, (X & ~4) + 1.
2046 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2047 bool Changed = SimplifyCommutative(I);
2048 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2050 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2051 // X + undef -> undef
2052 if (isa<UndefValue>(RHS))
2053 return ReplaceInstUsesWith(I, RHS);
2056 if (RHSC->isNullValue())
2057 return ReplaceInstUsesWith(I, LHS);
2059 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2060 // X + (signbit) --> X ^ signbit
2061 const APInt& Val = CI->getValue();
2062 uint32_t BitWidth = Val.getBitWidth();
2063 if (Val == APInt::getSignBit(BitWidth))
2064 return BinaryOperator::CreateXor(LHS, RHS);
2066 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2067 // (X & 254)+1 -> (X&254)|1
2068 if (SimplifyDemandedInstructionBits(I))
2071 // zext(bool) + C -> bool ? C + 1 : C
2072 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2073 if (ZI->getSrcTy() == Type::Int1Ty)
2074 return SelectInst::Create(ZI->getOperand(0), AddOne(CI, Context), CI);
2077 if (isa<PHINode>(LHS))
2078 if (Instruction *NV = FoldOpIntoPhi(I))
2081 ConstantInt *XorRHS = 0;
2083 if (isa<ConstantInt>(RHSC) &&
2084 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)), *Context)) {
2085 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2086 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2088 uint32_t Size = TySizeBits / 2;
2089 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2090 APInt CFF80Val(-C0080Val);
2092 if (TySizeBits > Size) {
2093 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2094 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2095 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2096 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2097 // This is a sign extend if the top bits are known zero.
2098 if (!MaskedValueIsZero(XorLHS,
2099 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2100 Size = 0; // Not a sign ext, but can't be any others either.
2105 C0080Val = APIntOps::lshr(C0080Val, Size);
2106 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2107 } while (Size >= 1);
2109 // FIXME: This shouldn't be necessary. When the backends can handle types
2110 // with funny bit widths then this switch statement should be removed. It
2111 // is just here to get the size of the "middle" type back up to something
2112 // that the back ends can handle.
2113 const Type *MiddleType = 0;
2116 case 32: MiddleType = Type::Int32Ty; break;
2117 case 16: MiddleType = Type::Int16Ty; break;
2118 case 8: MiddleType = Type::Int8Ty; break;
2121 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2122 InsertNewInstBefore(NewTrunc, I);
2123 return new SExtInst(NewTrunc, I.getType(), I.getName());
2128 if (I.getType() == Type::Int1Ty)
2129 return BinaryOperator::CreateXor(LHS, RHS);
2132 if (I.getType()->isInteger()) {
2133 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS, Context), Context))
2136 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2137 if (RHSI->getOpcode() == Instruction::Sub)
2138 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2139 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2141 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2142 if (LHSI->getOpcode() == Instruction::Sub)
2143 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2144 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2149 // -A + -B --> -(A + B)
2150 if (Value *LHSV = dyn_castNegVal(LHS, Context)) {
2151 if (LHS->getType()->isIntOrIntVector()) {
2152 if (Value *RHSV = dyn_castNegVal(RHS, Context)) {
2153 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2154 InsertNewInstBefore(NewAdd, I);
2155 return BinaryOperator::CreateNeg(*Context, NewAdd);
2159 return BinaryOperator::CreateSub(RHS, LHSV);
2163 if (!isa<Constant>(RHS))
2164 if (Value *V = dyn_castNegVal(RHS, Context))
2165 return BinaryOperator::CreateSub(LHS, V);
2169 if (Value *X = dyn_castFoldableMul(LHS, C2, Context)) {
2170 if (X == RHS) // X*C + X --> X * (C+1)
2171 return BinaryOperator::CreateMul(RHS, AddOne(C2, Context));
2173 // X*C1 + X*C2 --> X * (C1+C2)
2175 if (X == dyn_castFoldableMul(RHS, C1, Context))
2176 return BinaryOperator::CreateMul(X, ConstantExpr::getAdd(C1, C2));
2179 // X + X*C --> X * (C+1)
2180 if (dyn_castFoldableMul(RHS, C2, Context) == LHS)
2181 return BinaryOperator::CreateMul(LHS, AddOne(C2, Context));
2183 // X + ~X --> -1 since ~X = -X-1
2184 if (dyn_castNotVal(LHS, Context) == RHS ||
2185 dyn_castNotVal(RHS, Context) == LHS)
2186 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2189 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2190 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2)), *Context))
2191 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2, Context), Context))
2194 // A+B --> A|B iff A and B have no bits set in common.
2195 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2196 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2197 APInt LHSKnownOne(IT->getBitWidth(), 0);
2198 APInt LHSKnownZero(IT->getBitWidth(), 0);
2199 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2200 if (LHSKnownZero != 0) {
2201 APInt RHSKnownOne(IT->getBitWidth(), 0);
2202 APInt RHSKnownZero(IT->getBitWidth(), 0);
2203 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2205 // No bits in common -> bitwise or.
2206 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2207 return BinaryOperator::CreateOr(LHS, RHS);
2211 // W*X + Y*Z --> W * (X+Z) iff W == Y
2212 if (I.getType()->isIntOrIntVector()) {
2213 Value *W, *X, *Y, *Z;
2214 if (match(LHS, m_Mul(m_Value(W), m_Value(X)), *Context) &&
2215 match(RHS, m_Mul(m_Value(Y), m_Value(Z)), *Context)) {
2219 } else if (Y == X) {
2221 } else if (X == Z) {
2228 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2229 LHS->getName()), I);
2230 return BinaryOperator::CreateMul(W, NewAdd);
2235 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2237 if (match(LHS, m_Not(m_Value(X)), *Context)) // ~X + C --> (C-1) - X
2238 return BinaryOperator::CreateSub(SubOne(CRHS, Context), X);
2240 // (X & FF00) + xx00 -> (X+xx00) & FF00
2241 if (LHS->hasOneUse() &&
2242 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)), *Context)) {
2243 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
2244 if (Anded == CRHS) {
2245 // See if all bits from the first bit set in the Add RHS up are included
2246 // in the mask. First, get the rightmost bit.
2247 const APInt& AddRHSV = CRHS->getValue();
2249 // Form a mask of all bits from the lowest bit added through the top.
2250 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2252 // See if the and mask includes all of these bits.
2253 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2255 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2256 // Okay, the xform is safe. Insert the new add pronto.
2257 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2258 LHS->getName()), I);
2259 return BinaryOperator::CreateAnd(NewAdd, C2);
2264 // Try to fold constant add into select arguments.
2265 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2266 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2270 // add (select X 0 (sub n A)) A --> select X A n
2272 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2275 SI = dyn_cast<SelectInst>(RHS);
2278 if (SI && SI->hasOneUse()) {
2279 Value *TV = SI->getTrueValue();
2280 Value *FV = SI->getFalseValue();
2283 // Can we fold the add into the argument of the select?
2284 // We check both true and false select arguments for a matching subtract.
2285 if (match(FV, m_Zero(), *Context) &&
2286 match(TV, m_Sub(m_Value(N), m_Specific(A)), *Context))
2287 // Fold the add into the true select value.
2288 return SelectInst::Create(SI->getCondition(), N, A);
2289 if (match(TV, m_Zero(), *Context) &&
2290 match(FV, m_Sub(m_Value(N), m_Specific(A)), *Context))
2291 // Fold the add into the false select value.
2292 return SelectInst::Create(SI->getCondition(), A, N);
2296 // Check for (add (sext x), y), see if we can merge this into an
2297 // integer add followed by a sext.
2298 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2299 // (add (sext x), cst) --> (sext (add x, cst'))
2300 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2302 ConstantExpr::getTrunc(RHSC, LHSConv->getOperand(0)->getType());
2303 if (LHSConv->hasOneUse() &&
2304 ConstantExpr::getSExt(CI, I.getType()) == RHSC &&
2305 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2306 // Insert the new, smaller add.
2307 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2309 InsertNewInstBefore(NewAdd, I);
2310 return new SExtInst(NewAdd, I.getType());
2314 // (add (sext x), (sext y)) --> (sext (add int x, y))
2315 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2316 // Only do this if x/y have the same type, if at last one of them has a
2317 // single use (so we don't increase the number of sexts), and if the
2318 // integer add will not overflow.
2319 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2320 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2321 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2322 RHSConv->getOperand(0))) {
2323 // Insert the new integer add.
2324 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2325 RHSConv->getOperand(0),
2327 InsertNewInstBefore(NewAdd, I);
2328 return new SExtInst(NewAdd, I.getType());
2333 return Changed ? &I : 0;
2336 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2337 bool Changed = SimplifyCommutative(I);
2338 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2340 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2342 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2343 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2344 (I.getType())->getValueAPF()))
2345 return ReplaceInstUsesWith(I, LHS);
2348 if (isa<PHINode>(LHS))
2349 if (Instruction *NV = FoldOpIntoPhi(I))
2354 // -A + -B --> -(A + B)
2355 if (Value *LHSV = dyn_castFNegVal(LHS, Context))
2356 return BinaryOperator::CreateFSub(RHS, LHSV);
2359 if (!isa<Constant>(RHS))
2360 if (Value *V = dyn_castFNegVal(RHS, Context))
2361 return BinaryOperator::CreateFSub(LHS, V);
2363 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2364 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2365 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2366 return ReplaceInstUsesWith(I, LHS);
2368 // Check for (add double (sitofp x), y), see if we can merge this into an
2369 // integer add followed by a promotion.
2370 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2371 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2372 // ... if the constant fits in the integer value. This is useful for things
2373 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2374 // requires a constant pool load, and generally allows the add to be better
2376 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2378 ConstantExpr::getFPToSI(CFP, LHSConv->getOperand(0)->getType());
2379 if (LHSConv->hasOneUse() &&
2380 ConstantExpr::getSIToFP(CI, I.getType()) == CFP &&
2381 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2382 // Insert the new integer add.
2383 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2385 InsertNewInstBefore(NewAdd, I);
2386 return new SIToFPInst(NewAdd, I.getType());
2390 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2391 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2392 // Only do this if x/y have the same type, if at last one of them has a
2393 // single use (so we don't increase the number of int->fp conversions),
2394 // and if the integer add will not overflow.
2395 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2396 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2397 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2398 RHSConv->getOperand(0))) {
2399 // Insert the new integer add.
2400 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2401 RHSConv->getOperand(0),
2403 InsertNewInstBefore(NewAdd, I);
2404 return new SIToFPInst(NewAdd, I.getType());
2409 return Changed ? &I : 0;
2412 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2413 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2415 if (Op0 == Op1) // sub X, X -> 0
2416 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2418 // If this is a 'B = x-(-A)', change to B = x+A...
2419 if (Value *V = dyn_castNegVal(Op1, Context))
2420 return BinaryOperator::CreateAdd(Op0, V);
2422 if (isa<UndefValue>(Op0))
2423 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2424 if (isa<UndefValue>(Op1))
2425 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2427 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2428 // Replace (-1 - A) with (~A)...
2429 if (C->isAllOnesValue())
2430 return BinaryOperator::CreateNot(*Context, Op1);
2432 // C - ~X == X + (1+C)
2434 if (match(Op1, m_Not(m_Value(X)), *Context))
2435 return BinaryOperator::CreateAdd(X, AddOne(C, Context));
2437 // -(X >>u 31) -> (X >>s 31)
2438 // -(X >>s 31) -> (X >>u 31)
2440 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2441 if (SI->getOpcode() == Instruction::LShr) {
2442 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2443 // Check to see if we are shifting out everything but the sign bit.
2444 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2445 SI->getType()->getPrimitiveSizeInBits()-1) {
2446 // Ok, the transformation is safe. Insert AShr.
2447 return BinaryOperator::Create(Instruction::AShr,
2448 SI->getOperand(0), CU, SI->getName());
2452 else if (SI->getOpcode() == Instruction::AShr) {
2453 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2454 // Check to see if we are shifting out everything but the sign bit.
2455 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2456 SI->getType()->getPrimitiveSizeInBits()-1) {
2457 // Ok, the transformation is safe. Insert LShr.
2458 return BinaryOperator::CreateLShr(
2459 SI->getOperand(0), CU, SI->getName());
2466 // Try to fold constant sub into select arguments.
2467 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2468 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2471 // C - zext(bool) -> bool ? C - 1 : C
2472 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2473 if (ZI->getSrcTy() == Type::Int1Ty)
2474 return SelectInst::Create(ZI->getOperand(0), SubOne(C, Context), C);
2477 if (I.getType() == Type::Int1Ty)
2478 return BinaryOperator::CreateXor(Op0, Op1);
2480 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2481 if (Op1I->getOpcode() == Instruction::Add) {
2482 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2483 return BinaryOperator::CreateNeg(*Context, Op1I->getOperand(1),
2485 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2486 return BinaryOperator::CreateNeg(*Context, Op1I->getOperand(0),
2488 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2489 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2490 // C1-(X+C2) --> (C1-C2)-X
2491 return BinaryOperator::CreateSub(
2492 ConstantExpr::getSub(CI1, CI2), Op1I->getOperand(0));
2496 if (Op1I->hasOneUse()) {
2497 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2498 // is not used by anyone else...
2500 if (Op1I->getOpcode() == Instruction::Sub) {
2501 // Swap the two operands of the subexpr...
2502 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2503 Op1I->setOperand(0, IIOp1);
2504 Op1I->setOperand(1, IIOp0);
2506 // Create the new top level add instruction...
2507 return BinaryOperator::CreateAdd(Op0, Op1);
2510 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2512 if (Op1I->getOpcode() == Instruction::And &&
2513 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2514 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2517 InsertNewInstBefore(BinaryOperator::CreateNot(*Context,
2518 OtherOp, "B.not"), I);
2519 return BinaryOperator::CreateAnd(Op0, NewNot);
2522 // 0 - (X sdiv C) -> (X sdiv -C)
2523 if (Op1I->getOpcode() == Instruction::SDiv)
2524 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2526 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2527 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2528 ConstantExpr::getNeg(DivRHS));
2530 // X - X*C --> X * (1-C)
2531 ConstantInt *C2 = 0;
2532 if (dyn_castFoldableMul(Op1I, C2, Context) == Op0) {
2534 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1),
2536 return BinaryOperator::CreateMul(Op0, CP1);
2541 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2542 if (Op0I->getOpcode() == Instruction::Add) {
2543 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2544 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2545 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2546 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2547 } else if (Op0I->getOpcode() == Instruction::Sub) {
2548 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2549 return BinaryOperator::CreateNeg(*Context, Op0I->getOperand(1),
2555 if (Value *X = dyn_castFoldableMul(Op0, C1, Context)) {
2556 if (X == Op1) // X*C - X --> X * (C-1)
2557 return BinaryOperator::CreateMul(Op1, SubOne(C1, Context));
2559 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2560 if (X == dyn_castFoldableMul(Op1, C2, Context))
2561 return BinaryOperator::CreateMul(X, ConstantExpr::getSub(C1, C2));
2566 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2567 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2569 // If this is a 'B = x-(-A)', change to B = x+A...
2570 if (Value *V = dyn_castFNegVal(Op1, Context))
2571 return BinaryOperator::CreateFAdd(Op0, V);
2573 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2574 if (Op1I->getOpcode() == Instruction::FAdd) {
2575 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2576 return BinaryOperator::CreateFNeg(*Context, Op1I->getOperand(1),
2578 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2579 return BinaryOperator::CreateFNeg(*Context, Op1I->getOperand(0),
2587 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2588 /// comparison only checks the sign bit. If it only checks the sign bit, set
2589 /// TrueIfSigned if the result of the comparison is true when the input value is
2591 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2592 bool &TrueIfSigned) {
2594 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2595 TrueIfSigned = true;
2596 return RHS->isZero();
2597 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2598 TrueIfSigned = true;
2599 return RHS->isAllOnesValue();
2600 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2601 TrueIfSigned = false;
2602 return RHS->isAllOnesValue();
2603 case ICmpInst::ICMP_UGT:
2604 // True if LHS u> RHS and RHS == high-bit-mask - 1
2605 TrueIfSigned = true;
2606 return RHS->getValue() ==
2607 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2608 case ICmpInst::ICMP_UGE:
2609 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2610 TrueIfSigned = true;
2611 return RHS->getValue().isSignBit();
2617 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2618 bool Changed = SimplifyCommutative(I);
2619 Value *Op0 = I.getOperand(0);
2621 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2622 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2624 // Simplify mul instructions with a constant RHS...
2625 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2626 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2628 // ((X << C1)*C2) == (X * (C2 << C1))
2629 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2630 if (SI->getOpcode() == Instruction::Shl)
2631 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2632 return BinaryOperator::CreateMul(SI->getOperand(0),
2633 ConstantExpr::getShl(CI, ShOp));
2636 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2637 if (CI->equalsInt(1)) // X * 1 == X
2638 return ReplaceInstUsesWith(I, Op0);
2639 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2640 return BinaryOperator::CreateNeg(*Context, Op0, I.getName());
2642 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2643 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2644 return BinaryOperator::CreateShl(Op0,
2645 ConstantInt::get(Op0->getType(), Val.logBase2()));
2647 } else if (isa<VectorType>(Op1->getType())) {
2648 if (Op1->isNullValue())
2649 return ReplaceInstUsesWith(I, Op1);
2651 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2652 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2653 return BinaryOperator::CreateNeg(*Context, Op0, I.getName());
2655 // As above, vector X*splat(1.0) -> X in all defined cases.
2656 if (Constant *Splat = Op1V->getSplatValue()) {
2657 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2658 if (CI->equalsInt(1))
2659 return ReplaceInstUsesWith(I, Op0);
2664 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2665 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2666 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2667 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2668 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2670 InsertNewInstBefore(Add, I);
2671 Value *C1C2 = ConstantExpr::getMul(Op1,
2672 cast<Constant>(Op0I->getOperand(1)));
2673 return BinaryOperator::CreateAdd(Add, C1C2);
2677 // Try to fold constant mul into select arguments.
2678 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2679 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2682 if (isa<PHINode>(Op0))
2683 if (Instruction *NV = FoldOpIntoPhi(I))
2687 if (Value *Op0v = dyn_castNegVal(Op0, Context)) // -X * -Y = X*Y
2688 if (Value *Op1v = dyn_castNegVal(I.getOperand(1), Context))
2689 return BinaryOperator::CreateMul(Op0v, Op1v);
2691 // (X / Y) * Y = X - (X % Y)
2692 // (X / Y) * -Y = (X % Y) - X
2694 Value *Op1 = I.getOperand(1);
2695 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2697 (BO->getOpcode() != Instruction::UDiv &&
2698 BO->getOpcode() != Instruction::SDiv)) {
2700 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2702 Value *Neg = dyn_castNegVal(Op1, Context);
2703 if (BO && BO->hasOneUse() &&
2704 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2705 (BO->getOpcode() == Instruction::UDiv ||
2706 BO->getOpcode() == Instruction::SDiv)) {
2707 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2710 if (BO->getOpcode() == Instruction::UDiv)
2711 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2713 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2715 InsertNewInstBefore(Rem, I);
2719 return BinaryOperator::CreateSub(Op0BO, Rem);
2721 return BinaryOperator::CreateSub(Rem, Op0BO);
2725 if (I.getType() == Type::Int1Ty)
2726 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2728 // If one of the operands of the multiply is a cast from a boolean value, then
2729 // we know the bool is either zero or one, so this is a 'masking' multiply.
2730 // See if we can simplify things based on how the boolean was originally
2732 CastInst *BoolCast = 0;
2733 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2734 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2737 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2738 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2741 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2742 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2743 const Type *SCOpTy = SCIOp0->getType();
2746 // If the icmp is true iff the sign bit of X is set, then convert this
2747 // multiply into a shift/and combination.
2748 if (isa<ConstantInt>(SCIOp1) &&
2749 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2751 // Shift the X value right to turn it into "all signbits".
2752 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2753 SCOpTy->getPrimitiveSizeInBits()-1);
2755 InsertNewInstBefore(
2756 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2757 BoolCast->getOperand(0)->getName()+
2760 // If the multiply type is not the same as the source type, sign extend
2761 // or truncate to the multiply type.
2762 if (I.getType() != V->getType()) {
2763 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2764 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2765 Instruction::CastOps opcode =
2766 (SrcBits == DstBits ? Instruction::BitCast :
2767 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2768 V = InsertCastBefore(opcode, V, I.getType(), I);
2771 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2772 return BinaryOperator::CreateAnd(V, OtherOp);
2777 return Changed ? &I : 0;
2780 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2781 bool Changed = SimplifyCommutative(I);
2782 Value *Op0 = I.getOperand(0);
2784 // Simplify mul instructions with a constant RHS...
2785 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2786 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2787 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2788 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2789 if (Op1F->isExactlyValue(1.0))
2790 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2791 } else if (isa<VectorType>(Op1->getType())) {
2792 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2793 // As above, vector X*splat(1.0) -> X in all defined cases.
2794 if (Constant *Splat = Op1V->getSplatValue()) {
2795 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2796 if (F->isExactlyValue(1.0))
2797 return ReplaceInstUsesWith(I, Op0);
2802 // Try to fold constant mul into select arguments.
2803 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2804 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2807 if (isa<PHINode>(Op0))
2808 if (Instruction *NV = FoldOpIntoPhi(I))
2812 if (Value *Op0v = dyn_castFNegVal(Op0, Context)) // -X * -Y = X*Y
2813 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1), Context))
2814 return BinaryOperator::CreateFMul(Op0v, Op1v);
2816 return Changed ? &I : 0;
2819 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2821 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2822 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2824 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2825 int NonNullOperand = -1;
2826 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2827 if (ST->isNullValue())
2829 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2830 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2831 if (ST->isNullValue())
2834 if (NonNullOperand == -1)
2837 Value *SelectCond = SI->getOperand(0);
2839 // Change the div/rem to use 'Y' instead of the select.
2840 I.setOperand(1, SI->getOperand(NonNullOperand));
2842 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2843 // problem. However, the select, or the condition of the select may have
2844 // multiple uses. Based on our knowledge that the operand must be non-zero,
2845 // propagate the known value for the select into other uses of it, and
2846 // propagate a known value of the condition into its other users.
2848 // If the select and condition only have a single use, don't bother with this,
2850 if (SI->use_empty() && SelectCond->hasOneUse())
2853 // Scan the current block backward, looking for other uses of SI.
2854 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2856 while (BBI != BBFront) {
2858 // If we found a call to a function, we can't assume it will return, so
2859 // information from below it cannot be propagated above it.
2860 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2863 // Replace uses of the select or its condition with the known values.
2864 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2867 *I = SI->getOperand(NonNullOperand);
2869 } else if (*I == SelectCond) {
2870 *I = NonNullOperand == 1 ? ConstantInt::getTrue(*Context) :
2871 ConstantInt::getFalse(*Context);
2876 // If we past the instruction, quit looking for it.
2879 if (&*BBI == SelectCond)
2882 // If we ran out of things to eliminate, break out of the loop.
2883 if (SelectCond == 0 && SI == 0)
2891 /// This function implements the transforms on div instructions that work
2892 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2893 /// used by the visitors to those instructions.
2894 /// @brief Transforms common to all three div instructions
2895 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2896 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2898 // undef / X -> 0 for integer.
2899 // undef / X -> undef for FP (the undef could be a snan).
2900 if (isa<UndefValue>(Op0)) {
2901 if (Op0->getType()->isFPOrFPVector())
2902 return ReplaceInstUsesWith(I, Op0);
2903 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2906 // X / undef -> undef
2907 if (isa<UndefValue>(Op1))
2908 return ReplaceInstUsesWith(I, Op1);
2913 /// This function implements the transforms common to both integer division
2914 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2915 /// division instructions.
2916 /// @brief Common integer divide transforms
2917 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2918 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2920 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2922 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2923 Constant *CI = ConstantInt::get(Ty->getElementType(), 1);
2924 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2925 return ReplaceInstUsesWith(I, ConstantVector::get(Elts));
2928 Constant *CI = ConstantInt::get(I.getType(), 1);
2929 return ReplaceInstUsesWith(I, CI);
2932 if (Instruction *Common = commonDivTransforms(I))
2935 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2936 // This does not apply for fdiv.
2937 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2940 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2942 if (RHS->equalsInt(1))
2943 return ReplaceInstUsesWith(I, Op0);
2945 // (X / C1) / C2 -> X / (C1*C2)
2946 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2947 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2948 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2949 if (MultiplyOverflows(RHS, LHSRHS,
2950 I.getOpcode()==Instruction::SDiv, Context))
2951 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2953 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2954 ConstantExpr::getMul(RHS, LHSRHS));
2957 if (!RHS->isZero()) { // avoid X udiv 0
2958 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2959 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2961 if (isa<PHINode>(Op0))
2962 if (Instruction *NV = FoldOpIntoPhi(I))
2967 // 0 / X == 0, we don't need to preserve faults!
2968 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2969 if (LHS->equalsInt(0))
2970 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2972 // It can't be division by zero, hence it must be division by one.
2973 if (I.getType() == Type::Int1Ty)
2974 return ReplaceInstUsesWith(I, Op0);
2976 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2977 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
2980 return ReplaceInstUsesWith(I, Op0);
2986 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2987 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2989 // Handle the integer div common cases
2990 if (Instruction *Common = commonIDivTransforms(I))
2993 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2994 // X udiv C^2 -> X >> C
2995 // Check to see if this is an unsigned division with an exact power of 2,
2996 // if so, convert to a right shift.
2997 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2998 return BinaryOperator::CreateLShr(Op0,
2999 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3001 // X udiv C, where C >= signbit
3002 if (C->getValue().isNegative()) {
3003 Value *IC = InsertNewInstBefore(new ICmpInst(*Context,
3004 ICmpInst::ICMP_ULT, Op0, C),
3006 return SelectInst::Create(IC, Constant::getNullValue(I.getType()),
3007 ConstantInt::get(I.getType(), 1));
3011 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3012 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3013 if (RHSI->getOpcode() == Instruction::Shl &&
3014 isa<ConstantInt>(RHSI->getOperand(0))) {
3015 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3016 if (C1.isPowerOf2()) {
3017 Value *N = RHSI->getOperand(1);
3018 const Type *NTy = N->getType();
3019 if (uint32_t C2 = C1.logBase2()) {
3020 Constant *C2V = ConstantInt::get(NTy, C2);
3021 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3023 return BinaryOperator::CreateLShr(Op0, N);
3028 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3029 // where C1&C2 are powers of two.
3030 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3031 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3032 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3033 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3034 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3035 // Compute the shift amounts
3036 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3037 // Construct the "on true" case of the select
3038 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3039 Instruction *TSI = BinaryOperator::CreateLShr(
3040 Op0, TC, SI->getName()+".t");
3041 TSI = InsertNewInstBefore(TSI, I);
3043 // Construct the "on false" case of the select
3044 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3045 Instruction *FSI = BinaryOperator::CreateLShr(
3046 Op0, FC, SI->getName()+".f");
3047 FSI = InsertNewInstBefore(FSI, I);
3049 // construct the select instruction and return it.
3050 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3056 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3057 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3059 // Handle the integer div common cases
3060 if (Instruction *Common = commonIDivTransforms(I))
3063 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3065 if (RHS->isAllOnesValue())
3066 return BinaryOperator::CreateNeg(*Context, Op0);
3068 // sdiv X, C --> ashr X, log2(C)
3069 if (cast<SDivOperator>(&I)->isExact() &&
3070 RHS->getValue().isNonNegative() &&
3071 RHS->getValue().isPowerOf2()) {
3072 Value *ShAmt = llvm::ConstantInt::get(RHS->getType(),
3073 RHS->getValue().exactLogBase2());
3074 return BinaryOperator::CreateAShr(Op0, ShAmt, I.getName());
3078 // If the sign bits of both operands are zero (i.e. we can prove they are
3079 // unsigned inputs), turn this into a udiv.
3080 if (I.getType()->isInteger()) {
3081 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3082 if (MaskedValueIsZero(Op0, Mask)) {
3083 if (MaskedValueIsZero(Op1, Mask)) {
3084 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3085 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3087 ConstantInt *ShiftedInt;
3088 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value()), *Context) &&
3089 ShiftedInt->getValue().isPowerOf2()) {
3090 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3091 // Safe because the only negative value (1 << Y) can take on is
3092 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3093 // the sign bit set.
3094 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3102 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3103 return commonDivTransforms(I);
3106 /// This function implements the transforms on rem instructions that work
3107 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3108 /// is used by the visitors to those instructions.
3109 /// @brief Transforms common to all three rem instructions
3110 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3111 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3113 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3114 if (I.getType()->isFPOrFPVector())
3115 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3116 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3118 if (isa<UndefValue>(Op1))
3119 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3121 // Handle cases involving: rem X, (select Cond, Y, Z)
3122 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3128 /// This function implements the transforms common to both integer remainder
3129 /// instructions (urem and srem). It is called by the visitors to those integer
3130 /// remainder instructions.
3131 /// @brief Common integer remainder transforms
3132 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3133 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3135 if (Instruction *common = commonRemTransforms(I))
3138 // 0 % X == 0 for integer, we don't need to preserve faults!
3139 if (Constant *LHS = dyn_cast<Constant>(Op0))
3140 if (LHS->isNullValue())
3141 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3143 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3144 // X % 0 == undef, we don't need to preserve faults!
3145 if (RHS->equalsInt(0))
3146 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3148 if (RHS->equalsInt(1)) // X % 1 == 0
3149 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3151 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3152 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3153 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3155 } else if (isa<PHINode>(Op0I)) {
3156 if (Instruction *NV = FoldOpIntoPhi(I))
3160 // See if we can fold away this rem instruction.
3161 if (SimplifyDemandedInstructionBits(I))
3169 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3170 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3172 if (Instruction *common = commonIRemTransforms(I))
3175 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3176 // X urem C^2 -> X and C
3177 // Check to see if this is an unsigned remainder with an exact power of 2,
3178 // if so, convert to a bitwise and.
3179 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3180 if (C->getValue().isPowerOf2())
3181 return BinaryOperator::CreateAnd(Op0, SubOne(C, Context));
3184 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3185 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3186 if (RHSI->getOpcode() == Instruction::Shl &&
3187 isa<ConstantInt>(RHSI->getOperand(0))) {
3188 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3189 Constant *N1 = Constant::getAllOnesValue(I.getType());
3190 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3192 return BinaryOperator::CreateAnd(Op0, Add);
3197 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3198 // where C1&C2 are powers of two.
3199 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3200 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3201 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3202 // STO == 0 and SFO == 0 handled above.
3203 if ((STO->getValue().isPowerOf2()) &&
3204 (SFO->getValue().isPowerOf2())) {
3205 Value *TrueAnd = InsertNewInstBefore(
3206 BinaryOperator::CreateAnd(Op0, SubOne(STO, Context),
3207 SI->getName()+".t"), I);
3208 Value *FalseAnd = InsertNewInstBefore(
3209 BinaryOperator::CreateAnd(Op0, SubOne(SFO, Context),
3210 SI->getName()+".f"), I);
3211 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3219 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3220 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3222 // Handle the integer rem common cases
3223 if (Instruction *common = commonIRemTransforms(I))
3226 if (Value *RHSNeg = dyn_castNegVal(Op1, Context))
3227 if (!isa<Constant>(RHSNeg) ||
3228 (isa<ConstantInt>(RHSNeg) &&
3229 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3231 AddUsesToWorkList(I);
3232 I.setOperand(1, RHSNeg);
3236 // If the sign bits of both operands are zero (i.e. we can prove they are
3237 // unsigned inputs), turn this into a urem.
3238 if (I.getType()->isInteger()) {
3239 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3240 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3241 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3242 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3246 // If it's a constant vector, flip any negative values positive.
3247 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3248 unsigned VWidth = RHSV->getNumOperands();
3250 bool hasNegative = false;
3251 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3252 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3253 if (RHS->getValue().isNegative())
3257 std::vector<Constant *> Elts(VWidth);
3258 for (unsigned i = 0; i != VWidth; ++i) {
3259 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3260 if (RHS->getValue().isNegative())
3261 Elts[i] = cast<ConstantInt>(ConstantExpr::getNeg(RHS));
3267 Constant *NewRHSV = ConstantVector::get(Elts);
3268 if (NewRHSV != RHSV) {
3269 AddUsesToWorkList(I);
3270 I.setOperand(1, NewRHSV);
3279 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3280 return commonRemTransforms(I);
3283 // isOneBitSet - Return true if there is exactly one bit set in the specified
3285 static bool isOneBitSet(const ConstantInt *CI) {
3286 return CI->getValue().isPowerOf2();
3289 // isHighOnes - Return true if the constant is of the form 1+0+.
3290 // This is the same as lowones(~X).
3291 static bool isHighOnes(const ConstantInt *CI) {
3292 return (~CI->getValue() + 1).isPowerOf2();
3295 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3296 /// are carefully arranged to allow folding of expressions such as:
3298 /// (A < B) | (A > B) --> (A != B)
3300 /// Note that this is only valid if the first and second predicates have the
3301 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3303 /// Three bits are used to represent the condition, as follows:
3308 /// <=> Value Definition
3309 /// 000 0 Always false
3316 /// 111 7 Always true
3318 static unsigned getICmpCode(const ICmpInst *ICI) {
3319 switch (ICI->getPredicate()) {
3321 case ICmpInst::ICMP_UGT: return 1; // 001
3322 case ICmpInst::ICMP_SGT: return 1; // 001
3323 case ICmpInst::ICMP_EQ: return 2; // 010
3324 case ICmpInst::ICMP_UGE: return 3; // 011
3325 case ICmpInst::ICMP_SGE: return 3; // 011
3326 case ICmpInst::ICMP_ULT: return 4; // 100
3327 case ICmpInst::ICMP_SLT: return 4; // 100
3328 case ICmpInst::ICMP_NE: return 5; // 101
3329 case ICmpInst::ICMP_ULE: return 6; // 110
3330 case ICmpInst::ICMP_SLE: return 6; // 110
3333 llvm_unreachable("Invalid ICmp predicate!");
3338 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3339 /// predicate into a three bit mask. It also returns whether it is an ordered
3340 /// predicate by reference.
3341 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3344 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3345 case FCmpInst::FCMP_UNO: return 0; // 000
3346 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3347 case FCmpInst::FCMP_UGT: return 1; // 001
3348 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3349 case FCmpInst::FCMP_UEQ: return 2; // 010
3350 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3351 case FCmpInst::FCMP_UGE: return 3; // 011
3352 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3353 case FCmpInst::FCMP_ULT: return 4; // 100
3354 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3355 case FCmpInst::FCMP_UNE: return 5; // 101
3356 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3357 case FCmpInst::FCMP_ULE: return 6; // 110
3360 // Not expecting FCMP_FALSE and FCMP_TRUE;
3361 llvm_unreachable("Unexpected FCmp predicate!");
3366 /// getICmpValue - This is the complement of getICmpCode, which turns an
3367 /// opcode and two operands into either a constant true or false, or a brand
3368 /// new ICmp instruction. The sign is passed in to determine which kind
3369 /// of predicate to use in the new icmp instruction.
3370 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3371 LLVMContext *Context) {
3373 default: llvm_unreachable("Illegal ICmp code!");
3374 case 0: return ConstantInt::getFalse(*Context);
3377 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, LHS, RHS);
3379 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, LHS, RHS);
3380 case 2: return new ICmpInst(*Context, ICmpInst::ICMP_EQ, LHS, RHS);
3383 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHS, RHS);
3385 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHS, RHS);
3388 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHS, RHS);
3390 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHS, RHS);
3391 case 5: return new ICmpInst(*Context, ICmpInst::ICMP_NE, LHS, RHS);
3394 return new ICmpInst(*Context, ICmpInst::ICMP_SLE, LHS, RHS);
3396 return new ICmpInst(*Context, ICmpInst::ICMP_ULE, LHS, RHS);
3397 case 7: return ConstantInt::getTrue(*Context);
3401 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3402 /// opcode and two operands into either a FCmp instruction. isordered is passed
3403 /// in to determine which kind of predicate to use in the new fcmp instruction.
3404 static Value *getFCmpValue(bool isordered, unsigned code,
3405 Value *LHS, Value *RHS, LLVMContext *Context) {
3407 default: llvm_unreachable("Illegal FCmp code!");
3410 return new FCmpInst(*Context, FCmpInst::FCMP_ORD, LHS, RHS);
3412 return new FCmpInst(*Context, FCmpInst::FCMP_UNO, LHS, RHS);
3415 return new FCmpInst(*Context, FCmpInst::FCMP_OGT, LHS, RHS);
3417 return new FCmpInst(*Context, FCmpInst::FCMP_UGT, LHS, RHS);
3420 return new FCmpInst(*Context, FCmpInst::FCMP_OEQ, LHS, RHS);
3422 return new FCmpInst(*Context, FCmpInst::FCMP_UEQ, LHS, RHS);
3425 return new FCmpInst(*Context, FCmpInst::FCMP_OGE, LHS, RHS);
3427 return new FCmpInst(*Context, FCmpInst::FCMP_UGE, LHS, RHS);
3430 return new FCmpInst(*Context, FCmpInst::FCMP_OLT, LHS, RHS);
3432 return new FCmpInst(*Context, FCmpInst::FCMP_ULT, LHS, RHS);
3435 return new FCmpInst(*Context, FCmpInst::FCMP_ONE, LHS, RHS);
3437 return new FCmpInst(*Context, FCmpInst::FCMP_UNE, LHS, RHS);
3440 return new FCmpInst(*Context, FCmpInst::FCMP_OLE, LHS, RHS);
3442 return new FCmpInst(*Context, FCmpInst::FCMP_ULE, LHS, RHS);
3443 case 7: return ConstantInt::getTrue(*Context);
3447 /// PredicatesFoldable - Return true if both predicates match sign or if at
3448 /// least one of them is an equality comparison (which is signless).
3449 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3450 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3451 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3452 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3456 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3457 struct FoldICmpLogical {
3460 ICmpInst::Predicate pred;
3461 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3462 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3463 pred(ICI->getPredicate()) {}
3464 bool shouldApply(Value *V) const {
3465 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3466 if (PredicatesFoldable(pred, ICI->getPredicate()))
3467 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3468 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3471 Instruction *apply(Instruction &Log) const {
3472 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3473 if (ICI->getOperand(0) != LHS) {
3474 assert(ICI->getOperand(1) == LHS);
3475 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3478 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3479 unsigned LHSCode = getICmpCode(ICI);
3480 unsigned RHSCode = getICmpCode(RHSICI);
3482 switch (Log.getOpcode()) {
3483 case Instruction::And: Code = LHSCode & RHSCode; break;
3484 case Instruction::Or: Code = LHSCode | RHSCode; break;
3485 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3486 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3489 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3490 ICmpInst::isSignedPredicate(ICI->getPredicate());
3492 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3493 if (Instruction *I = dyn_cast<Instruction>(RV))
3495 // Otherwise, it's a constant boolean value...
3496 return IC.ReplaceInstUsesWith(Log, RV);
3499 } // end anonymous namespace
3501 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3502 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3503 // guaranteed to be a binary operator.
3504 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3506 ConstantInt *AndRHS,
3507 BinaryOperator &TheAnd) {
3508 Value *X = Op->getOperand(0);
3509 Constant *Together = 0;
3511 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
3513 switch (Op->getOpcode()) {
3514 case Instruction::Xor:
3515 if (Op->hasOneUse()) {
3516 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3517 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3518 InsertNewInstBefore(And, TheAnd);
3520 return BinaryOperator::CreateXor(And, Together);
3523 case Instruction::Or:
3524 if (Together == AndRHS) // (X | C) & C --> C
3525 return ReplaceInstUsesWith(TheAnd, AndRHS);
3527 if (Op->hasOneUse() && Together != OpRHS) {
3528 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3529 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3530 InsertNewInstBefore(Or, TheAnd);
3532 return BinaryOperator::CreateAnd(Or, AndRHS);
3535 case Instruction::Add:
3536 if (Op->hasOneUse()) {
3537 // Adding a one to a single bit bit-field should be turned into an XOR
3538 // of the bit. First thing to check is to see if this AND is with a
3539 // single bit constant.
3540 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3542 // If there is only one bit set...
3543 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3544 // Ok, at this point, we know that we are masking the result of the
3545 // ADD down to exactly one bit. If the constant we are adding has
3546 // no bits set below this bit, then we can eliminate the ADD.
3547 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3549 // Check to see if any bits below the one bit set in AndRHSV are set.
3550 if ((AddRHS & (AndRHSV-1)) == 0) {
3551 // If not, the only thing that can effect the output of the AND is
3552 // the bit specified by AndRHSV. If that bit is set, the effect of
3553 // the XOR is to toggle the bit. If it is clear, then the ADD has
3555 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3556 TheAnd.setOperand(0, X);
3559 // Pull the XOR out of the AND.
3560 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3561 InsertNewInstBefore(NewAnd, TheAnd);
3562 NewAnd->takeName(Op);
3563 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3570 case Instruction::Shl: {
3571 // We know that the AND will not produce any of the bits shifted in, so if
3572 // the anded constant includes them, clear them now!
3574 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3575 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3576 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3577 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShlMask);
3579 if (CI->getValue() == ShlMask) {
3580 // Masking out bits that the shift already masks
3581 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3582 } else if (CI != AndRHS) { // Reducing bits set in and.
3583 TheAnd.setOperand(1, CI);
3588 case Instruction::LShr:
3590 // We know that the AND will not produce any of the bits shifted in, so if
3591 // the anded constant includes them, clear them now! This only applies to
3592 // unsigned shifts, because a signed shr may bring in set bits!
3594 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3595 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3596 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3597 ConstantInt *CI = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3599 if (CI->getValue() == ShrMask) {
3600 // Masking out bits that the shift already masks.
3601 return ReplaceInstUsesWith(TheAnd, Op);
3602 } else if (CI != AndRHS) {
3603 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3608 case Instruction::AShr:
3610 // See if this is shifting in some sign extension, then masking it out
3612 if (Op->hasOneUse()) {
3613 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3614 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3615 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3616 Constant *C = ConstantInt::get(*Context, AndRHS->getValue() & ShrMask);
3617 if (C == AndRHS) { // Masking out bits shifted in.
3618 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3619 // Make the argument unsigned.
3620 Value *ShVal = Op->getOperand(0);
3621 ShVal = InsertNewInstBefore(
3622 BinaryOperator::CreateLShr(ShVal, OpRHS,
3623 Op->getName()), TheAnd);
3624 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3633 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3634 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3635 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3636 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3637 /// insert new instructions.
3638 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3639 bool isSigned, bool Inside,
3641 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3642 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3643 "Lo is not <= Hi in range emission code!");
3646 if (Lo == Hi) // Trivially false.
3647 return new ICmpInst(*Context, ICmpInst::ICMP_NE, V, V);
3649 // V >= Min && V < Hi --> V < Hi
3650 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3651 ICmpInst::Predicate pred = (isSigned ?
3652 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3653 return new ICmpInst(*Context, pred, V, Hi);
3656 // Emit V-Lo <u Hi-Lo
3657 Constant *NegLo = ConstantExpr::getNeg(Lo);
3658 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3659 InsertNewInstBefore(Add, IB);
3660 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3661 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, UpperBound);
3664 if (Lo == Hi) // Trivially true.
3665 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, V, V);
3667 // V < Min || V >= Hi -> V > Hi-1
3668 Hi = SubOne(cast<ConstantInt>(Hi), Context);
3669 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3670 ICmpInst::Predicate pred = (isSigned ?
3671 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3672 return new ICmpInst(*Context, pred, V, Hi);
3675 // Emit V-Lo >u Hi-1-Lo
3676 // Note that Hi has already had one subtracted from it, above.
3677 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3678 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3679 InsertNewInstBefore(Add, IB);
3680 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3681 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add, LowerBound);
3684 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3685 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3686 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3687 // not, since all 1s are not contiguous.
3688 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3689 const APInt& V = Val->getValue();
3690 uint32_t BitWidth = Val->getType()->getBitWidth();
3691 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3693 // look for the first zero bit after the run of ones
3694 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3695 // look for the first non-zero bit
3696 ME = V.getActiveBits();
3700 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3701 /// where isSub determines whether the operator is a sub. If we can fold one of
3702 /// the following xforms:
3704 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3705 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3706 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3708 /// return (A +/- B).
3710 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3711 ConstantInt *Mask, bool isSub,
3713 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3714 if (!LHSI || LHSI->getNumOperands() != 2 ||
3715 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3717 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3719 switch (LHSI->getOpcode()) {
3721 case Instruction::And:
3722 if (ConstantExpr::getAnd(N, Mask) == Mask) {
3723 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3724 if ((Mask->getValue().countLeadingZeros() +
3725 Mask->getValue().countPopulation()) ==
3726 Mask->getValue().getBitWidth())
3729 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3730 // part, we don't need any explicit masks to take them out of A. If that
3731 // is all N is, ignore it.
3732 uint32_t MB = 0, ME = 0;
3733 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3734 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3735 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3736 if (MaskedValueIsZero(RHS, Mask))
3741 case Instruction::Or:
3742 case Instruction::Xor:
3743 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3744 if ((Mask->getValue().countLeadingZeros() +
3745 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3746 && ConstantExpr::getAnd(N, Mask)->isNullValue())
3753 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3755 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3756 return InsertNewInstBefore(New, I);
3759 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3760 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3761 ICmpInst *LHS, ICmpInst *RHS) {
3763 ConstantInt *LHSCst, *RHSCst;
3764 ICmpInst::Predicate LHSCC, RHSCC;
3766 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3767 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3768 m_ConstantInt(LHSCst)), *Context) ||
3769 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3770 m_ConstantInt(RHSCst)), *Context))
3773 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3774 // where C is a power of 2
3775 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3776 LHSCst->getValue().isPowerOf2()) {
3777 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3778 InsertNewInstBefore(NewOr, I);
3779 return new ICmpInst(*Context, LHSCC, NewOr, LHSCst);
3782 // From here on, we only handle:
3783 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3784 if (Val != Val2) return 0;
3786 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3787 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3788 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3789 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3790 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3793 // We can't fold (ugt x, C) & (sgt x, C2).
3794 if (!PredicatesFoldable(LHSCC, RHSCC))
3797 // Ensure that the larger constant is on the RHS.
3799 if (ICmpInst::isSignedPredicate(LHSCC) ||
3800 (ICmpInst::isEquality(LHSCC) &&
3801 ICmpInst::isSignedPredicate(RHSCC)))
3802 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3804 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3807 std::swap(LHS, RHS);
3808 std::swap(LHSCst, RHSCst);
3809 std::swap(LHSCC, RHSCC);
3812 // At this point, we know we have have two icmp instructions
3813 // comparing a value against two constants and and'ing the result
3814 // together. Because of the above check, we know that we only have
3815 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3816 // (from the FoldICmpLogical check above), that the two constants
3817 // are not equal and that the larger constant is on the RHS
3818 assert(LHSCst != RHSCst && "Compares not folded above?");
3821 default: llvm_unreachable("Unknown integer condition code!");
3822 case ICmpInst::ICMP_EQ:
3824 default: llvm_unreachable("Unknown integer condition code!");
3825 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3826 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3827 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3828 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3829 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3830 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3831 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3832 return ReplaceInstUsesWith(I, LHS);
3834 case ICmpInst::ICMP_NE:
3836 default: llvm_unreachable("Unknown integer condition code!");
3837 case ICmpInst::ICMP_ULT:
3838 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X u< 14) -> X < 13
3839 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Val, LHSCst);
3840 break; // (X != 13 & X u< 15) -> no change
3841 case ICmpInst::ICMP_SLT:
3842 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X s< 14) -> X < 13
3843 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Val, LHSCst);
3844 break; // (X != 13 & X s< 15) -> no change
3845 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3846 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3847 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3848 return ReplaceInstUsesWith(I, RHS);
3849 case ICmpInst::ICMP_NE:
3850 if (LHSCst == SubOne(RHSCst, Context)){// (X != 13 & X != 14) -> X-13 >u 1
3851 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3852 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3853 Val->getName()+".off");
3854 InsertNewInstBefore(Add, I);
3855 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add,
3856 ConstantInt::get(Add->getType(), 1));
3858 break; // (X != 13 & X != 15) -> no change
3861 case ICmpInst::ICMP_ULT:
3863 default: llvm_unreachable("Unknown integer condition code!");
3864 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3865 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3866 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3867 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3869 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3870 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3871 return ReplaceInstUsesWith(I, LHS);
3872 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3876 case ICmpInst::ICMP_SLT:
3878 default: llvm_unreachable("Unknown integer condition code!");
3879 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3880 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3881 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3882 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3884 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3885 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3886 return ReplaceInstUsesWith(I, LHS);
3887 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3891 case ICmpInst::ICMP_UGT:
3893 default: llvm_unreachable("Unknown integer condition code!");
3894 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3895 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3896 return ReplaceInstUsesWith(I, RHS);
3897 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3899 case ICmpInst::ICMP_NE:
3900 if (RHSCst == AddOne(LHSCst, Context)) // (X u> 13 & X != 14) -> X u> 14
3901 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3902 break; // (X u> 13 & X != 15) -> no change
3903 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3904 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3905 RHSCst, false, true, I);
3906 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3910 case ICmpInst::ICMP_SGT:
3912 default: llvm_unreachable("Unknown integer condition code!");
3913 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3914 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3915 return ReplaceInstUsesWith(I, RHS);
3916 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3918 case ICmpInst::ICMP_NE:
3919 if (RHSCst == AddOne(LHSCst, Context)) // (X s> 13 & X != 14) -> X s> 14
3920 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3921 break; // (X s> 13 & X != 15) -> no change
3922 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3923 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3924 RHSCst, true, true, I);
3925 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3934 Instruction *InstCombiner::FoldAndOfFCmps(Instruction &I, FCmpInst *LHS,
3937 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
3938 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
3939 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
3940 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
3941 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
3942 // If either of the constants are nans, then the whole thing returns
3944 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
3945 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3946 return new FCmpInst(*Context, FCmpInst::FCMP_ORD,
3947 LHS->getOperand(0), RHS->getOperand(0));
3950 // Handle vector zeros. This occurs because the canonical form of
3951 // "fcmp ord x,x" is "fcmp ord x, 0".
3952 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
3953 isa<ConstantAggregateZero>(RHS->getOperand(1)))
3954 return new FCmpInst(*Context, FCmpInst::FCMP_ORD,
3955 LHS->getOperand(0), RHS->getOperand(0));
3959 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
3960 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
3961 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
3964 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
3965 // Swap RHS operands to match LHS.
3966 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
3967 std::swap(Op1LHS, Op1RHS);
3970 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
3971 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
3973 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC, Op0LHS, Op0RHS);
3975 if (Op0CC == FCmpInst::FCMP_FALSE || Op1CC == FCmpInst::FCMP_FALSE)
3976 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
3977 if (Op0CC == FCmpInst::FCMP_TRUE)
3978 return ReplaceInstUsesWith(I, RHS);
3979 if (Op1CC == FCmpInst::FCMP_TRUE)
3980 return ReplaceInstUsesWith(I, LHS);
3984 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
3985 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
3987 std::swap(LHS, RHS);
3988 std::swap(Op0Pred, Op1Pred);
3989 std::swap(Op0Ordered, Op1Ordered);
3992 // uno && ueq -> uno && (uno || eq) -> ueq
3993 // ord && olt -> ord && (ord && lt) -> olt
3994 if (Op0Ordered == Op1Ordered)
3995 return ReplaceInstUsesWith(I, RHS);
3997 // uno && oeq -> uno && (ord && eq) -> false
3998 // uno && ord -> false
4000 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
4001 // ord && ueq -> ord && (uno || eq) -> oeq
4002 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4003 Op0LHS, Op0RHS, Context));
4011 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
4012 bool Changed = SimplifyCommutative(I);
4013 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4015 if (isa<UndefValue>(Op1)) // X & undef -> 0
4016 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4020 return ReplaceInstUsesWith(I, Op1);
4022 // See if we can simplify any instructions used by the instruction whose sole
4023 // purpose is to compute bits we don't care about.
4024 if (SimplifyDemandedInstructionBits(I))
4026 if (isa<VectorType>(I.getType())) {
4027 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4028 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
4029 return ReplaceInstUsesWith(I, I.getOperand(0));
4030 } else if (isa<ConstantAggregateZero>(Op1)) {
4031 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
4035 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
4036 const APInt& AndRHSMask = AndRHS->getValue();
4037 APInt NotAndRHS(~AndRHSMask);
4039 // Optimize a variety of ((val OP C1) & C2) combinations...
4040 if (isa<BinaryOperator>(Op0)) {
4041 Instruction *Op0I = cast<Instruction>(Op0);
4042 Value *Op0LHS = Op0I->getOperand(0);
4043 Value *Op0RHS = Op0I->getOperand(1);
4044 switch (Op0I->getOpcode()) {
4045 case Instruction::Xor:
4046 case Instruction::Or:
4047 // If the mask is only needed on one incoming arm, push it up.
4048 if (Op0I->hasOneUse()) {
4049 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4050 // Not masking anything out for the LHS, move to RHS.
4051 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4052 Op0RHS->getName()+".masked");
4053 InsertNewInstBefore(NewRHS, I);
4054 return BinaryOperator::Create(
4055 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4057 if (!isa<Constant>(Op0RHS) &&
4058 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4059 // Not masking anything out for the RHS, move to LHS.
4060 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4061 Op0LHS->getName()+".masked");
4062 InsertNewInstBefore(NewLHS, I);
4063 return BinaryOperator::Create(
4064 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4069 case Instruction::Add:
4070 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4071 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4072 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4073 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4074 return BinaryOperator::CreateAnd(V, AndRHS);
4075 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4076 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4079 case Instruction::Sub:
4080 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4081 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4082 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4083 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4084 return BinaryOperator::CreateAnd(V, AndRHS);
4086 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4087 // has 1's for all bits that the subtraction with A might affect.
4088 if (Op0I->hasOneUse()) {
4089 uint32_t BitWidth = AndRHSMask.getBitWidth();
4090 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4091 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4093 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4094 if (!(A && A->isZero()) && // avoid infinite recursion.
4095 MaskedValueIsZero(Op0LHS, Mask)) {
4096 Instruction *NewNeg = BinaryOperator::CreateNeg(*Context, Op0RHS);
4097 InsertNewInstBefore(NewNeg, I);
4098 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4103 case Instruction::Shl:
4104 case Instruction::LShr:
4105 // (1 << x) & 1 --> zext(x == 0)
4106 // (1 >> x) & 1 --> zext(x == 0)
4107 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4108 Instruction *NewICmp = new ICmpInst(*Context, ICmpInst::ICMP_EQ,
4109 Op0RHS, Constant::getNullValue(I.getType()));
4110 InsertNewInstBefore(NewICmp, I);
4111 return new ZExtInst(NewICmp, I.getType());
4116 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4117 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4119 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4120 // If this is an integer truncation or change from signed-to-unsigned, and
4121 // if the source is an and/or with immediate, transform it. This
4122 // frequently occurs for bitfield accesses.
4123 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4124 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4125 CastOp->getNumOperands() == 2)
4126 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4127 if (CastOp->getOpcode() == Instruction::And) {
4128 // Change: and (cast (and X, C1) to T), C2
4129 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4130 // This will fold the two constants together, which may allow
4131 // other simplifications.
4132 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4133 CastOp->getOperand(0), I.getType(),
4134 CastOp->getName()+".shrunk");
4135 NewCast = InsertNewInstBefore(NewCast, I);
4136 // trunc_or_bitcast(C1)&C2
4138 ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4139 C3 = ConstantExpr::getAnd(C3, AndRHS);
4140 return BinaryOperator::CreateAnd(NewCast, C3);
4141 } else if (CastOp->getOpcode() == Instruction::Or) {
4142 // Change: and (cast (or X, C1) to T), C2
4143 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4145 ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
4146 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS)
4148 return ReplaceInstUsesWith(I, AndRHS);
4154 // Try to fold constant and into select arguments.
4155 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4156 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4158 if (isa<PHINode>(Op0))
4159 if (Instruction *NV = FoldOpIntoPhi(I))
4163 Value *Op0NotVal = dyn_castNotVal(Op0, Context);
4164 Value *Op1NotVal = dyn_castNotVal(Op1, Context);
4166 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4167 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4169 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4170 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4171 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4172 I.getName()+".demorgan");
4173 InsertNewInstBefore(Or, I);
4174 return BinaryOperator::CreateNot(*Context, Or);
4178 Value *A = 0, *B = 0, *C = 0, *D = 0;
4179 if (match(Op0, m_Or(m_Value(A), m_Value(B)), *Context)) {
4180 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4181 return ReplaceInstUsesWith(I, Op1);
4183 // (A|B) & ~(A&B) -> A^B
4184 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4185 if ((A == C && B == D) || (A == D && B == C))
4186 return BinaryOperator::CreateXor(A, B);
4190 if (match(Op1, m_Or(m_Value(A), m_Value(B)), *Context)) {
4191 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4192 return ReplaceInstUsesWith(I, Op0);
4194 // ~(A&B) & (A|B) -> A^B
4195 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4196 if ((A == C && B == D) || (A == D && B == C))
4197 return BinaryOperator::CreateXor(A, B);
4201 if (Op0->hasOneUse() &&
4202 match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4203 if (A == Op1) { // (A^B)&A -> A&(A^B)
4204 I.swapOperands(); // Simplify below
4205 std::swap(Op0, Op1);
4206 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4207 cast<BinaryOperator>(Op0)->swapOperands();
4208 I.swapOperands(); // Simplify below
4209 std::swap(Op0, Op1);
4213 if (Op1->hasOneUse() &&
4214 match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4215 if (B == Op0) { // B&(A^B) -> B&(B^A)
4216 cast<BinaryOperator>(Op1)->swapOperands();
4219 if (A == Op0) { // A&(A^B) -> A & ~B
4220 Instruction *NotB = BinaryOperator::CreateNot(*Context, B, "tmp");
4221 InsertNewInstBefore(NotB, I);
4222 return BinaryOperator::CreateAnd(A, NotB);
4226 // (A&((~A)|B)) -> A&B
4227 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A)), *Context) ||
4228 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1))), *Context))
4229 return BinaryOperator::CreateAnd(A, Op1);
4230 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A)), *Context) ||
4231 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0))), *Context))
4232 return BinaryOperator::CreateAnd(A, Op0);
4235 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4236 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4237 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4240 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4241 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4245 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4246 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4247 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4248 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4249 const Type *SrcTy = Op0C->getOperand(0)->getType();
4250 if (SrcTy == Op1C->getOperand(0)->getType() &&
4251 SrcTy->isIntOrIntVector() &&
4252 // Only do this if the casts both really cause code to be generated.
4253 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4255 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4257 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4258 Op1C->getOperand(0),
4260 InsertNewInstBefore(NewOp, I);
4261 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4265 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4266 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4267 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4268 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4269 SI0->getOperand(1) == SI1->getOperand(1) &&
4270 (SI0->hasOneUse() || SI1->hasOneUse())) {
4271 Instruction *NewOp =
4272 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4274 SI0->getName()), I);
4275 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4276 SI1->getOperand(1));
4280 // If and'ing two fcmp, try combine them into one.
4281 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4282 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4283 if (Instruction *Res = FoldAndOfFCmps(I, LHS, RHS))
4287 return Changed ? &I : 0;
4290 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4291 /// capable of providing pieces of a bswap. The subexpression provides pieces
4292 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4293 /// the expression came from the corresponding "byte swapped" byte in some other
4294 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4295 /// we know that the expression deposits the low byte of %X into the high byte
4296 /// of the bswap result and that all other bytes are zero. This expression is
4297 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4300 /// This function returns true if the match was unsuccessful and false if so.
4301 /// On entry to the function the "OverallLeftShift" is a signed integer value
4302 /// indicating the number of bytes that the subexpression is later shifted. For
4303 /// example, if the expression is later right shifted by 16 bits, the
4304 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4305 /// byte of ByteValues is actually being set.
4307 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4308 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4309 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4310 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4311 /// always in the local (OverallLeftShift) coordinate space.
4313 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4314 SmallVector<Value*, 8> &ByteValues) {
4315 if (Instruction *I = dyn_cast<Instruction>(V)) {
4316 // If this is an or instruction, it may be an inner node of the bswap.
4317 if (I->getOpcode() == Instruction::Or) {
4318 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4320 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4324 // If this is a logical shift by a constant multiple of 8, recurse with
4325 // OverallLeftShift and ByteMask adjusted.
4326 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4328 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4329 // Ensure the shift amount is defined and of a byte value.
4330 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4333 unsigned ByteShift = ShAmt >> 3;
4334 if (I->getOpcode() == Instruction::Shl) {
4335 // X << 2 -> collect(X, +2)
4336 OverallLeftShift += ByteShift;
4337 ByteMask >>= ByteShift;
4339 // X >>u 2 -> collect(X, -2)
4340 OverallLeftShift -= ByteShift;
4341 ByteMask <<= ByteShift;
4342 ByteMask &= (~0U >> (32-ByteValues.size()));
4345 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4346 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4348 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4352 // If this is a logical 'and' with a mask that clears bytes, clear the
4353 // corresponding bytes in ByteMask.
4354 if (I->getOpcode() == Instruction::And &&
4355 isa<ConstantInt>(I->getOperand(1))) {
4356 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4357 unsigned NumBytes = ByteValues.size();
4358 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4359 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4361 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4362 // If this byte is masked out by a later operation, we don't care what
4364 if ((ByteMask & (1 << i)) == 0)
4367 // If the AndMask is all zeros for this byte, clear the bit.
4368 APInt MaskB = AndMask & Byte;
4370 ByteMask &= ~(1U << i);
4374 // If the AndMask is not all ones for this byte, it's not a bytezap.
4378 // Otherwise, this byte is kept.
4381 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4386 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4387 // the input value to the bswap. Some observations: 1) if more than one byte
4388 // is demanded from this input, then it could not be successfully assembled
4389 // into a byteswap. At least one of the two bytes would not be aligned with
4390 // their ultimate destination.
4391 if (!isPowerOf2_32(ByteMask)) return true;
4392 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4394 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4395 // is demanded, it needs to go into byte 0 of the result. This means that the
4396 // byte needs to be shifted until it lands in the right byte bucket. The
4397 // shift amount depends on the position: if the byte is coming from the high
4398 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4399 // low part, it must be shifted left.
4400 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4401 if (InputByteNo < ByteValues.size()/2) {
4402 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4405 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4409 // If the destination byte value is already defined, the values are or'd
4410 // together, which isn't a bswap (unless it's an or of the same bits).
4411 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4413 ByteValues[DestByteNo] = V;
4417 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4418 /// If so, insert the new bswap intrinsic and return it.
4419 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4420 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4421 if (!ITy || ITy->getBitWidth() % 16 ||
4422 // ByteMask only allows up to 32-byte values.
4423 ITy->getBitWidth() > 32*8)
4424 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4426 /// ByteValues - For each byte of the result, we keep track of which value
4427 /// defines each byte.
4428 SmallVector<Value*, 8> ByteValues;
4429 ByteValues.resize(ITy->getBitWidth()/8);
4431 // Try to find all the pieces corresponding to the bswap.
4432 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4433 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4436 // Check to see if all of the bytes come from the same value.
4437 Value *V = ByteValues[0];
4438 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4440 // Check to make sure that all of the bytes come from the same value.
4441 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4442 if (ByteValues[i] != V)
4444 const Type *Tys[] = { ITy };
4445 Module *M = I.getParent()->getParent()->getParent();
4446 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4447 return CallInst::Create(F, V);
4450 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4451 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4452 /// we can simplify this expression to "cond ? C : D or B".
4453 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4455 LLVMContext *Context) {
4456 // If A is not a select of -1/0, this cannot match.
4458 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond)), *Context))
4461 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4462 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4463 return SelectInst::Create(Cond, C, B);
4464 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4465 return SelectInst::Create(Cond, C, B);
4466 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4467 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4468 return SelectInst::Create(Cond, C, D);
4469 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4470 return SelectInst::Create(Cond, C, D);
4474 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4475 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4476 ICmpInst *LHS, ICmpInst *RHS) {
4478 ConstantInt *LHSCst, *RHSCst;
4479 ICmpInst::Predicate LHSCC, RHSCC;
4481 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4482 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4483 m_ConstantInt(LHSCst)), *Context) ||
4484 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4485 m_ConstantInt(RHSCst)), *Context))
4488 // From here on, we only handle:
4489 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4490 if (Val != Val2) return 0;
4492 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4493 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4494 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4495 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4496 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4499 // We can't fold (ugt x, C) | (sgt x, C2).
4500 if (!PredicatesFoldable(LHSCC, RHSCC))
4503 // Ensure that the larger constant is on the RHS.
4505 if (ICmpInst::isSignedPredicate(LHSCC) ||
4506 (ICmpInst::isEquality(LHSCC) &&
4507 ICmpInst::isSignedPredicate(RHSCC)))
4508 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4510 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4513 std::swap(LHS, RHS);
4514 std::swap(LHSCst, RHSCst);
4515 std::swap(LHSCC, RHSCC);
4518 // At this point, we know we have have two icmp instructions
4519 // comparing a value against two constants and or'ing the result
4520 // together. Because of the above check, we know that we only have
4521 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4522 // FoldICmpLogical check above), that the two constants are not
4524 assert(LHSCst != RHSCst && "Compares not folded above?");
4527 default: llvm_unreachable("Unknown integer condition code!");
4528 case ICmpInst::ICMP_EQ:
4530 default: llvm_unreachable("Unknown integer condition code!");
4531 case ICmpInst::ICMP_EQ:
4532 if (LHSCst == SubOne(RHSCst, Context)) {
4533 // (X == 13 | X == 14) -> X-13 <u 2
4534 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4535 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4536 Val->getName()+".off");
4537 InsertNewInstBefore(Add, I);
4538 AddCST = ConstantExpr::getSub(AddOne(RHSCst, Context), LHSCst);
4539 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, AddCST);
4541 break; // (X == 13 | X == 15) -> no change
4542 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4543 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4545 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4546 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4547 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4548 return ReplaceInstUsesWith(I, RHS);
4551 case ICmpInst::ICMP_NE:
4553 default: llvm_unreachable("Unknown integer condition code!");
4554 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4555 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4556 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4557 return ReplaceInstUsesWith(I, LHS);
4558 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4559 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4560 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4561 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4564 case ICmpInst::ICMP_ULT:
4566 default: llvm_unreachable("Unknown integer condition code!");
4567 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4569 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4570 // If RHSCst is [us]MAXINT, it is always false. Not handling
4571 // this can cause overflow.
4572 if (RHSCst->isMaxValue(false))
4573 return ReplaceInstUsesWith(I, LHS);
4574 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4576 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4578 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4579 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4580 return ReplaceInstUsesWith(I, RHS);
4581 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4585 case ICmpInst::ICMP_SLT:
4587 default: llvm_unreachable("Unknown integer condition code!");
4588 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4590 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4591 // If RHSCst is [us]MAXINT, it is always false. Not handling
4592 // this can cause overflow.
4593 if (RHSCst->isMaxValue(true))
4594 return ReplaceInstUsesWith(I, LHS);
4595 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4597 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4599 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4600 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4601 return ReplaceInstUsesWith(I, RHS);
4602 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4606 case ICmpInst::ICMP_UGT:
4608 default: llvm_unreachable("Unknown integer condition code!");
4609 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4610 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4611 return ReplaceInstUsesWith(I, LHS);
4612 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4614 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4615 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4616 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4617 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4621 case ICmpInst::ICMP_SGT:
4623 default: llvm_unreachable("Unknown integer condition code!");
4624 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4625 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4626 return ReplaceInstUsesWith(I, LHS);
4627 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4629 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4630 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4631 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4632 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4640 Instruction *InstCombiner::FoldOrOfFCmps(Instruction &I, FCmpInst *LHS,
4642 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4643 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4644 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4645 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4646 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4647 // If either of the constants are nans, then the whole thing returns
4649 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4650 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4652 // Otherwise, no need to compare the two constants, compare the
4654 return new FCmpInst(*Context, FCmpInst::FCMP_UNO,
4655 LHS->getOperand(0), RHS->getOperand(0));
4658 // Handle vector zeros. This occurs because the canonical form of
4659 // "fcmp uno x,x" is "fcmp uno x, 0".
4660 if (isa<ConstantAggregateZero>(LHS->getOperand(1)) &&
4661 isa<ConstantAggregateZero>(RHS->getOperand(1)))
4662 return new FCmpInst(*Context, FCmpInst::FCMP_UNO,
4663 LHS->getOperand(0), RHS->getOperand(0));
4668 Value *Op0LHS = LHS->getOperand(0), *Op0RHS = LHS->getOperand(1);
4669 Value *Op1LHS = RHS->getOperand(0), *Op1RHS = RHS->getOperand(1);
4670 FCmpInst::Predicate Op0CC = LHS->getPredicate(), Op1CC = RHS->getPredicate();
4672 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4673 // Swap RHS operands to match LHS.
4674 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4675 std::swap(Op1LHS, Op1RHS);
4677 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4678 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4680 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4682 if (Op0CC == FCmpInst::FCMP_TRUE || Op1CC == FCmpInst::FCMP_TRUE)
4683 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
4684 if (Op0CC == FCmpInst::FCMP_FALSE)
4685 return ReplaceInstUsesWith(I, RHS);
4686 if (Op1CC == FCmpInst::FCMP_FALSE)
4687 return ReplaceInstUsesWith(I, LHS);
4690 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4691 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4692 if (Op0Ordered == Op1Ordered) {
4693 // If both are ordered or unordered, return a new fcmp with
4694 // or'ed predicates.
4695 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4696 Op0LHS, Op0RHS, Context);
4697 if (Instruction *I = dyn_cast<Instruction>(RV))
4699 // Otherwise, it's a constant boolean value...
4700 return ReplaceInstUsesWith(I, RV);
4706 /// FoldOrWithConstants - This helper function folds:
4708 /// ((A | B) & C1) | (B & C2)
4714 /// when the XOR of the two constants is "all ones" (-1).
4715 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4716 Value *A, Value *B, Value *C) {
4717 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4721 ConstantInt *CI2 = 0;
4722 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)), *Context)) return 0;
4724 APInt Xor = CI1->getValue() ^ CI2->getValue();
4725 if (!Xor.isAllOnesValue()) return 0;
4727 if (V1 == A || V1 == B) {
4728 Instruction *NewOp =
4729 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4730 return BinaryOperator::CreateOr(NewOp, V1);
4736 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4737 bool Changed = SimplifyCommutative(I);
4738 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4740 if (isa<UndefValue>(Op1)) // X | undef -> -1
4741 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4745 return ReplaceInstUsesWith(I, Op0);
4747 // See if we can simplify any instructions used by the instruction whose sole
4748 // purpose is to compute bits we don't care about.
4749 if (SimplifyDemandedInstructionBits(I))
4751 if (isa<VectorType>(I.getType())) {
4752 if (isa<ConstantAggregateZero>(Op1)) {
4753 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4754 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4755 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4756 return ReplaceInstUsesWith(I, I.getOperand(1));
4761 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4762 ConstantInt *C1 = 0; Value *X = 0;
4763 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4764 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1)), *Context) &&
4766 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4767 InsertNewInstBefore(Or, I);
4769 return BinaryOperator::CreateAnd(Or,
4770 ConstantInt::get(*Context, RHS->getValue() | C1->getValue()));
4773 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4774 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1)), *Context) &&
4776 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4777 InsertNewInstBefore(Or, I);
4779 return BinaryOperator::CreateXor(Or,
4780 ConstantInt::get(*Context, C1->getValue() & ~RHS->getValue()));
4783 // Try to fold constant and into select arguments.
4784 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4785 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4787 if (isa<PHINode>(Op0))
4788 if (Instruction *NV = FoldOpIntoPhi(I))
4792 Value *A = 0, *B = 0;
4793 ConstantInt *C1 = 0, *C2 = 0;
4795 if (match(Op0, m_And(m_Value(A), m_Value(B)), *Context))
4796 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4797 return ReplaceInstUsesWith(I, Op1);
4798 if (match(Op1, m_And(m_Value(A), m_Value(B)), *Context))
4799 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4800 return ReplaceInstUsesWith(I, Op0);
4802 // (A | B) | C and A | (B | C) -> bswap if possible.
4803 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4804 if (match(Op0, m_Or(m_Value(), m_Value()), *Context) ||
4805 match(Op1, m_Or(m_Value(), m_Value()), *Context) ||
4806 (match(Op0, m_Shift(m_Value(), m_Value()), *Context) &&
4807 match(Op1, m_Shift(m_Value(), m_Value()), *Context))) {
4808 if (Instruction *BSwap = MatchBSwap(I))
4812 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4813 if (Op0->hasOneUse() &&
4814 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4815 MaskedValueIsZero(Op1, C1->getValue())) {
4816 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4817 InsertNewInstBefore(NOr, I);
4819 return BinaryOperator::CreateXor(NOr, C1);
4822 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4823 if (Op1->hasOneUse() &&
4824 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4825 MaskedValueIsZero(Op0, C1->getValue())) {
4826 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4827 InsertNewInstBefore(NOr, I);
4829 return BinaryOperator::CreateXor(NOr, C1);
4833 Value *C = 0, *D = 0;
4834 if (match(Op0, m_And(m_Value(A), m_Value(C)), *Context) &&
4835 match(Op1, m_And(m_Value(B), m_Value(D)), *Context)) {
4836 Value *V1 = 0, *V2 = 0, *V3 = 0;
4837 C1 = dyn_cast<ConstantInt>(C);
4838 C2 = dyn_cast<ConstantInt>(D);
4839 if (C1 && C2) { // (A & C1)|(B & C2)
4840 // If we have: ((V + N) & C1) | (V & C2)
4841 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4842 // replace with V+N.
4843 if (C1->getValue() == ~C2->getValue()) {
4844 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4845 match(A, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4846 // Add commutes, try both ways.
4847 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4848 return ReplaceInstUsesWith(I, A);
4849 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4850 return ReplaceInstUsesWith(I, A);
4852 // Or commutes, try both ways.
4853 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4854 match(B, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4855 // Add commutes, try both ways.
4856 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4857 return ReplaceInstUsesWith(I, B);
4858 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4859 return ReplaceInstUsesWith(I, B);
4862 V1 = 0; V2 = 0; V3 = 0;
4865 // Check to see if we have any common things being and'ed. If so, find the
4866 // terms for V1 & (V2|V3).
4867 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4868 if (A == B) // (A & C)|(A & D) == A & (C|D)
4869 V1 = A, V2 = C, V3 = D;
4870 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4871 V1 = A, V2 = B, V3 = C;
4872 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4873 V1 = C, V2 = A, V3 = D;
4874 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4875 V1 = C, V2 = A, V3 = B;
4879 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4880 return BinaryOperator::CreateAnd(V1, Or);
4884 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4885 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4887 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4889 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4891 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4894 // ((A&~B)|(~A&B)) -> A^B
4895 if ((match(C, m_Not(m_Specific(D)), *Context) &&
4896 match(B, m_Not(m_Specific(A)), *Context)))
4897 return BinaryOperator::CreateXor(A, D);
4898 // ((~B&A)|(~A&B)) -> A^B
4899 if ((match(A, m_Not(m_Specific(D)), *Context) &&
4900 match(B, m_Not(m_Specific(C)), *Context)))
4901 return BinaryOperator::CreateXor(C, D);
4902 // ((A&~B)|(B&~A)) -> A^B
4903 if ((match(C, m_Not(m_Specific(B)), *Context) &&
4904 match(D, m_Not(m_Specific(A)), *Context)))
4905 return BinaryOperator::CreateXor(A, B);
4906 // ((~B&A)|(B&~A)) -> A^B
4907 if ((match(A, m_Not(m_Specific(B)), *Context) &&
4908 match(D, m_Not(m_Specific(C)), *Context)))
4909 return BinaryOperator::CreateXor(C, B);
4912 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4913 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4914 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4915 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4916 SI0->getOperand(1) == SI1->getOperand(1) &&
4917 (SI0->hasOneUse() || SI1->hasOneUse())) {
4918 Instruction *NewOp =
4919 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4921 SI0->getName()), I);
4922 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4923 SI1->getOperand(1));
4927 // ((A|B)&1)|(B&-2) -> (A&1) | B
4928 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4929 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4930 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4931 if (Ret) return Ret;
4933 // (B&-2)|((A|B)&1) -> (A&1) | B
4934 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4935 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4936 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4937 if (Ret) return Ret;
4940 if (match(Op0, m_Not(m_Value(A)), *Context)) { // ~A | Op1
4941 if (A == Op1) // ~A | A == -1
4942 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4946 // Note, A is still live here!
4947 if (match(Op1, m_Not(m_Value(B)), *Context)) { // Op0 | ~B
4949 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4951 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4952 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4953 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4954 I.getName()+".demorgan"), I);
4955 return BinaryOperator::CreateNot(*Context, And);
4959 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4960 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4961 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4964 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4965 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4969 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4970 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4971 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4972 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4973 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4974 !isa<ICmpInst>(Op1C->getOperand(0))) {
4975 const Type *SrcTy = Op0C->getOperand(0)->getType();
4976 if (SrcTy == Op1C->getOperand(0)->getType() &&
4977 SrcTy->isIntOrIntVector() &&
4978 // Only do this if the casts both really cause code to be
4980 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4982 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4984 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4985 Op1C->getOperand(0),
4987 InsertNewInstBefore(NewOp, I);
4988 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4995 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4996 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4997 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1)))
4998 if (Instruction *Res = FoldOrOfFCmps(I, LHS, RHS))
5002 return Changed ? &I : 0;
5007 // XorSelf - Implements: X ^ X --> 0
5010 XorSelf(Value *rhs) : RHS(rhs) {}
5011 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5012 Instruction *apply(BinaryOperator &Xor) const {
5019 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5020 bool Changed = SimplifyCommutative(I);
5021 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5023 if (isa<UndefValue>(Op1)) {
5024 if (isa<UndefValue>(Op0))
5025 // Handle undef ^ undef -> 0 special case. This is a common
5027 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5028 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5031 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5032 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1), Context)) {
5033 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5034 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5037 // See if we can simplify any instructions used by the instruction whose sole
5038 // purpose is to compute bits we don't care about.
5039 if (SimplifyDemandedInstructionBits(I))
5041 if (isa<VectorType>(I.getType()))
5042 if (isa<ConstantAggregateZero>(Op1))
5043 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5045 // Is this a ~ operation?
5046 if (Value *NotOp = dyn_castNotVal(&I, Context)) {
5047 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5048 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5049 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5050 if (Op0I->getOpcode() == Instruction::And ||
5051 Op0I->getOpcode() == Instruction::Or) {
5052 if (dyn_castNotVal(Op0I->getOperand(1), Context)) Op0I->swapOperands();
5053 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0), Context)) {
5055 BinaryOperator::CreateNot(*Context, Op0I->getOperand(1),
5056 Op0I->getOperand(1)->getName()+".not");
5057 InsertNewInstBefore(NotY, I);
5058 if (Op0I->getOpcode() == Instruction::And)
5059 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5061 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5068 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5069 if (RHS == ConstantInt::getTrue(*Context) && Op0->hasOneUse()) {
5070 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5071 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5072 return new ICmpInst(*Context, ICI->getInversePredicate(),
5073 ICI->getOperand(0), ICI->getOperand(1));
5075 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5076 return new FCmpInst(*Context, FCI->getInversePredicate(),
5077 FCI->getOperand(0), FCI->getOperand(1));
5080 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5081 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5082 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5083 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5084 Instruction::CastOps Opcode = Op0C->getOpcode();
5085 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
5086 if (RHS == ConstantExpr::getCast(Opcode,
5087 ConstantInt::getTrue(*Context),
5088 Op0C->getDestTy())) {
5089 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
5091 CI->getOpcode(), CI->getInversePredicate(),
5092 CI->getOperand(0), CI->getOperand(1)), I);
5093 NewCI->takeName(CI);
5094 return CastInst::Create(Opcode, NewCI, Op0C->getType());
5101 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5102 // ~(c-X) == X-c-1 == X+(-c-1)
5103 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5104 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5105 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
5106 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
5107 ConstantInt::get(I.getType(), 1));
5108 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5111 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5112 if (Op0I->getOpcode() == Instruction::Add) {
5113 // ~(X-c) --> (-c-1)-X
5114 if (RHS->isAllOnesValue()) {
5115 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
5116 return BinaryOperator::CreateSub(
5117 ConstantExpr::getSub(NegOp0CI,
5118 ConstantInt::get(I.getType(), 1)),
5119 Op0I->getOperand(0));
5120 } else if (RHS->getValue().isSignBit()) {
5121 // (X + C) ^ signbit -> (X + C + signbit)
5122 Constant *C = ConstantInt::get(*Context,
5123 RHS->getValue() + Op0CI->getValue());
5124 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5127 } else if (Op0I->getOpcode() == Instruction::Or) {
5128 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5129 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5130 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
5131 // Anything in both C1 and C2 is known to be zero, remove it from
5133 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
5134 NewRHS = ConstantExpr::getAnd(NewRHS,
5135 ConstantExpr::getNot(CommonBits));
5136 AddToWorkList(Op0I);
5137 I.setOperand(0, Op0I->getOperand(0));
5138 I.setOperand(1, NewRHS);
5145 // Try to fold constant and into select arguments.
5146 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5147 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5149 if (isa<PHINode>(Op0))
5150 if (Instruction *NV = FoldOpIntoPhi(I))
5154 if (Value *X = dyn_castNotVal(Op0, Context)) // ~A ^ A == -1
5156 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5158 if (Value *X = dyn_castNotVal(Op1, Context)) // A ^ ~A == -1
5160 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
5163 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5166 if (match(Op1I, m_Or(m_Value(A), m_Value(B)), *Context)) {
5167 if (A == Op0) { // B^(B|A) == (A|B)^B
5168 Op1I->swapOperands();
5170 std::swap(Op0, Op1);
5171 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5172 I.swapOperands(); // Simplified below.
5173 std::swap(Op0, Op1);
5175 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)), *Context)) {
5176 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5177 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)), *Context)) {
5178 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5179 } else if (match(Op1I, m_And(m_Value(A), m_Value(B)), *Context) &&
5181 if (A == Op0) { // A^(A&B) -> A^(B&A)
5182 Op1I->swapOperands();
5185 if (B == Op0) { // A^(B&A) -> (B&A)^A
5186 I.swapOperands(); // Simplified below.
5187 std::swap(Op0, Op1);
5192 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5195 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5196 Op0I->hasOneUse()) {
5197 if (A == Op1) // (B|A)^B == (A|B)^B
5199 if (B == Op1) { // (A|B)^B == A & ~B
5201 InsertNewInstBefore(BinaryOperator::CreateNot(*Context,
5203 return BinaryOperator::CreateAnd(A, NotB);
5205 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)), *Context)) {
5206 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5207 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)), *Context)) {
5208 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5209 } else if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5211 if (A == Op1) // (A&B)^A -> (B&A)^A
5213 if (B == Op1 && // (B&A)^A == ~B & A
5214 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5216 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, A, "tmp"), I);
5217 return BinaryOperator::CreateAnd(N, Op1);
5222 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5223 if (Op0I && Op1I && Op0I->isShift() &&
5224 Op0I->getOpcode() == Op1I->getOpcode() &&
5225 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5226 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5227 Instruction *NewOp =
5228 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5229 Op1I->getOperand(0),
5230 Op0I->getName()), I);
5231 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5232 Op1I->getOperand(1));
5236 Value *A, *B, *C, *D;
5237 // (A & B)^(A | B) -> A ^ B
5238 if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5239 match(Op1I, m_Or(m_Value(C), m_Value(D)), *Context)) {
5240 if ((A == C && B == D) || (A == D && B == C))
5241 return BinaryOperator::CreateXor(A, B);
5243 // (A | B)^(A & B) -> A ^ B
5244 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5245 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5246 if ((A == C && B == D) || (A == D && B == C))
5247 return BinaryOperator::CreateXor(A, B);
5251 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5252 match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5253 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5254 // (X & Y)^(X & Y) -> (Y^Z) & X
5255 Value *X = 0, *Y = 0, *Z = 0;
5257 X = A, Y = B, Z = D;
5259 X = A, Y = B, Z = C;
5261 X = B, Y = A, Z = D;
5263 X = B, Y = A, Z = C;
5266 Instruction *NewOp =
5267 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5268 return BinaryOperator::CreateAnd(NewOp, X);
5273 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5274 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5275 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
5278 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5279 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5280 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5281 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5282 const Type *SrcTy = Op0C->getOperand(0)->getType();
5283 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5284 // Only do this if the casts both really cause code to be generated.
5285 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5287 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5289 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5290 Op1C->getOperand(0),
5292 InsertNewInstBefore(NewOp, I);
5293 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5298 return Changed ? &I : 0;
5301 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5302 LLVMContext *Context) {
5303 return cast<ConstantInt>(ConstantExpr::getExtractElement(V, Idx));
5306 static bool HasAddOverflow(ConstantInt *Result,
5307 ConstantInt *In1, ConstantInt *In2,
5310 if (In2->getValue().isNegative())
5311 return Result->getValue().sgt(In1->getValue());
5313 return Result->getValue().slt(In1->getValue());
5315 return Result->getValue().ult(In1->getValue());
5318 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5319 /// overflowed for this type.
5320 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5321 Constant *In2, LLVMContext *Context,
5322 bool IsSigned = false) {
5323 Result = ConstantExpr::getAdd(In1, In2);
5325 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5326 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5327 Constant *Idx = ConstantInt::get(Type::Int32Ty, i);
5328 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5329 ExtractElement(In1, Idx, Context),
5330 ExtractElement(In2, Idx, Context),
5337 return HasAddOverflow(cast<ConstantInt>(Result),
5338 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5342 static bool HasSubOverflow(ConstantInt *Result,
5343 ConstantInt *In1, ConstantInt *In2,
5346 if (In2->getValue().isNegative())
5347 return Result->getValue().slt(In1->getValue());
5349 return Result->getValue().sgt(In1->getValue());
5351 return Result->getValue().ugt(In1->getValue());
5354 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5355 /// overflowed for this type.
5356 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5357 Constant *In2, LLVMContext *Context,
5358 bool IsSigned = false) {
5359 Result = ConstantExpr::getSub(In1, In2);
5361 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5362 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5363 Constant *Idx = ConstantInt::get(Type::Int32Ty, i);
5364 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5365 ExtractElement(In1, Idx, Context),
5366 ExtractElement(In2, Idx, Context),
5373 return HasSubOverflow(cast<ConstantInt>(Result),
5374 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5378 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5379 /// code necessary to compute the offset from the base pointer (without adding
5380 /// in the base pointer). Return the result as a signed integer of intptr size.
5381 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5382 TargetData &TD = *IC.getTargetData();
5383 gep_type_iterator GTI = gep_type_begin(GEP);
5384 const Type *IntPtrTy = TD.getIntPtrType();
5385 LLVMContext *Context = IC.getContext();
5386 Value *Result = Constant::getNullValue(IntPtrTy);
5388 // Build a mask for high order bits.
5389 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5390 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5392 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5395 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5396 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5397 if (OpC->isZero()) continue;
5399 // Handle a struct index, which adds its field offset to the pointer.
5400 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5401 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5403 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5405 ConstantInt::get(*Context,
5406 RC->getValue() + APInt(IntPtrWidth, Size));
5408 Result = IC.InsertNewInstBefore(
5409 BinaryOperator::CreateAdd(Result,
5410 ConstantInt::get(IntPtrTy, Size),
5411 GEP->getName()+".offs"), I);
5415 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5417 ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5418 Scale = ConstantExpr::getMul(OC, Scale);
5419 if (Constant *RC = dyn_cast<Constant>(Result))
5420 Result = ConstantExpr::getAdd(RC, Scale);
5422 // Emit an add instruction.
5423 Result = IC.InsertNewInstBefore(
5424 BinaryOperator::CreateAdd(Result, Scale,
5425 GEP->getName()+".offs"), I);
5429 // Convert to correct type.
5430 if (Op->getType() != IntPtrTy) {
5431 if (Constant *OpC = dyn_cast<Constant>(Op))
5432 Op = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true);
5434 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5436 Op->getName()+".c"), I);
5439 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
5440 if (Constant *OpC = dyn_cast<Constant>(Op))
5441 Op = ConstantExpr::getMul(OpC, Scale);
5442 else // We'll let instcombine(mul) convert this to a shl if possible.
5443 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5444 GEP->getName()+".idx"), I);
5447 // Emit an add instruction.
5448 if (isa<Constant>(Op) && isa<Constant>(Result))
5449 Result = ConstantExpr::getAdd(cast<Constant>(Op),
5450 cast<Constant>(Result));
5452 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5453 GEP->getName()+".offs"), I);
5459 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5460 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5461 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5462 /// be complex, and scales are involved. The above expression would also be
5463 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5464 /// This later form is less amenable to optimization though, and we are allowed
5465 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5467 /// If we can't emit an optimized form for this expression, this returns null.
5469 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5471 TargetData &TD = *IC.getTargetData();
5472 gep_type_iterator GTI = gep_type_begin(GEP);
5474 // Check to see if this gep only has a single variable index. If so, and if
5475 // any constant indices are a multiple of its scale, then we can compute this
5476 // in terms of the scale of the variable index. For example, if the GEP
5477 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5478 // because the expression will cross zero at the same point.
5479 unsigned i, e = GEP->getNumOperands();
5481 for (i = 1; i != e; ++i, ++GTI) {
5482 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5483 // Compute the aggregate offset of constant indices.
5484 if (CI->isZero()) continue;
5486 // Handle a struct index, which adds its field offset to the pointer.
5487 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5488 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5490 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5491 Offset += Size*CI->getSExtValue();
5494 // Found our variable index.
5499 // If there are no variable indices, we must have a constant offset, just
5500 // evaluate it the general way.
5501 if (i == e) return 0;
5503 Value *VariableIdx = GEP->getOperand(i);
5504 // Determine the scale factor of the variable element. For example, this is
5505 // 4 if the variable index is into an array of i32.
5506 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5508 // Verify that there are no other variable indices. If so, emit the hard way.
5509 for (++i, ++GTI; i != e; ++i, ++GTI) {
5510 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5513 // Compute the aggregate offset of constant indices.
5514 if (CI->isZero()) continue;
5516 // Handle a struct index, which adds its field offset to the pointer.
5517 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5518 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5520 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5521 Offset += Size*CI->getSExtValue();
5525 // Okay, we know we have a single variable index, which must be a
5526 // pointer/array/vector index. If there is no offset, life is simple, return
5528 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5530 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5531 // we don't need to bother extending: the extension won't affect where the
5532 // computation crosses zero.
5533 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5534 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5535 VariableIdx->getName(), &I);
5539 // Otherwise, there is an index. The computation we will do will be modulo
5540 // the pointer size, so get it.
5541 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5543 Offset &= PtrSizeMask;
5544 VariableScale &= PtrSizeMask;
5546 // To do this transformation, any constant index must be a multiple of the
5547 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5548 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5549 // multiple of the variable scale.
5550 int64_t NewOffs = Offset / (int64_t)VariableScale;
5551 if (Offset != NewOffs*(int64_t)VariableScale)
5554 // Okay, we can do this evaluation. Start by converting the index to intptr.
5555 const Type *IntPtrTy = TD.getIntPtrType();
5556 if (VariableIdx->getType() != IntPtrTy)
5557 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5559 VariableIdx->getName(), &I);
5560 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5561 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5565 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5566 /// else. At this point we know that the GEP is on the LHS of the comparison.
5567 Instruction *InstCombiner::FoldGEPICmp(GEPOperator *GEPLHS, Value *RHS,
5568 ICmpInst::Predicate Cond,
5570 // Look through bitcasts.
5571 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5572 RHS = BCI->getOperand(0);
5574 Value *PtrBase = GEPLHS->getOperand(0);
5575 if (TD && PtrBase == RHS && GEPLHS->isInBounds()) {
5576 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5577 // This transformation (ignoring the base and scales) is valid because we
5578 // know pointers can't overflow since the gep is inbounds. See if we can
5579 // output an optimized form.
5580 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5582 // If not, synthesize the offset the hard way.
5584 Offset = EmitGEPOffset(GEPLHS, I, *this);
5585 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), Offset,
5586 Constant::getNullValue(Offset->getType()));
5587 } else if (GEPOperator *GEPRHS = dyn_cast<GEPOperator>(RHS)) {
5588 // If the base pointers are different, but the indices are the same, just
5589 // compare the base pointer.
5590 if (PtrBase != GEPRHS->getOperand(0)) {
5591 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5592 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5593 GEPRHS->getOperand(0)->getType();
5595 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5596 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5597 IndicesTheSame = false;
5601 // If all indices are the same, just compare the base pointers.
5603 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond),
5604 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5606 // Otherwise, the base pointers are different and the indices are
5607 // different, bail out.
5611 // If one of the GEPs has all zero indices, recurse.
5612 bool AllZeros = true;
5613 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5614 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5615 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5620 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5621 ICmpInst::getSwappedPredicate(Cond), I);
5623 // If the other GEP has all zero indices, recurse.
5625 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5626 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5627 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5632 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5634 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5635 // If the GEPs only differ by one index, compare it.
5636 unsigned NumDifferences = 0; // Keep track of # differences.
5637 unsigned DiffOperand = 0; // The operand that differs.
5638 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5639 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5640 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5641 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5642 // Irreconcilable differences.
5646 if (NumDifferences++) break;
5651 if (NumDifferences == 0) // SAME GEP?
5652 return ReplaceInstUsesWith(I, // No comparison is needed here.
5653 ConstantInt::get(Type::Int1Ty,
5654 ICmpInst::isTrueWhenEqual(Cond)));
5656 else if (NumDifferences == 1) {
5657 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5658 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5659 // Make sure we do a signed comparison here.
5660 return new ICmpInst(*Context,
5661 ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5665 // Only lower this if the icmp is the only user of the GEP or if we expect
5666 // the result to fold to a constant!
5668 (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5669 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5670 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5671 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5672 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5673 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), L, R);
5679 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5681 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5684 if (!isa<ConstantFP>(RHSC)) return 0;
5685 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5687 // Get the width of the mantissa. We don't want to hack on conversions that
5688 // might lose information from the integer, e.g. "i64 -> float"
5689 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5690 if (MantissaWidth == -1) return 0; // Unknown.
5692 // Check to see that the input is converted from an integer type that is small
5693 // enough that preserves all bits. TODO: check here for "known" sign bits.
5694 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5695 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5697 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5698 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5702 // If the conversion would lose info, don't hack on this.
5703 if ((int)InputSize > MantissaWidth)
5706 // Otherwise, we can potentially simplify the comparison. We know that it
5707 // will always come through as an integer value and we know the constant is
5708 // not a NAN (it would have been previously simplified).
5709 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5711 ICmpInst::Predicate Pred;
5712 switch (I.getPredicate()) {
5713 default: llvm_unreachable("Unexpected predicate!");
5714 case FCmpInst::FCMP_UEQ:
5715 case FCmpInst::FCMP_OEQ:
5716 Pred = ICmpInst::ICMP_EQ;
5718 case FCmpInst::FCMP_UGT:
5719 case FCmpInst::FCMP_OGT:
5720 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5722 case FCmpInst::FCMP_UGE:
5723 case FCmpInst::FCMP_OGE:
5724 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5726 case FCmpInst::FCMP_ULT:
5727 case FCmpInst::FCMP_OLT:
5728 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5730 case FCmpInst::FCMP_ULE:
5731 case FCmpInst::FCMP_OLE:
5732 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5734 case FCmpInst::FCMP_UNE:
5735 case FCmpInst::FCMP_ONE:
5736 Pred = ICmpInst::ICMP_NE;
5738 case FCmpInst::FCMP_ORD:
5739 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5740 case FCmpInst::FCMP_UNO:
5741 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5744 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5746 // Now we know that the APFloat is a normal number, zero or inf.
5748 // See if the FP constant is too large for the integer. For example,
5749 // comparing an i8 to 300.0.
5750 unsigned IntWidth = IntTy->getScalarSizeInBits();
5753 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5754 // and large values.
5755 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5756 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5757 APFloat::rmNearestTiesToEven);
5758 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5759 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5760 Pred == ICmpInst::ICMP_SLE)
5761 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5762 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5765 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5766 // +INF and large values.
5767 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5768 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5769 APFloat::rmNearestTiesToEven);
5770 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5771 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5772 Pred == ICmpInst::ICMP_ULE)
5773 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5774 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5779 // See if the RHS value is < SignedMin.
5780 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5781 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5782 APFloat::rmNearestTiesToEven);
5783 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5784 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5785 Pred == ICmpInst::ICMP_SGE)
5786 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5787 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5791 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5792 // [0, UMAX], but it may still be fractional. See if it is fractional by
5793 // casting the FP value to the integer value and back, checking for equality.
5794 // Don't do this for zero, because -0.0 is not fractional.
5795 Constant *RHSInt = LHSUnsigned
5796 ? ConstantExpr::getFPToUI(RHSC, IntTy)
5797 : ConstantExpr::getFPToSI(RHSC, IntTy);
5798 if (!RHS.isZero()) {
5799 bool Equal = LHSUnsigned
5800 ? ConstantExpr::getUIToFP(RHSInt, RHSC->getType()) == RHSC
5801 : ConstantExpr::getSIToFP(RHSInt, RHSC->getType()) == RHSC;
5803 // If we had a comparison against a fractional value, we have to adjust
5804 // the compare predicate and sometimes the value. RHSC is rounded towards
5805 // zero at this point.
5807 default: llvm_unreachable("Unexpected integer comparison!");
5808 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5809 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5810 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5811 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5812 case ICmpInst::ICMP_ULE:
5813 // (float)int <= 4.4 --> int <= 4
5814 // (float)int <= -4.4 --> false
5815 if (RHS.isNegative())
5816 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5818 case ICmpInst::ICMP_SLE:
5819 // (float)int <= 4.4 --> int <= 4
5820 // (float)int <= -4.4 --> int < -4
5821 if (RHS.isNegative())
5822 Pred = ICmpInst::ICMP_SLT;
5824 case ICmpInst::ICMP_ULT:
5825 // (float)int < -4.4 --> false
5826 // (float)int < 4.4 --> int <= 4
5827 if (RHS.isNegative())
5828 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5829 Pred = ICmpInst::ICMP_ULE;
5831 case ICmpInst::ICMP_SLT:
5832 // (float)int < -4.4 --> int < -4
5833 // (float)int < 4.4 --> int <= 4
5834 if (!RHS.isNegative())
5835 Pred = ICmpInst::ICMP_SLE;
5837 case ICmpInst::ICMP_UGT:
5838 // (float)int > 4.4 --> int > 4
5839 // (float)int > -4.4 --> true
5840 if (RHS.isNegative())
5841 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5843 case ICmpInst::ICMP_SGT:
5844 // (float)int > 4.4 --> int > 4
5845 // (float)int > -4.4 --> int >= -4
5846 if (RHS.isNegative())
5847 Pred = ICmpInst::ICMP_SGE;
5849 case ICmpInst::ICMP_UGE:
5850 // (float)int >= -4.4 --> true
5851 // (float)int >= 4.4 --> int > 4
5852 if (!RHS.isNegative())
5853 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5854 Pred = ICmpInst::ICMP_UGT;
5856 case ICmpInst::ICMP_SGE:
5857 // (float)int >= -4.4 --> int >= -4
5858 // (float)int >= 4.4 --> int > 4
5859 if (!RHS.isNegative())
5860 Pred = ICmpInst::ICMP_SGT;
5866 // Lower this FP comparison into an appropriate integer version of the
5868 return new ICmpInst(*Context, Pred, LHSI->getOperand(0), RHSInt);
5871 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5872 bool Changed = SimplifyCompare(I);
5873 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5875 // Fold trivial predicates.
5876 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5877 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5878 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5879 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5881 // Simplify 'fcmp pred X, X'
5883 switch (I.getPredicate()) {
5884 default: llvm_unreachable("Unknown predicate!");
5885 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5886 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5887 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5888 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5889 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5890 case FCmpInst::FCMP_OLT: // True if ordered and less than
5891 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5892 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5894 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5895 case FCmpInst::FCMP_ULT: // True if unordered or less than
5896 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5897 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5898 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5899 I.setPredicate(FCmpInst::FCMP_UNO);
5900 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5903 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5904 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5905 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5906 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5907 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5908 I.setPredicate(FCmpInst::FCMP_ORD);
5909 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5914 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5915 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5917 // Handle fcmp with constant RHS
5918 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5919 // If the constant is a nan, see if we can fold the comparison based on it.
5920 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5921 if (CFP->getValueAPF().isNaN()) {
5922 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5923 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
5924 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5925 "Comparison must be either ordered or unordered!");
5926 // True if unordered.
5927 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
5931 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5932 switch (LHSI->getOpcode()) {
5933 case Instruction::PHI:
5934 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5935 // block. If in the same block, we're encouraging jump threading. If
5936 // not, we are just pessimizing the code by making an i1 phi.
5937 if (LHSI->getParent() == I.getParent())
5938 if (Instruction *NV = FoldOpIntoPhi(I))
5941 case Instruction::SIToFP:
5942 case Instruction::UIToFP:
5943 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5946 case Instruction::Select:
5947 // If either operand of the select is a constant, we can fold the
5948 // comparison into the select arms, which will cause one to be
5949 // constant folded and the select turned into a bitwise or.
5950 Value *Op1 = 0, *Op2 = 0;
5951 if (LHSI->hasOneUse()) {
5952 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5953 // Fold the known value into the constant operand.
5954 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5955 // Insert a new FCmp of the other select operand.
5956 Op2 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5957 LHSI->getOperand(2), RHSC,
5959 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5960 // Fold the known value into the constant operand.
5961 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5962 // Insert a new FCmp of the other select operand.
5963 Op1 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5964 LHSI->getOperand(1), RHSC,
5970 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5975 return Changed ? &I : 0;
5978 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5979 bool Changed = SimplifyCompare(I);
5980 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5981 const Type *Ty = Op0->getType();
5985 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5986 I.isTrueWhenEqual()));
5988 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5989 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5991 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5992 // addresses never equal each other! We already know that Op0 != Op1.
5993 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5994 isa<ConstantPointerNull>(Op0)) &&
5995 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5996 isa<ConstantPointerNull>(Op1)))
5997 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5998 !I.isTrueWhenEqual()));
6000 // icmp's with boolean values can always be turned into bitwise operations
6001 if (Ty == Type::Int1Ty) {
6002 switch (I.getPredicate()) {
6003 default: llvm_unreachable("Invalid icmp instruction!");
6004 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
6005 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
6006 InsertNewInstBefore(Xor, I);
6007 return BinaryOperator::CreateNot(*Context, Xor);
6009 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6010 return BinaryOperator::CreateXor(Op0, Op1);
6012 case ICmpInst::ICMP_UGT:
6013 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6015 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6016 Instruction *Not = BinaryOperator::CreateNot(*Context,
6017 Op0, I.getName()+"tmp");
6018 InsertNewInstBefore(Not, I);
6019 return BinaryOperator::CreateAnd(Not, Op1);
6021 case ICmpInst::ICMP_SGT:
6022 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6024 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6025 Instruction *Not = BinaryOperator::CreateNot(*Context,
6026 Op1, I.getName()+"tmp");
6027 InsertNewInstBefore(Not, I);
6028 return BinaryOperator::CreateAnd(Not, Op0);
6030 case ICmpInst::ICMP_UGE:
6031 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6033 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6034 Instruction *Not = BinaryOperator::CreateNot(*Context,
6035 Op0, I.getName()+"tmp");
6036 InsertNewInstBefore(Not, I);
6037 return BinaryOperator::CreateOr(Not, Op1);
6039 case ICmpInst::ICMP_SGE:
6040 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6042 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6043 Instruction *Not = BinaryOperator::CreateNot(*Context,
6044 Op1, I.getName()+"tmp");
6045 InsertNewInstBefore(Not, I);
6046 return BinaryOperator::CreateOr(Not, Op0);
6051 unsigned BitWidth = 0;
6053 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6054 else if (Ty->isIntOrIntVector())
6055 BitWidth = Ty->getScalarSizeInBits();
6057 bool isSignBit = false;
6059 // See if we are doing a comparison with a constant.
6060 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6061 Value *A = 0, *B = 0;
6063 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6064 if (I.isEquality() && CI->isNullValue() &&
6065 match(Op0, m_Sub(m_Value(A), m_Value(B)), *Context)) {
6066 // (icmp cond A B) if cond is equality
6067 return new ICmpInst(*Context, I.getPredicate(), A, B);
6070 // If we have an icmp le or icmp ge instruction, turn it into the
6071 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6072 // them being folded in the code below.
6073 switch (I.getPredicate()) {
6075 case ICmpInst::ICMP_ULE:
6076 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6077 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6078 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Op0,
6079 AddOne(CI, Context));
6080 case ICmpInst::ICMP_SLE:
6081 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6082 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6083 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6084 AddOne(CI, Context));
6085 case ICmpInst::ICMP_UGE:
6086 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6087 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6088 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Op0,
6089 SubOne(CI, Context));
6090 case ICmpInst::ICMP_SGE:
6091 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6092 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6093 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6094 SubOne(CI, Context));
6097 // If this comparison is a normal comparison, it demands all
6098 // bits, if it is a sign bit comparison, it only demands the sign bit.
6100 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6103 // See if we can fold the comparison based on range information we can get
6104 // by checking whether bits are known to be zero or one in the input.
6105 if (BitWidth != 0) {
6106 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6107 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6109 if (SimplifyDemandedBits(I.getOperandUse(0),
6110 isSignBit ? APInt::getSignBit(BitWidth)
6111 : APInt::getAllOnesValue(BitWidth),
6112 Op0KnownZero, Op0KnownOne, 0))
6114 if (SimplifyDemandedBits(I.getOperandUse(1),
6115 APInt::getAllOnesValue(BitWidth),
6116 Op1KnownZero, Op1KnownOne, 0))
6119 // Given the known and unknown bits, compute a range that the LHS could be
6120 // in. Compute the Min, Max and RHS values based on the known bits. For the
6121 // EQ and NE we use unsigned values.
6122 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6123 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6124 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6125 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6127 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6130 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6132 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6136 // If Min and Max are known to be the same, then SimplifyDemandedBits
6137 // figured out that the LHS is a constant. Just constant fold this now so
6138 // that code below can assume that Min != Max.
6139 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6140 return new ICmpInst(*Context, I.getPredicate(),
6141 ConstantInt::get(*Context, Op0Min), Op1);
6142 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6143 return new ICmpInst(*Context, I.getPredicate(), Op0,
6144 ConstantInt::get(*Context, Op1Min));
6146 // Based on the range information we know about the LHS, see if we can
6147 // simplify this comparison. For example, (x&4) < 8 is always true.
6148 switch (I.getPredicate()) {
6149 default: llvm_unreachable("Unknown icmp opcode!");
6150 case ICmpInst::ICMP_EQ:
6151 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6152 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6154 case ICmpInst::ICMP_NE:
6155 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6156 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6158 case ICmpInst::ICMP_ULT:
6159 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6160 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6161 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6162 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6163 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6164 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6165 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6166 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6167 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6168 SubOne(CI, Context));
6170 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6171 if (CI->isMinValue(true))
6172 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6173 Constant::getAllOnesValue(Op0->getType()));
6176 case ICmpInst::ICMP_UGT:
6177 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6178 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6179 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6180 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6182 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6183 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6184 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6185 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6186 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6187 AddOne(CI, Context));
6189 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6190 if (CI->isMaxValue(true))
6191 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6192 Constant::getNullValue(Op0->getType()));
6195 case ICmpInst::ICMP_SLT:
6196 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6197 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6198 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6199 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6200 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6201 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6202 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6203 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6204 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6205 SubOne(CI, Context));
6208 case ICmpInst::ICMP_SGT:
6209 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6210 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6211 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6212 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6214 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6215 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6216 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6217 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6218 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6219 AddOne(CI, Context));
6222 case ICmpInst::ICMP_SGE:
6223 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6224 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6225 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6226 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6227 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6229 case ICmpInst::ICMP_SLE:
6230 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6231 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6232 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6233 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6234 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6236 case ICmpInst::ICMP_UGE:
6237 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6238 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6239 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6240 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6241 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6243 case ICmpInst::ICMP_ULE:
6244 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6245 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6246 return ReplaceInstUsesWith(I, ConstantInt::getTrue(*Context));
6247 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6248 return ReplaceInstUsesWith(I, ConstantInt::getFalse(*Context));
6252 // Turn a signed comparison into an unsigned one if both operands
6253 // are known to have the same sign.
6254 if (I.isSignedPredicate() &&
6255 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6256 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6257 return new ICmpInst(*Context, I.getUnsignedPredicate(), Op0, Op1);
6260 // Test if the ICmpInst instruction is used exclusively by a select as
6261 // part of a minimum or maximum operation. If so, refrain from doing
6262 // any other folding. This helps out other analyses which understand
6263 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6264 // and CodeGen. And in this case, at least one of the comparison
6265 // operands has at least one user besides the compare (the select),
6266 // which would often largely negate the benefit of folding anyway.
6268 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6269 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6270 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6273 // See if we are doing a comparison between a constant and an instruction that
6274 // can be folded into the comparison.
6275 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6276 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6277 // instruction, see if that instruction also has constants so that the
6278 // instruction can be folded into the icmp
6279 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6280 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6284 // Handle icmp with constant (but not simple integer constant) RHS
6285 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6286 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6287 switch (LHSI->getOpcode()) {
6288 case Instruction::GetElementPtr:
6289 if (RHSC->isNullValue()) {
6290 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6291 bool isAllZeros = true;
6292 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6293 if (!isa<Constant>(LHSI->getOperand(i)) ||
6294 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6299 return new ICmpInst(*Context, I.getPredicate(), LHSI->getOperand(0),
6300 Constant::getNullValue(LHSI->getOperand(0)->getType()));
6304 case Instruction::PHI:
6305 // Only fold icmp into the PHI if the phi and fcmp are in the same
6306 // block. If in the same block, we're encouraging jump threading. If
6307 // not, we are just pessimizing the code by making an i1 phi.
6308 if (LHSI->getParent() == I.getParent())
6309 if (Instruction *NV = FoldOpIntoPhi(I))
6312 case Instruction::Select: {
6313 // If either operand of the select is a constant, we can fold the
6314 // comparison into the select arms, which will cause one to be
6315 // constant folded and the select turned into a bitwise or.
6316 Value *Op1 = 0, *Op2 = 0;
6317 if (LHSI->hasOneUse()) {
6318 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6319 // Fold the known value into the constant operand.
6320 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6321 // Insert a new ICmp of the other select operand.
6322 Op2 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6323 LHSI->getOperand(2), RHSC,
6325 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6326 // Fold the known value into the constant operand.
6327 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
6328 // Insert a new ICmp of the other select operand.
6329 Op1 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6330 LHSI->getOperand(1), RHSC,
6336 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6339 case Instruction::Malloc:
6340 // If we have (malloc != null), and if the malloc has a single use, we
6341 // can assume it is successful and remove the malloc.
6342 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6343 AddToWorkList(LHSI);
6344 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
6345 !I.isTrueWhenEqual()));
6351 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6352 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op0))
6353 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6355 if (GEPOperator *GEP = dyn_cast<GEPOperator>(Op1))
6356 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6357 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6360 // Test to see if the operands of the icmp are casted versions of other
6361 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6363 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6364 if (isa<PointerType>(Op0->getType()) &&
6365 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6366 // We keep moving the cast from the left operand over to the right
6367 // operand, where it can often be eliminated completely.
6368 Op0 = CI->getOperand(0);
6370 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6371 // so eliminate it as well.
6372 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6373 Op1 = CI2->getOperand(0);
6375 // If Op1 is a constant, we can fold the cast into the constant.
6376 if (Op0->getType() != Op1->getType()) {
6377 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6378 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
6380 // Otherwise, cast the RHS right before the icmp
6381 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6384 return new ICmpInst(*Context, I.getPredicate(), Op0, Op1);
6388 if (isa<CastInst>(Op0)) {
6389 // Handle the special case of: icmp (cast bool to X), <cst>
6390 // This comes up when you have code like
6393 // For generality, we handle any zero-extension of any operand comparison
6394 // with a constant or another cast from the same type.
6395 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6396 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6400 // See if it's the same type of instruction on the left and right.
6401 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6402 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6403 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6404 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6405 switch (Op0I->getOpcode()) {
6407 case Instruction::Add:
6408 case Instruction::Sub:
6409 case Instruction::Xor:
6410 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6411 return new ICmpInst(*Context, I.getPredicate(), Op0I->getOperand(0),
6412 Op1I->getOperand(0));
6413 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6414 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6415 if (CI->getValue().isSignBit()) {
6416 ICmpInst::Predicate Pred = I.isSignedPredicate()
6417 ? I.getUnsignedPredicate()
6418 : I.getSignedPredicate();
6419 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6420 Op1I->getOperand(0));
6423 if (CI->getValue().isMaxSignedValue()) {
6424 ICmpInst::Predicate Pred = I.isSignedPredicate()
6425 ? I.getUnsignedPredicate()
6426 : I.getSignedPredicate();
6427 Pred = I.getSwappedPredicate(Pred);
6428 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6429 Op1I->getOperand(0));
6433 case Instruction::Mul:
6434 if (!I.isEquality())
6437 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6438 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6439 // Mask = -1 >> count-trailing-zeros(Cst).
6440 if (!CI->isZero() && !CI->isOne()) {
6441 const APInt &AP = CI->getValue();
6442 ConstantInt *Mask = ConstantInt::get(*Context,
6443 APInt::getLowBitsSet(AP.getBitWidth(),
6445 AP.countTrailingZeros()));
6446 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6448 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6450 InsertNewInstBefore(And1, I);
6451 InsertNewInstBefore(And2, I);
6452 return new ICmpInst(*Context, I.getPredicate(), And1, And2);
6461 // ~x < ~y --> y < x
6463 if (match(Op0, m_Not(m_Value(A)), *Context) &&
6464 match(Op1, m_Not(m_Value(B)), *Context))
6465 return new ICmpInst(*Context, I.getPredicate(), B, A);
6468 if (I.isEquality()) {
6469 Value *A, *B, *C, *D;
6471 // -x == -y --> x == y
6472 if (match(Op0, m_Neg(m_Value(A)), *Context) &&
6473 match(Op1, m_Neg(m_Value(B)), *Context))
6474 return new ICmpInst(*Context, I.getPredicate(), A, B);
6476 if (match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
6477 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6478 Value *OtherVal = A == Op1 ? B : A;
6479 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6480 Constant::getNullValue(A->getType()));
6483 if (match(Op1, m_Xor(m_Value(C), m_Value(D)), *Context)) {
6484 // A^c1 == C^c2 --> A == C^(c1^c2)
6485 ConstantInt *C1, *C2;
6486 if (match(B, m_ConstantInt(C1), *Context) &&
6487 match(D, m_ConstantInt(C2), *Context) && Op1->hasOneUse()) {
6489 ConstantInt::get(*Context, C1->getValue() ^ C2->getValue());
6490 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6491 return new ICmpInst(*Context, I.getPredicate(), A,
6492 InsertNewInstBefore(Xor, I));
6495 // A^B == A^D -> B == D
6496 if (A == C) return new ICmpInst(*Context, I.getPredicate(), B, D);
6497 if (A == D) return new ICmpInst(*Context, I.getPredicate(), B, C);
6498 if (B == C) return new ICmpInst(*Context, I.getPredicate(), A, D);
6499 if (B == D) return new ICmpInst(*Context, I.getPredicate(), A, C);
6503 if (match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context) &&
6504 (A == Op0 || B == Op0)) {
6505 // A == (A^B) -> B == 0
6506 Value *OtherVal = A == Op0 ? B : A;
6507 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6508 Constant::getNullValue(A->getType()));
6511 // (A-B) == A -> B == 0
6512 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B)), *Context))
6513 return new ICmpInst(*Context, I.getPredicate(), B,
6514 Constant::getNullValue(B->getType()));
6516 // A == (A-B) -> B == 0
6517 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B)), *Context))
6518 return new ICmpInst(*Context, I.getPredicate(), B,
6519 Constant::getNullValue(B->getType()));
6521 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6522 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6523 match(Op0, m_And(m_Value(A), m_Value(B)), *Context) &&
6524 match(Op1, m_And(m_Value(C), m_Value(D)), *Context)) {
6525 Value *X = 0, *Y = 0, *Z = 0;
6528 X = B; Y = D; Z = A;
6529 } else if (A == D) {
6530 X = B; Y = C; Z = A;
6531 } else if (B == C) {
6532 X = A; Y = D; Z = B;
6533 } else if (B == D) {
6534 X = A; Y = C; Z = B;
6537 if (X) { // Build (X^Y) & Z
6538 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6539 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6540 I.setOperand(0, Op1);
6541 I.setOperand(1, Constant::getNullValue(Op1->getType()));
6546 return Changed ? &I : 0;
6550 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6551 /// and CmpRHS are both known to be integer constants.
6552 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6553 ConstantInt *DivRHS) {
6554 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6555 const APInt &CmpRHSV = CmpRHS->getValue();
6557 // FIXME: If the operand types don't match the type of the divide
6558 // then don't attempt this transform. The code below doesn't have the
6559 // logic to deal with a signed divide and an unsigned compare (and
6560 // vice versa). This is because (x /s C1) <s C2 produces different
6561 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6562 // (x /u C1) <u C2. Simply casting the operands and result won't
6563 // work. :( The if statement below tests that condition and bails
6565 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6566 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6568 if (DivRHS->isZero())
6569 return 0; // The ProdOV computation fails on divide by zero.
6570 if (DivIsSigned && DivRHS->isAllOnesValue())
6571 return 0; // The overflow computation also screws up here
6572 if (DivRHS->isOne())
6573 return 0; // Not worth bothering, and eliminates some funny cases
6576 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6577 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6578 // C2 (CI). By solving for X we can turn this into a range check
6579 // instead of computing a divide.
6580 Constant *Prod = ConstantExpr::getMul(CmpRHS, DivRHS);
6582 // Determine if the product overflows by seeing if the product is
6583 // not equal to the divide. Make sure we do the same kind of divide
6584 // as in the LHS instruction that we're folding.
6585 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
6586 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
6588 // Get the ICmp opcode
6589 ICmpInst::Predicate Pred = ICI.getPredicate();
6591 // Figure out the interval that is being checked. For example, a comparison
6592 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6593 // Compute this interval based on the constants involved and the signedness of
6594 // the compare/divide. This computes a half-open interval, keeping track of
6595 // whether either value in the interval overflows. After analysis each
6596 // overflow variable is set to 0 if it's corresponding bound variable is valid
6597 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6598 int LoOverflow = 0, HiOverflow = 0;
6599 Constant *LoBound = 0, *HiBound = 0;
6601 if (!DivIsSigned) { // udiv
6602 // e.g. X/5 op 3 --> [15, 20)
6604 HiOverflow = LoOverflow = ProdOV;
6606 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6607 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6608 if (CmpRHSV == 0) { // (X / pos) op 0
6609 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6610 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS,
6613 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6614 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6615 HiOverflow = LoOverflow = ProdOV;
6617 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6618 } else { // (X / pos) op neg
6619 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6620 HiBound = AddOne(Prod, Context);
6621 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6623 ConstantInt* DivNeg =
6624 cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6625 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6629 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6630 if (CmpRHSV == 0) { // (X / neg) op 0
6631 // e.g. X/-5 op 0 --> [-4, 5)
6632 LoBound = AddOne(DivRHS, Context);
6633 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
6634 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6635 HiOverflow = 1; // [INTMIN+1, overflow)
6636 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6638 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6639 // e.g. X/-5 op 3 --> [-19, -14)
6640 HiBound = AddOne(Prod, Context);
6641 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6643 LoOverflow = AddWithOverflow(LoBound, HiBound,
6644 DivRHS, Context, true) ? -1 : 0;
6645 } else { // (X / neg) op neg
6646 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6647 LoOverflow = HiOverflow = ProdOV;
6649 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6652 // Dividing by a negative swaps the condition. LT <-> GT
6653 Pred = ICmpInst::getSwappedPredicate(Pred);
6656 Value *X = DivI->getOperand(0);
6658 default: llvm_unreachable("Unhandled icmp opcode!");
6659 case ICmpInst::ICMP_EQ:
6660 if (LoOverflow && HiOverflow)
6661 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6662 else if (HiOverflow)
6663 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6664 ICmpInst::ICMP_UGE, X, LoBound);
6665 else if (LoOverflow)
6666 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6667 ICmpInst::ICMP_ULT, X, HiBound);
6669 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6670 case ICmpInst::ICMP_NE:
6671 if (LoOverflow && HiOverflow)
6672 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6673 else if (HiOverflow)
6674 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6675 ICmpInst::ICMP_ULT, X, LoBound);
6676 else if (LoOverflow)
6677 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6678 ICmpInst::ICMP_UGE, X, HiBound);
6680 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6681 case ICmpInst::ICMP_ULT:
6682 case ICmpInst::ICMP_SLT:
6683 if (LoOverflow == +1) // Low bound is greater than input range.
6684 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6685 if (LoOverflow == -1) // Low bound is less than input range.
6686 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6687 return new ICmpInst(*Context, Pred, X, LoBound);
6688 case ICmpInst::ICMP_UGT:
6689 case ICmpInst::ICMP_SGT:
6690 if (HiOverflow == +1) // High bound greater than input range.
6691 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6692 else if (HiOverflow == -1) // High bound less than input range.
6693 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6694 if (Pred == ICmpInst::ICMP_UGT)
6695 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, X, HiBound);
6697 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, X, HiBound);
6702 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6704 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6707 const APInt &RHSV = RHS->getValue();
6709 switch (LHSI->getOpcode()) {
6710 case Instruction::Trunc:
6711 if (ICI.isEquality() && LHSI->hasOneUse()) {
6712 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6713 // of the high bits truncated out of x are known.
6714 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6715 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6716 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6717 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6718 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6720 // If all the high bits are known, we can do this xform.
6721 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6722 // Pull in the high bits from known-ones set.
6723 APInt NewRHS(RHS->getValue());
6724 NewRHS.zext(SrcBits);
6726 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6727 ConstantInt::get(*Context, NewRHS));
6732 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6733 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6734 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6736 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6737 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6738 Value *CompareVal = LHSI->getOperand(0);
6740 // If the sign bit of the XorCST is not set, there is no change to
6741 // the operation, just stop using the Xor.
6742 if (!XorCST->getValue().isNegative()) {
6743 ICI.setOperand(0, CompareVal);
6744 AddToWorkList(LHSI);
6748 // Was the old condition true if the operand is positive?
6749 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6751 // If so, the new one isn't.
6752 isTrueIfPositive ^= true;
6754 if (isTrueIfPositive)
6755 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, CompareVal,
6756 SubOne(RHS, Context));
6758 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, CompareVal,
6759 AddOne(RHS, Context));
6762 if (LHSI->hasOneUse()) {
6763 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6764 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6765 const APInt &SignBit = XorCST->getValue();
6766 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6767 ? ICI.getUnsignedPredicate()
6768 : ICI.getSignedPredicate();
6769 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6770 ConstantInt::get(*Context, RHSV ^ SignBit));
6773 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6774 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6775 const APInt &NotSignBit = XorCST->getValue();
6776 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6777 ? ICI.getUnsignedPredicate()
6778 : ICI.getSignedPredicate();
6779 Pred = ICI.getSwappedPredicate(Pred);
6780 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6781 ConstantInt::get(*Context, RHSV ^ NotSignBit));
6786 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6787 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6788 LHSI->getOperand(0)->hasOneUse()) {
6789 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6791 // If the LHS is an AND of a truncating cast, we can widen the
6792 // and/compare to be the input width without changing the value
6793 // produced, eliminating a cast.
6794 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6795 // We can do this transformation if either the AND constant does not
6796 // have its sign bit set or if it is an equality comparison.
6797 // Extending a relational comparison when we're checking the sign
6798 // bit would not work.
6799 if (Cast->hasOneUse() &&
6800 (ICI.isEquality() ||
6801 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6803 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6804 APInt NewCST = AndCST->getValue();
6805 NewCST.zext(BitWidth);
6807 NewCI.zext(BitWidth);
6808 Instruction *NewAnd =
6809 BinaryOperator::CreateAnd(Cast->getOperand(0),
6810 ConstantInt::get(*Context, NewCST), LHSI->getName());
6811 InsertNewInstBefore(NewAnd, ICI);
6812 return new ICmpInst(*Context, ICI.getPredicate(), NewAnd,
6813 ConstantInt::get(*Context, NewCI));
6817 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6818 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6819 // happens a LOT in code produced by the C front-end, for bitfield
6821 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6822 if (Shift && !Shift->isShift())
6826 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6827 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6828 const Type *AndTy = AndCST->getType(); // Type of the and.
6830 // We can fold this as long as we can't shift unknown bits
6831 // into the mask. This can only happen with signed shift
6832 // rights, as they sign-extend.
6834 bool CanFold = Shift->isLogicalShift();
6836 // To test for the bad case of the signed shr, see if any
6837 // of the bits shifted in could be tested after the mask.
6838 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6839 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6841 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6842 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6843 AndCST->getValue()) == 0)
6849 if (Shift->getOpcode() == Instruction::Shl)
6850 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6852 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6854 // Check to see if we are shifting out any of the bits being
6856 if (ConstantExpr::get(Shift->getOpcode(),
6857 NewCst, ShAmt) != RHS) {
6858 // If we shifted bits out, the fold is not going to work out.
6859 // As a special case, check to see if this means that the
6860 // result is always true or false now.
6861 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6862 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
6863 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6864 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
6866 ICI.setOperand(1, NewCst);
6867 Constant *NewAndCST;
6868 if (Shift->getOpcode() == Instruction::Shl)
6869 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6871 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6872 LHSI->setOperand(1, NewAndCST);
6873 LHSI->setOperand(0, Shift->getOperand(0));
6874 AddToWorkList(Shift); // Shift is dead.
6875 AddUsesToWorkList(ICI);
6881 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6882 // preferable because it allows the C<<Y expression to be hoisted out
6883 // of a loop if Y is invariant and X is not.
6884 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6885 ICI.isEquality() && !Shift->isArithmeticShift() &&
6886 !isa<Constant>(Shift->getOperand(0))) {
6889 if (Shift->getOpcode() == Instruction::LShr) {
6890 NS = BinaryOperator::CreateShl(AndCST,
6891 Shift->getOperand(1), "tmp");
6893 // Insert a logical shift.
6894 NS = BinaryOperator::CreateLShr(AndCST,
6895 Shift->getOperand(1), "tmp");
6897 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6899 // Compute X & (C << Y).
6900 Instruction *NewAnd =
6901 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6902 InsertNewInstBefore(NewAnd, ICI);
6904 ICI.setOperand(0, NewAnd);
6910 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6911 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6914 uint32_t TypeBits = RHSV.getBitWidth();
6916 // Check that the shift amount is in range. If not, don't perform
6917 // undefined shifts. When the shift is visited it will be
6919 if (ShAmt->uge(TypeBits))
6922 if (ICI.isEquality()) {
6923 // If we are comparing against bits always shifted out, the
6924 // comparison cannot succeed.
6926 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt),
6928 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6929 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6930 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6931 return ReplaceInstUsesWith(ICI, Cst);
6934 if (LHSI->hasOneUse()) {
6935 // Otherwise strength reduce the shift into an and.
6936 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6938 ConstantInt::get(*Context, APInt::getLowBitsSet(TypeBits,
6939 TypeBits-ShAmtVal));
6942 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6943 Mask, LHSI->getName()+".mask");
6944 Value *And = InsertNewInstBefore(AndI, ICI);
6945 return new ICmpInst(*Context, ICI.getPredicate(), And,
6946 ConstantInt::get(*Context, RHSV.lshr(ShAmtVal)));
6950 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6951 bool TrueIfSigned = false;
6952 if (LHSI->hasOneUse() &&
6953 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6954 // (X << 31) <s 0 --> (X&1) != 0
6955 Constant *Mask = ConstantInt::get(*Context, APInt(TypeBits, 1) <<
6956 (TypeBits-ShAmt->getZExtValue()-1));
6958 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6959 Mask, LHSI->getName()+".mask");
6960 Value *And = InsertNewInstBefore(AndI, ICI);
6962 return new ICmpInst(*Context,
6963 TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6964 And, Constant::getNullValue(And->getType()));
6969 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6970 case Instruction::AShr: {
6971 // Only handle equality comparisons of shift-by-constant.
6972 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6973 if (!ShAmt || !ICI.isEquality()) break;
6975 // Check that the shift amount is in range. If not, don't perform
6976 // undefined shifts. When the shift is visited it will be
6978 uint32_t TypeBits = RHSV.getBitWidth();
6979 if (ShAmt->uge(TypeBits))
6982 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6984 // If we are comparing against bits always shifted out, the
6985 // comparison cannot succeed.
6986 APInt Comp = RHSV << ShAmtVal;
6987 if (LHSI->getOpcode() == Instruction::LShr)
6988 Comp = Comp.lshr(ShAmtVal);
6990 Comp = Comp.ashr(ShAmtVal);
6992 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6993 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6994 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6995 return ReplaceInstUsesWith(ICI, Cst);
6998 // Otherwise, check to see if the bits shifted out are known to be zero.
6999 // If so, we can compare against the unshifted value:
7000 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
7001 if (LHSI->hasOneUse() &&
7002 MaskedValueIsZero(LHSI->getOperand(0),
7003 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
7004 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
7005 ConstantExpr::getShl(RHS, ShAmt));
7008 if (LHSI->hasOneUse()) {
7009 // Otherwise strength reduce the shift into an and.
7010 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7011 Constant *Mask = ConstantInt::get(*Context, Val);
7014 BinaryOperator::CreateAnd(LHSI->getOperand(0),
7015 Mask, LHSI->getName()+".mask");
7016 Value *And = InsertNewInstBefore(AndI, ICI);
7017 return new ICmpInst(*Context, ICI.getPredicate(), And,
7018 ConstantExpr::getShl(RHS, ShAmt));
7023 case Instruction::SDiv:
7024 case Instruction::UDiv:
7025 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7026 // Fold this div into the comparison, producing a range check.
7027 // Determine, based on the divide type, what the range is being
7028 // checked. If there is an overflow on the low or high side, remember
7029 // it, otherwise compute the range [low, hi) bounding the new value.
7030 // See: InsertRangeTest above for the kinds of replacements possible.
7031 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7032 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7037 case Instruction::Add:
7038 // Fold: icmp pred (add, X, C1), C2
7040 if (!ICI.isEquality()) {
7041 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7043 const APInt &LHSV = LHSC->getValue();
7045 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7048 if (ICI.isSignedPredicate()) {
7049 if (CR.getLower().isSignBit()) {
7050 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7051 ConstantInt::get(*Context, CR.getUpper()));
7052 } else if (CR.getUpper().isSignBit()) {
7053 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7054 ConstantInt::get(*Context, CR.getLower()));
7057 if (CR.getLower().isMinValue()) {
7058 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7059 ConstantInt::get(*Context, CR.getUpper()));
7060 } else if (CR.getUpper().isMinValue()) {
7061 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7062 ConstantInt::get(*Context, CR.getLower()));
7069 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7070 if (ICI.isEquality()) {
7071 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7073 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7074 // the second operand is a constant, simplify a bit.
7075 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7076 switch (BO->getOpcode()) {
7077 case Instruction::SRem:
7078 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7079 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7080 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7081 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7082 Instruction *NewRem =
7083 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
7085 InsertNewInstBefore(NewRem, ICI);
7086 return new ICmpInst(*Context, ICI.getPredicate(), NewRem,
7087 Constant::getNullValue(BO->getType()));
7091 case Instruction::Add:
7092 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7093 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7094 if (BO->hasOneUse())
7095 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7096 ConstantExpr::getSub(RHS, BOp1C));
7097 } else if (RHSV == 0) {
7098 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7099 // efficiently invertible, or if the add has just this one use.
7100 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7102 if (Value *NegVal = dyn_castNegVal(BOp1, Context))
7103 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, NegVal);
7104 else if (Value *NegVal = dyn_castNegVal(BOp0, Context))
7105 return new ICmpInst(*Context, ICI.getPredicate(), NegVal, BOp1);
7106 else if (BO->hasOneUse()) {
7107 Instruction *Neg = BinaryOperator::CreateNeg(*Context, BOp1);
7108 InsertNewInstBefore(Neg, ICI);
7110 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, Neg);
7114 case Instruction::Xor:
7115 // For the xor case, we can xor two constants together, eliminating
7116 // the explicit xor.
7117 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7118 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7119 ConstantExpr::getXor(RHS, BOC));
7122 case Instruction::Sub:
7123 // Replace (([sub|xor] A, B) != 0) with (A != B)
7125 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7129 case Instruction::Or:
7130 // If bits are being or'd in that are not present in the constant we
7131 // are comparing against, then the comparison could never succeed!
7132 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7133 Constant *NotCI = ConstantExpr::getNot(RHS);
7134 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
7135 return ReplaceInstUsesWith(ICI,
7136 ConstantInt::get(Type::Int1Ty,
7141 case Instruction::And:
7142 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7143 // If bits are being compared against that are and'd out, then the
7144 // comparison can never succeed!
7145 if ((RHSV & ~BOC->getValue()) != 0)
7146 return ReplaceInstUsesWith(ICI,
7147 ConstantInt::get(Type::Int1Ty,
7150 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7151 if (RHS == BOC && RHSV.isPowerOf2())
7152 return new ICmpInst(*Context, isICMP_NE ? ICmpInst::ICMP_EQ :
7153 ICmpInst::ICMP_NE, LHSI,
7154 Constant::getNullValue(RHS->getType()));
7156 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7157 if (BOC->getValue().isSignBit()) {
7158 Value *X = BO->getOperand(0);
7159 Constant *Zero = Constant::getNullValue(X->getType());
7160 ICmpInst::Predicate pred = isICMP_NE ?
7161 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7162 return new ICmpInst(*Context, pred, X, Zero);
7165 // ((X & ~7) == 0) --> X < 8
7166 if (RHSV == 0 && isHighOnes(BOC)) {
7167 Value *X = BO->getOperand(0);
7168 Constant *NegX = ConstantExpr::getNeg(BOC);
7169 ICmpInst::Predicate pred = isICMP_NE ?
7170 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7171 return new ICmpInst(*Context, pred, X, NegX);
7176 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7177 // Handle icmp {eq|ne} <intrinsic>, intcst.
7178 if (II->getIntrinsicID() == Intrinsic::bswap) {
7180 ICI.setOperand(0, II->getOperand(1));
7181 ICI.setOperand(1, ConstantInt::get(*Context, RHSV.byteSwap()));
7189 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7190 /// We only handle extending casts so far.
7192 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7193 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7194 Value *LHSCIOp = LHSCI->getOperand(0);
7195 const Type *SrcTy = LHSCIOp->getType();
7196 const Type *DestTy = LHSCI->getType();
7199 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7200 // integer type is the same size as the pointer type.
7201 if (TD && LHSCI->getOpcode() == Instruction::PtrToInt &&
7202 TD->getPointerSizeInBits() ==
7203 cast<IntegerType>(DestTy)->getBitWidth()) {
7205 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7206 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
7207 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7208 RHSOp = RHSC->getOperand(0);
7209 // If the pointer types don't match, insert a bitcast.
7210 if (LHSCIOp->getType() != RHSOp->getType())
7211 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
7215 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSOp);
7218 // The code below only handles extension cast instructions, so far.
7220 if (LHSCI->getOpcode() != Instruction::ZExt &&
7221 LHSCI->getOpcode() != Instruction::SExt)
7224 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7225 bool isSignedCmp = ICI.isSignedPredicate();
7227 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7228 // Not an extension from the same type?
7229 RHSCIOp = CI->getOperand(0);
7230 if (RHSCIOp->getType() != LHSCIOp->getType())
7233 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7234 // and the other is a zext), then we can't handle this.
7235 if (CI->getOpcode() != LHSCI->getOpcode())
7238 // Deal with equality cases early.
7239 if (ICI.isEquality())
7240 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7242 // A signed comparison of sign extended values simplifies into a
7243 // signed comparison.
7244 if (isSignedCmp && isSignedExt)
7245 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7247 // The other three cases all fold into an unsigned comparison.
7248 return new ICmpInst(*Context, ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7251 // If we aren't dealing with a constant on the RHS, exit early
7252 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7256 // Compute the constant that would happen if we truncated to SrcTy then
7257 // reextended to DestTy.
7258 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
7259 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(),
7262 // If the re-extended constant didn't change...
7264 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7265 // For example, we might have:
7266 // %A = sext i16 %X to i32
7267 // %B = icmp ugt i32 %A, 1330
7268 // It is incorrect to transform this into
7269 // %B = icmp ugt i16 %X, 1330
7270 // because %A may have negative value.
7272 // However, we allow this when the compare is EQ/NE, because they are
7274 if (isSignedExt == isSignedCmp || ICI.isEquality())
7275 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, Res1);
7279 // The re-extended constant changed so the constant cannot be represented
7280 // in the shorter type. Consequently, we cannot emit a simple comparison.
7282 // First, handle some easy cases. We know the result cannot be equal at this
7283 // point so handle the ICI.isEquality() cases
7284 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7285 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse(*Context));
7286 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7287 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue(*Context));
7289 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7290 // should have been folded away previously and not enter in here.
7293 // We're performing a signed comparison.
7294 if (cast<ConstantInt>(CI)->getValue().isNegative())
7295 Result = ConstantInt::getFalse(*Context); // X < (small) --> false
7297 Result = ConstantInt::getTrue(*Context); // X < (large) --> true
7299 // We're performing an unsigned comparison.
7301 // We're performing an unsigned comp with a sign extended value.
7302 // This is true if the input is >= 0. [aka >s -1]
7303 Constant *NegOne = Constant::getAllOnesValue(SrcTy);
7304 Result = InsertNewInstBefore(new ICmpInst(*Context, ICmpInst::ICMP_SGT,
7305 LHSCIOp, NegOne, ICI.getName()), ICI);
7307 // Unsigned extend & unsigned compare -> always true.
7308 Result = ConstantInt::getTrue(*Context);
7312 // Finally, return the value computed.
7313 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7314 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7315 return ReplaceInstUsesWith(ICI, Result);
7317 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7318 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7319 "ICmp should be folded!");
7320 if (Constant *CI = dyn_cast<Constant>(Result))
7321 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
7322 return BinaryOperator::CreateNot(*Context, Result);
7325 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7326 return commonShiftTransforms(I);
7329 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7330 return commonShiftTransforms(I);
7333 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7334 if (Instruction *R = commonShiftTransforms(I))
7337 Value *Op0 = I.getOperand(0);
7339 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7340 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7341 if (CSI->isAllOnesValue())
7342 return ReplaceInstUsesWith(I, CSI);
7344 // See if we can turn a signed shr into an unsigned shr.
7345 if (MaskedValueIsZero(Op0,
7346 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7347 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7349 // Arithmetic shifting an all-sign-bit value is a no-op.
7350 unsigned NumSignBits = ComputeNumSignBits(Op0);
7351 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7352 return ReplaceInstUsesWith(I, Op0);
7357 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7358 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7359 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7361 // shl X, 0 == X and shr X, 0 == X
7362 // shl 0, X == 0 and shr 0, X == 0
7363 if (Op1 == Constant::getNullValue(Op1->getType()) ||
7364 Op0 == Constant::getNullValue(Op0->getType()))
7365 return ReplaceInstUsesWith(I, Op0);
7367 if (isa<UndefValue>(Op0)) {
7368 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7369 return ReplaceInstUsesWith(I, Op0);
7370 else // undef << X -> 0, undef >>u X -> 0
7371 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7373 if (isa<UndefValue>(Op1)) {
7374 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7375 return ReplaceInstUsesWith(I, Op0);
7376 else // X << undef, X >>u undef -> 0
7377 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7380 // See if we can fold away this shift.
7381 if (SimplifyDemandedInstructionBits(I))
7384 // Try to fold constant and into select arguments.
7385 if (isa<Constant>(Op0))
7386 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7387 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7390 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7391 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7396 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7397 BinaryOperator &I) {
7398 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7400 // See if we can simplify any instructions used by the instruction whose sole
7401 // purpose is to compute bits we don't care about.
7402 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7404 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7407 if (Op1->uge(TypeBits)) {
7408 if (I.getOpcode() != Instruction::AShr)
7409 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
7411 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
7416 // ((X*C1) << C2) == (X * (C1 << C2))
7417 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7418 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7419 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7420 return BinaryOperator::CreateMul(BO->getOperand(0),
7421 ConstantExpr::getShl(BOOp, Op1));
7423 // Try to fold constant and into select arguments.
7424 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7425 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7427 if (isa<PHINode>(Op0))
7428 if (Instruction *NV = FoldOpIntoPhi(I))
7431 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7432 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7433 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7434 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7435 // place. Don't try to do this transformation in this case. Also, we
7436 // require that the input operand is a shift-by-constant so that we have
7437 // confidence that the shifts will get folded together. We could do this
7438 // xform in more cases, but it is unlikely to be profitable.
7439 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7440 isa<ConstantInt>(TrOp->getOperand(1))) {
7441 // Okay, we'll do this xform. Make the shift of shift.
7442 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
7443 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7445 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7447 // For logical shifts, the truncation has the effect of making the high
7448 // part of the register be zeros. Emulate this by inserting an AND to
7449 // clear the top bits as needed. This 'and' will usually be zapped by
7450 // other xforms later if dead.
7451 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7452 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7453 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7455 // The mask we constructed says what the trunc would do if occurring
7456 // between the shifts. We want to know the effect *after* the second
7457 // shift. We know that it is a logical shift by a constant, so adjust the
7458 // mask as appropriate.
7459 if (I.getOpcode() == Instruction::Shl)
7460 MaskV <<= Op1->getZExtValue();
7462 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7463 MaskV = MaskV.lshr(Op1->getZExtValue());
7467 BinaryOperator::CreateAnd(NSh, ConstantInt::get(*Context, MaskV),
7469 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7471 // Return the value truncated to the interesting size.
7472 return new TruncInst(And, I.getType());
7476 if (Op0->hasOneUse()) {
7477 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7478 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7481 switch (Op0BO->getOpcode()) {
7483 case Instruction::Add:
7484 case Instruction::And:
7485 case Instruction::Or:
7486 case Instruction::Xor: {
7487 // These operators commute.
7488 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7489 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7490 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7491 m_Specific(Op1)), *Context)){
7492 Instruction *YS = BinaryOperator::CreateShl(
7493 Op0BO->getOperand(0), Op1,
7495 InsertNewInstBefore(YS, I); // (Y << C)
7497 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7498 Op0BO->getOperand(1)->getName());
7499 InsertNewInstBefore(X, I); // (X + (Y << C))
7500 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7501 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7502 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7505 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7506 Value *Op0BOOp1 = Op0BO->getOperand(1);
7507 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7509 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7510 m_ConstantInt(CC)), *Context) &&
7511 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7512 Instruction *YS = BinaryOperator::CreateShl(
7513 Op0BO->getOperand(0), Op1,
7515 InsertNewInstBefore(YS, I); // (Y << C)
7517 BinaryOperator::CreateAnd(V1,
7518 ConstantExpr::getShl(CC, Op1),
7519 V1->getName()+".mask");
7520 InsertNewInstBefore(XM, I); // X & (CC << C)
7522 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7527 case Instruction::Sub: {
7528 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7529 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7530 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7531 m_Specific(Op1)), *Context)){
7532 Instruction *YS = BinaryOperator::CreateShl(
7533 Op0BO->getOperand(1), Op1,
7535 InsertNewInstBefore(YS, I); // (Y << C)
7537 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7538 Op0BO->getOperand(0)->getName());
7539 InsertNewInstBefore(X, I); // (X + (Y << C))
7540 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7541 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context,
7542 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7545 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7546 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7547 match(Op0BO->getOperand(0),
7548 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7549 m_ConstantInt(CC)), *Context) && V2 == Op1 &&
7550 cast<BinaryOperator>(Op0BO->getOperand(0))
7551 ->getOperand(0)->hasOneUse()) {
7552 Instruction *YS = BinaryOperator::CreateShl(
7553 Op0BO->getOperand(1), Op1,
7555 InsertNewInstBefore(YS, I); // (Y << C)
7557 BinaryOperator::CreateAnd(V1,
7558 ConstantExpr::getShl(CC, Op1),
7559 V1->getName()+".mask");
7560 InsertNewInstBefore(XM, I); // X & (CC << C)
7562 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7570 // If the operand is an bitwise operator with a constant RHS, and the
7571 // shift is the only use, we can pull it out of the shift.
7572 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7573 bool isValid = true; // Valid only for And, Or, Xor
7574 bool highBitSet = false; // Transform if high bit of constant set?
7576 switch (Op0BO->getOpcode()) {
7577 default: isValid = false; break; // Do not perform transform!
7578 case Instruction::Add:
7579 isValid = isLeftShift;
7581 case Instruction::Or:
7582 case Instruction::Xor:
7585 case Instruction::And:
7590 // If this is a signed shift right, and the high bit is modified
7591 // by the logical operation, do not perform the transformation.
7592 // The highBitSet boolean indicates the value of the high bit of
7593 // the constant which would cause it to be modified for this
7596 if (isValid && I.getOpcode() == Instruction::AShr)
7597 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7600 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
7602 Instruction *NewShift =
7603 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7604 InsertNewInstBefore(NewShift, I);
7605 NewShift->takeName(Op0BO);
7607 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7614 // Find out if this is a shift of a shift by a constant.
7615 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7616 if (ShiftOp && !ShiftOp->isShift())
7619 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7620 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7621 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7622 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7623 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7624 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7625 Value *X = ShiftOp->getOperand(0);
7627 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7629 const IntegerType *Ty = cast<IntegerType>(I.getType());
7631 // Check for (X << c1) << c2 and (X >> c1) >> c2
7632 if (I.getOpcode() == ShiftOp->getOpcode()) {
7633 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7635 if (AmtSum >= TypeBits) {
7636 if (I.getOpcode() != Instruction::AShr)
7637 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7638 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7641 return BinaryOperator::Create(I.getOpcode(), X,
7642 ConstantInt::get(Ty, AmtSum));
7643 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7644 I.getOpcode() == Instruction::AShr) {
7645 if (AmtSum >= TypeBits)
7646 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
7648 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7649 return BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, AmtSum));
7650 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7651 I.getOpcode() == Instruction::LShr) {
7652 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7653 if (AmtSum >= TypeBits)
7654 AmtSum = TypeBits-1;
7656 Instruction *Shift =
7657 BinaryOperator::CreateAShr(X, ConstantInt::get(Ty, AmtSum));
7658 InsertNewInstBefore(Shift, I);
7660 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7661 return BinaryOperator::CreateAnd(Shift, ConstantInt::get(*Context, Mask));
7664 // Okay, if we get here, one shift must be left, and the other shift must be
7665 // right. See if the amounts are equal.
7666 if (ShiftAmt1 == ShiftAmt2) {
7667 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7668 if (I.getOpcode() == Instruction::Shl) {
7669 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7670 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7672 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7673 if (I.getOpcode() == Instruction::LShr) {
7674 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7675 return BinaryOperator::CreateAnd(X, ConstantInt::get(*Context, Mask));
7677 // We can simplify ((X << C) >>s C) into a trunc + sext.
7678 // NOTE: we could do this for any C, but that would make 'unusual' integer
7679 // types. For now, just stick to ones well-supported by the code
7681 const Type *SExtType = 0;
7682 switch (Ty->getBitWidth() - ShiftAmt1) {
7689 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
7694 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7695 InsertNewInstBefore(NewTrunc, I);
7696 return new SExtInst(NewTrunc, Ty);
7698 // Otherwise, we can't handle it yet.
7699 } else if (ShiftAmt1 < ShiftAmt2) {
7700 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7702 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7703 if (I.getOpcode() == Instruction::Shl) {
7704 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7705 ShiftOp->getOpcode() == Instruction::AShr);
7706 Instruction *Shift =
7707 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7708 InsertNewInstBefore(Shift, I);
7710 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7711 return BinaryOperator::CreateAnd(Shift,
7712 ConstantInt::get(*Context, Mask));
7715 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7716 if (I.getOpcode() == Instruction::LShr) {
7717 assert(ShiftOp->getOpcode() == Instruction::Shl);
7718 Instruction *Shift =
7719 BinaryOperator::CreateLShr(X, ConstantInt::get(Ty, ShiftDiff));
7720 InsertNewInstBefore(Shift, I);
7722 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7723 return BinaryOperator::CreateAnd(Shift,
7724 ConstantInt::get(*Context, Mask));
7727 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7729 assert(ShiftAmt2 < ShiftAmt1);
7730 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7732 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7733 if (I.getOpcode() == Instruction::Shl) {
7734 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7735 ShiftOp->getOpcode() == Instruction::AShr);
7736 Instruction *Shift =
7737 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7738 ConstantInt::get(Ty, ShiftDiff));
7739 InsertNewInstBefore(Shift, I);
7741 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7742 return BinaryOperator::CreateAnd(Shift,
7743 ConstantInt::get(*Context, Mask));
7746 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7747 if (I.getOpcode() == Instruction::LShr) {
7748 assert(ShiftOp->getOpcode() == Instruction::Shl);
7749 Instruction *Shift =
7750 BinaryOperator::CreateShl(X, ConstantInt::get(Ty, ShiftDiff));
7751 InsertNewInstBefore(Shift, I);
7753 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7754 return BinaryOperator::CreateAnd(Shift,
7755 ConstantInt::get(*Context, Mask));
7758 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7765 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7766 /// expression. If so, decompose it, returning some value X, such that Val is
7769 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7770 int &Offset, LLVMContext *Context) {
7771 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7772 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7773 Offset = CI->getZExtValue();
7775 return ConstantInt::get(Type::Int32Ty, 0);
7776 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7777 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7778 if (I->getOpcode() == Instruction::Shl) {
7779 // This is a value scaled by '1 << the shift amt'.
7780 Scale = 1U << RHS->getZExtValue();
7782 return I->getOperand(0);
7783 } else if (I->getOpcode() == Instruction::Mul) {
7784 // This value is scaled by 'RHS'.
7785 Scale = RHS->getZExtValue();
7787 return I->getOperand(0);
7788 } else if (I->getOpcode() == Instruction::Add) {
7789 // We have X+C. Check to see if we really have (X*C2)+C1,
7790 // where C1 is divisible by C2.
7793 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7795 Offset += RHS->getZExtValue();
7802 // Otherwise, we can't look past this.
7809 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7810 /// try to eliminate the cast by moving the type information into the alloc.
7811 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7812 AllocationInst &AI) {
7813 const PointerType *PTy = cast<PointerType>(CI.getType());
7815 // Remove any uses of AI that are dead.
7816 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7818 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7819 Instruction *User = cast<Instruction>(*UI++);
7820 if (isInstructionTriviallyDead(User)) {
7821 while (UI != E && *UI == User)
7822 ++UI; // If this instruction uses AI more than once, don't break UI.
7825 DOUT << "IC: DCE: " << *User;
7826 EraseInstFromFunction(*User);
7830 // This requires TargetData to get the alloca alignment and size information.
7833 // Get the type really allocated and the type casted to.
7834 const Type *AllocElTy = AI.getAllocatedType();
7835 const Type *CastElTy = PTy->getElementType();
7836 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7838 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7839 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7840 if (CastElTyAlign < AllocElTyAlign) return 0;
7842 // If the allocation has multiple uses, only promote it if we are strictly
7843 // increasing the alignment of the resultant allocation. If we keep it the
7844 // same, we open the door to infinite loops of various kinds. (A reference
7845 // from a dbg.declare doesn't count as a use for this purpose.)
7846 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7847 CastElTyAlign == AllocElTyAlign) return 0;
7849 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7850 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7851 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7853 // See if we can satisfy the modulus by pulling a scale out of the array
7855 unsigned ArraySizeScale;
7857 Value *NumElements = // See if the array size is a decomposable linear expr.
7858 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7859 ArrayOffset, Context);
7861 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7863 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7864 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7866 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7871 // If the allocation size is constant, form a constant mul expression
7872 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7873 if (isa<ConstantInt>(NumElements))
7874 Amt = ConstantExpr::getMul(cast<ConstantInt>(NumElements),
7875 cast<ConstantInt>(Amt));
7876 // otherwise multiply the amount and the number of elements
7878 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7879 Amt = InsertNewInstBefore(Tmp, AI);
7883 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7884 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7885 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7886 Amt = InsertNewInstBefore(Tmp, AI);
7889 AllocationInst *New;
7890 if (isa<MallocInst>(AI))
7891 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7893 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7894 InsertNewInstBefore(New, AI);
7897 // If the allocation has one real use plus a dbg.declare, just remove the
7899 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7900 EraseInstFromFunction(*DI);
7902 // If the allocation has multiple real uses, insert a cast and change all
7903 // things that used it to use the new cast. This will also hack on CI, but it
7905 else if (!AI.hasOneUse()) {
7906 AddUsesToWorkList(AI);
7907 // New is the allocation instruction, pointer typed. AI is the original
7908 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7909 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7910 InsertNewInstBefore(NewCast, AI);
7911 AI.replaceAllUsesWith(NewCast);
7913 return ReplaceInstUsesWith(CI, New);
7916 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7917 /// and return it as type Ty without inserting any new casts and without
7918 /// changing the computed value. This is used by code that tries to decide
7919 /// whether promoting or shrinking integer operations to wider or smaller types
7920 /// will allow us to eliminate a truncate or extend.
7922 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7923 /// extension operation if Ty is larger.
7925 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7926 /// should return true if trunc(V) can be computed by computing V in the smaller
7927 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7928 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7929 /// efficiently truncated.
7931 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7932 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7933 /// the final result.
7934 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7936 int &NumCastsRemoved){
7937 // We can always evaluate constants in another type.
7938 if (isa<Constant>(V))
7941 Instruction *I = dyn_cast<Instruction>(V);
7942 if (!I) return false;
7944 const Type *OrigTy = V->getType();
7946 // If this is an extension or truncate, we can often eliminate it.
7947 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7948 // If this is a cast from the destination type, we can trivially eliminate
7949 // it, and this will remove a cast overall.
7950 if (I->getOperand(0)->getType() == Ty) {
7951 // If the first operand is itself a cast, and is eliminable, do not count
7952 // this as an eliminable cast. We would prefer to eliminate those two
7954 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7960 // We can't extend or shrink something that has multiple uses: doing so would
7961 // require duplicating the instruction in general, which isn't profitable.
7962 if (!I->hasOneUse()) return false;
7964 unsigned Opc = I->getOpcode();
7966 case Instruction::Add:
7967 case Instruction::Sub:
7968 case Instruction::Mul:
7969 case Instruction::And:
7970 case Instruction::Or:
7971 case Instruction::Xor:
7972 // These operators can all arbitrarily be extended or truncated.
7973 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7975 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7978 case Instruction::UDiv:
7979 case Instruction::URem: {
7980 // UDiv and URem can be truncated if all the truncated bits are zero.
7981 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7982 uint32_t BitWidth = Ty->getScalarSizeInBits();
7983 if (BitWidth < OrigBitWidth) {
7984 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7985 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7986 MaskedValueIsZero(I->getOperand(1), Mask)) {
7987 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7989 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7995 case Instruction::Shl:
7996 // If we are truncating the result of this SHL, and if it's a shift of a
7997 // constant amount, we can always perform a SHL in a smaller type.
7998 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7999 uint32_t BitWidth = Ty->getScalarSizeInBits();
8000 if (BitWidth < OrigTy->getScalarSizeInBits() &&
8001 CI->getLimitedValue(BitWidth) < BitWidth)
8002 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8006 case Instruction::LShr:
8007 // If this is a truncate of a logical shr, we can truncate it to a smaller
8008 // lshr iff we know that the bits we would otherwise be shifting in are
8010 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8011 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8012 uint32_t BitWidth = Ty->getScalarSizeInBits();
8013 if (BitWidth < OrigBitWidth &&
8014 MaskedValueIsZero(I->getOperand(0),
8015 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
8016 CI->getLimitedValue(BitWidth) < BitWidth) {
8017 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8022 case Instruction::ZExt:
8023 case Instruction::SExt:
8024 case Instruction::Trunc:
8025 // If this is the same kind of case as our original (e.g. zext+zext), we
8026 // can safely replace it. Note that replacing it does not reduce the number
8027 // of casts in the input.
8031 // sext (zext ty1), ty2 -> zext ty2
8032 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8035 case Instruction::Select: {
8036 SelectInst *SI = cast<SelectInst>(I);
8037 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8039 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8042 case Instruction::PHI: {
8043 // We can change a phi if we can change all operands.
8044 PHINode *PN = cast<PHINode>(I);
8045 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8046 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8052 // TODO: Can handle more cases here.
8059 /// EvaluateInDifferentType - Given an expression that
8060 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8061 /// evaluate the expression.
8062 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8064 if (Constant *C = dyn_cast<Constant>(V))
8065 return ConstantExpr::getIntegerCast(C, Ty,
8066 isSigned /*Sext or ZExt*/);
8068 // Otherwise, it must be an instruction.
8069 Instruction *I = cast<Instruction>(V);
8070 Instruction *Res = 0;
8071 unsigned Opc = I->getOpcode();
8073 case Instruction::Add:
8074 case Instruction::Sub:
8075 case Instruction::Mul:
8076 case Instruction::And:
8077 case Instruction::Or:
8078 case Instruction::Xor:
8079 case Instruction::AShr:
8080 case Instruction::LShr:
8081 case Instruction::Shl:
8082 case Instruction::UDiv:
8083 case Instruction::URem: {
8084 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8085 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8086 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8089 case Instruction::Trunc:
8090 case Instruction::ZExt:
8091 case Instruction::SExt:
8092 // If the source type of the cast is the type we're trying for then we can
8093 // just return the source. There's no need to insert it because it is not
8095 if (I->getOperand(0)->getType() == Ty)
8096 return I->getOperand(0);
8098 // Otherwise, must be the same type of cast, so just reinsert a new one.
8099 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8102 case Instruction::Select: {
8103 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8104 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8105 Res = SelectInst::Create(I->getOperand(0), True, False);
8108 case Instruction::PHI: {
8109 PHINode *OPN = cast<PHINode>(I);
8110 PHINode *NPN = PHINode::Create(Ty);
8111 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8112 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8113 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8119 // TODO: Can handle more cases here.
8120 llvm_unreachable("Unreachable!");
8125 return InsertNewInstBefore(Res, *I);
8128 /// @brief Implement the transforms common to all CastInst visitors.
8129 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8130 Value *Src = CI.getOperand(0);
8132 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8133 // eliminate it now.
8134 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8135 if (Instruction::CastOps opc =
8136 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8137 // The first cast (CSrc) is eliminable so we need to fix up or replace
8138 // the second cast (CI). CSrc will then have a good chance of being dead.
8139 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8143 // If we are casting a select then fold the cast into the select
8144 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8145 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8148 // If we are casting a PHI then fold the cast into the PHI
8149 if (isa<PHINode>(Src))
8150 if (Instruction *NV = FoldOpIntoPhi(CI))
8156 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8157 /// or not there is a sequence of GEP indices into the type that will land us at
8158 /// the specified offset. If so, fill them into NewIndices and return the
8159 /// resultant element type, otherwise return null.
8160 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8161 SmallVectorImpl<Value*> &NewIndices,
8162 const TargetData *TD,
8163 LLVMContext *Context) {
8165 if (!Ty->isSized()) return 0;
8167 // Start with the index over the outer type. Note that the type size
8168 // might be zero (even if the offset isn't zero) if the indexed type
8169 // is something like [0 x {int, int}]
8170 const Type *IntPtrTy = TD->getIntPtrType();
8171 int64_t FirstIdx = 0;
8172 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8173 FirstIdx = Offset/TySize;
8174 Offset -= FirstIdx*TySize;
8176 // Handle hosts where % returns negative instead of values [0..TySize).
8180 assert(Offset >= 0);
8182 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8185 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
8187 // Index into the types. If we fail, set OrigBase to null.
8189 // Indexing into tail padding between struct/array elements.
8190 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8193 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8194 const StructLayout *SL = TD->getStructLayout(STy);
8195 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8196 "Offset must stay within the indexed type");
8198 unsigned Elt = SL->getElementContainingOffset(Offset);
8199 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
8201 Offset -= SL->getElementOffset(Elt);
8202 Ty = STy->getElementType(Elt);
8203 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8204 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8205 assert(EltSize && "Cannot index into a zero-sized array");
8206 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
8208 Ty = AT->getElementType();
8210 // Otherwise, we can't index into the middle of this atomic type, bail.
8218 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8219 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8220 Value *Src = CI.getOperand(0);
8222 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8223 // If casting the result of a getelementptr instruction with no offset, turn
8224 // this into a cast of the original pointer!
8225 if (GEP->hasAllZeroIndices()) {
8226 // Changing the cast operand is usually not a good idea but it is safe
8227 // here because the pointer operand is being replaced with another
8228 // pointer operand so the opcode doesn't need to change.
8230 CI.setOperand(0, GEP->getOperand(0));
8234 // If the GEP has a single use, and the base pointer is a bitcast, and the
8235 // GEP computes a constant offset, see if we can convert these three
8236 // instructions into fewer. This typically happens with unions and other
8237 // non-type-safe code.
8238 if (TD && GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8239 if (GEP->hasAllConstantIndices()) {
8240 // We are guaranteed to get a constant from EmitGEPOffset.
8241 ConstantInt *OffsetV =
8242 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8243 int64_t Offset = OffsetV->getSExtValue();
8245 // Get the base pointer input of the bitcast, and the type it points to.
8246 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8247 const Type *GEPIdxTy =
8248 cast<PointerType>(OrigBase->getType())->getElementType();
8249 SmallVector<Value*, 8> NewIndices;
8250 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8251 // If we were able to index down into an element, create the GEP
8252 // and bitcast the result. This eliminates one bitcast, potentially
8254 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
8256 NewIndices.end(), "");
8257 InsertNewInstBefore(NGEP, CI);
8258 NGEP->takeName(GEP);
8259 if (cast<GEPOperator>(GEP)->isInBounds())
8260 cast<GEPOperator>(NGEP)->setIsInBounds(true);
8262 if (isa<BitCastInst>(CI))
8263 return new BitCastInst(NGEP, CI.getType());
8264 assert(isa<PtrToIntInst>(CI));
8265 return new PtrToIntInst(NGEP, CI.getType());
8271 return commonCastTransforms(CI);
8274 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8275 /// type like i42. We don't want to introduce operations on random non-legal
8276 /// integer types where they don't already exist in the code. In the future,
8277 /// we should consider making this based off target-data, so that 32-bit targets
8278 /// won't get i64 operations etc.
8279 static bool isSafeIntegerType(const Type *Ty) {
8280 switch (Ty->getPrimitiveSizeInBits()) {
8291 /// commonIntCastTransforms - This function implements the common transforms
8292 /// for trunc, zext, and sext.
8293 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8294 if (Instruction *Result = commonCastTransforms(CI))
8297 Value *Src = CI.getOperand(0);
8298 const Type *SrcTy = Src->getType();
8299 const Type *DestTy = CI.getType();
8300 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8301 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8303 // See if we can simplify any instructions used by the LHS whose sole
8304 // purpose is to compute bits we don't care about.
8305 if (SimplifyDemandedInstructionBits(CI))
8308 // If the source isn't an instruction or has more than one use then we
8309 // can't do anything more.
8310 Instruction *SrcI = dyn_cast<Instruction>(Src);
8311 if (!SrcI || !Src->hasOneUse())
8314 // Attempt to propagate the cast into the instruction for int->int casts.
8315 int NumCastsRemoved = 0;
8316 // Only do this if the dest type is a simple type, don't convert the
8317 // expression tree to something weird like i93 unless the source is also
8319 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8320 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8321 CanEvaluateInDifferentType(SrcI, DestTy,
8322 CI.getOpcode(), NumCastsRemoved)) {
8323 // If this cast is a truncate, evaluting in a different type always
8324 // eliminates the cast, so it is always a win. If this is a zero-extension,
8325 // we need to do an AND to maintain the clear top-part of the computation,
8326 // so we require that the input have eliminated at least one cast. If this
8327 // is a sign extension, we insert two new casts (to do the extension) so we
8328 // require that two casts have been eliminated.
8329 bool DoXForm = false;
8330 bool JustReplace = false;
8331 switch (CI.getOpcode()) {
8333 // All the others use floating point so we shouldn't actually
8334 // get here because of the check above.
8335 llvm_unreachable("Unknown cast type");
8336 case Instruction::Trunc:
8339 case Instruction::ZExt: {
8340 DoXForm = NumCastsRemoved >= 1;
8341 if (!DoXForm && 0) {
8342 // If it's unnecessary to issue an AND to clear the high bits, it's
8343 // always profitable to do this xform.
8344 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8345 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8346 if (MaskedValueIsZero(TryRes, Mask))
8347 return ReplaceInstUsesWith(CI, TryRes);
8349 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8350 if (TryI->use_empty())
8351 EraseInstFromFunction(*TryI);
8355 case Instruction::SExt: {
8356 DoXForm = NumCastsRemoved >= 2;
8357 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8358 // If we do not have to emit the truncate + sext pair, then it's always
8359 // profitable to do this xform.
8361 // It's not safe to eliminate the trunc + sext pair if one of the
8362 // eliminated cast is a truncate. e.g.
8363 // t2 = trunc i32 t1 to i16
8364 // t3 = sext i16 t2 to i32
8367 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8368 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8369 if (NumSignBits > (DestBitSize - SrcBitSize))
8370 return ReplaceInstUsesWith(CI, TryRes);
8372 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8373 if (TryI->use_empty())
8374 EraseInstFromFunction(*TryI);
8381 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
8383 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8384 CI.getOpcode() == Instruction::SExt);
8386 // Just replace this cast with the result.
8387 return ReplaceInstUsesWith(CI, Res);
8389 assert(Res->getType() == DestTy);
8390 switch (CI.getOpcode()) {
8391 default: llvm_unreachable("Unknown cast type!");
8392 case Instruction::Trunc:
8393 // Just replace this cast with the result.
8394 return ReplaceInstUsesWith(CI, Res);
8395 case Instruction::ZExt: {
8396 assert(SrcBitSize < DestBitSize && "Not a zext?");
8398 // If the high bits are already zero, just replace this cast with the
8400 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8401 if (MaskedValueIsZero(Res, Mask))
8402 return ReplaceInstUsesWith(CI, Res);
8404 // We need to emit an AND to clear the high bits.
8405 Constant *C = ConstantInt::get(*Context,
8406 APInt::getLowBitsSet(DestBitSize, SrcBitSize));
8407 return BinaryOperator::CreateAnd(Res, C);
8409 case Instruction::SExt: {
8410 // If the high bits are already filled with sign bit, just replace this
8411 // cast with the result.
8412 unsigned NumSignBits = ComputeNumSignBits(Res);
8413 if (NumSignBits > (DestBitSize - SrcBitSize))
8414 return ReplaceInstUsesWith(CI, Res);
8416 // We need to emit a cast to truncate, then a cast to sext.
8417 return CastInst::Create(Instruction::SExt,
8418 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8425 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8426 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8428 switch (SrcI->getOpcode()) {
8429 case Instruction::Add:
8430 case Instruction::Mul:
8431 case Instruction::And:
8432 case Instruction::Or:
8433 case Instruction::Xor:
8434 // If we are discarding information, rewrite.
8435 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8436 // Don't insert two casts unless at least one can be eliminated.
8437 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8438 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8439 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8440 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8441 return BinaryOperator::Create(
8442 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8446 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8447 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8448 SrcI->getOpcode() == Instruction::Xor &&
8449 Op1 == ConstantInt::getTrue(*Context) &&
8450 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8451 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8452 return BinaryOperator::CreateXor(New,
8453 ConstantInt::get(CI.getType(), 1));
8457 case Instruction::Shl: {
8458 // Canonicalize trunc inside shl, if we can.
8459 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8460 if (CI && DestBitSize < SrcBitSize &&
8461 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8462 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8463 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8464 return BinaryOperator::CreateShl(Op0c, Op1c);
8472 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8473 if (Instruction *Result = commonIntCastTransforms(CI))
8476 Value *Src = CI.getOperand(0);
8477 const Type *Ty = CI.getType();
8478 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8479 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8481 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8482 if (DestBitWidth == 1) {
8483 Constant *One = ConstantInt::get(Src->getType(), 1);
8484 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8485 Value *Zero = Constant::getNullValue(Src->getType());
8486 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Src, Zero);
8489 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8490 ConstantInt *ShAmtV = 0;
8492 if (Src->hasOneUse() &&
8493 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)), *Context)) {
8494 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8496 // Get a mask for the bits shifting in.
8497 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8498 if (MaskedValueIsZero(ShiftOp, Mask)) {
8499 if (ShAmt >= DestBitWidth) // All zeros.
8500 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
8502 // Okay, we can shrink this. Truncate the input, then return a new
8504 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8505 Value *V2 = ConstantExpr::getTrunc(ShAmtV, Ty);
8506 return BinaryOperator::CreateLShr(V1, V2);
8513 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8514 /// in order to eliminate the icmp.
8515 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8517 // If we are just checking for a icmp eq of a single bit and zext'ing it
8518 // to an integer, then shift the bit to the appropriate place and then
8519 // cast to integer to avoid the comparison.
8520 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8521 const APInt &Op1CV = Op1C->getValue();
8523 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8524 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8525 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8526 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8527 if (!DoXform) return ICI;
8529 Value *In = ICI->getOperand(0);
8530 Value *Sh = ConstantInt::get(In->getType(),
8531 In->getType()->getScalarSizeInBits()-1);
8532 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8533 In->getName()+".lobit"),
8535 if (In->getType() != CI.getType())
8536 In = CastInst::CreateIntegerCast(In, CI.getType(),
8537 false/*ZExt*/, "tmp", &CI);
8539 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8540 Constant *One = ConstantInt::get(In->getType(), 1);
8541 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8542 In->getName()+".not"),
8546 return ReplaceInstUsesWith(CI, In);
8551 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8552 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8553 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8554 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8555 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8556 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8557 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8558 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8559 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8560 // This only works for EQ and NE
8561 ICI->isEquality()) {
8562 // If Op1C some other power of two, convert:
8563 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8564 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8565 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8566 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8568 APInt KnownZeroMask(~KnownZero);
8569 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8570 if (!DoXform) return ICI;
8572 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8573 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8574 // (X&4) == 2 --> false
8575 // (X&4) != 2 --> true
8576 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
8577 Res = ConstantExpr::getZExt(Res, CI.getType());
8578 return ReplaceInstUsesWith(CI, Res);
8581 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8582 Value *In = ICI->getOperand(0);
8584 // Perform a logical shr by shiftamt.
8585 // Insert the shift to put the result in the low bit.
8586 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8587 ConstantInt::get(In->getType(), ShiftAmt),
8588 In->getName()+".lobit"), CI);
8591 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8592 Constant *One = ConstantInt::get(In->getType(), 1);
8593 In = BinaryOperator::CreateXor(In, One, "tmp");
8594 InsertNewInstBefore(cast<Instruction>(In), CI);
8597 if (CI.getType() == In->getType())
8598 return ReplaceInstUsesWith(CI, In);
8600 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8608 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8609 // If one of the common conversion will work ..
8610 if (Instruction *Result = commonIntCastTransforms(CI))
8613 Value *Src = CI.getOperand(0);
8615 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8616 // types and if the sizes are just right we can convert this into a logical
8617 // 'and' which will be much cheaper than the pair of casts.
8618 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8619 // Get the sizes of the types involved. We know that the intermediate type
8620 // will be smaller than A or C, but don't know the relation between A and C.
8621 Value *A = CSrc->getOperand(0);
8622 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8623 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8624 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8625 // If we're actually extending zero bits, then if
8626 // SrcSize < DstSize: zext(a & mask)
8627 // SrcSize == DstSize: a & mask
8628 // SrcSize > DstSize: trunc(a) & mask
8629 if (SrcSize < DstSize) {
8630 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8631 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
8633 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8634 InsertNewInstBefore(And, CI);
8635 return new ZExtInst(And, CI.getType());
8636 } else if (SrcSize == DstSize) {
8637 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8638 return BinaryOperator::CreateAnd(A, ConstantInt::get(A->getType(),
8640 } else if (SrcSize > DstSize) {
8641 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8642 InsertNewInstBefore(Trunc, CI);
8643 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8644 return BinaryOperator::CreateAnd(Trunc,
8645 ConstantInt::get(Trunc->getType(),
8650 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8651 return transformZExtICmp(ICI, CI);
8653 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8654 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8655 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8656 // of the (zext icmp) will be transformed.
8657 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8658 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8659 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8660 (transformZExtICmp(LHS, CI, false) ||
8661 transformZExtICmp(RHS, CI, false))) {
8662 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8663 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8664 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8668 // zext(trunc(t) & C) -> (t & zext(C)).
8669 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8670 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8671 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8672 Value *TI0 = TI->getOperand(0);
8673 if (TI0->getType() == CI.getType())
8675 BinaryOperator::CreateAnd(TI0,
8676 ConstantExpr::getZExt(C, CI.getType()));
8679 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8680 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8681 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8682 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8683 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8684 And->getOperand(1) == C)
8685 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8686 Value *TI0 = TI->getOperand(0);
8687 if (TI0->getType() == CI.getType()) {
8688 Constant *ZC = ConstantExpr::getZExt(C, CI.getType());
8689 Instruction *NewAnd = BinaryOperator::CreateAnd(TI0, ZC, "tmp");
8690 InsertNewInstBefore(NewAnd, *And);
8691 return BinaryOperator::CreateXor(NewAnd, ZC);
8698 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8699 if (Instruction *I = commonIntCastTransforms(CI))
8702 Value *Src = CI.getOperand(0);
8704 // Canonicalize sign-extend from i1 to a select.
8705 if (Src->getType() == Type::Int1Ty)
8706 return SelectInst::Create(Src,
8707 Constant::getAllOnesValue(CI.getType()),
8708 Constant::getNullValue(CI.getType()));
8710 // See if the value being truncated is already sign extended. If so, just
8711 // eliminate the trunc/sext pair.
8712 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8713 Value *Op = cast<User>(Src)->getOperand(0);
8714 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8715 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8716 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8717 unsigned NumSignBits = ComputeNumSignBits(Op);
8719 if (OpBits == DestBits) {
8720 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8721 // bits, it is already ready.
8722 if (NumSignBits > DestBits-MidBits)
8723 return ReplaceInstUsesWith(CI, Op);
8724 } else if (OpBits < DestBits) {
8725 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8726 // bits, just sext from i32.
8727 if (NumSignBits > OpBits-MidBits)
8728 return new SExtInst(Op, CI.getType(), "tmp");
8730 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8731 // bits, just truncate to i32.
8732 if (NumSignBits > OpBits-MidBits)
8733 return new TruncInst(Op, CI.getType(), "tmp");
8737 // If the input is a shl/ashr pair of a same constant, then this is a sign
8738 // extension from a smaller value. If we could trust arbitrary bitwidth
8739 // integers, we could turn this into a truncate to the smaller bit and then
8740 // use a sext for the whole extension. Since we don't, look deeper and check
8741 // for a truncate. If the source and dest are the same type, eliminate the
8742 // trunc and extend and just do shifts. For example, turn:
8743 // %a = trunc i32 %i to i8
8744 // %b = shl i8 %a, 6
8745 // %c = ashr i8 %b, 6
8746 // %d = sext i8 %c to i32
8748 // %a = shl i32 %i, 30
8749 // %d = ashr i32 %a, 30
8751 ConstantInt *BA = 0, *CA = 0;
8752 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8753 m_ConstantInt(CA)), *Context) &&
8754 BA == CA && isa<TruncInst>(A)) {
8755 Value *I = cast<TruncInst>(A)->getOperand(0);
8756 if (I->getType() == CI.getType()) {
8757 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8758 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8759 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8760 Constant *ShAmtV = ConstantInt::get(CI.getType(), ShAmt);
8761 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8763 return BinaryOperator::CreateAShr(I, ShAmtV);
8770 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8771 /// in the specified FP type without changing its value.
8772 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8773 LLVMContext *Context) {
8775 APFloat F = CFP->getValueAPF();
8776 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8778 return ConstantFP::get(*Context, F);
8782 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8783 /// through it until we get the source value.
8784 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8785 if (Instruction *I = dyn_cast<Instruction>(V))
8786 if (I->getOpcode() == Instruction::FPExt)
8787 return LookThroughFPExtensions(I->getOperand(0), Context);
8789 // If this value is a constant, return the constant in the smallest FP type
8790 // that can accurately represent it. This allows us to turn
8791 // (float)((double)X+2.0) into x+2.0f.
8792 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8793 if (CFP->getType() == Type::PPC_FP128Ty)
8794 return V; // No constant folding of this.
8795 // See if the value can be truncated to float and then reextended.
8796 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8798 if (CFP->getType() == Type::DoubleTy)
8799 return V; // Won't shrink.
8800 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8802 // Don't try to shrink to various long double types.
8808 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8809 if (Instruction *I = commonCastTransforms(CI))
8812 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8813 // smaller than the destination type, we can eliminate the truncate by doing
8814 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8815 // many builtins (sqrt, etc).
8816 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8817 if (OpI && OpI->hasOneUse()) {
8818 switch (OpI->getOpcode()) {
8820 case Instruction::FAdd:
8821 case Instruction::FSub:
8822 case Instruction::FMul:
8823 case Instruction::FDiv:
8824 case Instruction::FRem:
8825 const Type *SrcTy = OpI->getType();
8826 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8827 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8828 if (LHSTrunc->getType() != SrcTy &&
8829 RHSTrunc->getType() != SrcTy) {
8830 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8831 // If the source types were both smaller than the destination type of
8832 // the cast, do this xform.
8833 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8834 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8835 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8837 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8839 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8848 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8849 return commonCastTransforms(CI);
8852 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8853 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8855 return commonCastTransforms(FI);
8857 // fptoui(uitofp(X)) --> X
8858 // fptoui(sitofp(X)) --> X
8859 // This is safe if the intermediate type has enough bits in its mantissa to
8860 // accurately represent all values of X. For example, do not do this with
8861 // i64->float->i64. This is also safe for sitofp case, because any negative
8862 // 'X' value would cause an undefined result for the fptoui.
8863 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8864 OpI->getOperand(0)->getType() == FI.getType() &&
8865 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8866 OpI->getType()->getFPMantissaWidth())
8867 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8869 return commonCastTransforms(FI);
8872 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8873 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8875 return commonCastTransforms(FI);
8877 // fptosi(sitofp(X)) --> X
8878 // fptosi(uitofp(X)) --> X
8879 // This is safe if the intermediate type has enough bits in its mantissa to
8880 // accurately represent all values of X. For example, do not do this with
8881 // i64->float->i64. This is also safe for sitofp case, because any negative
8882 // 'X' value would cause an undefined result for the fptoui.
8883 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8884 OpI->getOperand(0)->getType() == FI.getType() &&
8885 (int)FI.getType()->getScalarSizeInBits() <=
8886 OpI->getType()->getFPMantissaWidth())
8887 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8889 return commonCastTransforms(FI);
8892 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8893 return commonCastTransforms(CI);
8896 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8897 return commonCastTransforms(CI);
8900 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8901 // If the destination integer type is smaller than the intptr_t type for
8902 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8903 // trunc to be exposed to other transforms. Don't do this for extending
8904 // ptrtoint's, because we don't know if the target sign or zero extends its
8907 CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8908 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8909 TD->getIntPtrType(),
8911 return new TruncInst(P, CI.getType());
8914 return commonPointerCastTransforms(CI);
8917 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8918 // If the source integer type is larger than the intptr_t type for
8919 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8920 // allows the trunc to be exposed to other transforms. Don't do this for
8921 // extending inttoptr's, because we don't know if the target sign or zero
8922 // extends to pointers.
8924 CI.getOperand(0)->getType()->getScalarSizeInBits() >
8925 TD->getPointerSizeInBits()) {
8926 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8927 TD->getIntPtrType(),
8929 return new IntToPtrInst(P, CI.getType());
8932 if (Instruction *I = commonCastTransforms(CI))
8938 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8939 // If the operands are integer typed then apply the integer transforms,
8940 // otherwise just apply the common ones.
8941 Value *Src = CI.getOperand(0);
8942 const Type *SrcTy = Src->getType();
8943 const Type *DestTy = CI.getType();
8945 if (isa<PointerType>(SrcTy)) {
8946 if (Instruction *I = commonPointerCastTransforms(CI))
8949 if (Instruction *Result = commonCastTransforms(CI))
8954 // Get rid of casts from one type to the same type. These are useless and can
8955 // be replaced by the operand.
8956 if (DestTy == Src->getType())
8957 return ReplaceInstUsesWith(CI, Src);
8959 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8960 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8961 const Type *DstElTy = DstPTy->getElementType();
8962 const Type *SrcElTy = SrcPTy->getElementType();
8964 // If the address spaces don't match, don't eliminate the bitcast, which is
8965 // required for changing types.
8966 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
8969 // If we are casting a malloc or alloca to a pointer to a type of the same
8970 // size, rewrite the allocation instruction to allocate the "right" type.
8971 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
8972 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
8975 // If the source and destination are pointers, and this cast is equivalent
8976 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
8977 // This can enhance SROA and other transforms that want type-safe pointers.
8978 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
8979 unsigned NumZeros = 0;
8980 while (SrcElTy != DstElTy &&
8981 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
8982 SrcElTy->getNumContainedTypes() /* not "{}" */) {
8983 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
8987 // If we found a path from the src to dest, create the getelementptr now.
8988 if (SrcElTy == DstElTy) {
8989 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
8990 Instruction *GEP = GetElementPtrInst::Create(Src,
8991 Idxs.begin(), Idxs.end(), "",
8992 ((Instruction*) NULL));
8993 cast<GEPOperator>(GEP)->setIsInBounds(true);
8998 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
8999 if (DestVTy->getNumElements() == 1) {
9000 if (!isa<VectorType>(SrcTy)) {
9001 Value *Elem = InsertCastBefore(Instruction::BitCast, Src,
9002 DestVTy->getElementType(), CI);
9003 return InsertElementInst::Create(UndefValue::get(DestTy), Elem,
9004 Constant::getNullValue(Type::Int32Ty));
9006 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
9010 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
9011 if (SrcVTy->getNumElements() == 1) {
9012 if (!isa<VectorType>(DestTy)) {
9014 ExtractElementInst::Create(Src, Constant::getNullValue(Type::Int32Ty));
9015 InsertNewInstBefore(Elem, CI);
9016 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
9021 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9022 if (SVI->hasOneUse()) {
9023 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9024 // a bitconvert to a vector with the same # elts.
9025 if (isa<VectorType>(DestTy) &&
9026 cast<VectorType>(DestTy)->getNumElements() ==
9027 SVI->getType()->getNumElements() &&
9028 SVI->getType()->getNumElements() ==
9029 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9031 // If either of the operands is a cast from CI.getType(), then
9032 // evaluating the shuffle in the casted destination's type will allow
9033 // us to eliminate at least one cast.
9034 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9035 Tmp->getOperand(0)->getType() == DestTy) ||
9036 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9037 Tmp->getOperand(0)->getType() == DestTy)) {
9038 Value *LHS = InsertCastBefore(Instruction::BitCast,
9039 SVI->getOperand(0), DestTy, CI);
9040 Value *RHS = InsertCastBefore(Instruction::BitCast,
9041 SVI->getOperand(1), DestTy, CI);
9042 // Return a new shuffle vector. Use the same element ID's, as we
9043 // know the vector types match #elts.
9044 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9052 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9054 /// %D = select %cond, %C, %A
9056 /// %C = select %cond, %B, 0
9059 /// Assuming that the specified instruction is an operand to the select, return
9060 /// a bitmask indicating which operands of this instruction are foldable if they
9061 /// equal the other incoming value of the select.
9063 static unsigned GetSelectFoldableOperands(Instruction *I) {
9064 switch (I->getOpcode()) {
9065 case Instruction::Add:
9066 case Instruction::Mul:
9067 case Instruction::And:
9068 case Instruction::Or:
9069 case Instruction::Xor:
9070 return 3; // Can fold through either operand.
9071 case Instruction::Sub: // Can only fold on the amount subtracted.
9072 case Instruction::Shl: // Can only fold on the shift amount.
9073 case Instruction::LShr:
9074 case Instruction::AShr:
9077 return 0; // Cannot fold
9081 /// GetSelectFoldableConstant - For the same transformation as the previous
9082 /// function, return the identity constant that goes into the select.
9083 static Constant *GetSelectFoldableConstant(Instruction *I,
9084 LLVMContext *Context) {
9085 switch (I->getOpcode()) {
9086 default: llvm_unreachable("This cannot happen!");
9087 case Instruction::Add:
9088 case Instruction::Sub:
9089 case Instruction::Or:
9090 case Instruction::Xor:
9091 case Instruction::Shl:
9092 case Instruction::LShr:
9093 case Instruction::AShr:
9094 return Constant::getNullValue(I->getType());
9095 case Instruction::And:
9096 return Constant::getAllOnesValue(I->getType());
9097 case Instruction::Mul:
9098 return ConstantInt::get(I->getType(), 1);
9102 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9103 /// have the same opcode and only one use each. Try to simplify this.
9104 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9106 if (TI->getNumOperands() == 1) {
9107 // If this is a non-volatile load or a cast from the same type,
9110 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9113 return 0; // unknown unary op.
9116 // Fold this by inserting a select from the input values.
9117 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9118 FI->getOperand(0), SI.getName()+".v");
9119 InsertNewInstBefore(NewSI, SI);
9120 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9124 // Only handle binary operators here.
9125 if (!isa<BinaryOperator>(TI))
9128 // Figure out if the operations have any operands in common.
9129 Value *MatchOp, *OtherOpT, *OtherOpF;
9131 if (TI->getOperand(0) == FI->getOperand(0)) {
9132 MatchOp = TI->getOperand(0);
9133 OtherOpT = TI->getOperand(1);
9134 OtherOpF = FI->getOperand(1);
9135 MatchIsOpZero = true;
9136 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9137 MatchOp = TI->getOperand(1);
9138 OtherOpT = TI->getOperand(0);
9139 OtherOpF = FI->getOperand(0);
9140 MatchIsOpZero = false;
9141 } else if (!TI->isCommutative()) {
9143 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9144 MatchOp = TI->getOperand(0);
9145 OtherOpT = TI->getOperand(1);
9146 OtherOpF = FI->getOperand(0);
9147 MatchIsOpZero = true;
9148 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9149 MatchOp = TI->getOperand(1);
9150 OtherOpT = TI->getOperand(0);
9151 OtherOpF = FI->getOperand(1);
9152 MatchIsOpZero = true;
9157 // If we reach here, they do have operations in common.
9158 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9159 OtherOpF, SI.getName()+".v");
9160 InsertNewInstBefore(NewSI, SI);
9162 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9164 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9166 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9168 llvm_unreachable("Shouldn't get here");
9172 static bool isSelect01(Constant *C1, Constant *C2) {
9173 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9176 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9179 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9182 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9183 /// facilitate further optimization.
9184 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9186 // See the comment above GetSelectFoldableOperands for a description of the
9187 // transformation we are doing here.
9188 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9189 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9190 !isa<Constant>(FalseVal)) {
9191 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9192 unsigned OpToFold = 0;
9193 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9195 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9200 Constant *C = GetSelectFoldableConstant(TVI, Context);
9201 Value *OOp = TVI->getOperand(2-OpToFold);
9202 // Avoid creating select between 2 constants unless it's selecting
9204 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9205 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9206 InsertNewInstBefore(NewSel, SI);
9207 NewSel->takeName(TVI);
9208 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9209 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9210 llvm_unreachable("Unknown instruction!!");
9217 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9218 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9219 !isa<Constant>(TrueVal)) {
9220 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9221 unsigned OpToFold = 0;
9222 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9224 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9229 Constant *C = GetSelectFoldableConstant(FVI, Context);
9230 Value *OOp = FVI->getOperand(2-OpToFold);
9231 // Avoid creating select between 2 constants unless it's selecting
9233 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9234 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9235 InsertNewInstBefore(NewSel, SI);
9236 NewSel->takeName(FVI);
9237 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9238 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9239 llvm_unreachable("Unknown instruction!!");
9249 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9250 /// ICmpInst as its first operand.
9252 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9254 bool Changed = false;
9255 ICmpInst::Predicate Pred = ICI->getPredicate();
9256 Value *CmpLHS = ICI->getOperand(0);
9257 Value *CmpRHS = ICI->getOperand(1);
9258 Value *TrueVal = SI.getTrueValue();
9259 Value *FalseVal = SI.getFalseValue();
9261 // Check cases where the comparison is with a constant that
9262 // can be adjusted to fit the min/max idiom. We may edit ICI in
9263 // place here, so make sure the select is the only user.
9264 if (ICI->hasOneUse())
9265 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9268 case ICmpInst::ICMP_ULT:
9269 case ICmpInst::ICMP_SLT: {
9270 // X < MIN ? T : F --> F
9271 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9272 return ReplaceInstUsesWith(SI, FalseVal);
9273 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9274 Constant *AdjustedRHS = SubOne(CI, Context);
9275 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9276 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9277 Pred = ICmpInst::getSwappedPredicate(Pred);
9278 CmpRHS = AdjustedRHS;
9279 std::swap(FalseVal, TrueVal);
9280 ICI->setPredicate(Pred);
9281 ICI->setOperand(1, CmpRHS);
9282 SI.setOperand(1, TrueVal);
9283 SI.setOperand(2, FalseVal);
9288 case ICmpInst::ICMP_UGT:
9289 case ICmpInst::ICMP_SGT: {
9290 // X > MAX ? T : F --> F
9291 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9292 return ReplaceInstUsesWith(SI, FalseVal);
9293 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9294 Constant *AdjustedRHS = AddOne(CI, Context);
9295 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9296 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9297 Pred = ICmpInst::getSwappedPredicate(Pred);
9298 CmpRHS = AdjustedRHS;
9299 std::swap(FalseVal, TrueVal);
9300 ICI->setPredicate(Pred);
9301 ICI->setOperand(1, CmpRHS);
9302 SI.setOperand(1, TrueVal);
9303 SI.setOperand(2, FalseVal);
9310 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9311 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9312 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9313 if (match(TrueVal, m_ConstantInt<-1>(), *Context) &&
9314 match(FalseVal, m_ConstantInt<0>(), *Context))
9315 Pred = ICI->getPredicate();
9316 else if (match(TrueVal, m_ConstantInt<0>(), *Context) &&
9317 match(FalseVal, m_ConstantInt<-1>(), *Context))
9318 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9320 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9321 // If we are just checking for a icmp eq of a single bit and zext'ing it
9322 // to an integer, then shift the bit to the appropriate place and then
9323 // cast to integer to avoid the comparison.
9324 const APInt &Op1CV = CI->getValue();
9326 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9327 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9328 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9329 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9330 Value *In = ICI->getOperand(0);
9331 Value *Sh = ConstantInt::get(In->getType(),
9332 In->getType()->getScalarSizeInBits()-1);
9333 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9334 In->getName()+".lobit"),
9336 if (In->getType() != SI.getType())
9337 In = CastInst::CreateIntegerCast(In, SI.getType(),
9338 true/*SExt*/, "tmp", ICI);
9340 if (Pred == ICmpInst::ICMP_SGT)
9341 In = InsertNewInstBefore(BinaryOperator::CreateNot(*Context, In,
9342 In->getName()+".not"), *ICI);
9344 return ReplaceInstUsesWith(SI, In);
9349 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9350 // Transform (X == Y) ? X : Y -> Y
9351 if (Pred == ICmpInst::ICMP_EQ)
9352 return ReplaceInstUsesWith(SI, FalseVal);
9353 // Transform (X != Y) ? X : Y -> X
9354 if (Pred == ICmpInst::ICMP_NE)
9355 return ReplaceInstUsesWith(SI, TrueVal);
9356 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9358 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9359 // Transform (X == Y) ? Y : X -> X
9360 if (Pred == ICmpInst::ICMP_EQ)
9361 return ReplaceInstUsesWith(SI, FalseVal);
9362 // Transform (X != Y) ? Y : X -> Y
9363 if (Pred == ICmpInst::ICMP_NE)
9364 return ReplaceInstUsesWith(SI, TrueVal);
9365 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9368 /// NOTE: if we wanted to, this is where to detect integer ABS
9370 return Changed ? &SI : 0;
9373 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9374 Value *CondVal = SI.getCondition();
9375 Value *TrueVal = SI.getTrueValue();
9376 Value *FalseVal = SI.getFalseValue();
9378 // select true, X, Y -> X
9379 // select false, X, Y -> Y
9380 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9381 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9383 // select C, X, X -> X
9384 if (TrueVal == FalseVal)
9385 return ReplaceInstUsesWith(SI, TrueVal);
9387 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9388 return ReplaceInstUsesWith(SI, FalseVal);
9389 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9390 return ReplaceInstUsesWith(SI, TrueVal);
9391 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9392 if (isa<Constant>(TrueVal))
9393 return ReplaceInstUsesWith(SI, TrueVal);
9395 return ReplaceInstUsesWith(SI, FalseVal);
9398 if (SI.getType() == Type::Int1Ty) {
9399 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9400 if (C->getZExtValue()) {
9401 // Change: A = select B, true, C --> A = or B, C
9402 return BinaryOperator::CreateOr(CondVal, FalseVal);
9404 // Change: A = select B, false, C --> A = and !B, C
9406 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9407 "not."+CondVal->getName()), SI);
9408 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9410 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9411 if (C->getZExtValue() == false) {
9412 // Change: A = select B, C, false --> A = and B, C
9413 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9415 // Change: A = select B, C, true --> A = or !B, C
9417 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9418 "not."+CondVal->getName()), SI);
9419 return BinaryOperator::CreateOr(NotCond, TrueVal);
9423 // select a, b, a -> a&b
9424 // select a, a, b -> a|b
9425 if (CondVal == TrueVal)
9426 return BinaryOperator::CreateOr(CondVal, FalseVal);
9427 else if (CondVal == FalseVal)
9428 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9431 // Selecting between two integer constants?
9432 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9433 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9434 // select C, 1, 0 -> zext C to int
9435 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9436 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9437 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9438 // select C, 0, 1 -> zext !C to int
9440 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9441 "not."+CondVal->getName()), SI);
9442 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9445 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9446 // If one of the constants is zero (we know they can't both be) and we
9447 // have an icmp instruction with zero, and we have an 'and' with the
9448 // non-constant value, eliminate this whole mess. This corresponds to
9449 // cases like this: ((X & 27) ? 27 : 0)
9450 if (TrueValC->isZero() || FalseValC->isZero())
9451 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9452 cast<Constant>(IC->getOperand(1))->isNullValue())
9453 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9454 if (ICA->getOpcode() == Instruction::And &&
9455 isa<ConstantInt>(ICA->getOperand(1)) &&
9456 (ICA->getOperand(1) == TrueValC ||
9457 ICA->getOperand(1) == FalseValC) &&
9458 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9459 // Okay, now we know that everything is set up, we just don't
9460 // know whether we have a icmp_ne or icmp_eq and whether the
9461 // true or false val is the zero.
9462 bool ShouldNotVal = !TrueValC->isZero();
9463 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9466 V = InsertNewInstBefore(BinaryOperator::Create(
9467 Instruction::Xor, V, ICA->getOperand(1)), SI);
9468 return ReplaceInstUsesWith(SI, V);
9473 // See if we are selecting two values based on a comparison of the two values.
9474 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9475 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9476 // Transform (X == Y) ? X : Y -> Y
9477 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9478 // This is not safe in general for floating point:
9479 // consider X== -0, Y== +0.
9480 // It becomes safe if either operand is a nonzero constant.
9481 ConstantFP *CFPt, *CFPf;
9482 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9483 !CFPt->getValueAPF().isZero()) ||
9484 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9485 !CFPf->getValueAPF().isZero()))
9486 return ReplaceInstUsesWith(SI, FalseVal);
9488 // Transform (X != Y) ? X : Y -> X
9489 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9490 return ReplaceInstUsesWith(SI, TrueVal);
9491 // NOTE: if we wanted to, this is where to detect MIN/MAX
9493 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9494 // Transform (X == Y) ? Y : X -> X
9495 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9496 // This is not safe in general for floating point:
9497 // consider X== -0, Y== +0.
9498 // It becomes safe if either operand is a nonzero constant.
9499 ConstantFP *CFPt, *CFPf;
9500 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9501 !CFPt->getValueAPF().isZero()) ||
9502 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9503 !CFPf->getValueAPF().isZero()))
9504 return ReplaceInstUsesWith(SI, FalseVal);
9506 // Transform (X != Y) ? Y : X -> Y
9507 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9508 return ReplaceInstUsesWith(SI, TrueVal);
9509 // NOTE: if we wanted to, this is where to detect MIN/MAX
9511 // NOTE: if we wanted to, this is where to detect ABS
9514 // See if we are selecting two values based on a comparison of the two values.
9515 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9516 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9519 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9520 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9521 if (TI->hasOneUse() && FI->hasOneUse()) {
9522 Instruction *AddOp = 0, *SubOp = 0;
9524 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9525 if (TI->getOpcode() == FI->getOpcode())
9526 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9529 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9530 // even legal for FP.
9531 if ((TI->getOpcode() == Instruction::Sub &&
9532 FI->getOpcode() == Instruction::Add) ||
9533 (TI->getOpcode() == Instruction::FSub &&
9534 FI->getOpcode() == Instruction::FAdd)) {
9535 AddOp = FI; SubOp = TI;
9536 } else if ((FI->getOpcode() == Instruction::Sub &&
9537 TI->getOpcode() == Instruction::Add) ||
9538 (FI->getOpcode() == Instruction::FSub &&
9539 TI->getOpcode() == Instruction::FAdd)) {
9540 AddOp = TI; SubOp = FI;
9544 Value *OtherAddOp = 0;
9545 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9546 OtherAddOp = AddOp->getOperand(1);
9547 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9548 OtherAddOp = AddOp->getOperand(0);
9552 // So at this point we know we have (Y -> OtherAddOp):
9553 // select C, (add X, Y), (sub X, Z)
9554 Value *NegVal; // Compute -Z
9555 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9556 NegVal = ConstantExpr::getNeg(C);
9558 NegVal = InsertNewInstBefore(
9559 BinaryOperator::CreateNeg(*Context, SubOp->getOperand(1),
9563 Value *NewTrueOp = OtherAddOp;
9564 Value *NewFalseOp = NegVal;
9566 std::swap(NewTrueOp, NewFalseOp);
9567 Instruction *NewSel =
9568 SelectInst::Create(CondVal, NewTrueOp,
9569 NewFalseOp, SI.getName() + ".p");
9571 NewSel = InsertNewInstBefore(NewSel, SI);
9572 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9577 // See if we can fold the select into one of our operands.
9578 if (SI.getType()->isInteger()) {
9579 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9584 if (BinaryOperator::isNot(CondVal)) {
9585 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9586 SI.setOperand(1, FalseVal);
9587 SI.setOperand(2, TrueVal);
9594 /// EnforceKnownAlignment - If the specified pointer points to an object that
9595 /// we control, modify the object's alignment to PrefAlign. This isn't
9596 /// often possible though. If alignment is important, a more reliable approach
9597 /// is to simply align all global variables and allocation instructions to
9598 /// their preferred alignment from the beginning.
9600 static unsigned EnforceKnownAlignment(Value *V,
9601 unsigned Align, unsigned PrefAlign) {
9603 User *U = dyn_cast<User>(V);
9604 if (!U) return Align;
9606 switch (Operator::getOpcode(U)) {
9608 case Instruction::BitCast:
9609 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9610 case Instruction::GetElementPtr: {
9611 // If all indexes are zero, it is just the alignment of the base pointer.
9612 bool AllZeroOperands = true;
9613 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9614 if (!isa<Constant>(*i) ||
9615 !cast<Constant>(*i)->isNullValue()) {
9616 AllZeroOperands = false;
9620 if (AllZeroOperands) {
9621 // Treat this like a bitcast.
9622 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9628 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9629 // If there is a large requested alignment and we can, bump up the alignment
9631 if (!GV->isDeclaration()) {
9632 if (GV->getAlignment() >= PrefAlign)
9633 Align = GV->getAlignment();
9635 GV->setAlignment(PrefAlign);
9639 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9640 // If there is a requested alignment and if this is an alloca, round up. We
9641 // don't do this for malloc, because some systems can't respect the request.
9642 if (isa<AllocaInst>(AI)) {
9643 if (AI->getAlignment() >= PrefAlign)
9644 Align = AI->getAlignment();
9646 AI->setAlignment(PrefAlign);
9655 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9656 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9657 /// and it is more than the alignment of the ultimate object, see if we can
9658 /// increase the alignment of the ultimate object, making this check succeed.
9659 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9660 unsigned PrefAlign) {
9661 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9662 sizeof(PrefAlign) * CHAR_BIT;
9663 APInt Mask = APInt::getAllOnesValue(BitWidth);
9664 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9665 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9666 unsigned TrailZ = KnownZero.countTrailingOnes();
9667 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9669 if (PrefAlign > Align)
9670 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9672 // We don't need to make any adjustment.
9676 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9677 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9678 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9679 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9680 unsigned CopyAlign = MI->getAlignment();
9682 if (CopyAlign < MinAlign) {
9683 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9688 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9690 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9691 if (MemOpLength == 0) return 0;
9693 // Source and destination pointer types are always "i8*" for intrinsic. See
9694 // if the size is something we can handle with a single primitive load/store.
9695 // A single load+store correctly handles overlapping memory in the memmove
9697 unsigned Size = MemOpLength->getZExtValue();
9698 if (Size == 0) return MI; // Delete this mem transfer.
9700 if (Size > 8 || (Size&(Size-1)))
9701 return 0; // If not 1/2/4/8 bytes, exit.
9703 // Use an integer load+store unless we can find something better.
9705 PointerType::getUnqual(IntegerType::get(Size<<3));
9707 // Memcpy forces the use of i8* for the source and destination. That means
9708 // that if you're using memcpy to move one double around, you'll get a cast
9709 // from double* to i8*. We'd much rather use a double load+store rather than
9710 // an i64 load+store, here because this improves the odds that the source or
9711 // dest address will be promotable. See if we can find a better type than the
9712 // integer datatype.
9713 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9714 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9715 if (TD && SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9716 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9717 // down through these levels if so.
9718 while (!SrcETy->isSingleValueType()) {
9719 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9720 if (STy->getNumElements() == 1)
9721 SrcETy = STy->getElementType(0);
9724 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9725 if (ATy->getNumElements() == 1)
9726 SrcETy = ATy->getElementType();
9733 if (SrcETy->isSingleValueType())
9734 NewPtrTy = PointerType::getUnqual(SrcETy);
9739 // If the memcpy/memmove provides better alignment info than we can
9741 SrcAlign = std::max(SrcAlign, CopyAlign);
9742 DstAlign = std::max(DstAlign, CopyAlign);
9744 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9745 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9746 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9747 InsertNewInstBefore(L, *MI);
9748 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9750 // Set the size of the copy to 0, it will be deleted on the next iteration.
9751 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
9755 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9756 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9757 if (MI->getAlignment() < Alignment) {
9758 MI->setAlignment(ConstantInt::get(MI->getAlignmentType(),
9763 // Extract the length and alignment and fill if they are constant.
9764 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9765 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9766 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9768 uint64_t Len = LenC->getZExtValue();
9769 Alignment = MI->getAlignment();
9771 // If the length is zero, this is a no-op
9772 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9774 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9775 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9776 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
9778 Value *Dest = MI->getDest();
9779 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
9781 // Alignment 0 is identity for alignment 1 for memset, but not store.
9782 if (Alignment == 0) Alignment = 1;
9784 // Extract the fill value and store.
9785 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9786 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill),
9787 Dest, false, Alignment), *MI);
9789 // Set the size of the copy to 0, it will be deleted on the next iteration.
9790 MI->setLength(Constant::getNullValue(LenC->getType()));
9798 /// visitCallInst - CallInst simplification. This mostly only handles folding
9799 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9800 /// the heavy lifting.
9802 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9803 // If the caller function is nounwind, mark the call as nounwind, even if the
9805 if (CI.getParent()->getParent()->doesNotThrow() &&
9806 !CI.doesNotThrow()) {
9807 CI.setDoesNotThrow();
9813 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9814 if (!II) return visitCallSite(&CI);
9816 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9818 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9819 bool Changed = false;
9821 // memmove/cpy/set of zero bytes is a noop.
9822 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9823 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9825 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9826 if (CI->getZExtValue() == 1) {
9827 // Replace the instruction with just byte operations. We would
9828 // transform other cases to loads/stores, but we don't know if
9829 // alignment is sufficient.
9833 // If we have a memmove and the source operation is a constant global,
9834 // then the source and dest pointers can't alias, so we can change this
9835 // into a call to memcpy.
9836 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9837 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9838 if (GVSrc->isConstant()) {
9839 Module *M = CI.getParent()->getParent()->getParent();
9840 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9842 Tys[0] = CI.getOperand(3)->getType();
9844 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9848 // memmove(x,x,size) -> noop.
9849 if (MMI->getSource() == MMI->getDest())
9850 return EraseInstFromFunction(CI);
9853 // If we can determine a pointer alignment that is bigger than currently
9854 // set, update the alignment.
9855 if (isa<MemTransferInst>(MI)) {
9856 if (Instruction *I = SimplifyMemTransfer(MI))
9858 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9859 if (Instruction *I = SimplifyMemSet(MSI))
9863 if (Changed) return II;
9866 switch (II->getIntrinsicID()) {
9868 case Intrinsic::bswap:
9869 // bswap(bswap(x)) -> x
9870 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9871 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9872 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9874 case Intrinsic::ppc_altivec_lvx:
9875 case Intrinsic::ppc_altivec_lvxl:
9876 case Intrinsic::x86_sse_loadu_ps:
9877 case Intrinsic::x86_sse2_loadu_pd:
9878 case Intrinsic::x86_sse2_loadu_dq:
9879 // Turn PPC lvx -> load if the pointer is known aligned.
9880 // Turn X86 loadups -> load if the pointer is known aligned.
9881 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9882 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9883 PointerType::getUnqual(II->getType()),
9885 return new LoadInst(Ptr);
9888 case Intrinsic::ppc_altivec_stvx:
9889 case Intrinsic::ppc_altivec_stvxl:
9890 // Turn stvx -> store if the pointer is known aligned.
9891 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9892 const Type *OpPtrTy =
9893 PointerType::getUnqual(II->getOperand(1)->getType());
9894 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9895 return new StoreInst(II->getOperand(1), Ptr);
9898 case Intrinsic::x86_sse_storeu_ps:
9899 case Intrinsic::x86_sse2_storeu_pd:
9900 case Intrinsic::x86_sse2_storeu_dq:
9901 // Turn X86 storeu -> store if the pointer is known aligned.
9902 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9903 const Type *OpPtrTy =
9904 PointerType::getUnqual(II->getOperand(2)->getType());
9905 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9906 return new StoreInst(II->getOperand(2), Ptr);
9910 case Intrinsic::x86_sse_cvttss2si: {
9911 // These intrinsics only demands the 0th element of its input vector. If
9912 // we can simplify the input based on that, do so now.
9914 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9915 APInt DemandedElts(VWidth, 1);
9916 APInt UndefElts(VWidth, 0);
9917 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9919 II->setOperand(1, V);
9925 case Intrinsic::ppc_altivec_vperm:
9926 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9927 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9928 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9930 // Check that all of the elements are integer constants or undefs.
9931 bool AllEltsOk = true;
9932 for (unsigned i = 0; i != 16; ++i) {
9933 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9934 !isa<UndefValue>(Mask->getOperand(i))) {
9941 // Cast the input vectors to byte vectors.
9942 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9943 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9944 Value *Result = UndefValue::get(Op0->getType());
9946 // Only extract each element once.
9947 Value *ExtractedElts[32];
9948 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9950 for (unsigned i = 0; i != 16; ++i) {
9951 if (isa<UndefValue>(Mask->getOperand(i)))
9953 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9954 Idx &= 31; // Match the hardware behavior.
9956 if (ExtractedElts[Idx] == 0) {
9958 ExtractElementInst::Create(Idx < 16 ? Op0 : Op1,
9959 ConstantInt::get(Type::Int32Ty, Idx&15, false), "tmp");
9960 InsertNewInstBefore(Elt, CI);
9961 ExtractedElts[Idx] = Elt;
9964 // Insert this value into the result vector.
9965 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9966 ConstantInt::get(Type::Int32Ty, i, false),
9968 InsertNewInstBefore(cast<Instruction>(Result), CI);
9970 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9975 case Intrinsic::stackrestore: {
9976 // If the save is right next to the restore, remove the restore. This can
9977 // happen when variable allocas are DCE'd.
9978 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9979 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9980 BasicBlock::iterator BI = SS;
9982 return EraseInstFromFunction(CI);
9986 // Scan down this block to see if there is another stack restore in the
9987 // same block without an intervening call/alloca.
9988 BasicBlock::iterator BI = II;
9989 TerminatorInst *TI = II->getParent()->getTerminator();
9990 bool CannotRemove = false;
9991 for (++BI; &*BI != TI; ++BI) {
9992 if (isa<AllocaInst>(BI)) {
9993 CannotRemove = true;
9996 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
9997 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
9998 // If there is a stackrestore below this one, remove this one.
9999 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10000 return EraseInstFromFunction(CI);
10001 // Otherwise, ignore the intrinsic.
10003 // If we found a non-intrinsic call, we can't remove the stack
10005 CannotRemove = true;
10011 // If the stack restore is in a return/unwind block and if there are no
10012 // allocas or calls between the restore and the return, nuke the restore.
10013 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10014 return EraseInstFromFunction(CI);
10019 return visitCallSite(II);
10022 // InvokeInst simplification
10024 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10025 return visitCallSite(&II);
10028 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10029 /// passed through the varargs area, we can eliminate the use of the cast.
10030 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10031 const CastInst * const CI,
10032 const TargetData * const TD,
10034 if (!CI->isLosslessCast())
10037 // The size of ByVal arguments is derived from the type, so we
10038 // can't change to a type with a different size. If the size were
10039 // passed explicitly we could avoid this check.
10040 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10043 const Type* SrcTy =
10044 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10045 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10046 if (!SrcTy->isSized() || !DstTy->isSized())
10048 if (!TD || TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10053 // visitCallSite - Improvements for call and invoke instructions.
10055 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10056 bool Changed = false;
10058 // If the callee is a constexpr cast of a function, attempt to move the cast
10059 // to the arguments of the call/invoke.
10060 if (transformConstExprCastCall(CS)) return 0;
10062 Value *Callee = CS.getCalledValue();
10064 if (Function *CalleeF = dyn_cast<Function>(Callee))
10065 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10066 Instruction *OldCall = CS.getInstruction();
10067 // If the call and callee calling conventions don't match, this call must
10068 // be unreachable, as the call is undefined.
10069 new StoreInst(ConstantInt::getTrue(*Context),
10070 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
10072 if (!OldCall->use_empty())
10073 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
10074 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10075 return EraseInstFromFunction(*OldCall);
10079 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10080 // This instruction is not reachable, just remove it. We insert a store to
10081 // undef so that we know that this code is not reachable, despite the fact
10082 // that we can't modify the CFG here.
10083 new StoreInst(ConstantInt::getTrue(*Context),
10084 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
10085 CS.getInstruction());
10087 if (!CS.getInstruction()->use_empty())
10088 CS.getInstruction()->
10089 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
10091 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10092 // Don't break the CFG, insert a dummy cond branch.
10093 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10094 ConstantInt::getTrue(*Context), II);
10096 return EraseInstFromFunction(*CS.getInstruction());
10099 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10100 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10101 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10102 return transformCallThroughTrampoline(CS);
10104 const PointerType *PTy = cast<PointerType>(Callee->getType());
10105 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10106 if (FTy->isVarArg()) {
10107 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10108 // See if we can optimize any arguments passed through the varargs area of
10110 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10111 E = CS.arg_end(); I != E; ++I, ++ix) {
10112 CastInst *CI = dyn_cast<CastInst>(*I);
10113 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10114 *I = CI->getOperand(0);
10120 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10121 // Inline asm calls cannot throw - mark them 'nounwind'.
10122 CS.setDoesNotThrow();
10126 return Changed ? CS.getInstruction() : 0;
10129 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10130 // attempt to move the cast to the arguments of the call/invoke.
10132 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10133 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10134 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10135 if (CE->getOpcode() != Instruction::BitCast ||
10136 !isa<Function>(CE->getOperand(0)))
10138 Function *Callee = cast<Function>(CE->getOperand(0));
10139 Instruction *Caller = CS.getInstruction();
10140 const AttrListPtr &CallerPAL = CS.getAttributes();
10142 // Okay, this is a cast from a function to a different type. Unless doing so
10143 // would cause a type conversion of one of our arguments, change this call to
10144 // be a direct call with arguments casted to the appropriate types.
10146 const FunctionType *FT = Callee->getFunctionType();
10147 const Type *OldRetTy = Caller->getType();
10148 const Type *NewRetTy = FT->getReturnType();
10150 if (isa<StructType>(NewRetTy))
10151 return false; // TODO: Handle multiple return values.
10153 // Check to see if we are changing the return type...
10154 if (OldRetTy != NewRetTy) {
10155 if (Callee->isDeclaration() &&
10156 // Conversion is ok if changing from one pointer type to another or from
10157 // a pointer to an integer of the same size.
10158 !((isa<PointerType>(OldRetTy) || !TD ||
10159 OldRetTy == TD->getIntPtrType()) &&
10160 (isa<PointerType>(NewRetTy) || !TD ||
10161 NewRetTy == TD->getIntPtrType())))
10162 return false; // Cannot transform this return value.
10164 if (!Caller->use_empty() &&
10165 // void -> non-void is handled specially
10166 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
10167 return false; // Cannot transform this return value.
10169 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10170 Attributes RAttrs = CallerPAL.getRetAttributes();
10171 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10172 return false; // Attribute not compatible with transformed value.
10175 // If the callsite is an invoke instruction, and the return value is used by
10176 // a PHI node in a successor, we cannot change the return type of the call
10177 // because there is no place to put the cast instruction (without breaking
10178 // the critical edge). Bail out in this case.
10179 if (!Caller->use_empty())
10180 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10181 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10183 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10184 if (PN->getParent() == II->getNormalDest() ||
10185 PN->getParent() == II->getUnwindDest())
10189 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10190 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10192 CallSite::arg_iterator AI = CS.arg_begin();
10193 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10194 const Type *ParamTy = FT->getParamType(i);
10195 const Type *ActTy = (*AI)->getType();
10197 if (!CastInst::isCastable(ActTy, ParamTy))
10198 return false; // Cannot transform this parameter value.
10200 if (CallerPAL.getParamAttributes(i + 1)
10201 & Attribute::typeIncompatible(ParamTy))
10202 return false; // Attribute not compatible with transformed value.
10204 // Converting from one pointer type to another or between a pointer and an
10205 // integer of the same size is safe even if we do not have a body.
10206 bool isConvertible = ActTy == ParamTy ||
10207 (TD && ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
10208 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType())));
10209 if (Callee->isDeclaration() && !isConvertible) return false;
10212 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10213 Callee->isDeclaration())
10214 return false; // Do not delete arguments unless we have a function body.
10216 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10217 !CallerPAL.isEmpty())
10218 // In this case we have more arguments than the new function type, but we
10219 // won't be dropping them. Check that these extra arguments have attributes
10220 // that are compatible with being a vararg call argument.
10221 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10222 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10224 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10225 if (PAttrs & Attribute::VarArgsIncompatible)
10229 // Okay, we decided that this is a safe thing to do: go ahead and start
10230 // inserting cast instructions as necessary...
10231 std::vector<Value*> Args;
10232 Args.reserve(NumActualArgs);
10233 SmallVector<AttributeWithIndex, 8> attrVec;
10234 attrVec.reserve(NumCommonArgs);
10236 // Get any return attributes.
10237 Attributes RAttrs = CallerPAL.getRetAttributes();
10239 // If the return value is not being used, the type may not be compatible
10240 // with the existing attributes. Wipe out any problematic attributes.
10241 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10243 // Add the new return attributes.
10245 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10247 AI = CS.arg_begin();
10248 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10249 const Type *ParamTy = FT->getParamType(i);
10250 if ((*AI)->getType() == ParamTy) {
10251 Args.push_back(*AI);
10253 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10254 false, ParamTy, false);
10255 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
10256 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
10259 // Add any parameter attributes.
10260 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10261 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10264 // If the function takes more arguments than the call was taking, add them
10266 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10267 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
10269 // If we are removing arguments to the function, emit an obnoxious warning...
10270 if (FT->getNumParams() < NumActualArgs) {
10271 if (!FT->isVarArg()) {
10272 errs() << "WARNING: While resolving call to function '"
10273 << Callee->getName() << "' arguments were dropped!\n";
10275 // Add all of the arguments in their promoted form to the arg list...
10276 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10277 const Type *PTy = getPromotedType((*AI)->getType());
10278 if (PTy != (*AI)->getType()) {
10279 // Must promote to pass through va_arg area!
10280 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
10282 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
10283 InsertNewInstBefore(Cast, *Caller);
10284 Args.push_back(Cast);
10286 Args.push_back(*AI);
10289 // Add any parameter attributes.
10290 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10291 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10296 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10297 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10299 if (NewRetTy == Type::VoidTy)
10300 Caller->setName(""); // Void type should not have a name.
10302 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),
10306 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10307 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10308 Args.begin(), Args.end(),
10309 Caller->getName(), Caller);
10310 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10311 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10313 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10314 Caller->getName(), Caller);
10315 CallInst *CI = cast<CallInst>(Caller);
10316 if (CI->isTailCall())
10317 cast<CallInst>(NC)->setTailCall();
10318 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10319 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10322 // Insert a cast of the return type as necessary.
10324 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10325 if (NV->getType() != Type::VoidTy) {
10326 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10328 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10330 // If this is an invoke instruction, we should insert it after the first
10331 // non-phi, instruction in the normal successor block.
10332 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10333 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10334 InsertNewInstBefore(NC, *I);
10336 // Otherwise, it's a call, just insert cast right after the call instr
10337 InsertNewInstBefore(NC, *Caller);
10339 AddUsersToWorkList(*Caller);
10341 NV = UndefValue::get(Caller->getType());
10345 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10346 Caller->replaceAllUsesWith(NV);
10347 Caller->eraseFromParent();
10348 RemoveFromWorkList(Caller);
10352 // transformCallThroughTrampoline - Turn a call to a function created by the
10353 // init_trampoline intrinsic into a direct call to the underlying function.
10355 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10356 Value *Callee = CS.getCalledValue();
10357 const PointerType *PTy = cast<PointerType>(Callee->getType());
10358 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10359 const AttrListPtr &Attrs = CS.getAttributes();
10361 // If the call already has the 'nest' attribute somewhere then give up -
10362 // otherwise 'nest' would occur twice after splicing in the chain.
10363 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10366 IntrinsicInst *Tramp =
10367 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10369 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10370 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10371 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10373 const AttrListPtr &NestAttrs = NestF->getAttributes();
10374 if (!NestAttrs.isEmpty()) {
10375 unsigned NestIdx = 1;
10376 const Type *NestTy = 0;
10377 Attributes NestAttr = Attribute::None;
10379 // Look for a parameter marked with the 'nest' attribute.
10380 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10381 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10382 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10383 // Record the parameter type and any other attributes.
10385 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10390 Instruction *Caller = CS.getInstruction();
10391 std::vector<Value*> NewArgs;
10392 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10394 SmallVector<AttributeWithIndex, 8> NewAttrs;
10395 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10397 // Insert the nest argument into the call argument list, which may
10398 // mean appending it. Likewise for attributes.
10400 // Add any result attributes.
10401 if (Attributes Attr = Attrs.getRetAttributes())
10402 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10406 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10408 if (Idx == NestIdx) {
10409 // Add the chain argument and attributes.
10410 Value *NestVal = Tramp->getOperand(3);
10411 if (NestVal->getType() != NestTy)
10412 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10413 NewArgs.push_back(NestVal);
10414 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10420 // Add the original argument and attributes.
10421 NewArgs.push_back(*I);
10422 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10424 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10430 // Add any function attributes.
10431 if (Attributes Attr = Attrs.getFnAttributes())
10432 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10434 // The trampoline may have been bitcast to a bogus type (FTy).
10435 // Handle this by synthesizing a new function type, equal to FTy
10436 // with the chain parameter inserted.
10438 std::vector<const Type*> NewTypes;
10439 NewTypes.reserve(FTy->getNumParams()+1);
10441 // Insert the chain's type into the list of parameter types, which may
10442 // mean appending it.
10445 FunctionType::param_iterator I = FTy->param_begin(),
10446 E = FTy->param_end();
10449 if (Idx == NestIdx)
10450 // Add the chain's type.
10451 NewTypes.push_back(NestTy);
10456 // Add the original type.
10457 NewTypes.push_back(*I);
10463 // Replace the trampoline call with a direct call. Let the generic
10464 // code sort out any function type mismatches.
10465 FunctionType *NewFTy = FunctionType::get(FTy->getReturnType(), NewTypes,
10467 Constant *NewCallee =
10468 NestF->getType() == PointerType::getUnqual(NewFTy) ?
10469 NestF : ConstantExpr::getBitCast(NestF,
10470 PointerType::getUnqual(NewFTy));
10471 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),
10474 Instruction *NewCaller;
10475 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10476 NewCaller = InvokeInst::Create(NewCallee,
10477 II->getNormalDest(), II->getUnwindDest(),
10478 NewArgs.begin(), NewArgs.end(),
10479 Caller->getName(), Caller);
10480 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10481 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10483 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10484 Caller->getName(), Caller);
10485 if (cast<CallInst>(Caller)->isTailCall())
10486 cast<CallInst>(NewCaller)->setTailCall();
10487 cast<CallInst>(NewCaller)->
10488 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10489 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10491 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10492 Caller->replaceAllUsesWith(NewCaller);
10493 Caller->eraseFromParent();
10494 RemoveFromWorkList(Caller);
10499 // Replace the trampoline call with a direct call. Since there is no 'nest'
10500 // parameter, there is no need to adjust the argument list. Let the generic
10501 // code sort out any function type mismatches.
10502 Constant *NewCallee =
10503 NestF->getType() == PTy ? NestF :
10504 ConstantExpr::getBitCast(NestF, PTy);
10505 CS.setCalledFunction(NewCallee);
10506 return CS.getInstruction();
10509 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10510 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10511 /// and a single binop.
10512 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10513 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10514 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10515 unsigned Opc = FirstInst->getOpcode();
10516 Value *LHSVal = FirstInst->getOperand(0);
10517 Value *RHSVal = FirstInst->getOperand(1);
10519 const Type *LHSType = LHSVal->getType();
10520 const Type *RHSType = RHSVal->getType();
10522 // Scan to see if all operands are the same opcode, all have one use, and all
10523 // kill their operands (i.e. the operands have one use).
10524 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10525 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10526 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10527 // Verify type of the LHS matches so we don't fold cmp's of different
10528 // types or GEP's with different index types.
10529 I->getOperand(0)->getType() != LHSType ||
10530 I->getOperand(1)->getType() != RHSType)
10533 // If they are CmpInst instructions, check their predicates
10534 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10535 if (cast<CmpInst>(I)->getPredicate() !=
10536 cast<CmpInst>(FirstInst)->getPredicate())
10539 // Keep track of which operand needs a phi node.
10540 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10541 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10544 // Otherwise, this is safe to transform!
10546 Value *InLHS = FirstInst->getOperand(0);
10547 Value *InRHS = FirstInst->getOperand(1);
10548 PHINode *NewLHS = 0, *NewRHS = 0;
10550 NewLHS = PHINode::Create(LHSType,
10551 FirstInst->getOperand(0)->getName() + ".pn");
10552 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10553 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10554 InsertNewInstBefore(NewLHS, PN);
10559 NewRHS = PHINode::Create(RHSType,
10560 FirstInst->getOperand(1)->getName() + ".pn");
10561 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10562 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10563 InsertNewInstBefore(NewRHS, PN);
10567 // Add all operands to the new PHIs.
10568 if (NewLHS || NewRHS) {
10569 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10570 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10572 Value *NewInLHS = InInst->getOperand(0);
10573 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10576 Value *NewInRHS = InInst->getOperand(1);
10577 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10582 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10583 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10584 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10585 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10589 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10590 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10592 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10593 FirstInst->op_end());
10594 // This is true if all GEP bases are allocas and if all indices into them are
10596 bool AllBasePointersAreAllocas = true;
10598 // Scan to see if all operands are the same opcode, all have one use, and all
10599 // kill their operands (i.e. the operands have one use).
10600 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10601 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10602 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10603 GEP->getNumOperands() != FirstInst->getNumOperands())
10606 // Keep track of whether or not all GEPs are of alloca pointers.
10607 if (AllBasePointersAreAllocas &&
10608 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10609 !GEP->hasAllConstantIndices()))
10610 AllBasePointersAreAllocas = false;
10612 // Compare the operand lists.
10613 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10614 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10617 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10618 // if one of the PHIs has a constant for the index. The index may be
10619 // substantially cheaper to compute for the constants, so making it a
10620 // variable index could pessimize the path. This also handles the case
10621 // for struct indices, which must always be constant.
10622 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10623 isa<ConstantInt>(GEP->getOperand(op)))
10626 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10628 FixedOperands[op] = 0; // Needs a PHI.
10632 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10633 // bother doing this transformation. At best, this will just save a bit of
10634 // offset calculation, but all the predecessors will have to materialize the
10635 // stack address into a register anyway. We'd actually rather *clone* the
10636 // load up into the predecessors so that we have a load of a gep of an alloca,
10637 // which can usually all be folded into the load.
10638 if (AllBasePointersAreAllocas)
10641 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10642 // that is variable.
10643 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10645 bool HasAnyPHIs = false;
10646 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10647 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10648 Value *FirstOp = FirstInst->getOperand(i);
10649 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10650 FirstOp->getName()+".pn");
10651 InsertNewInstBefore(NewPN, PN);
10653 NewPN->reserveOperandSpace(e);
10654 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10655 OperandPhis[i] = NewPN;
10656 FixedOperands[i] = NewPN;
10661 // Add all operands to the new PHIs.
10663 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10664 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10665 BasicBlock *InBB = PN.getIncomingBlock(i);
10667 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10668 if (PHINode *OpPhi = OperandPhis[op])
10669 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10673 Value *Base = FixedOperands[0];
10674 GetElementPtrInst *GEP =
10675 GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10676 FixedOperands.end());
10677 if (cast<GEPOperator>(FirstInst)->isInBounds())
10678 cast<GEPOperator>(GEP)->setIsInBounds(true);
10683 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10684 /// sink the load out of the block that defines it. This means that it must be
10685 /// obvious the value of the load is not changed from the point of the load to
10686 /// the end of the block it is in.
10688 /// Finally, it is safe, but not profitable, to sink a load targetting a
10689 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10691 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10692 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10694 for (++BBI; BBI != E; ++BBI)
10695 if (BBI->mayWriteToMemory())
10698 // Check for non-address taken alloca. If not address-taken already, it isn't
10699 // profitable to do this xform.
10700 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10701 bool isAddressTaken = false;
10702 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10704 if (isa<LoadInst>(UI)) continue;
10705 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10706 // If storing TO the alloca, then the address isn't taken.
10707 if (SI->getOperand(1) == AI) continue;
10709 isAddressTaken = true;
10713 if (!isAddressTaken && AI->isStaticAlloca())
10717 // If this load is a load from a GEP with a constant offset from an alloca,
10718 // then we don't want to sink it. In its present form, it will be
10719 // load [constant stack offset]. Sinking it will cause us to have to
10720 // materialize the stack addresses in each predecessor in a register only to
10721 // do a shared load from register in the successor.
10722 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10723 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10724 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10731 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10732 // operator and they all are only used by the PHI, PHI together their
10733 // inputs, and do the operation once, to the result of the PHI.
10734 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10735 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10737 // Scan the instruction, looking for input operations that can be folded away.
10738 // If all input operands to the phi are the same instruction (e.g. a cast from
10739 // the same type or "+42") we can pull the operation through the PHI, reducing
10740 // code size and simplifying code.
10741 Constant *ConstantOp = 0;
10742 const Type *CastSrcTy = 0;
10743 bool isVolatile = false;
10744 if (isa<CastInst>(FirstInst)) {
10745 CastSrcTy = FirstInst->getOperand(0)->getType();
10746 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10747 // Can fold binop, compare or shift here if the RHS is a constant,
10748 // otherwise call FoldPHIArgBinOpIntoPHI.
10749 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10750 if (ConstantOp == 0)
10751 return FoldPHIArgBinOpIntoPHI(PN);
10752 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10753 isVolatile = LI->isVolatile();
10754 // We can't sink the load if the loaded value could be modified between the
10755 // load and the PHI.
10756 if (LI->getParent() != PN.getIncomingBlock(0) ||
10757 !isSafeAndProfitableToSinkLoad(LI))
10760 // If the PHI is of volatile loads and the load block has multiple
10761 // successors, sinking it would remove a load of the volatile value from
10762 // the path through the other successor.
10764 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10767 } else if (isa<GetElementPtrInst>(FirstInst)) {
10768 return FoldPHIArgGEPIntoPHI(PN);
10770 return 0; // Cannot fold this operation.
10773 // Check to see if all arguments are the same operation.
10774 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10775 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10776 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10777 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10780 if (I->getOperand(0)->getType() != CastSrcTy)
10781 return 0; // Cast operation must match.
10782 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10783 // We can't sink the load if the loaded value could be modified between
10784 // the load and the PHI.
10785 if (LI->isVolatile() != isVolatile ||
10786 LI->getParent() != PN.getIncomingBlock(i) ||
10787 !isSafeAndProfitableToSinkLoad(LI))
10790 // If the PHI is of volatile loads and the load block has multiple
10791 // successors, sinking it would remove a load of the volatile value from
10792 // the path through the other successor.
10794 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10797 } else if (I->getOperand(1) != ConstantOp) {
10802 // Okay, they are all the same operation. Create a new PHI node of the
10803 // correct type, and PHI together all of the LHS's of the instructions.
10804 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10805 PN.getName()+".in");
10806 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10808 Value *InVal = FirstInst->getOperand(0);
10809 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10811 // Add all operands to the new PHI.
10812 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10813 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10814 if (NewInVal != InVal)
10816 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10821 // The new PHI unions all of the same values together. This is really
10822 // common, so we handle it intelligently here for compile-time speed.
10826 InsertNewInstBefore(NewPN, PN);
10830 // Insert and return the new operation.
10831 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10832 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10833 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10834 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10835 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10836 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10837 PhiVal, ConstantOp);
10838 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10840 // If this was a volatile load that we are merging, make sure to loop through
10841 // and mark all the input loads as non-volatile. If we don't do this, we will
10842 // insert a new volatile load and the old ones will not be deletable.
10844 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10845 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10847 return new LoadInst(PhiVal, "", isVolatile);
10850 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10852 static bool DeadPHICycle(PHINode *PN,
10853 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10854 if (PN->use_empty()) return true;
10855 if (!PN->hasOneUse()) return false;
10857 // Remember this node, and if we find the cycle, return.
10858 if (!PotentiallyDeadPHIs.insert(PN))
10861 // Don't scan crazily complex things.
10862 if (PotentiallyDeadPHIs.size() == 16)
10865 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10866 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10871 /// PHIsEqualValue - Return true if this phi node is always equal to
10872 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10873 /// z = some value; x = phi (y, z); y = phi (x, z)
10874 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10875 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10876 // See if we already saw this PHI node.
10877 if (!ValueEqualPHIs.insert(PN))
10880 // Don't scan crazily complex things.
10881 if (ValueEqualPHIs.size() == 16)
10884 // Scan the operands to see if they are either phi nodes or are equal to
10886 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10887 Value *Op = PN->getIncomingValue(i);
10888 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10889 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10891 } else if (Op != NonPhiInVal)
10899 // PHINode simplification
10901 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10902 // If LCSSA is around, don't mess with Phi nodes
10903 if (MustPreserveLCSSA) return 0;
10905 if (Value *V = PN.hasConstantValue())
10906 return ReplaceInstUsesWith(PN, V);
10908 // If all PHI operands are the same operation, pull them through the PHI,
10909 // reducing code size.
10910 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10911 isa<Instruction>(PN.getIncomingValue(1)) &&
10912 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10913 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10914 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10915 // than themselves more than once.
10916 PN.getIncomingValue(0)->hasOneUse())
10917 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10920 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10921 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10922 // PHI)... break the cycle.
10923 if (PN.hasOneUse()) {
10924 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10925 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10926 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10927 PotentiallyDeadPHIs.insert(&PN);
10928 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10929 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10932 // If this phi has a single use, and if that use just computes a value for
10933 // the next iteration of a loop, delete the phi. This occurs with unused
10934 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10935 // common case here is good because the only other things that catch this
10936 // are induction variable analysis (sometimes) and ADCE, which is only run
10938 if (PHIUser->hasOneUse() &&
10939 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10940 PHIUser->use_back() == &PN) {
10941 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
10945 // We sometimes end up with phi cycles that non-obviously end up being the
10946 // same value, for example:
10947 // z = some value; x = phi (y, z); y = phi (x, z)
10948 // where the phi nodes don't necessarily need to be in the same block. Do a
10949 // quick check to see if the PHI node only contains a single non-phi value, if
10950 // so, scan to see if the phi cycle is actually equal to that value.
10952 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10953 // Scan for the first non-phi operand.
10954 while (InValNo != NumOperandVals &&
10955 isa<PHINode>(PN.getIncomingValue(InValNo)))
10958 if (InValNo != NumOperandVals) {
10959 Value *NonPhiInVal = PN.getOperand(InValNo);
10961 // Scan the rest of the operands to see if there are any conflicts, if so
10962 // there is no need to recursively scan other phis.
10963 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10964 Value *OpVal = PN.getIncomingValue(InValNo);
10965 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10969 // If we scanned over all operands, then we have one unique value plus
10970 // phi values. Scan PHI nodes to see if they all merge in each other or
10972 if (InValNo == NumOperandVals) {
10973 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10974 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10975 return ReplaceInstUsesWith(PN, NonPhiInVal);
10982 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10983 Instruction *InsertPoint,
10984 InstCombiner *IC) {
10985 unsigned PtrSize = DTy->getScalarSizeInBits();
10986 unsigned VTySize = V->getType()->getScalarSizeInBits();
10987 // We must cast correctly to the pointer type. Ensure that we
10988 // sign extend the integer value if it is smaller as this is
10989 // used for address computation.
10990 Instruction::CastOps opcode =
10991 (VTySize < PtrSize ? Instruction::SExt :
10992 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10993 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10997 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10998 Value *PtrOp = GEP.getOperand(0);
10999 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
11000 // If so, eliminate the noop.
11001 if (GEP.getNumOperands() == 1)
11002 return ReplaceInstUsesWith(GEP, PtrOp);
11004 if (isa<UndefValue>(GEP.getOperand(0)))
11005 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
11007 bool HasZeroPointerIndex = false;
11008 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
11009 HasZeroPointerIndex = C->isNullValue();
11011 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11012 return ReplaceInstUsesWith(GEP, PtrOp);
11014 // Eliminate unneeded casts for indices.
11015 bool MadeChange = false;
11017 gep_type_iterator GTI = gep_type_begin(GEP);
11018 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
11019 i != e; ++i, ++GTI) {
11020 if (TD && isa<SequentialType>(*GTI)) {
11021 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
11022 if (CI->getOpcode() == Instruction::ZExt ||
11023 CI->getOpcode() == Instruction::SExt) {
11024 const Type *SrcTy = CI->getOperand(0)->getType();
11025 // We can eliminate a cast from i32 to i64 iff the target
11026 // is a 32-bit pointer target.
11027 if (SrcTy->getScalarSizeInBits() >= TD->getPointerSizeInBits()) {
11029 *i = CI->getOperand(0);
11033 // If we are using a wider index than needed for this platform, shrink it
11034 // to what we need. If narrower, sign-extend it to what we need.
11035 // If the incoming value needs a cast instruction,
11036 // insert it. This explicit cast can make subsequent optimizations more
11039 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
11040 if (Constant *C = dyn_cast<Constant>(Op)) {
11041 *i = ConstantExpr::getTrunc(C, TD->getIntPtrType());
11044 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
11049 } else if (TD->getTypeSizeInBits(Op->getType())
11050 < TD->getPointerSizeInBits()) {
11051 if (Constant *C = dyn_cast<Constant>(Op)) {
11052 *i = ConstantExpr::getSExt(C, TD->getIntPtrType());
11055 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
11063 if (MadeChange) return &GEP;
11065 // Combine Indices - If the source pointer to this getelementptr instruction
11066 // is a getelementptr instruction, combine the indices of the two
11067 // getelementptr instructions into a single instruction.
11069 SmallVector<Value*, 8> SrcGEPOperands;
11070 bool BothInBounds = cast<GEPOperator>(&GEP)->isInBounds();
11071 if (GEPOperator *Src = dyn_cast<GEPOperator>(PtrOp)) {
11072 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
11073 if (!Src->isInBounds())
11074 BothInBounds = false;
11077 if (!SrcGEPOperands.empty()) {
11078 // Note that if our source is a gep chain itself that we wait for that
11079 // chain to be resolved before we perform this transformation. This
11080 // avoids us creating a TON of code in some cases.
11082 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
11083 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
11084 return 0; // Wait until our source is folded to completion.
11086 SmallVector<Value*, 8> Indices;
11088 // Find out whether the last index in the source GEP is a sequential idx.
11089 bool EndsWithSequential = false;
11090 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
11091 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
11092 EndsWithSequential = !isa<StructType>(*I);
11094 // Can we combine the two pointer arithmetics offsets?
11095 if (EndsWithSequential) {
11096 // Replace: gep (gep %P, long B), long A, ...
11097 // With: T = long A+B; gep %P, T, ...
11099 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
11100 if (SO1 == Constant::getNullValue(SO1->getType())) {
11102 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
11105 // If they aren't the same type, convert both to an integer of the
11106 // target's pointer size.
11107 if (SO1->getType() != GO1->getType()) {
11108 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
11110 ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
11111 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
11113 ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
11115 unsigned PS = TD->getPointerSizeInBits();
11116 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
11117 // Convert GO1 to SO1's type.
11118 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
11120 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
11121 // Convert SO1 to GO1's type.
11122 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
11124 const Type *PT = TD->getIntPtrType();
11125 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
11126 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
11130 if (isa<Constant>(SO1) && isa<Constant>(GO1))
11131 Sum = ConstantExpr::getAdd(cast<Constant>(SO1),
11132 cast<Constant>(GO1));
11134 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11135 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
11139 // Recycle the GEP we already have if possible.
11140 if (SrcGEPOperands.size() == 2) {
11141 GEP.setOperand(0, SrcGEPOperands[0]);
11142 GEP.setOperand(1, Sum);
11145 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11146 SrcGEPOperands.end()-1);
11147 Indices.push_back(Sum);
11148 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
11150 } else if (isa<Constant>(*GEP.idx_begin()) &&
11151 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11152 SrcGEPOperands.size() != 1) {
11153 // Otherwise we can do the fold if the first index of the GEP is a zero
11154 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11155 SrcGEPOperands.end());
11156 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
11159 if (!Indices.empty()) {
11160 GetElementPtrInst *NewGEP = GetElementPtrInst::Create(SrcGEPOperands[0],
11165 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11169 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
11170 // GEP of global variable. If all of the indices for this GEP are
11171 // constants, we can promote this to a constexpr instead of an instruction.
11173 // Scan for nonconstants...
11174 SmallVector<Constant*, 8> Indices;
11175 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
11176 for (; I != E && isa<Constant>(*I); ++I)
11177 Indices.push_back(cast<Constant>(*I));
11179 if (I == E) { // If they are all constants...
11180 Constant *CE = ConstantExpr::getGetElementPtr(GV,
11181 &Indices[0],Indices.size());
11183 // Replace all uses of the GEP with the new constexpr...
11184 return ReplaceInstUsesWith(GEP, CE);
11186 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
11187 if (!isa<PointerType>(X->getType())) {
11188 // Not interesting. Source pointer must be a cast from pointer.
11189 } else if (HasZeroPointerIndex) {
11190 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11191 // into : GEP [10 x i8]* X, i32 0, ...
11193 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11194 // into : GEP i8* X, ...
11196 // This occurs when the program declares an array extern like "int X[];"
11197 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11198 const PointerType *XTy = cast<PointerType>(X->getType());
11199 if (const ArrayType *CATy =
11200 dyn_cast<ArrayType>(CPTy->getElementType())) {
11201 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11202 if (CATy->getElementType() == XTy->getElementType()) {
11203 // -> GEP i8* X, ...
11204 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11205 GetElementPtrInst *NewGEP =
11206 GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11208 if (cast<GEPOperator>(&GEP)->isInBounds())
11209 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11211 } else if (const ArrayType *XATy =
11212 dyn_cast<ArrayType>(XTy->getElementType())) {
11213 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11214 if (CATy->getElementType() == XATy->getElementType()) {
11215 // -> GEP [10 x i8]* X, i32 0, ...
11216 // At this point, we know that the cast source type is a pointer
11217 // to an array of the same type as the destination pointer
11218 // array. Because the array type is never stepped over (there
11219 // is a leading zero) we can fold the cast into this GEP.
11220 GEP.setOperand(0, X);
11225 } else if (GEP.getNumOperands() == 2) {
11226 // Transform things like:
11227 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11228 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11229 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11230 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11231 if (TD && isa<ArrayType>(SrcElTy) &&
11232 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11233 TD->getTypeAllocSize(ResElTy)) {
11235 Idx[0] = Constant::getNullValue(Type::Int32Ty);
11236 Idx[1] = GEP.getOperand(1);
11237 GetElementPtrInst *NewGEP =
11238 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11239 if (cast<GEPOperator>(&GEP)->isInBounds())
11240 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11241 Value *V = InsertNewInstBefore(NewGEP, GEP);
11242 // V and GEP are both pointer types --> BitCast
11243 return new BitCastInst(V, GEP.getType());
11246 // Transform things like:
11247 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11248 // (where tmp = 8*tmp2) into:
11249 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11251 if (TD && isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
11252 uint64_t ArrayEltSize =
11253 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11255 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11256 // allow either a mul, shift, or constant here.
11258 ConstantInt *Scale = 0;
11259 if (ArrayEltSize == 1) {
11260 NewIdx = GEP.getOperand(1);
11262 ConstantInt::get(cast<IntegerType>(NewIdx->getType()), 1);
11263 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11264 NewIdx = ConstantInt::get(CI->getType(), 1);
11266 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11267 if (Inst->getOpcode() == Instruction::Shl &&
11268 isa<ConstantInt>(Inst->getOperand(1))) {
11269 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11270 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11271 Scale = ConstantInt::get(cast<IntegerType>(Inst->getType()),
11273 NewIdx = Inst->getOperand(0);
11274 } else if (Inst->getOpcode() == Instruction::Mul &&
11275 isa<ConstantInt>(Inst->getOperand(1))) {
11276 Scale = cast<ConstantInt>(Inst->getOperand(1));
11277 NewIdx = Inst->getOperand(0);
11281 // If the index will be to exactly the right offset with the scale taken
11282 // out, perform the transformation. Note, we don't know whether Scale is
11283 // signed or not. We'll use unsigned version of division/modulo
11284 // operation after making sure Scale doesn't have the sign bit set.
11285 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11286 Scale->getZExtValue() % ArrayEltSize == 0) {
11287 Scale = ConstantInt::get(Scale->getType(),
11288 Scale->getZExtValue() / ArrayEltSize);
11289 if (Scale->getZExtValue() != 1) {
11291 ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
11293 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
11294 NewIdx = InsertNewInstBefore(Sc, GEP);
11297 // Insert the new GEP instruction.
11299 Idx[0] = Constant::getNullValue(Type::Int32Ty);
11301 Instruction *NewGEP =
11302 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11303 if (cast<GEPOperator>(&GEP)->isInBounds())
11304 cast<GEPOperator>(NewGEP)->setIsInBounds(true);
11305 NewGEP = InsertNewInstBefore(NewGEP, GEP);
11306 // The NewGEP must be pointer typed, so must the old one -> BitCast
11307 return new BitCastInst(NewGEP, GEP.getType());
11313 /// See if we can simplify:
11314 /// X = bitcast A to B*
11315 /// Y = gep X, <...constant indices...>
11316 /// into a gep of the original struct. This is important for SROA and alias
11317 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11318 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11320 !isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11321 // Determine how much the GEP moves the pointer. We are guaranteed to get
11322 // a constant back from EmitGEPOffset.
11323 ConstantInt *OffsetV =
11324 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11325 int64_t Offset = OffsetV->getSExtValue();
11327 // If this GEP instruction doesn't move the pointer, just replace the GEP
11328 // with a bitcast of the real input to the dest type.
11330 // If the bitcast is of an allocation, and the allocation will be
11331 // converted to match the type of the cast, don't touch this.
11332 if (isa<AllocationInst>(BCI->getOperand(0))) {
11333 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11334 if (Instruction *I = visitBitCast(*BCI)) {
11337 BCI->getParent()->getInstList().insert(BCI, I);
11338 ReplaceInstUsesWith(*BCI, I);
11343 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11346 // Otherwise, if the offset is non-zero, we need to find out if there is a
11347 // field at Offset in 'A's type. If so, we can pull the cast through the
11349 SmallVector<Value*, 8> NewIndices;
11351 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11352 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11353 Instruction *NGEP =
11354 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11356 if (NGEP->getType() == GEP.getType()) return NGEP;
11357 if (cast<GEPOperator>(&GEP)->isInBounds())
11358 cast<GEPOperator>(NGEP)->setIsInBounds(true);
11359 InsertNewInstBefore(NGEP, GEP);
11360 NGEP->takeName(&GEP);
11361 return new BitCastInst(NGEP, GEP.getType());
11369 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11370 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11371 if (AI.isArrayAllocation()) { // Check C != 1
11372 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11373 const Type *NewTy =
11374 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
11375 AllocationInst *New = 0;
11377 // Create and insert the replacement instruction...
11378 if (isa<MallocInst>(AI))
11379 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11381 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11382 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11385 InsertNewInstBefore(New, AI);
11387 // Scan to the end of the allocation instructions, to skip over a block of
11388 // allocas if possible...also skip interleaved debug info
11390 BasicBlock::iterator It = New;
11391 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11393 // Now that I is pointing to the first non-allocation-inst in the block,
11394 // insert our getelementptr instruction...
11396 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
11400 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11401 New->getName()+".sub", It);
11402 cast<GEPOperator>(V)->setIsInBounds(true);
11404 // Now make everything use the getelementptr instead of the original
11406 return ReplaceInstUsesWith(AI, V);
11407 } else if (isa<UndefValue>(AI.getArraySize())) {
11408 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11412 if (TD && isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11413 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11414 // Note that we only do this for alloca's, because malloc should allocate
11415 // and return a unique pointer, even for a zero byte allocation.
11416 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11417 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
11419 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11420 if (AI.getAlignment() == 0)
11421 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11427 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11428 Value *Op = FI.getOperand(0);
11430 // free undef -> unreachable.
11431 if (isa<UndefValue>(Op)) {
11432 // Insert a new store to null because we cannot modify the CFG here.
11433 new StoreInst(ConstantInt::getTrue(*Context),
11434 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
11435 return EraseInstFromFunction(FI);
11438 // If we have 'free null' delete the instruction. This can happen in stl code
11439 // when lots of inlining happens.
11440 if (isa<ConstantPointerNull>(Op))
11441 return EraseInstFromFunction(FI);
11443 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11444 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11445 FI.setOperand(0, CI->getOperand(0));
11449 // Change free (gep X, 0,0,0,0) into free(X)
11450 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11451 if (GEPI->hasAllZeroIndices()) {
11452 AddToWorkList(GEPI);
11453 FI.setOperand(0, GEPI->getOperand(0));
11458 // Change free(malloc) into nothing, if the malloc has a single use.
11459 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11460 if (MI->hasOneUse()) {
11461 EraseInstFromFunction(FI);
11462 return EraseInstFromFunction(*MI);
11469 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11470 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11471 const TargetData *TD) {
11472 User *CI = cast<User>(LI.getOperand(0));
11473 Value *CastOp = CI->getOperand(0);
11474 LLVMContext *Context = IC.getContext();
11477 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11478 // Instead of loading constant c string, use corresponding integer value
11479 // directly if string length is small enough.
11481 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11482 unsigned len = Str.length();
11483 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11484 unsigned numBits = Ty->getPrimitiveSizeInBits();
11485 // Replace LI with immediate integer store.
11486 if ((numBits >> 3) == len + 1) {
11487 APInt StrVal(numBits, 0);
11488 APInt SingleChar(numBits, 0);
11489 if (TD->isLittleEndian()) {
11490 for (signed i = len-1; i >= 0; i--) {
11491 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11492 StrVal = (StrVal << 8) | SingleChar;
11495 for (unsigned i = 0; i < len; i++) {
11496 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11497 StrVal = (StrVal << 8) | SingleChar;
11499 // Append NULL at the end.
11501 StrVal = (StrVal << 8) | SingleChar;
11503 Value *NL = ConstantInt::get(*Context, StrVal);
11504 return IC.ReplaceInstUsesWith(LI, NL);
11510 const PointerType *DestTy = cast<PointerType>(CI->getType());
11511 const Type *DestPTy = DestTy->getElementType();
11512 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11514 // If the address spaces don't match, don't eliminate the cast.
11515 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11518 const Type *SrcPTy = SrcTy->getElementType();
11520 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11521 isa<VectorType>(DestPTy)) {
11522 // If the source is an array, the code below will not succeed. Check to
11523 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11525 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11526 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11527 if (ASrcTy->getNumElements() != 0) {
11529 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
11530 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
11531 SrcTy = cast<PointerType>(CastOp->getType());
11532 SrcPTy = SrcTy->getElementType();
11535 if (IC.getTargetData() &&
11536 (SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11537 isa<VectorType>(SrcPTy)) &&
11538 // Do not allow turning this into a load of an integer, which is then
11539 // casted to a pointer, this pessimizes pointer analysis a lot.
11540 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11541 IC.getTargetData()->getTypeSizeInBits(SrcPTy) ==
11542 IC.getTargetData()->getTypeSizeInBits(DestPTy)) {
11544 // Okay, we are casting from one integer or pointer type to another of
11545 // the same size. Instead of casting the pointer before the load, cast
11546 // the result of the loaded value.
11547 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11549 LI.isVolatile()),LI);
11550 // Now cast the result of the load.
11551 return new BitCastInst(NewLoad, LI.getType());
11558 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11559 Value *Op = LI.getOperand(0);
11561 // Attempt to improve the alignment.
11563 unsigned KnownAlign =
11564 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11566 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11567 LI.getAlignment()))
11568 LI.setAlignment(KnownAlign);
11571 // load (cast X) --> cast (load X) iff safe
11572 if (isa<CastInst>(Op))
11573 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11576 // None of the following transforms are legal for volatile loads.
11577 if (LI.isVolatile()) return 0;
11579 // Do really simple store-to-load forwarding and load CSE, to catch cases
11580 // where there are several consequtive memory accesses to the same location,
11581 // separated by a few arithmetic operations.
11582 BasicBlock::iterator BBI = &LI;
11583 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11584 return ReplaceInstUsesWith(LI, AvailableVal);
11586 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11587 const Value *GEPI0 = GEPI->getOperand(0);
11588 // TODO: Consider a target hook for valid address spaces for this xform.
11589 if (isa<ConstantPointerNull>(GEPI0) &&
11590 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11591 // Insert a new store to null instruction before the load to indicate
11592 // that this code is not reachable. We do this instead of inserting
11593 // an unreachable instruction directly because we cannot modify the
11595 new StoreInst(UndefValue::get(LI.getType()),
11596 Constant::getNullValue(Op->getType()), &LI);
11597 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11601 if (Constant *C = dyn_cast<Constant>(Op)) {
11602 // load null/undef -> undef
11603 // TODO: Consider a target hook for valid address spaces for this xform.
11604 if (isa<UndefValue>(C) || (C->isNullValue() &&
11605 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11606 // Insert a new store to null instruction before the load to indicate that
11607 // this code is not reachable. We do this instead of inserting an
11608 // unreachable instruction directly because we cannot modify the CFG.
11609 new StoreInst(UndefValue::get(LI.getType()),
11610 Constant::getNullValue(Op->getType()), &LI);
11611 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11614 // Instcombine load (constant global) into the value loaded.
11615 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11616 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11617 return ReplaceInstUsesWith(LI, GV->getInitializer());
11619 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11620 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11621 if (CE->getOpcode() == Instruction::GetElementPtr) {
11622 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11623 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11625 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11627 return ReplaceInstUsesWith(LI, V);
11628 if (CE->getOperand(0)->isNullValue()) {
11629 // Insert a new store to null instruction before the load to indicate
11630 // that this code is not reachable. We do this instead of inserting
11631 // an unreachable instruction directly because we cannot modify the
11633 new StoreInst(UndefValue::get(LI.getType()),
11634 Constant::getNullValue(Op->getType()), &LI);
11635 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11638 } else if (CE->isCast()) {
11639 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11645 // If this load comes from anywhere in a constant global, and if the global
11646 // is all undef or zero, we know what it loads.
11647 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11648 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11649 if (GV->getInitializer()->isNullValue())
11650 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
11651 else if (isa<UndefValue>(GV->getInitializer()))
11652 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
11656 if (Op->hasOneUse()) {
11657 // Change select and PHI nodes to select values instead of addresses: this
11658 // helps alias analysis out a lot, allows many others simplifications, and
11659 // exposes redundancy in the code.
11661 // Note that we cannot do the transformation unless we know that the
11662 // introduced loads cannot trap! Something like this is valid as long as
11663 // the condition is always false: load (select bool %C, int* null, int* %G),
11664 // but it would not be valid if we transformed it to load from null
11665 // unconditionally.
11667 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11668 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11669 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11670 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11671 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11672 SI->getOperand(1)->getName()+".val"), LI);
11673 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11674 SI->getOperand(2)->getName()+".val"), LI);
11675 return SelectInst::Create(SI->getCondition(), V1, V2);
11678 // load (select (cond, null, P)) -> load P
11679 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11680 if (C->isNullValue()) {
11681 LI.setOperand(0, SI->getOperand(2));
11685 // load (select (cond, P, null)) -> load P
11686 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11687 if (C->isNullValue()) {
11688 LI.setOperand(0, SI->getOperand(1));
11696 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11697 /// when possible. This makes it generally easy to do alias analysis and/or
11698 /// SROA/mem2reg of the memory object.
11699 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11700 User *CI = cast<User>(SI.getOperand(1));
11701 Value *CastOp = CI->getOperand(0);
11703 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11704 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11705 if (SrcTy == 0) return 0;
11707 const Type *SrcPTy = SrcTy->getElementType();
11709 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11712 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11713 /// to its first element. This allows us to handle things like:
11714 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11715 /// on 32-bit hosts.
11716 SmallVector<Value*, 4> NewGEPIndices;
11718 // If the source is an array, the code below will not succeed. Check to
11719 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11721 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11722 // Index through pointer.
11723 Constant *Zero = Constant::getNullValue(Type::Int32Ty);
11724 NewGEPIndices.push_back(Zero);
11727 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11728 if (!STy->getNumElements()) /* Struct can be empty {} */
11730 NewGEPIndices.push_back(Zero);
11731 SrcPTy = STy->getElementType(0);
11732 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11733 NewGEPIndices.push_back(Zero);
11734 SrcPTy = ATy->getElementType();
11740 SrcTy = PointerType::get(SrcPTy, SrcTy->getAddressSpace());
11743 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11746 // If the pointers point into different address spaces or if they point to
11747 // values with different sizes, we can't do the transformation.
11748 if (!IC.getTargetData() ||
11749 SrcTy->getAddressSpace() !=
11750 cast<PointerType>(CI->getType())->getAddressSpace() ||
11751 IC.getTargetData()->getTypeSizeInBits(SrcPTy) !=
11752 IC.getTargetData()->getTypeSizeInBits(DestPTy))
11755 // Okay, we are casting from one integer or pointer type to another of
11756 // the same size. Instead of casting the pointer before
11757 // the store, cast the value to be stored.
11759 Value *SIOp0 = SI.getOperand(0);
11760 Instruction::CastOps opcode = Instruction::BitCast;
11761 const Type* CastSrcTy = SIOp0->getType();
11762 const Type* CastDstTy = SrcPTy;
11763 if (isa<PointerType>(CastDstTy)) {
11764 if (CastSrcTy->isInteger())
11765 opcode = Instruction::IntToPtr;
11766 } else if (isa<IntegerType>(CastDstTy)) {
11767 if (isa<PointerType>(SIOp0->getType()))
11768 opcode = Instruction::PtrToInt;
11771 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11772 // emit a GEP to index into its first field.
11773 if (!NewGEPIndices.empty()) {
11774 if (Constant *C = dyn_cast<Constant>(CastOp))
11775 CastOp = ConstantExpr::getGetElementPtr(C, &NewGEPIndices[0],
11776 NewGEPIndices.size());
11778 CastOp = IC.InsertNewInstBefore(
11779 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11780 NewGEPIndices.end()), SI);
11781 cast<GEPOperator>(CastOp)->setIsInBounds(true);
11784 if (Constant *C = dyn_cast<Constant>(SIOp0))
11785 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
11787 NewCast = IC.InsertNewInstBefore(
11788 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11790 return new StoreInst(NewCast, CastOp);
11793 /// equivalentAddressValues - Test if A and B will obviously have the same
11794 /// value. This includes recognizing that %t0 and %t1 will have the same
11795 /// value in code like this:
11796 /// %t0 = getelementptr \@a, 0, 3
11797 /// store i32 0, i32* %t0
11798 /// %t1 = getelementptr \@a, 0, 3
11799 /// %t2 = load i32* %t1
11801 static bool equivalentAddressValues(Value *A, Value *B) {
11802 // Test if the values are trivially equivalent.
11803 if (A == B) return true;
11805 // Test if the values come form identical arithmetic instructions.
11806 if (isa<BinaryOperator>(A) ||
11807 isa<CastInst>(A) ||
11809 isa<GetElementPtrInst>(A))
11810 if (Instruction *BI = dyn_cast<Instruction>(B))
11811 if (cast<Instruction>(A)->isIdenticalTo(BI))
11814 // Otherwise they may not be equivalent.
11818 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11819 // return the llvm.dbg.declare.
11820 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11821 if (!V->hasNUses(2))
11823 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11825 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11827 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11828 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11835 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11836 Value *Val = SI.getOperand(0);
11837 Value *Ptr = SI.getOperand(1);
11839 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11840 EraseInstFromFunction(SI);
11845 // If the RHS is an alloca with a single use, zapify the store, making the
11847 // If the RHS is an alloca with a two uses, the other one being a
11848 // llvm.dbg.declare, zapify the store and the declare, making the
11849 // alloca dead. We must do this to prevent declare's from affecting
11851 if (!SI.isVolatile()) {
11852 if (Ptr->hasOneUse()) {
11853 if (isa<AllocaInst>(Ptr)) {
11854 EraseInstFromFunction(SI);
11858 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11859 if (isa<AllocaInst>(GEP->getOperand(0))) {
11860 if (GEP->getOperand(0)->hasOneUse()) {
11861 EraseInstFromFunction(SI);
11865 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11866 EraseInstFromFunction(*DI);
11867 EraseInstFromFunction(SI);
11874 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11875 EraseInstFromFunction(*DI);
11876 EraseInstFromFunction(SI);
11882 // Attempt to improve the alignment.
11884 unsigned KnownAlign =
11885 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11887 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11888 SI.getAlignment()))
11889 SI.setAlignment(KnownAlign);
11892 // Do really simple DSE, to catch cases where there are several consecutive
11893 // stores to the same location, separated by a few arithmetic operations. This
11894 // situation often occurs with bitfield accesses.
11895 BasicBlock::iterator BBI = &SI;
11896 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11899 // Don't count debug info directives, lest they affect codegen,
11900 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11901 // It is necessary for correctness to skip those that feed into a
11902 // llvm.dbg.declare, as these are not present when debugging is off.
11903 if (isa<DbgInfoIntrinsic>(BBI) ||
11904 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11909 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11910 // Prev store isn't volatile, and stores to the same location?
11911 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11912 SI.getOperand(1))) {
11915 EraseInstFromFunction(*PrevSI);
11921 // If this is a load, we have to stop. However, if the loaded value is from
11922 // the pointer we're loading and is producing the pointer we're storing,
11923 // then *this* store is dead (X = load P; store X -> P).
11924 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11925 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11926 !SI.isVolatile()) {
11927 EraseInstFromFunction(SI);
11931 // Otherwise, this is a load from some other location. Stores before it
11932 // may not be dead.
11936 // Don't skip over loads or things that can modify memory.
11937 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11942 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11944 // store X, null -> turns into 'unreachable' in SimplifyCFG
11945 if (isa<ConstantPointerNull>(Ptr) &&
11946 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11947 if (!isa<UndefValue>(Val)) {
11948 SI.setOperand(0, UndefValue::get(Val->getType()));
11949 if (Instruction *U = dyn_cast<Instruction>(Val))
11950 AddToWorkList(U); // Dropped a use.
11953 return 0; // Do not modify these!
11956 // store undef, Ptr -> noop
11957 if (isa<UndefValue>(Val)) {
11958 EraseInstFromFunction(SI);
11963 // If the pointer destination is a cast, see if we can fold the cast into the
11965 if (isa<CastInst>(Ptr))
11966 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11968 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11970 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11974 // If this store is the last instruction in the basic block (possibly
11975 // excepting debug info instructions and the pointer bitcasts that feed
11976 // into them), and if the block ends with an unconditional branch, try
11977 // to move it to the successor block.
11981 } while (isa<DbgInfoIntrinsic>(BBI) ||
11982 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11983 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11984 if (BI->isUnconditional())
11985 if (SimplifyStoreAtEndOfBlock(SI))
11986 return 0; // xform done!
11991 /// SimplifyStoreAtEndOfBlock - Turn things like:
11992 /// if () { *P = v1; } else { *P = v2 }
11993 /// into a phi node with a store in the successor.
11995 /// Simplify things like:
11996 /// *P = v1; if () { *P = v2; }
11997 /// into a phi node with a store in the successor.
11999 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
12000 BasicBlock *StoreBB = SI.getParent();
12002 // Check to see if the successor block has exactly two incoming edges. If
12003 // so, see if the other predecessor contains a store to the same location.
12004 // if so, insert a PHI node (if needed) and move the stores down.
12005 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
12007 // Determine whether Dest has exactly two predecessors and, if so, compute
12008 // the other predecessor.
12009 pred_iterator PI = pred_begin(DestBB);
12010 BasicBlock *OtherBB = 0;
12011 if (*PI != StoreBB)
12014 if (PI == pred_end(DestBB))
12017 if (*PI != StoreBB) {
12022 if (++PI != pred_end(DestBB))
12025 // Bail out if all the relevant blocks aren't distinct (this can happen,
12026 // for example, if SI is in an infinite loop)
12027 if (StoreBB == DestBB || OtherBB == DestBB)
12030 // Verify that the other block ends in a branch and is not otherwise empty.
12031 BasicBlock::iterator BBI = OtherBB->getTerminator();
12032 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
12033 if (!OtherBr || BBI == OtherBB->begin())
12036 // If the other block ends in an unconditional branch, check for the 'if then
12037 // else' case. there is an instruction before the branch.
12038 StoreInst *OtherStore = 0;
12039 if (OtherBr->isUnconditional()) {
12041 // Skip over debugging info.
12042 while (isa<DbgInfoIntrinsic>(BBI) ||
12043 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12044 if (BBI==OtherBB->begin())
12048 // If this isn't a store, or isn't a store to the same location, bail out.
12049 OtherStore = dyn_cast<StoreInst>(BBI);
12050 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
12053 // Otherwise, the other block ended with a conditional branch. If one of the
12054 // destinations is StoreBB, then we have the if/then case.
12055 if (OtherBr->getSuccessor(0) != StoreBB &&
12056 OtherBr->getSuccessor(1) != StoreBB)
12059 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12060 // if/then triangle. See if there is a store to the same ptr as SI that
12061 // lives in OtherBB.
12063 // Check to see if we find the matching store.
12064 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12065 if (OtherStore->getOperand(1) != SI.getOperand(1))
12069 // If we find something that may be using or overwriting the stored
12070 // value, or if we run out of instructions, we can't do the xform.
12071 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12072 BBI == OtherBB->begin())
12076 // In order to eliminate the store in OtherBr, we have to
12077 // make sure nothing reads or overwrites the stored value in
12079 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12080 // FIXME: This should really be AA driven.
12081 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12086 // Insert a PHI node now if we need it.
12087 Value *MergedVal = OtherStore->getOperand(0);
12088 if (MergedVal != SI.getOperand(0)) {
12089 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12090 PN->reserveOperandSpace(2);
12091 PN->addIncoming(SI.getOperand(0), SI.getParent());
12092 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12093 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12096 // Advance to a place where it is safe to insert the new store and
12098 BBI = DestBB->getFirstNonPHI();
12099 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12100 OtherStore->isVolatile()), *BBI);
12102 // Nuke the old stores.
12103 EraseInstFromFunction(SI);
12104 EraseInstFromFunction(*OtherStore);
12110 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12111 // Change br (not X), label True, label False to: br X, label False, True
12113 BasicBlock *TrueDest;
12114 BasicBlock *FalseDest;
12115 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest), *Context) &&
12116 !isa<Constant>(X)) {
12117 // Swap Destinations and condition...
12118 BI.setCondition(X);
12119 BI.setSuccessor(0, FalseDest);
12120 BI.setSuccessor(1, TrueDest);
12124 // Cannonicalize fcmp_one -> fcmp_oeq
12125 FCmpInst::Predicate FPred; Value *Y;
12126 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12127 TrueDest, FalseDest), *Context))
12128 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12129 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
12130 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
12131 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
12132 Instruction *NewSCC = new FCmpInst(I, NewPred, X, Y, "");
12133 NewSCC->takeName(I);
12134 // Swap Destinations and condition...
12135 BI.setCondition(NewSCC);
12136 BI.setSuccessor(0, FalseDest);
12137 BI.setSuccessor(1, TrueDest);
12138 RemoveFromWorkList(I);
12139 I->eraseFromParent();
12140 AddToWorkList(NewSCC);
12144 // Cannonicalize icmp_ne -> icmp_eq
12145 ICmpInst::Predicate IPred;
12146 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12147 TrueDest, FalseDest), *Context))
12148 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12149 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12150 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
12151 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
12152 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
12153 Instruction *NewSCC = new ICmpInst(I, NewPred, X, Y, "");
12154 NewSCC->takeName(I);
12155 // Swap Destinations and condition...
12156 BI.setCondition(NewSCC);
12157 BI.setSuccessor(0, FalseDest);
12158 BI.setSuccessor(1, TrueDest);
12159 RemoveFromWorkList(I);
12160 I->eraseFromParent();;
12161 AddToWorkList(NewSCC);
12168 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12169 Value *Cond = SI.getCondition();
12170 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12171 if (I->getOpcode() == Instruction::Add)
12172 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12173 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12174 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12176 ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
12178 SI.setOperand(0, I->getOperand(0));
12186 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12187 Value *Agg = EV.getAggregateOperand();
12189 if (!EV.hasIndices())
12190 return ReplaceInstUsesWith(EV, Agg);
12192 if (Constant *C = dyn_cast<Constant>(Agg)) {
12193 if (isa<UndefValue>(C))
12194 return ReplaceInstUsesWith(EV, UndefValue::get(EV.getType()));
12196 if (isa<ConstantAggregateZero>(C))
12197 return ReplaceInstUsesWith(EV, Constant::getNullValue(EV.getType()));
12199 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12200 // Extract the element indexed by the first index out of the constant
12201 Value *V = C->getOperand(*EV.idx_begin());
12202 if (EV.getNumIndices() > 1)
12203 // Extract the remaining indices out of the constant indexed by the
12205 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12207 return ReplaceInstUsesWith(EV, V);
12209 return 0; // Can't handle other constants
12211 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12212 // We're extracting from an insertvalue instruction, compare the indices
12213 const unsigned *exti, *exte, *insi, *inse;
12214 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12215 exte = EV.idx_end(), inse = IV->idx_end();
12216 exti != exte && insi != inse;
12218 if (*insi != *exti)
12219 // The insert and extract both reference distinctly different elements.
12220 // This means the extract is not influenced by the insert, and we can
12221 // replace the aggregate operand of the extract with the aggregate
12222 // operand of the insert. i.e., replace
12223 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12224 // %E = extractvalue { i32, { i32 } } %I, 0
12226 // %E = extractvalue { i32, { i32 } } %A, 0
12227 return ExtractValueInst::Create(IV->getAggregateOperand(),
12228 EV.idx_begin(), EV.idx_end());
12230 if (exti == exte && insi == inse)
12231 // Both iterators are at the end: Index lists are identical. Replace
12232 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12233 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12235 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12236 if (exti == exte) {
12237 // The extract list is a prefix of the insert list. i.e. replace
12238 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12239 // %E = extractvalue { i32, { i32 } } %I, 1
12241 // %X = extractvalue { i32, { i32 } } %A, 1
12242 // %E = insertvalue { i32 } %X, i32 42, 0
12243 // by switching the order of the insert and extract (though the
12244 // insertvalue should be left in, since it may have other uses).
12245 Value *NewEV = InsertNewInstBefore(
12246 ExtractValueInst::Create(IV->getAggregateOperand(),
12247 EV.idx_begin(), EV.idx_end()),
12249 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12253 // The insert list is a prefix of the extract list
12254 // We can simply remove the common indices from the extract and make it
12255 // operate on the inserted value instead of the insertvalue result.
12257 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12258 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12260 // %E extractvalue { i32 } { i32 42 }, 0
12261 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12264 // Can't simplify extracts from other values. Note that nested extracts are
12265 // already simplified implicitely by the above (extract ( extract (insert) )
12266 // will be translated into extract ( insert ( extract ) ) first and then just
12267 // the value inserted, if appropriate).
12271 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12272 /// is to leave as a vector operation.
12273 static bool CheapToScalarize(Value *V, bool isConstant) {
12274 if (isa<ConstantAggregateZero>(V))
12276 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12277 if (isConstant) return true;
12278 // If all elts are the same, we can extract.
12279 Constant *Op0 = C->getOperand(0);
12280 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12281 if (C->getOperand(i) != Op0)
12285 Instruction *I = dyn_cast<Instruction>(V);
12286 if (!I) return false;
12288 // Insert element gets simplified to the inserted element or is deleted if
12289 // this is constant idx extract element and its a constant idx insertelt.
12290 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12291 isa<ConstantInt>(I->getOperand(2)))
12293 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12295 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12296 if (BO->hasOneUse() &&
12297 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12298 CheapToScalarize(BO->getOperand(1), isConstant)))
12300 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12301 if (CI->hasOneUse() &&
12302 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12303 CheapToScalarize(CI->getOperand(1), isConstant)))
12309 /// Read and decode a shufflevector mask.
12311 /// It turns undef elements into values that are larger than the number of
12312 /// elements in the input.
12313 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12314 unsigned NElts = SVI->getType()->getNumElements();
12315 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12316 return std::vector<unsigned>(NElts, 0);
12317 if (isa<UndefValue>(SVI->getOperand(2)))
12318 return std::vector<unsigned>(NElts, 2*NElts);
12320 std::vector<unsigned> Result;
12321 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12322 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12323 if (isa<UndefValue>(*i))
12324 Result.push_back(NElts*2); // undef -> 8
12326 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12330 /// FindScalarElement - Given a vector and an element number, see if the scalar
12331 /// value is already around as a register, for example if it were inserted then
12332 /// extracted from the vector.
12333 static Value *FindScalarElement(Value *V, unsigned EltNo,
12334 LLVMContext *Context) {
12335 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12336 const VectorType *PTy = cast<VectorType>(V->getType());
12337 unsigned Width = PTy->getNumElements();
12338 if (EltNo >= Width) // Out of range access.
12339 return UndefValue::get(PTy->getElementType());
12341 if (isa<UndefValue>(V))
12342 return UndefValue::get(PTy->getElementType());
12343 else if (isa<ConstantAggregateZero>(V))
12344 return Constant::getNullValue(PTy->getElementType());
12345 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12346 return CP->getOperand(EltNo);
12347 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12348 // If this is an insert to a variable element, we don't know what it is.
12349 if (!isa<ConstantInt>(III->getOperand(2)))
12351 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12353 // If this is an insert to the element we are looking for, return the
12355 if (EltNo == IIElt)
12356 return III->getOperand(1);
12358 // Otherwise, the insertelement doesn't modify the value, recurse on its
12360 return FindScalarElement(III->getOperand(0), EltNo, Context);
12361 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12362 unsigned LHSWidth =
12363 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12364 unsigned InEl = getShuffleMask(SVI)[EltNo];
12365 if (InEl < LHSWidth)
12366 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12367 else if (InEl < LHSWidth*2)
12368 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12370 return UndefValue::get(PTy->getElementType());
12373 // Otherwise, we don't know.
12377 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12378 // If vector val is undef, replace extract with scalar undef.
12379 if (isa<UndefValue>(EI.getOperand(0)))
12380 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12382 // If vector val is constant 0, replace extract with scalar 0.
12383 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12384 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
12386 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12387 // If vector val is constant with all elements the same, replace EI with
12388 // that element. When the elements are not identical, we cannot replace yet
12389 // (we do that below, but only when the index is constant).
12390 Constant *op0 = C->getOperand(0);
12391 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12392 if (C->getOperand(i) != op0) {
12397 return ReplaceInstUsesWith(EI, op0);
12400 // If extracting a specified index from the vector, see if we can recursively
12401 // find a previously computed scalar that was inserted into the vector.
12402 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12403 unsigned IndexVal = IdxC->getZExtValue();
12404 unsigned VectorWidth =
12405 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12407 // If this is extracting an invalid index, turn this into undef, to avoid
12408 // crashing the code below.
12409 if (IndexVal >= VectorWidth)
12410 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12412 // This instruction only demands the single element from the input vector.
12413 // If the input vector has a single use, simplify it based on this use
12415 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12416 APInt UndefElts(VectorWidth, 0);
12417 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12418 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12419 DemandedMask, UndefElts)) {
12420 EI.setOperand(0, V);
12425 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12426 return ReplaceInstUsesWith(EI, Elt);
12428 // If the this extractelement is directly using a bitcast from a vector of
12429 // the same number of elements, see if we can find the source element from
12430 // it. In this case, we will end up needing to bitcast the scalars.
12431 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12432 if (const VectorType *VT =
12433 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12434 if (VT->getNumElements() == VectorWidth)
12435 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12436 IndexVal, Context))
12437 return new BitCastInst(Elt, EI.getType());
12441 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12442 if (I->hasOneUse()) {
12443 // Push extractelement into predecessor operation if legal and
12444 // profitable to do so
12445 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12446 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12447 if (CheapToScalarize(BO, isConstantElt)) {
12448 ExtractElementInst *newEI0 =
12449 ExtractElementInst::Create(BO->getOperand(0), EI.getOperand(1),
12450 EI.getName()+".lhs");
12451 ExtractElementInst *newEI1 =
12452 ExtractElementInst::Create(BO->getOperand(1), EI.getOperand(1),
12453 EI.getName()+".rhs");
12454 InsertNewInstBefore(newEI0, EI);
12455 InsertNewInstBefore(newEI1, EI);
12456 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12458 } else if (isa<LoadInst>(I)) {
12460 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12461 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12462 PointerType::get(EI.getType(), AS),EI);
12463 GetElementPtrInst *GEP =
12464 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12465 cast<GEPOperator>(GEP)->setIsInBounds(true);
12466 InsertNewInstBefore(GEP, EI);
12467 return new LoadInst(GEP);
12470 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12471 // Extracting the inserted element?
12472 if (IE->getOperand(2) == EI.getOperand(1))
12473 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12474 // If the inserted and extracted elements are constants, they must not
12475 // be the same value, extract from the pre-inserted value instead.
12476 if (isa<Constant>(IE->getOperand(2)) &&
12477 isa<Constant>(EI.getOperand(1))) {
12478 AddUsesToWorkList(EI);
12479 EI.setOperand(0, IE->getOperand(0));
12482 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12483 // If this is extracting an element from a shufflevector, figure out where
12484 // it came from and extract from the appropriate input element instead.
12485 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12486 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12488 unsigned LHSWidth =
12489 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12491 if (SrcIdx < LHSWidth)
12492 Src = SVI->getOperand(0);
12493 else if (SrcIdx < LHSWidth*2) {
12494 SrcIdx -= LHSWidth;
12495 Src = SVI->getOperand(1);
12497 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
12499 return ExtractElementInst::Create(Src,
12500 ConstantInt::get(Type::Int32Ty, SrcIdx, false));
12503 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12508 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12509 /// elements from either LHS or RHS, return the shuffle mask and true.
12510 /// Otherwise, return false.
12511 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12512 std::vector<Constant*> &Mask,
12513 LLVMContext *Context) {
12514 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12515 "Invalid CollectSingleShuffleElements");
12516 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12518 if (isa<UndefValue>(V)) {
12519 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12521 } else if (V == LHS) {
12522 for (unsigned i = 0; i != NumElts; ++i)
12523 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12525 } else if (V == RHS) {
12526 for (unsigned i = 0; i != NumElts; ++i)
12527 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
12529 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12530 // If this is an insert of an extract from some other vector, include it.
12531 Value *VecOp = IEI->getOperand(0);
12532 Value *ScalarOp = IEI->getOperand(1);
12533 Value *IdxOp = IEI->getOperand(2);
12535 if (!isa<ConstantInt>(IdxOp))
12537 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12539 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12540 // Okay, we can handle this if the vector we are insertinting into is
12541 // transitively ok.
12542 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12543 // If so, update the mask to reflect the inserted undef.
12544 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
12547 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12548 if (isa<ConstantInt>(EI->getOperand(1)) &&
12549 EI->getOperand(0)->getType() == V->getType()) {
12550 unsigned ExtractedIdx =
12551 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12553 // This must be extracting from either LHS or RHS.
12554 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12555 // Okay, we can handle this if the vector we are insertinting into is
12556 // transitively ok.
12557 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12558 // If so, update the mask to reflect the inserted value.
12559 if (EI->getOperand(0) == LHS) {
12560 Mask[InsertedIdx % NumElts] =
12561 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12563 assert(EI->getOperand(0) == RHS);
12564 Mask[InsertedIdx % NumElts] =
12565 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
12574 // TODO: Handle shufflevector here!
12579 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12580 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12581 /// that computes V and the LHS value of the shuffle.
12582 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12583 Value *&RHS, LLVMContext *Context) {
12584 assert(isa<VectorType>(V->getType()) &&
12585 (RHS == 0 || V->getType() == RHS->getType()) &&
12586 "Invalid shuffle!");
12587 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12589 if (isa<UndefValue>(V)) {
12590 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
12592 } else if (isa<ConstantAggregateZero>(V)) {
12593 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
12595 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12596 // If this is an insert of an extract from some other vector, include it.
12597 Value *VecOp = IEI->getOperand(0);
12598 Value *ScalarOp = IEI->getOperand(1);
12599 Value *IdxOp = IEI->getOperand(2);
12601 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12602 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12603 EI->getOperand(0)->getType() == V->getType()) {
12604 unsigned ExtractedIdx =
12605 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12606 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12608 // Either the extracted from or inserted into vector must be RHSVec,
12609 // otherwise we'd end up with a shuffle of three inputs.
12610 if (EI->getOperand(0) == RHS || RHS == 0) {
12611 RHS = EI->getOperand(0);
12612 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12613 Mask[InsertedIdx % NumElts] =
12614 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
12618 if (VecOp == RHS) {
12619 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12621 // Everything but the extracted element is replaced with the RHS.
12622 for (unsigned i = 0; i != NumElts; ++i) {
12623 if (i != InsertedIdx)
12624 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
12629 // If this insertelement is a chain that comes from exactly these two
12630 // vectors, return the vector and the effective shuffle.
12631 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12633 return EI->getOperand(0);
12638 // TODO: Handle shufflevector here!
12640 // Otherwise, can't do anything fancy. Return an identity vector.
12641 for (unsigned i = 0; i != NumElts; ++i)
12642 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
12646 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12647 Value *VecOp = IE.getOperand(0);
12648 Value *ScalarOp = IE.getOperand(1);
12649 Value *IdxOp = IE.getOperand(2);
12651 // Inserting an undef or into an undefined place, remove this.
12652 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12653 ReplaceInstUsesWith(IE, VecOp);
12655 // If the inserted element was extracted from some other vector, and if the
12656 // indexes are constant, try to turn this into a shufflevector operation.
12657 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12658 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12659 EI->getOperand(0)->getType() == IE.getType()) {
12660 unsigned NumVectorElts = IE.getType()->getNumElements();
12661 unsigned ExtractedIdx =
12662 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12663 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12665 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12666 return ReplaceInstUsesWith(IE, VecOp);
12668 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12669 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
12671 // If we are extracting a value from a vector, then inserting it right
12672 // back into the same place, just use the input vector.
12673 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12674 return ReplaceInstUsesWith(IE, VecOp);
12676 // We could theoretically do this for ANY input. However, doing so could
12677 // turn chains of insertelement instructions into a chain of shufflevector
12678 // instructions, and right now we do not merge shufflevectors. As such,
12679 // only do this in a situation where it is clear that there is benefit.
12680 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12681 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12682 // the values of VecOp, except then one read from EIOp0.
12683 // Build a new shuffle mask.
12684 std::vector<Constant*> Mask;
12685 if (isa<UndefValue>(VecOp))
12686 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
12688 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12689 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
12692 Mask[InsertedIdx] =
12693 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
12694 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12695 ConstantVector::get(Mask));
12698 // If this insertelement isn't used by some other insertelement, turn it
12699 // (and any insertelements it points to), into one big shuffle.
12700 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12701 std::vector<Constant*> Mask;
12703 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12704 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
12705 // We now have a shuffle of LHS, RHS, Mask.
12706 return new ShuffleVectorInst(LHS, RHS,
12707 ConstantVector::get(Mask));
12712 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12713 APInt UndefElts(VWidth, 0);
12714 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12715 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12722 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12723 Value *LHS = SVI.getOperand(0);
12724 Value *RHS = SVI.getOperand(1);
12725 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12727 bool MadeChange = false;
12729 // Undefined shuffle mask -> undefined value.
12730 if (isa<UndefValue>(SVI.getOperand(2)))
12731 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
12733 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12735 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12738 APInt UndefElts(VWidth, 0);
12739 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12740 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12741 LHS = SVI.getOperand(0);
12742 RHS = SVI.getOperand(1);
12746 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12747 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12748 if (LHS == RHS || isa<UndefValue>(LHS)) {
12749 if (isa<UndefValue>(LHS) && LHS == RHS) {
12750 // shuffle(undef,undef,mask) -> undef.
12751 return ReplaceInstUsesWith(SVI, LHS);
12754 // Remap any references to RHS to use LHS.
12755 std::vector<Constant*> Elts;
12756 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12757 if (Mask[i] >= 2*e)
12758 Elts.push_back(UndefValue::get(Type::Int32Ty));
12760 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12761 (Mask[i] < e && isa<UndefValue>(LHS))) {
12762 Mask[i] = 2*e; // Turn into undef.
12763 Elts.push_back(UndefValue::get(Type::Int32Ty));
12765 Mask[i] = Mask[i] % e; // Force to LHS.
12766 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
12770 SVI.setOperand(0, SVI.getOperand(1));
12771 SVI.setOperand(1, UndefValue::get(RHS->getType()));
12772 SVI.setOperand(2, ConstantVector::get(Elts));
12773 LHS = SVI.getOperand(0);
12774 RHS = SVI.getOperand(1);
12778 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12779 bool isLHSID = true, isRHSID = true;
12781 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12782 if (Mask[i] >= e*2) continue; // Ignore undef values.
12783 // Is this an identity shuffle of the LHS value?
12784 isLHSID &= (Mask[i] == i);
12786 // Is this an identity shuffle of the RHS value?
12787 isRHSID &= (Mask[i]-e == i);
12790 // Eliminate identity shuffles.
12791 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12792 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12794 // If the LHS is a shufflevector itself, see if we can combine it with this
12795 // one without producing an unusual shuffle. Here we are really conservative:
12796 // we are absolutely afraid of producing a shuffle mask not in the input
12797 // program, because the code gen may not be smart enough to turn a merged
12798 // shuffle into two specific shuffles: it may produce worse code. As such,
12799 // we only merge two shuffles if the result is one of the two input shuffle
12800 // masks. In this case, merging the shuffles just removes one instruction,
12801 // which we know is safe. This is good for things like turning:
12802 // (splat(splat)) -> splat.
12803 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12804 if (isa<UndefValue>(RHS)) {
12805 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12807 std::vector<unsigned> NewMask;
12808 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12809 if (Mask[i] >= 2*e)
12810 NewMask.push_back(2*e);
12812 NewMask.push_back(LHSMask[Mask[i]]);
12814 // If the result mask is equal to the src shuffle or this shuffle mask, do
12815 // the replacement.
12816 if (NewMask == LHSMask || NewMask == Mask) {
12817 unsigned LHSInNElts =
12818 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12819 std::vector<Constant*> Elts;
12820 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12821 if (NewMask[i] >= LHSInNElts*2) {
12822 Elts.push_back(UndefValue::get(Type::Int32Ty));
12824 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
12827 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12828 LHSSVI->getOperand(1),
12829 ConstantVector::get(Elts));
12834 return MadeChange ? &SVI : 0;
12840 /// TryToSinkInstruction - Try to move the specified instruction from its
12841 /// current block into the beginning of DestBlock, which can only happen if it's
12842 /// safe to move the instruction past all of the instructions between it and the
12843 /// end of its block.
12844 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12845 assert(I->hasOneUse() && "Invariants didn't hold!");
12847 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12848 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12851 // Do not sink alloca instructions out of the entry block.
12852 if (isa<AllocaInst>(I) && I->getParent() ==
12853 &DestBlock->getParent()->getEntryBlock())
12856 // We can only sink load instructions if there is nothing between the load and
12857 // the end of block that could change the value.
12858 if (I->mayReadFromMemory()) {
12859 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12861 if (Scan->mayWriteToMemory())
12865 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12867 CopyPrecedingStopPoint(I, InsertPos);
12868 I->moveBefore(InsertPos);
12874 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12875 /// all reachable code to the worklist.
12877 /// This has a couple of tricks to make the code faster and more powerful. In
12878 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12879 /// them to the worklist (this significantly speeds up instcombine on code where
12880 /// many instructions are dead or constant). Additionally, if we find a branch
12881 /// whose condition is a known constant, we only visit the reachable successors.
12883 static void AddReachableCodeToWorklist(BasicBlock *BB,
12884 SmallPtrSet<BasicBlock*, 64> &Visited,
12886 const TargetData *TD) {
12887 SmallVector<BasicBlock*, 256> Worklist;
12888 Worklist.push_back(BB);
12890 while (!Worklist.empty()) {
12891 BB = Worklist.back();
12892 Worklist.pop_back();
12894 // We have now visited this block! If we've already been here, ignore it.
12895 if (!Visited.insert(BB)) continue;
12897 DbgInfoIntrinsic *DBI_Prev = NULL;
12898 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12899 Instruction *Inst = BBI++;
12901 // DCE instruction if trivially dead.
12902 if (isInstructionTriviallyDead(Inst)) {
12904 DOUT << "IC: DCE: " << *Inst;
12905 Inst->eraseFromParent();
12909 // ConstantProp instruction if trivially constant.
12910 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12911 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12912 Inst->replaceAllUsesWith(C);
12914 Inst->eraseFromParent();
12918 // If there are two consecutive llvm.dbg.stoppoint calls then
12919 // it is likely that the optimizer deleted code in between these
12921 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12924 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12925 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12926 IC.RemoveFromWorkList(DBI_Prev);
12927 DBI_Prev->eraseFromParent();
12929 DBI_Prev = DBI_Next;
12934 IC.AddToWorkList(Inst);
12937 // Recursively visit successors. If this is a branch or switch on a
12938 // constant, only visit the reachable successor.
12939 TerminatorInst *TI = BB->getTerminator();
12940 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12941 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12942 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12943 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12944 Worklist.push_back(ReachableBB);
12947 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12948 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12949 // See if this is an explicit destination.
12950 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12951 if (SI->getCaseValue(i) == Cond) {
12952 BasicBlock *ReachableBB = SI->getSuccessor(i);
12953 Worklist.push_back(ReachableBB);
12957 // Otherwise it is the default destination.
12958 Worklist.push_back(SI->getSuccessor(0));
12963 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12964 Worklist.push_back(TI->getSuccessor(i));
12968 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12969 bool Changed = false;
12970 TD = getAnalysisIfAvailable<TargetData>();
12972 DEBUG(errs() << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12973 << F.getNameStr() << "\n");
12976 // Do a depth-first traversal of the function, populate the worklist with
12977 // the reachable instructions. Ignore blocks that are not reachable. Keep
12978 // track of which blocks we visit.
12979 SmallPtrSet<BasicBlock*, 64> Visited;
12980 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12982 // Do a quick scan over the function. If we find any blocks that are
12983 // unreachable, remove any instructions inside of them. This prevents
12984 // the instcombine code from having to deal with some bad special cases.
12985 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12986 if (!Visited.count(BB)) {
12987 Instruction *Term = BB->getTerminator();
12988 while (Term != BB->begin()) { // Remove instrs bottom-up
12989 BasicBlock::iterator I = Term; --I;
12991 DOUT << "IC: DCE: " << *I;
12992 // A debug intrinsic shouldn't force another iteration if we weren't
12993 // going to do one without it.
12994 if (!isa<DbgInfoIntrinsic>(I)) {
12998 if (!I->use_empty())
12999 I->replaceAllUsesWith(UndefValue::get(I->getType()));
13000 I->eraseFromParent();
13005 while (!Worklist.empty()) {
13006 Instruction *I = RemoveOneFromWorkList();
13007 if (I == 0) continue; // skip null values.
13009 // Check to see if we can DCE the instruction.
13010 if (isInstructionTriviallyDead(I)) {
13011 // Add operands to the worklist.
13012 if (I->getNumOperands() < 4)
13013 AddUsesToWorkList(*I);
13016 DOUT << "IC: DCE: " << *I;
13018 I->eraseFromParent();
13019 RemoveFromWorkList(I);
13024 // Instruction isn't dead, see if we can constant propagate it.
13025 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
13026 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
13028 // Add operands to the worklist.
13029 AddUsesToWorkList(*I);
13030 ReplaceInstUsesWith(*I, C);
13033 I->eraseFromParent();
13034 RemoveFromWorkList(I);
13040 // See if we can constant fold its operands.
13041 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
13042 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
13043 if (Constant *NewC = ConstantFoldConstantExpression(CE,
13044 F.getContext(), TD))
13051 // See if we can trivially sink this instruction to a successor basic block.
13052 if (I->hasOneUse()) {
13053 BasicBlock *BB = I->getParent();
13054 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
13055 if (UserParent != BB) {
13056 bool UserIsSuccessor = false;
13057 // See if the user is one of our successors.
13058 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13059 if (*SI == UserParent) {
13060 UserIsSuccessor = true;
13064 // If the user is one of our immediate successors, and if that successor
13065 // only has us as a predecessors (we'd have to split the critical edge
13066 // otherwise), we can keep going.
13067 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
13068 next(pred_begin(UserParent)) == pred_end(UserParent))
13069 // Okay, the CFG is simple enough, try to sink this instruction.
13070 Changed |= TryToSinkInstruction(I, UserParent);
13074 // Now that we have an instruction, try combining it to simplify it...
13078 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
13079 if (Instruction *Result = visit(*I)) {
13081 // Should we replace the old instruction with a new one?
13083 DOUT << "IC: Old = " << *I
13084 << " New = " << *Result;
13086 // Everything uses the new instruction now.
13087 I->replaceAllUsesWith(Result);
13089 // Push the new instruction and any users onto the worklist.
13090 AddToWorkList(Result);
13091 AddUsersToWorkList(*Result);
13093 // Move the name to the new instruction first.
13094 Result->takeName(I);
13096 // Insert the new instruction into the basic block...
13097 BasicBlock *InstParent = I->getParent();
13098 BasicBlock::iterator InsertPos = I;
13100 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13101 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13104 InstParent->getInstList().insert(InsertPos, Result);
13106 // Make sure that we reprocess all operands now that we reduced their
13108 AddUsesToWorkList(*I);
13110 // Instructions can end up on the worklist more than once. Make sure
13111 // we do not process an instruction that has been deleted.
13112 RemoveFromWorkList(I);
13114 // Erase the old instruction.
13115 InstParent->getInstList().erase(I);
13118 DOUT << "IC: Mod = " << OrigI
13119 << " New = " << *I;
13122 // If the instruction was modified, it's possible that it is now dead.
13123 // if so, remove it.
13124 if (isInstructionTriviallyDead(I)) {
13125 // Make sure we process all operands now that we are reducing their
13127 AddUsesToWorkList(*I);
13129 // Instructions may end up in the worklist more than once. Erase all
13130 // occurrences of this instruction.
13131 RemoveFromWorkList(I);
13132 I->eraseFromParent();
13135 AddUsersToWorkList(*I);
13142 assert(WorklistMap.empty() && "Worklist empty, but map not?");
13144 // Do an explicit clear, this shrinks the map if needed.
13145 WorklistMap.clear();
13150 bool InstCombiner::runOnFunction(Function &F) {
13151 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13152 Context = &F.getContext();
13154 bool EverMadeChange = false;
13156 // Iterate while there is work to do.
13157 unsigned Iteration = 0;
13158 while (DoOneIteration(F, Iteration++))
13159 EverMadeChange = true;
13160 return EverMadeChange;
13163 FunctionPass *llvm::createInstructionCombiningPass() {
13164 return new InstCombiner();