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/ADT/DenseMap.h"
59 #include "llvm/ADT/SmallVector.h"
60 #include "llvm/ADT/SmallPtrSet.h"
61 #include "llvm/ADT/Statistic.h"
62 #include "llvm/ADT/STLExtras.h"
67 using namespace llvm::PatternMatch;
69 STATISTIC(NumCombined , "Number of insts combined");
70 STATISTIC(NumConstProp, "Number of constant folds");
71 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
72 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
73 STATISTIC(NumSunkInst , "Number of instructions sunk");
76 class VISIBILITY_HIDDEN InstCombiner
77 : public FunctionPass,
78 public InstVisitor<InstCombiner, Instruction*> {
79 // Worklist of all of the instructions that need to be simplified.
80 SmallVector<Instruction*, 256> Worklist;
81 DenseMap<Instruction*, unsigned> WorklistMap;
83 bool MustPreserveLCSSA;
85 static char ID; // Pass identification, replacement for typeid
86 InstCombiner() : FunctionPass(&ID) {}
88 LLVMContext *getContext() { return Context; }
90 /// AddToWorkList - Add the specified instruction to the worklist if it
91 /// isn't already in it.
92 void AddToWorkList(Instruction *I) {
93 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
94 Worklist.push_back(I);
97 // RemoveFromWorkList - remove I from the worklist if it exists.
98 void RemoveFromWorkList(Instruction *I) {
99 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
100 if (It == WorklistMap.end()) return; // Not in worklist.
102 // Don't bother moving everything down, just null out the slot.
103 Worklist[It->second] = 0;
105 WorklistMap.erase(It);
108 Instruction *RemoveOneFromWorkList() {
109 Instruction *I = Worklist.back();
111 WorklistMap.erase(I);
116 /// AddUsersToWorkList - When an instruction is simplified, add all users of
117 /// the instruction to the work lists because they might get more simplified
120 void AddUsersToWorkList(Value &I) {
121 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
123 AddToWorkList(cast<Instruction>(*UI));
126 /// AddUsesToWorkList - When an instruction is simplified, add operands to
127 /// the work lists because they might get more simplified now.
129 void AddUsesToWorkList(Instruction &I) {
130 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
131 if (Instruction *Op = dyn_cast<Instruction>(*i))
135 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
136 /// dead. Add all of its operands to the worklist, turning them into
137 /// undef's to reduce the number of uses of those instructions.
139 /// Return the specified operand before it is turned into an undef.
141 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
142 Value *R = I.getOperand(op);
144 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
145 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
147 // Set the operand to undef to drop the use.
148 *i = Context->getUndef(Op->getType());
155 virtual bool runOnFunction(Function &F);
157 bool DoOneIteration(Function &F, unsigned ItNum);
159 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
160 AU.addRequired<TargetData>();
161 AU.addPreservedID(LCSSAID);
162 AU.setPreservesCFG();
165 TargetData &getTargetData() const { return *TD; }
167 // Visitation implementation - Implement instruction combining for different
168 // instruction types. The semantics are as follows:
170 // null - No change was made
171 // I - Change was made, I is still valid, I may be dead though
172 // otherwise - Change was made, replace I with returned instruction
174 Instruction *visitAdd(BinaryOperator &I);
175 Instruction *visitFAdd(BinaryOperator &I);
176 Instruction *visitSub(BinaryOperator &I);
177 Instruction *visitFSub(BinaryOperator &I);
178 Instruction *visitMul(BinaryOperator &I);
179 Instruction *visitFMul(BinaryOperator &I);
180 Instruction *visitURem(BinaryOperator &I);
181 Instruction *visitSRem(BinaryOperator &I);
182 Instruction *visitFRem(BinaryOperator &I);
183 bool SimplifyDivRemOfSelect(BinaryOperator &I);
184 Instruction *commonRemTransforms(BinaryOperator &I);
185 Instruction *commonIRemTransforms(BinaryOperator &I);
186 Instruction *commonDivTransforms(BinaryOperator &I);
187 Instruction *commonIDivTransforms(BinaryOperator &I);
188 Instruction *visitUDiv(BinaryOperator &I);
189 Instruction *visitSDiv(BinaryOperator &I);
190 Instruction *visitFDiv(BinaryOperator &I);
191 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
192 Instruction *visitAnd(BinaryOperator &I);
193 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
194 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
195 Value *A, Value *B, Value *C);
196 Instruction *visitOr (BinaryOperator &I);
197 Instruction *visitXor(BinaryOperator &I);
198 Instruction *visitShl(BinaryOperator &I);
199 Instruction *visitAShr(BinaryOperator &I);
200 Instruction *visitLShr(BinaryOperator &I);
201 Instruction *commonShiftTransforms(BinaryOperator &I);
202 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
204 Instruction *visitFCmpInst(FCmpInst &I);
205 Instruction *visitICmpInst(ICmpInst &I);
206 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
207 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
210 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
211 ConstantInt *DivRHS);
213 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
214 ICmpInst::Predicate Cond, Instruction &I);
215 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
217 Instruction *commonCastTransforms(CastInst &CI);
218 Instruction *commonIntCastTransforms(CastInst &CI);
219 Instruction *commonPointerCastTransforms(CastInst &CI);
220 Instruction *visitTrunc(TruncInst &CI);
221 Instruction *visitZExt(ZExtInst &CI);
222 Instruction *visitSExt(SExtInst &CI);
223 Instruction *visitFPTrunc(FPTruncInst &CI);
224 Instruction *visitFPExt(CastInst &CI);
225 Instruction *visitFPToUI(FPToUIInst &FI);
226 Instruction *visitFPToSI(FPToSIInst &FI);
227 Instruction *visitUIToFP(CastInst &CI);
228 Instruction *visitSIToFP(CastInst &CI);
229 Instruction *visitPtrToInt(PtrToIntInst &CI);
230 Instruction *visitIntToPtr(IntToPtrInst &CI);
231 Instruction *visitBitCast(BitCastInst &CI);
232 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
234 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
235 Instruction *visitSelectInst(SelectInst &SI);
236 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
237 Instruction *visitCallInst(CallInst &CI);
238 Instruction *visitInvokeInst(InvokeInst &II);
239 Instruction *visitPHINode(PHINode &PN);
240 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
241 Instruction *visitAllocationInst(AllocationInst &AI);
242 Instruction *visitFreeInst(FreeInst &FI);
243 Instruction *visitLoadInst(LoadInst &LI);
244 Instruction *visitStoreInst(StoreInst &SI);
245 Instruction *visitBranchInst(BranchInst &BI);
246 Instruction *visitSwitchInst(SwitchInst &SI);
247 Instruction *visitInsertElementInst(InsertElementInst &IE);
248 Instruction *visitExtractElementInst(ExtractElementInst &EI);
249 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
250 Instruction *visitExtractValueInst(ExtractValueInst &EV);
252 // visitInstruction - Specify what to return for unhandled instructions...
253 Instruction *visitInstruction(Instruction &I) { return 0; }
256 Instruction *visitCallSite(CallSite CS);
257 bool transformConstExprCastCall(CallSite CS);
258 Instruction *transformCallThroughTrampoline(CallSite CS);
259 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
260 bool DoXform = true);
261 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
262 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
266 // InsertNewInstBefore - insert an instruction New before instruction Old
267 // in the program. Add the new instruction to the worklist.
269 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
270 assert(New && New->getParent() == 0 &&
271 "New instruction already inserted into a basic block!");
272 BasicBlock *BB = Old.getParent();
273 BB->getInstList().insert(&Old, New); // Insert inst
278 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
279 /// This also adds the cast to the worklist. Finally, this returns the
281 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
283 if (V->getType() == Ty) return V;
285 if (Constant *CV = dyn_cast<Constant>(V))
286 return Context->getConstantExprCast(opc, CV, Ty);
288 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
293 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
294 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
298 // ReplaceInstUsesWith - This method is to be used when an instruction is
299 // found to be dead, replacable with another preexisting expression. Here
300 // we add all uses of I to the worklist, replace all uses of I with the new
301 // value, then return I, so that the inst combiner will know that I was
304 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
305 AddUsersToWorkList(I); // Add all modified instrs to worklist
307 I.replaceAllUsesWith(V);
310 // If we are replacing the instruction with itself, this must be in a
311 // segment of unreachable code, so just clobber the instruction.
312 I.replaceAllUsesWith(Context->getUndef(I.getType()));
317 // EraseInstFromFunction - When dealing with an instruction that has side
318 // effects or produces a void value, we can't rely on DCE to delete the
319 // instruction. Instead, visit methods should return the value returned by
321 Instruction *EraseInstFromFunction(Instruction &I) {
322 assert(I.use_empty() && "Cannot erase instruction that is used!");
323 AddUsesToWorkList(I);
324 RemoveFromWorkList(&I);
326 return 0; // Don't do anything with FI
329 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
330 APInt &KnownOne, unsigned Depth = 0) const {
331 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
334 bool MaskedValueIsZero(Value *V, const APInt &Mask,
335 unsigned Depth = 0) const {
336 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
338 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
339 return llvm::ComputeNumSignBits(Op, TD, Depth);
344 /// SimplifyCommutative - This performs a few simplifications for
345 /// commutative operators.
346 bool SimplifyCommutative(BinaryOperator &I);
348 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
349 /// most-complex to least-complex order.
350 bool SimplifyCompare(CmpInst &I);
352 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
353 /// based on the demanded bits.
354 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
355 APInt& KnownZero, APInt& KnownOne,
357 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
358 APInt& KnownZero, APInt& KnownOne,
361 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
362 /// SimplifyDemandedBits knows about. See if the instruction has any
363 /// properties that allow us to simplify its operands.
364 bool SimplifyDemandedInstructionBits(Instruction &Inst);
366 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
367 APInt& UndefElts, unsigned Depth = 0);
369 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
370 // PHI node as operand #0, see if we can fold the instruction into the PHI
371 // (which is only possible if all operands to the PHI are constants).
372 Instruction *FoldOpIntoPhi(Instruction &I);
374 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
375 // operator and they all are only used by the PHI, PHI together their
376 // inputs, and do the operation once, to the result of the PHI.
377 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
378 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
379 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
382 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
383 ConstantInt *AndRHS, BinaryOperator &TheAnd);
385 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
386 bool isSub, Instruction &I);
387 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
388 bool isSigned, bool Inside, Instruction &IB);
389 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
390 Instruction *MatchBSwap(BinaryOperator &I);
391 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
392 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
393 Instruction *SimplifyMemSet(MemSetInst *MI);
396 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
398 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
399 unsigned CastOpc, int &NumCastsRemoved);
400 unsigned GetOrEnforceKnownAlignment(Value *V,
401 unsigned PrefAlign = 0);
406 char InstCombiner::ID = 0;
407 static RegisterPass<InstCombiner>
408 X("instcombine", "Combine redundant instructions");
410 // getComplexity: Assign a complexity or rank value to LLVM Values...
411 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
412 static unsigned getComplexity(LLVMContext *Context, Value *V) {
413 if (isa<Instruction>(V)) {
414 if (BinaryOperator::isNeg(V) ||
415 BinaryOperator::isFNeg(V) ||
416 BinaryOperator::isNot(V))
420 if (isa<Argument>(V)) return 3;
421 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
424 // isOnlyUse - Return true if this instruction will be deleted if we stop using
426 static bool isOnlyUse(Value *V) {
427 return V->hasOneUse() || isa<Constant>(V);
430 // getPromotedType - Return the specified type promoted as it would be to pass
431 // though a va_arg area...
432 static const Type *getPromotedType(const Type *Ty) {
433 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
434 if (ITy->getBitWidth() < 32)
435 return Type::Int32Ty;
440 /// getBitCastOperand - If the specified operand is a CastInst, a constant
441 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
442 /// operand value, otherwise return null.
443 static Value *getBitCastOperand(Value *V) {
444 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
446 return I->getOperand(0);
447 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
448 // GetElementPtrInst?
449 if (GEP->hasAllZeroIndices())
450 return GEP->getOperand(0);
451 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
452 if (CE->getOpcode() == Instruction::BitCast)
453 // BitCast ConstantExp?
454 return CE->getOperand(0);
455 else if (CE->getOpcode() == Instruction::GetElementPtr) {
456 // GetElementPtr ConstantExp?
457 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
459 ConstantInt *CI = dyn_cast<ConstantInt>(I);
460 if (!CI || !CI->isZero())
461 // Any non-zero indices? Not cast-like.
464 // All-zero indices? This is just like casting.
465 return CE->getOperand(0);
471 /// This function is a wrapper around CastInst::isEliminableCastPair. It
472 /// simply extracts arguments and returns what that function returns.
473 static Instruction::CastOps
474 isEliminableCastPair(
475 const CastInst *CI, ///< The first cast instruction
476 unsigned opcode, ///< The opcode of the second cast instruction
477 const Type *DstTy, ///< The target type for the second cast instruction
478 TargetData *TD ///< The target data for pointer size
481 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
482 const Type *MidTy = CI->getType(); // B from above
484 // Get the opcodes of the two Cast instructions
485 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
486 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
488 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
489 DstTy, TD->getIntPtrType());
491 // We don't want to form an inttoptr or ptrtoint that converts to an integer
492 // type that differs from the pointer size.
493 if ((Res == Instruction::IntToPtr && SrcTy != TD->getIntPtrType()) ||
494 (Res == Instruction::PtrToInt && DstTy != TD->getIntPtrType()))
497 return Instruction::CastOps(Res);
500 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
501 /// in any code being generated. It does not require codegen if V is simple
502 /// enough or if the cast can be folded into other casts.
503 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
504 const Type *Ty, TargetData *TD) {
505 if (V->getType() == Ty || isa<Constant>(V)) return false;
507 // If this is another cast that can be eliminated, it isn't codegen either.
508 if (const CastInst *CI = dyn_cast<CastInst>(V))
509 if (isEliminableCastPair(CI, opcode, Ty, TD))
514 // SimplifyCommutative - This performs a few simplifications for commutative
517 // 1. Order operands such that they are listed from right (least complex) to
518 // left (most complex). This puts constants before unary operators before
521 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
522 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
524 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
525 bool Changed = false;
526 if (getComplexity(Context, I.getOperand(0)) <
527 getComplexity(Context, I.getOperand(1)))
528 Changed = !I.swapOperands();
530 if (!I.isAssociative()) return Changed;
531 Instruction::BinaryOps Opcode = I.getOpcode();
532 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
533 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
534 if (isa<Constant>(I.getOperand(1))) {
535 Constant *Folded = Context->getConstantExpr(I.getOpcode(),
536 cast<Constant>(I.getOperand(1)),
537 cast<Constant>(Op->getOperand(1)));
538 I.setOperand(0, Op->getOperand(0));
539 I.setOperand(1, Folded);
541 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
542 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
543 isOnlyUse(Op) && isOnlyUse(Op1)) {
544 Constant *C1 = cast<Constant>(Op->getOperand(1));
545 Constant *C2 = cast<Constant>(Op1->getOperand(1));
547 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
548 Constant *Folded = Context->getConstantExpr(I.getOpcode(), C1, C2);
549 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
553 I.setOperand(0, New);
554 I.setOperand(1, Folded);
561 /// SimplifyCompare - For a CmpInst this function just orders the operands
562 /// so that theyare listed from right (least complex) to left (most complex).
563 /// This puts constants before unary operators before binary operators.
564 bool InstCombiner::SimplifyCompare(CmpInst &I) {
565 if (getComplexity(Context, I.getOperand(0)) >=
566 getComplexity(Context, I.getOperand(1)))
569 // Compare instructions are not associative so there's nothing else we can do.
573 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
574 // if the LHS is a constant zero (which is the 'negate' form).
576 static inline Value *dyn_castNegVal(Value *V, LLVMContext *Context) {
577 if (BinaryOperator::isNeg(V))
578 return BinaryOperator::getNegArgument(V);
580 // Constants can be considered to be negated values if they can be folded.
581 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
582 return Context->getConstantExprNeg(C);
584 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
585 if (C->getType()->getElementType()->isInteger())
586 return Context->getConstantExprNeg(C);
591 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
592 // instruction if the LHS is a constant negative zero (which is the 'negate'
595 static inline Value *dyn_castFNegVal(Value *V, LLVMContext *Context) {
596 if (BinaryOperator::isFNeg(V))
597 return BinaryOperator::getFNegArgument(V);
599 // Constants can be considered to be negated values if they can be folded.
600 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
601 return Context->getConstantExprFNeg(C);
603 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
604 if (C->getType()->getElementType()->isFloatingPoint())
605 return Context->getConstantExprFNeg(C);
610 static inline Value *dyn_castNotVal(Value *V, LLVMContext *Context) {
611 if (BinaryOperator::isNot(V))
612 return BinaryOperator::getNotArgument(V);
614 // Constants can be considered to be not'ed values...
615 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
616 return Context->getConstantInt(~C->getValue());
620 // dyn_castFoldableMul - If this value is a multiply that can be folded into
621 // other computations (because it has a constant operand), return the
622 // non-constant operand of the multiply, and set CST to point to the multiplier.
623 // Otherwise, return null.
625 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST,
626 LLVMContext *Context) {
627 if (V->hasOneUse() && V->getType()->isInteger())
628 if (Instruction *I = dyn_cast<Instruction>(V)) {
629 if (I->getOpcode() == Instruction::Mul)
630 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
631 return I->getOperand(0);
632 if (I->getOpcode() == Instruction::Shl)
633 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
634 // The multiplier is really 1 << CST.
635 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
636 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
637 CST = Context->getConstantInt(APInt(BitWidth, 1).shl(CSTVal));
638 return I->getOperand(0);
644 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
645 /// expression, return it.
646 static User *dyn_castGetElementPtr(Value *V) {
647 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
648 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
649 if (CE->getOpcode() == Instruction::GetElementPtr)
650 return cast<User>(V);
654 /// AddOne - Add one to a ConstantInt
655 static Constant *AddOne(Constant *C, LLVMContext *Context) {
656 return Context->getConstantExprAdd(C,
657 Context->getConstantInt(C->getType(), 1));
659 /// SubOne - Subtract one from a ConstantInt
660 static Constant *SubOne(ConstantInt *C, LLVMContext *Context) {
661 return Context->getConstantExprSub(C,
662 Context->getConstantInt(C->getType(), 1));
664 /// MultiplyOverflows - True if the multiply can not be expressed in an int
666 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign,
667 LLVMContext *Context) {
668 uint32_t W = C1->getBitWidth();
669 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
678 APInt MulExt = LHSExt * RHSExt;
681 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
682 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
683 return MulExt.slt(Min) || MulExt.sgt(Max);
685 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
689 /// ShrinkDemandedConstant - Check to see if the specified operand of the
690 /// specified instruction is a constant integer. If so, check to see if there
691 /// are any bits set in the constant that are not demanded. If so, shrink the
692 /// constant and return true.
693 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
694 APInt Demanded, LLVMContext *Context) {
695 assert(I && "No instruction?");
696 assert(OpNo < I->getNumOperands() && "Operand index too large");
698 // If the operand is not a constant integer, nothing to do.
699 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
700 if (!OpC) return false;
702 // If there are no bits set that aren't demanded, nothing to do.
703 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
704 if ((~Demanded & OpC->getValue()) == 0)
707 // This instruction is producing bits that are not demanded. Shrink the RHS.
708 Demanded &= OpC->getValue();
709 I->setOperand(OpNo, Context->getConstantInt(Demanded));
713 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
714 // set of known zero and one bits, compute the maximum and minimum values that
715 // could have the specified known zero and known one bits, returning them in
717 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
718 const APInt& KnownOne,
719 APInt& Min, APInt& Max) {
720 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
721 KnownZero.getBitWidth() == Min.getBitWidth() &&
722 KnownZero.getBitWidth() == Max.getBitWidth() &&
723 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
724 APInt UnknownBits = ~(KnownZero|KnownOne);
726 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
727 // bit if it is unknown.
729 Max = KnownOne|UnknownBits;
731 if (UnknownBits.isNegative()) { // Sign bit is unknown
732 Min.set(Min.getBitWidth()-1);
733 Max.clear(Max.getBitWidth()-1);
737 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
738 // a set of known zero and one bits, compute the maximum and minimum values that
739 // could have the specified known zero and known one bits, returning them in
741 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
742 const APInt &KnownOne,
743 APInt &Min, APInt &Max) {
744 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
745 KnownZero.getBitWidth() == Min.getBitWidth() &&
746 KnownZero.getBitWidth() == Max.getBitWidth() &&
747 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
748 APInt UnknownBits = ~(KnownZero|KnownOne);
750 // The minimum value is when the unknown bits are all zeros.
752 // The maximum value is when the unknown bits are all ones.
753 Max = KnownOne|UnknownBits;
756 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
757 /// SimplifyDemandedBits knows about. See if the instruction has any
758 /// properties that allow us to simplify its operands.
759 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
760 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
761 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
762 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
764 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
765 KnownZero, KnownOne, 0);
766 if (V == 0) return false;
767 if (V == &Inst) return true;
768 ReplaceInstUsesWith(Inst, V);
772 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
773 /// specified instruction operand if possible, updating it in place. It returns
774 /// true if it made any change and false otherwise.
775 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
776 APInt &KnownZero, APInt &KnownOne,
778 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
779 KnownZero, KnownOne, Depth);
780 if (NewVal == 0) return false;
786 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
787 /// value based on the demanded bits. When this function is called, it is known
788 /// that only the bits set in DemandedMask of the result of V are ever used
789 /// downstream. Consequently, depending on the mask and V, it may be possible
790 /// to replace V with a constant or one of its operands. In such cases, this
791 /// function does the replacement and returns true. In all other cases, it
792 /// returns false after analyzing the expression and setting KnownOne and known
793 /// to be one in the expression. KnownZero contains all the bits that are known
794 /// to be zero in the expression. These are provided to potentially allow the
795 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
796 /// the expression. KnownOne and KnownZero always follow the invariant that
797 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
798 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
799 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
800 /// and KnownOne must all be the same.
802 /// This returns null if it did not change anything and it permits no
803 /// simplification. This returns V itself if it did some simplification of V's
804 /// operands based on the information about what bits are demanded. This returns
805 /// some other non-null value if it found out that V is equal to another value
806 /// in the context where the specified bits are demanded, but not for all users.
807 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
808 APInt &KnownZero, APInt &KnownOne,
810 assert(V != 0 && "Null pointer of Value???");
811 assert(Depth <= 6 && "Limit Search Depth");
812 uint32_t BitWidth = DemandedMask.getBitWidth();
813 const Type *VTy = V->getType();
814 assert((TD || !isa<PointerType>(VTy)) &&
815 "SimplifyDemandedBits needs to know bit widths!");
816 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
817 (!VTy->isIntOrIntVector() ||
818 VTy->getScalarSizeInBits() == BitWidth) &&
819 KnownZero.getBitWidth() == BitWidth &&
820 KnownOne.getBitWidth() == BitWidth &&
821 "Value *V, DemandedMask, KnownZero and KnownOne "
822 "must have same BitWidth");
823 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
824 // We know all of the bits for a constant!
825 KnownOne = CI->getValue() & DemandedMask;
826 KnownZero = ~KnownOne & DemandedMask;
829 if (isa<ConstantPointerNull>(V)) {
830 // We know all of the bits for a constant!
832 KnownZero = DemandedMask;
838 if (DemandedMask == 0) { // Not demanding any bits from V.
839 if (isa<UndefValue>(V))
841 return Context->getUndef(VTy);
844 if (Depth == 6) // Limit search depth.
847 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
848 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
850 Instruction *I = dyn_cast<Instruction>(V);
852 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
853 return 0; // Only analyze instructions.
856 // If there are multiple uses of this value and we aren't at the root, then
857 // we can't do any simplifications of the operands, because DemandedMask
858 // only reflects the bits demanded by *one* of the users.
859 if (Depth != 0 && !I->hasOneUse()) {
860 // Despite the fact that we can't simplify this instruction in all User's
861 // context, we can at least compute the knownzero/knownone bits, and we can
862 // do simplifications that apply to *just* the one user if we know that
863 // this instruction has a simpler value in that context.
864 if (I->getOpcode() == Instruction::And) {
865 // If either the LHS or the RHS are Zero, the result is zero.
866 ComputeMaskedBits(I->getOperand(1), DemandedMask,
867 RHSKnownZero, RHSKnownOne, Depth+1);
868 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
869 LHSKnownZero, LHSKnownOne, Depth+1);
871 // If all of the demanded bits are known 1 on one side, return the other.
872 // These bits cannot contribute to the result of the 'and' in this
874 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
875 (DemandedMask & ~LHSKnownZero))
876 return I->getOperand(0);
877 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
878 (DemandedMask & ~RHSKnownZero))
879 return I->getOperand(1);
881 // If all of the demanded bits in the inputs are known zeros, return zero.
882 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
883 return Context->getNullValue(VTy);
885 } else if (I->getOpcode() == Instruction::Or) {
886 // We can simplify (X|Y) -> X or Y in the user's context if we know that
887 // only bits from X or Y are demanded.
889 // If either the LHS or the RHS are One, the result is One.
890 ComputeMaskedBits(I->getOperand(1), DemandedMask,
891 RHSKnownZero, RHSKnownOne, Depth+1);
892 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
893 LHSKnownZero, LHSKnownOne, Depth+1);
895 // If all of the demanded bits are known zero on one side, return the
896 // other. These bits cannot contribute to the result of the 'or' in this
898 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
899 (DemandedMask & ~LHSKnownOne))
900 return I->getOperand(0);
901 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
902 (DemandedMask & ~RHSKnownOne))
903 return I->getOperand(1);
905 // If all of the potentially set bits on one side are known to be set on
906 // the other side, just use the 'other' side.
907 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
908 (DemandedMask & (~RHSKnownZero)))
909 return I->getOperand(0);
910 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
911 (DemandedMask & (~LHSKnownZero)))
912 return I->getOperand(1);
915 // Compute the KnownZero/KnownOne bits to simplify things downstream.
916 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
920 // If this is the root being simplified, allow it to have multiple uses,
921 // just set the DemandedMask to all bits so that we can try to simplify the
922 // operands. This allows visitTruncInst (for example) to simplify the
923 // operand of a trunc without duplicating all the logic below.
924 if (Depth == 0 && !V->hasOneUse())
925 DemandedMask = APInt::getAllOnesValue(BitWidth);
927 switch (I->getOpcode()) {
929 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
931 case Instruction::And:
932 // If either the LHS or the RHS are Zero, the result is zero.
933 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
934 RHSKnownZero, RHSKnownOne, Depth+1) ||
935 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
936 LHSKnownZero, LHSKnownOne, Depth+1))
938 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
939 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
941 // If all of the demanded bits are known 1 on one side, return the other.
942 // These bits cannot contribute to the result of the 'and'.
943 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
944 (DemandedMask & ~LHSKnownZero))
945 return I->getOperand(0);
946 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
947 (DemandedMask & ~RHSKnownZero))
948 return I->getOperand(1);
950 // If all of the demanded bits in the inputs are known zeros, return zero.
951 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
952 return Context->getNullValue(VTy);
954 // If the RHS is a constant, see if we can simplify it.
955 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero, Context))
958 // Output known-1 bits are only known if set in both the LHS & RHS.
959 RHSKnownOne &= LHSKnownOne;
960 // Output known-0 are known to be clear if zero in either the LHS | RHS.
961 RHSKnownZero |= LHSKnownZero;
963 case Instruction::Or:
964 // If either the LHS or the RHS are One, the result is One.
965 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
966 RHSKnownZero, RHSKnownOne, Depth+1) ||
967 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
968 LHSKnownZero, LHSKnownOne, Depth+1))
970 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
971 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
973 // If all of the demanded bits are known zero on one side, return the other.
974 // These bits cannot contribute to the result of the 'or'.
975 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
976 (DemandedMask & ~LHSKnownOne))
977 return I->getOperand(0);
978 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
979 (DemandedMask & ~RHSKnownOne))
980 return I->getOperand(1);
982 // If all of the potentially set bits on one side are known to be set on
983 // the other side, just use the 'other' side.
984 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
985 (DemandedMask & (~RHSKnownZero)))
986 return I->getOperand(0);
987 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
988 (DemandedMask & (~LHSKnownZero)))
989 return I->getOperand(1);
991 // If the RHS is a constant, see if we can simplify it.
992 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context))
995 // Output known-0 bits are only known if clear in both the LHS & RHS.
996 RHSKnownZero &= LHSKnownZero;
997 // Output known-1 are known to be set if set in either the LHS | RHS.
998 RHSKnownOne |= LHSKnownOne;
1000 case Instruction::Xor: {
1001 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1002 RHSKnownZero, RHSKnownOne, Depth+1) ||
1003 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1004 LHSKnownZero, LHSKnownOne, Depth+1))
1006 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1007 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1009 // If all of the demanded bits are known zero on one side, return the other.
1010 // These bits cannot contribute to the result of the 'xor'.
1011 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1012 return I->getOperand(0);
1013 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1014 return I->getOperand(1);
1016 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1017 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1018 (RHSKnownOne & LHSKnownOne);
1019 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1020 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1021 (RHSKnownOne & LHSKnownZero);
1023 // If all of the demanded bits are known to be zero on one side or the
1024 // other, turn this into an *inclusive* or.
1025 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1026 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1028 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1030 return InsertNewInstBefore(Or, *I);
1033 // If all of the demanded bits on one side are known, and all of the set
1034 // bits on that side are also known to be set on the other side, turn this
1035 // into an AND, as we know the bits will be cleared.
1036 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1037 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1039 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1040 Constant *AndC = Context->getConstantInt(~RHSKnownOne & DemandedMask);
1042 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1043 return InsertNewInstBefore(And, *I);
1047 // If the RHS is a constant, see if we can simplify it.
1048 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1049 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context))
1052 RHSKnownZero = KnownZeroOut;
1053 RHSKnownOne = KnownOneOut;
1056 case Instruction::Select:
1057 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1058 RHSKnownZero, RHSKnownOne, Depth+1) ||
1059 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1060 LHSKnownZero, LHSKnownOne, Depth+1))
1062 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1063 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1065 // If the operands are constants, see if we can simplify them.
1066 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context) ||
1067 ShrinkDemandedConstant(I, 2, DemandedMask, Context))
1070 // Only known if known in both the LHS and RHS.
1071 RHSKnownOne &= LHSKnownOne;
1072 RHSKnownZero &= LHSKnownZero;
1074 case Instruction::Trunc: {
1075 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1076 DemandedMask.zext(truncBf);
1077 RHSKnownZero.zext(truncBf);
1078 RHSKnownOne.zext(truncBf);
1079 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1080 RHSKnownZero, RHSKnownOne, Depth+1))
1082 DemandedMask.trunc(BitWidth);
1083 RHSKnownZero.trunc(BitWidth);
1084 RHSKnownOne.trunc(BitWidth);
1085 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1088 case Instruction::BitCast:
1089 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1090 return false; // vector->int or fp->int?
1092 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1093 if (const VectorType *SrcVTy =
1094 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1095 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1096 // Don't touch a bitcast between vectors of different element counts.
1099 // Don't touch a scalar-to-vector bitcast.
1101 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1102 // Don't touch a vector-to-scalar bitcast.
1105 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1106 RHSKnownZero, RHSKnownOne, Depth+1))
1108 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1110 case Instruction::ZExt: {
1111 // Compute the bits in the result that are not present in the input.
1112 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1114 DemandedMask.trunc(SrcBitWidth);
1115 RHSKnownZero.trunc(SrcBitWidth);
1116 RHSKnownOne.trunc(SrcBitWidth);
1117 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1118 RHSKnownZero, RHSKnownOne, Depth+1))
1120 DemandedMask.zext(BitWidth);
1121 RHSKnownZero.zext(BitWidth);
1122 RHSKnownOne.zext(BitWidth);
1123 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1124 // The top bits are known to be zero.
1125 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1128 case Instruction::SExt: {
1129 // Compute the bits in the result that are not present in the input.
1130 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1132 APInt InputDemandedBits = DemandedMask &
1133 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1135 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1136 // If any of the sign extended bits are demanded, we know that the sign
1138 if ((NewBits & DemandedMask) != 0)
1139 InputDemandedBits.set(SrcBitWidth-1);
1141 InputDemandedBits.trunc(SrcBitWidth);
1142 RHSKnownZero.trunc(SrcBitWidth);
1143 RHSKnownOne.trunc(SrcBitWidth);
1144 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1145 RHSKnownZero, RHSKnownOne, Depth+1))
1147 InputDemandedBits.zext(BitWidth);
1148 RHSKnownZero.zext(BitWidth);
1149 RHSKnownOne.zext(BitWidth);
1150 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1152 // If the sign bit of the input is known set or clear, then we know the
1153 // top bits of the result.
1155 // If the input sign bit is known zero, or if the NewBits are not demanded
1156 // convert this into a zero extension.
1157 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1158 // Convert to ZExt cast
1159 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1160 return InsertNewInstBefore(NewCast, *I);
1161 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1162 RHSKnownOne |= NewBits;
1166 case Instruction::Add: {
1167 // Figure out what the input bits are. If the top bits of the and result
1168 // are not demanded, then the add doesn't demand them from its input
1170 unsigned NLZ = DemandedMask.countLeadingZeros();
1172 // If there is a constant on the RHS, there are a variety of xformations
1174 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1175 // If null, this should be simplified elsewhere. Some of the xforms here
1176 // won't work if the RHS is zero.
1180 // If the top bit of the output is demanded, demand everything from the
1181 // input. Otherwise, we demand all the input bits except NLZ top bits.
1182 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1184 // Find information about known zero/one bits in the input.
1185 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1186 LHSKnownZero, LHSKnownOne, Depth+1))
1189 // If the RHS of the add has bits set that can't affect the input, reduce
1191 if (ShrinkDemandedConstant(I, 1, InDemandedBits, Context))
1194 // Avoid excess work.
1195 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1198 // Turn it into OR if input bits are zero.
1199 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1201 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1203 return InsertNewInstBefore(Or, *I);
1206 // We can say something about the output known-zero and known-one bits,
1207 // depending on potential carries from the input constant and the
1208 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1209 // bits set and the RHS constant is 0x01001, then we know we have a known
1210 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1212 // To compute this, we first compute the potential carry bits. These are
1213 // the bits which may be modified. I'm not aware of a better way to do
1215 const APInt &RHSVal = RHS->getValue();
1216 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1218 // Now that we know which bits have carries, compute the known-1/0 sets.
1220 // Bits are known one if they are known zero in one operand and one in the
1221 // other, and there is no input carry.
1222 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1223 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1225 // Bits are known zero if they are known zero in both operands and there
1226 // is no input carry.
1227 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1229 // If the high-bits of this ADD are not demanded, then it does not demand
1230 // the high bits of its LHS or RHS.
1231 if (DemandedMask[BitWidth-1] == 0) {
1232 // Right fill the mask of bits for this ADD to demand the most
1233 // significant bit and all those below it.
1234 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1235 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1236 LHSKnownZero, LHSKnownOne, Depth+1) ||
1237 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1238 LHSKnownZero, LHSKnownOne, Depth+1))
1244 case Instruction::Sub:
1245 // If the high-bits of this SUB are not demanded, then it does not demand
1246 // the high bits of its LHS or RHS.
1247 if (DemandedMask[BitWidth-1] == 0) {
1248 // Right fill the mask of bits for this SUB to demand the most
1249 // significant bit and all those below it.
1250 uint32_t NLZ = DemandedMask.countLeadingZeros();
1251 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1252 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1253 LHSKnownZero, LHSKnownOne, Depth+1) ||
1254 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1255 LHSKnownZero, LHSKnownOne, Depth+1))
1258 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1259 // the known zeros and ones.
1260 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1262 case Instruction::Shl:
1263 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1264 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1265 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1266 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1267 RHSKnownZero, RHSKnownOne, Depth+1))
1269 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1270 RHSKnownZero <<= ShiftAmt;
1271 RHSKnownOne <<= ShiftAmt;
1272 // low bits known zero.
1274 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1277 case Instruction::LShr:
1278 // For a logical shift right
1279 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1280 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1282 // Unsigned shift right.
1283 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1284 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1285 RHSKnownZero, RHSKnownOne, Depth+1))
1287 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1288 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1289 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1291 // Compute the new bits that are at the top now.
1292 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1293 RHSKnownZero |= HighBits; // high bits known zero.
1297 case Instruction::AShr:
1298 // If this is an arithmetic shift right and only the low-bit is set, we can
1299 // always convert this into a logical shr, even if the shift amount is
1300 // variable. The low bit of the shift cannot be an input sign bit unless
1301 // the shift amount is >= the size of the datatype, which is undefined.
1302 if (DemandedMask == 1) {
1303 // Perform the logical shift right.
1304 Instruction *NewVal = BinaryOperator::CreateLShr(
1305 I->getOperand(0), I->getOperand(1), I->getName());
1306 return InsertNewInstBefore(NewVal, *I);
1309 // If the sign bit is the only bit demanded by this ashr, then there is no
1310 // need to do it, the shift doesn't change the high bit.
1311 if (DemandedMask.isSignBit())
1312 return I->getOperand(0);
1314 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1315 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1317 // Signed shift right.
1318 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1319 // If any of the "high bits" are demanded, we should set the sign bit as
1321 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1322 DemandedMaskIn.set(BitWidth-1);
1323 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1324 RHSKnownZero, RHSKnownOne, Depth+1))
1326 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1327 // Compute the new bits that are at the top now.
1328 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1329 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1330 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1332 // Handle the sign bits.
1333 APInt SignBit(APInt::getSignBit(BitWidth));
1334 // Adjust to where it is now in the mask.
1335 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1337 // If the input sign bit is known to be zero, or if none of the top bits
1338 // are demanded, turn this into an unsigned shift right.
1339 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1340 (HighBits & ~DemandedMask) == HighBits) {
1341 // Perform the logical shift right.
1342 Instruction *NewVal = BinaryOperator::CreateLShr(
1343 I->getOperand(0), SA, I->getName());
1344 return InsertNewInstBefore(NewVal, *I);
1345 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1346 RHSKnownOne |= HighBits;
1350 case Instruction::SRem:
1351 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1352 APInt RA = Rem->getValue().abs();
1353 if (RA.isPowerOf2()) {
1354 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1355 return I->getOperand(0);
1357 APInt LowBits = RA - 1;
1358 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1359 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1360 LHSKnownZero, LHSKnownOne, Depth+1))
1363 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1364 LHSKnownZero |= ~LowBits;
1366 KnownZero |= LHSKnownZero & DemandedMask;
1368 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1372 case Instruction::URem: {
1373 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1374 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1375 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1376 KnownZero2, KnownOne2, Depth+1) ||
1377 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1378 KnownZero2, KnownOne2, Depth+1))
1381 unsigned Leaders = KnownZero2.countLeadingOnes();
1382 Leaders = std::max(Leaders,
1383 KnownZero2.countLeadingOnes());
1384 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1387 case Instruction::Call:
1388 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1389 switch (II->getIntrinsicID()) {
1391 case Intrinsic::bswap: {
1392 // If the only bits demanded come from one byte of the bswap result,
1393 // just shift the input byte into position to eliminate the bswap.
1394 unsigned NLZ = DemandedMask.countLeadingZeros();
1395 unsigned NTZ = DemandedMask.countTrailingZeros();
1397 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1398 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1399 // have 14 leading zeros, round to 8.
1402 // If we need exactly one byte, we can do this transformation.
1403 if (BitWidth-NLZ-NTZ == 8) {
1404 unsigned ResultBit = NTZ;
1405 unsigned InputBit = BitWidth-NTZ-8;
1407 // Replace this with either a left or right shift to get the byte into
1409 Instruction *NewVal;
1410 if (InputBit > ResultBit)
1411 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1412 Context->getConstantInt(I->getType(), InputBit-ResultBit));
1414 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1415 Context->getConstantInt(I->getType(), ResultBit-InputBit));
1416 NewVal->takeName(I);
1417 return InsertNewInstBefore(NewVal, *I);
1420 // TODO: Could compute known zero/one bits based on the input.
1425 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1429 // If the client is only demanding bits that we know, return the known
1431 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1432 Constant *C = Context->getConstantInt(RHSKnownOne);
1433 if (isa<PointerType>(V->getType()))
1434 C = Context->getConstantExprIntToPtr(C, V->getType());
1441 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1442 /// any number of elements. DemandedElts contains the set of elements that are
1443 /// actually used by the caller. This method analyzes which elements of the
1444 /// operand are undef and returns that information in UndefElts.
1446 /// If the information about demanded elements can be used to simplify the
1447 /// operation, the operation is simplified, then the resultant value is
1448 /// returned. This returns null if no change was made.
1449 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1452 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1453 APInt EltMask(APInt::getAllOnesValue(VWidth));
1454 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1456 if (isa<UndefValue>(V)) {
1457 // If the entire vector is undefined, just return this info.
1458 UndefElts = EltMask;
1460 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1461 UndefElts = EltMask;
1462 return Context->getUndef(V->getType());
1466 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1467 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1468 Constant *Undef = Context->getUndef(EltTy);
1470 std::vector<Constant*> Elts;
1471 for (unsigned i = 0; i != VWidth; ++i)
1472 if (!DemandedElts[i]) { // If not demanded, set to undef.
1473 Elts.push_back(Undef);
1475 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1476 Elts.push_back(Undef);
1478 } else { // Otherwise, defined.
1479 Elts.push_back(CP->getOperand(i));
1482 // If we changed the constant, return it.
1483 Constant *NewCP = Context->getConstantVector(Elts);
1484 return NewCP != CP ? NewCP : 0;
1485 } else if (isa<ConstantAggregateZero>(V)) {
1486 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1489 // Check if this is identity. If so, return 0 since we are not simplifying
1491 if (DemandedElts == ((1ULL << VWidth) -1))
1494 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1495 Constant *Zero = Context->getNullValue(EltTy);
1496 Constant *Undef = Context->getUndef(EltTy);
1497 std::vector<Constant*> Elts;
1498 for (unsigned i = 0; i != VWidth; ++i) {
1499 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1500 Elts.push_back(Elt);
1502 UndefElts = DemandedElts ^ EltMask;
1503 return Context->getConstantVector(Elts);
1506 // Limit search depth.
1510 // If multiple users are using the root value, procede with
1511 // simplification conservatively assuming that all elements
1513 if (!V->hasOneUse()) {
1514 // Quit if we find multiple users of a non-root value though.
1515 // They'll be handled when it's their turn to be visited by
1516 // the main instcombine process.
1518 // TODO: Just compute the UndefElts information recursively.
1521 // Conservatively assume that all elements are needed.
1522 DemandedElts = EltMask;
1525 Instruction *I = dyn_cast<Instruction>(V);
1526 if (!I) return 0; // Only analyze instructions.
1528 bool MadeChange = false;
1529 APInt UndefElts2(VWidth, 0);
1531 switch (I->getOpcode()) {
1534 case Instruction::InsertElement: {
1535 // If this is a variable index, we don't know which element it overwrites.
1536 // demand exactly the same input as we produce.
1537 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1539 // Note that we can't propagate undef elt info, because we don't know
1540 // which elt is getting updated.
1541 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1542 UndefElts2, Depth+1);
1543 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1547 // If this is inserting an element that isn't demanded, remove this
1549 unsigned IdxNo = Idx->getZExtValue();
1550 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1551 return AddSoonDeadInstToWorklist(*I, 0);
1553 // Otherwise, the element inserted overwrites whatever was there, so the
1554 // input demanded set is simpler than the output set.
1555 APInt DemandedElts2 = DemandedElts;
1556 DemandedElts2.clear(IdxNo);
1557 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1558 UndefElts, Depth+1);
1559 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1561 // The inserted element is defined.
1562 UndefElts.clear(IdxNo);
1565 case Instruction::ShuffleVector: {
1566 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1567 uint64_t LHSVWidth =
1568 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1569 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1570 for (unsigned i = 0; i < VWidth; i++) {
1571 if (DemandedElts[i]) {
1572 unsigned MaskVal = Shuffle->getMaskValue(i);
1573 if (MaskVal != -1u) {
1574 assert(MaskVal < LHSVWidth * 2 &&
1575 "shufflevector mask index out of range!");
1576 if (MaskVal < LHSVWidth)
1577 LeftDemanded.set(MaskVal);
1579 RightDemanded.set(MaskVal - LHSVWidth);
1584 APInt UndefElts4(LHSVWidth, 0);
1585 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1586 UndefElts4, Depth+1);
1587 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1589 APInt UndefElts3(LHSVWidth, 0);
1590 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1591 UndefElts3, Depth+1);
1592 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1594 bool NewUndefElts = false;
1595 for (unsigned i = 0; i < VWidth; i++) {
1596 unsigned MaskVal = Shuffle->getMaskValue(i);
1597 if (MaskVal == -1u) {
1599 } else if (MaskVal < LHSVWidth) {
1600 if (UndefElts4[MaskVal]) {
1601 NewUndefElts = true;
1605 if (UndefElts3[MaskVal - LHSVWidth]) {
1606 NewUndefElts = true;
1613 // Add additional discovered undefs.
1614 std::vector<Constant*> Elts;
1615 for (unsigned i = 0; i < VWidth; ++i) {
1617 Elts.push_back(Context->getUndef(Type::Int32Ty));
1619 Elts.push_back(Context->getConstantInt(Type::Int32Ty,
1620 Shuffle->getMaskValue(i)));
1622 I->setOperand(2, Context->getConstantVector(Elts));
1627 case Instruction::BitCast: {
1628 // Vector->vector casts only.
1629 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1631 unsigned InVWidth = VTy->getNumElements();
1632 APInt InputDemandedElts(InVWidth, 0);
1635 if (VWidth == InVWidth) {
1636 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1637 // elements as are demanded of us.
1639 InputDemandedElts = DemandedElts;
1640 } else if (VWidth > InVWidth) {
1644 // If there are more elements in the result than there are in the source,
1645 // then an input element is live if any of the corresponding output
1646 // elements are live.
1647 Ratio = VWidth/InVWidth;
1648 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1649 if (DemandedElts[OutIdx])
1650 InputDemandedElts.set(OutIdx/Ratio);
1656 // If there are more elements in the source than there are in the result,
1657 // then an input element is live if the corresponding output element is
1659 Ratio = InVWidth/VWidth;
1660 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1661 if (DemandedElts[InIdx/Ratio])
1662 InputDemandedElts.set(InIdx);
1665 // div/rem demand all inputs, because they don't want divide by zero.
1666 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1667 UndefElts2, Depth+1);
1669 I->setOperand(0, TmpV);
1673 UndefElts = UndefElts2;
1674 if (VWidth > InVWidth) {
1675 llvm_unreachable("Unimp");
1676 // If there are more elements in the result than there are in the source,
1677 // then an output element is undef if the corresponding input element is
1679 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1680 if (UndefElts2[OutIdx/Ratio])
1681 UndefElts.set(OutIdx);
1682 } else if (VWidth < InVWidth) {
1683 llvm_unreachable("Unimp");
1684 // If there are more elements in the source than there are in the result,
1685 // then a result element is undef if all of the corresponding input
1686 // elements are undef.
1687 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1688 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1689 if (!UndefElts2[InIdx]) // Not undef?
1690 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1694 case Instruction::And:
1695 case Instruction::Or:
1696 case Instruction::Xor:
1697 case Instruction::Add:
1698 case Instruction::Sub:
1699 case Instruction::Mul:
1700 // div/rem demand all inputs, because they don't want divide by zero.
1701 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1702 UndefElts, Depth+1);
1703 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1704 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1705 UndefElts2, Depth+1);
1706 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1708 // Output elements are undefined if both are undefined. Consider things
1709 // like undef&0. The result is known zero, not undef.
1710 UndefElts &= UndefElts2;
1713 case Instruction::Call: {
1714 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1716 switch (II->getIntrinsicID()) {
1719 // Binary vector operations that work column-wise. A dest element is a
1720 // function of the corresponding input elements from the two inputs.
1721 case Intrinsic::x86_sse_sub_ss:
1722 case Intrinsic::x86_sse_mul_ss:
1723 case Intrinsic::x86_sse_min_ss:
1724 case Intrinsic::x86_sse_max_ss:
1725 case Intrinsic::x86_sse2_sub_sd:
1726 case Intrinsic::x86_sse2_mul_sd:
1727 case Intrinsic::x86_sse2_min_sd:
1728 case Intrinsic::x86_sse2_max_sd:
1729 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1730 UndefElts, Depth+1);
1731 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1732 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1733 UndefElts2, Depth+1);
1734 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1736 // If only the low elt is demanded and this is a scalarizable intrinsic,
1737 // scalarize it now.
1738 if (DemandedElts == 1) {
1739 switch (II->getIntrinsicID()) {
1741 case Intrinsic::x86_sse_sub_ss:
1742 case Intrinsic::x86_sse_mul_ss:
1743 case Intrinsic::x86_sse2_sub_sd:
1744 case Intrinsic::x86_sse2_mul_sd:
1745 // TODO: Lower MIN/MAX/ABS/etc
1746 Value *LHS = II->getOperand(1);
1747 Value *RHS = II->getOperand(2);
1748 // Extract the element as scalars.
1749 LHS = InsertNewInstBefore(new ExtractElementInst(LHS,
1750 Context->getConstantInt(Type::Int32Ty, 0U, false), "tmp"), *II);
1751 RHS = InsertNewInstBefore(new ExtractElementInst(RHS,
1752 Context->getConstantInt(Type::Int32Ty, 0U, false), "tmp"), *II);
1754 switch (II->getIntrinsicID()) {
1755 default: llvm_unreachable("Case stmts out of sync!");
1756 case Intrinsic::x86_sse_sub_ss:
1757 case Intrinsic::x86_sse2_sub_sd:
1758 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1759 II->getName()), *II);
1761 case Intrinsic::x86_sse_mul_ss:
1762 case Intrinsic::x86_sse2_mul_sd:
1763 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1764 II->getName()), *II);
1769 InsertElementInst::Create(
1770 Context->getUndef(II->getType()), TmpV,
1771 Context->getConstantInt(Type::Int32Ty, 0U, false), II->getName());
1772 InsertNewInstBefore(New, *II);
1773 AddSoonDeadInstToWorklist(*II, 0);
1778 // Output elements are undefined if both are undefined. Consider things
1779 // like undef&0. The result is known zero, not undef.
1780 UndefElts &= UndefElts2;
1786 return MadeChange ? I : 0;
1790 /// AssociativeOpt - Perform an optimization on an associative operator. This
1791 /// function is designed to check a chain of associative operators for a
1792 /// potential to apply a certain optimization. Since the optimization may be
1793 /// applicable if the expression was reassociated, this checks the chain, then
1794 /// reassociates the expression as necessary to expose the optimization
1795 /// opportunity. This makes use of a special Functor, which must define
1796 /// 'shouldApply' and 'apply' methods.
1798 template<typename Functor>
1799 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F,
1800 LLVMContext *Context) {
1801 unsigned Opcode = Root.getOpcode();
1802 Value *LHS = Root.getOperand(0);
1804 // Quick check, see if the immediate LHS matches...
1805 if (F.shouldApply(LHS))
1806 return F.apply(Root);
1808 // Otherwise, if the LHS is not of the same opcode as the root, return.
1809 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1810 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1811 // Should we apply this transform to the RHS?
1812 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1814 // If not to the RHS, check to see if we should apply to the LHS...
1815 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1816 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1820 // If the functor wants to apply the optimization to the RHS of LHSI,
1821 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1823 // Now all of the instructions are in the current basic block, go ahead
1824 // and perform the reassociation.
1825 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1827 // First move the selected RHS to the LHS of the root...
1828 Root.setOperand(0, LHSI->getOperand(1));
1830 // Make what used to be the LHS of the root be the user of the root...
1831 Value *ExtraOperand = TmpLHSI->getOperand(1);
1832 if (&Root == TmpLHSI) {
1833 Root.replaceAllUsesWith(Context->getNullValue(TmpLHSI->getType()));
1836 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1837 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1838 BasicBlock::iterator ARI = &Root; ++ARI;
1839 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1842 // Now propagate the ExtraOperand down the chain of instructions until we
1844 while (TmpLHSI != LHSI) {
1845 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1846 // Move the instruction to immediately before the chain we are
1847 // constructing to avoid breaking dominance properties.
1848 NextLHSI->moveBefore(ARI);
1851 Value *NextOp = NextLHSI->getOperand(1);
1852 NextLHSI->setOperand(1, ExtraOperand);
1854 ExtraOperand = NextOp;
1857 // Now that the instructions are reassociated, have the functor perform
1858 // the transformation...
1859 return F.apply(Root);
1862 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1869 // AddRHS - Implements: X + X --> X << 1
1872 LLVMContext *Context;
1873 AddRHS(Value *rhs, LLVMContext *C) : RHS(rhs), Context(C) {}
1874 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1875 Instruction *apply(BinaryOperator &Add) const {
1876 return BinaryOperator::CreateShl(Add.getOperand(0),
1877 Context->getConstantInt(Add.getType(), 1));
1881 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1883 struct AddMaskingAnd {
1885 LLVMContext *Context;
1886 AddMaskingAnd(Constant *c, LLVMContext *C) : C2(c), Context(C) {}
1887 bool shouldApply(Value *LHS) const {
1889 return match(LHS, m_And(m_Value(), m_ConstantInt(C1)), *Context) &&
1890 Context->getConstantExprAnd(C1, C2)->isNullValue();
1892 Instruction *apply(BinaryOperator &Add) const {
1893 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1899 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1901 LLVMContext *Context = IC->getContext();
1903 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1904 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1907 // Figure out if the constant is the left or the right argument.
1908 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1909 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1911 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1913 return Context->getConstantExpr(I.getOpcode(), SOC, ConstOperand);
1914 return Context->getConstantExpr(I.getOpcode(), ConstOperand, SOC);
1917 Value *Op0 = SO, *Op1 = ConstOperand;
1919 std::swap(Op0, Op1);
1921 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1922 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1923 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1924 New = CmpInst::Create(*Context, CI->getOpcode(), CI->getPredicate(),
1925 Op0, Op1, SO->getName()+".cmp");
1927 llvm_unreachable("Unknown binary instruction type!");
1929 return IC->InsertNewInstBefore(New, I);
1932 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1933 // constant as the other operand, try to fold the binary operator into the
1934 // select arguments. This also works for Cast instructions, which obviously do
1935 // not have a second operand.
1936 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1938 // Don't modify shared select instructions
1939 if (!SI->hasOneUse()) return 0;
1940 Value *TV = SI->getOperand(1);
1941 Value *FV = SI->getOperand(2);
1943 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1944 // Bool selects with constant operands can be folded to logical ops.
1945 if (SI->getType() == Type::Int1Ty) return 0;
1947 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1948 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1950 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1957 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1958 /// node as operand #0, see if we can fold the instruction into the PHI (which
1959 /// is only possible if all operands to the PHI are constants).
1960 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1961 PHINode *PN = cast<PHINode>(I.getOperand(0));
1962 unsigned NumPHIValues = PN->getNumIncomingValues();
1963 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1965 // Check to see if all of the operands of the PHI are constants. If there is
1966 // one non-constant value, remember the BB it is. If there is more than one
1967 // or if *it* is a PHI, bail out.
1968 BasicBlock *NonConstBB = 0;
1969 for (unsigned i = 0; i != NumPHIValues; ++i)
1970 if (!isa<Constant>(PN->getIncomingValue(i))) {
1971 if (NonConstBB) return 0; // More than one non-const value.
1972 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1973 NonConstBB = PN->getIncomingBlock(i);
1975 // If the incoming non-constant value is in I's block, we have an infinite
1977 if (NonConstBB == I.getParent())
1981 // If there is exactly one non-constant value, we can insert a copy of the
1982 // operation in that block. However, if this is a critical edge, we would be
1983 // inserting the computation one some other paths (e.g. inside a loop). Only
1984 // do this if the pred block is unconditionally branching into the phi block.
1986 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1987 if (!BI || !BI->isUnconditional()) return 0;
1990 // Okay, we can do the transformation: create the new PHI node.
1991 PHINode *NewPN = PHINode::Create(I.getType(), "");
1992 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1993 InsertNewInstBefore(NewPN, *PN);
1994 NewPN->takeName(PN);
1996 // Next, add all of the operands to the PHI.
1997 if (I.getNumOperands() == 2) {
1998 Constant *C = cast<Constant>(I.getOperand(1));
1999 for (unsigned i = 0; i != NumPHIValues; ++i) {
2001 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2002 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2003 InV = Context->getConstantExprCompare(CI->getPredicate(), InC, C);
2005 InV = Context->getConstantExpr(I.getOpcode(), InC, C);
2007 assert(PN->getIncomingBlock(i) == NonConstBB);
2008 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2009 InV = BinaryOperator::Create(BO->getOpcode(),
2010 PN->getIncomingValue(i), C, "phitmp",
2011 NonConstBB->getTerminator());
2012 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2013 InV = CmpInst::Create(*Context, CI->getOpcode(),
2015 PN->getIncomingValue(i), C, "phitmp",
2016 NonConstBB->getTerminator());
2018 llvm_unreachable("Unknown binop!");
2020 AddToWorkList(cast<Instruction>(InV));
2022 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2025 CastInst *CI = cast<CastInst>(&I);
2026 const Type *RetTy = CI->getType();
2027 for (unsigned i = 0; i != NumPHIValues; ++i) {
2029 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2030 InV = Context->getConstantExprCast(CI->getOpcode(), InC, RetTy);
2032 assert(PN->getIncomingBlock(i) == NonConstBB);
2033 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2034 I.getType(), "phitmp",
2035 NonConstBB->getTerminator());
2036 AddToWorkList(cast<Instruction>(InV));
2038 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2041 return ReplaceInstUsesWith(I, NewPN);
2045 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2046 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2047 /// This basically requires proving that the add in the original type would not
2048 /// overflow to change the sign bit or have a carry out.
2049 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2050 // There are different heuristics we can use for this. Here are some simple
2053 // Add has the property that adding any two 2's complement numbers can only
2054 // have one carry bit which can change a sign. As such, if LHS and RHS each
2055 // have at least two sign bits, we know that the addition of the two values will
2056 // sign extend fine.
2057 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2061 // If one of the operands only has one non-zero bit, and if the other operand
2062 // has a known-zero bit in a more significant place than it (not including the
2063 // sign bit) the ripple may go up to and fill the zero, but won't change the
2064 // sign. For example, (X & ~4) + 1.
2072 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2073 bool Changed = SimplifyCommutative(I);
2074 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2076 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2077 // X + undef -> undef
2078 if (isa<UndefValue>(RHS))
2079 return ReplaceInstUsesWith(I, RHS);
2082 if (RHSC->isNullValue())
2083 return ReplaceInstUsesWith(I, LHS);
2085 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2086 // X + (signbit) --> X ^ signbit
2087 const APInt& Val = CI->getValue();
2088 uint32_t BitWidth = Val.getBitWidth();
2089 if (Val == APInt::getSignBit(BitWidth))
2090 return BinaryOperator::CreateXor(LHS, RHS);
2092 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2093 // (X & 254)+1 -> (X&254)|1
2094 if (SimplifyDemandedInstructionBits(I))
2097 // zext(bool) + C -> bool ? C + 1 : C
2098 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2099 if (ZI->getSrcTy() == Type::Int1Ty)
2100 return SelectInst::Create(ZI->getOperand(0), AddOne(CI, Context), CI);
2103 if (isa<PHINode>(LHS))
2104 if (Instruction *NV = FoldOpIntoPhi(I))
2107 ConstantInt *XorRHS = 0;
2109 if (isa<ConstantInt>(RHSC) &&
2110 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)), *Context)) {
2111 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2112 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2114 uint32_t Size = TySizeBits / 2;
2115 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2116 APInt CFF80Val(-C0080Val);
2118 if (TySizeBits > Size) {
2119 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2120 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2121 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2122 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2123 // This is a sign extend if the top bits are known zero.
2124 if (!MaskedValueIsZero(XorLHS,
2125 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2126 Size = 0; // Not a sign ext, but can't be any others either.
2131 C0080Val = APIntOps::lshr(C0080Val, Size);
2132 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2133 } while (Size >= 1);
2135 // FIXME: This shouldn't be necessary. When the backends can handle types
2136 // with funny bit widths then this switch statement should be removed. It
2137 // is just here to get the size of the "middle" type back up to something
2138 // that the back ends can handle.
2139 const Type *MiddleType = 0;
2142 case 32: MiddleType = Type::Int32Ty; break;
2143 case 16: MiddleType = Type::Int16Ty; break;
2144 case 8: MiddleType = Type::Int8Ty; break;
2147 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2148 InsertNewInstBefore(NewTrunc, I);
2149 return new SExtInst(NewTrunc, I.getType(), I.getName());
2154 if (I.getType() == Type::Int1Ty)
2155 return BinaryOperator::CreateXor(LHS, RHS);
2158 if (I.getType()->isInteger()) {
2159 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS, Context), Context))
2162 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2163 if (RHSI->getOpcode() == Instruction::Sub)
2164 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2165 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2167 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2168 if (LHSI->getOpcode() == Instruction::Sub)
2169 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2170 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2175 // -A + -B --> -(A + B)
2176 if (Value *LHSV = dyn_castNegVal(LHS, Context)) {
2177 if (LHS->getType()->isIntOrIntVector()) {
2178 if (Value *RHSV = dyn_castNegVal(RHS, Context)) {
2179 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2180 InsertNewInstBefore(NewAdd, I);
2181 return BinaryOperator::CreateNeg(*Context, NewAdd);
2185 return BinaryOperator::CreateSub(RHS, LHSV);
2189 if (!isa<Constant>(RHS))
2190 if (Value *V = dyn_castNegVal(RHS, Context))
2191 return BinaryOperator::CreateSub(LHS, V);
2195 if (Value *X = dyn_castFoldableMul(LHS, C2, Context)) {
2196 if (X == RHS) // X*C + X --> X * (C+1)
2197 return BinaryOperator::CreateMul(RHS, AddOne(C2, Context));
2199 // X*C1 + X*C2 --> X * (C1+C2)
2201 if (X == dyn_castFoldableMul(RHS, C1, Context))
2202 return BinaryOperator::CreateMul(X, Context->getConstantExprAdd(C1, C2));
2205 // X + X*C --> X * (C+1)
2206 if (dyn_castFoldableMul(RHS, C2, Context) == LHS)
2207 return BinaryOperator::CreateMul(LHS, AddOne(C2, Context));
2209 // X + ~X --> -1 since ~X = -X-1
2210 if (dyn_castNotVal(LHS, Context) == RHS ||
2211 dyn_castNotVal(RHS, Context) == LHS)
2212 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
2215 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2216 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2)), *Context))
2217 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2, Context), Context))
2220 // A+B --> A|B iff A and B have no bits set in common.
2221 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2222 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2223 APInt LHSKnownOne(IT->getBitWidth(), 0);
2224 APInt LHSKnownZero(IT->getBitWidth(), 0);
2225 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2226 if (LHSKnownZero != 0) {
2227 APInt RHSKnownOne(IT->getBitWidth(), 0);
2228 APInt RHSKnownZero(IT->getBitWidth(), 0);
2229 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2231 // No bits in common -> bitwise or.
2232 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2233 return BinaryOperator::CreateOr(LHS, RHS);
2237 // W*X + Y*Z --> W * (X+Z) iff W == Y
2238 if (I.getType()->isIntOrIntVector()) {
2239 Value *W, *X, *Y, *Z;
2240 if (match(LHS, m_Mul(m_Value(W), m_Value(X)), *Context) &&
2241 match(RHS, m_Mul(m_Value(Y), m_Value(Z)), *Context)) {
2245 } else if (Y == X) {
2247 } else if (X == Z) {
2254 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2255 LHS->getName()), I);
2256 return BinaryOperator::CreateMul(W, NewAdd);
2261 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2263 if (match(LHS, m_Not(m_Value(X)), *Context)) // ~X + C --> (C-1) - X
2264 return BinaryOperator::CreateSub(SubOne(CRHS, Context), X);
2266 // (X & FF00) + xx00 -> (X+xx00) & FF00
2267 if (LHS->hasOneUse() &&
2268 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)), *Context)) {
2269 Constant *Anded = Context->getConstantExprAnd(CRHS, C2);
2270 if (Anded == CRHS) {
2271 // See if all bits from the first bit set in the Add RHS up are included
2272 // in the mask. First, get the rightmost bit.
2273 const APInt& AddRHSV = CRHS->getValue();
2275 // Form a mask of all bits from the lowest bit added through the top.
2276 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2278 // See if the and mask includes all of these bits.
2279 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2281 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2282 // Okay, the xform is safe. Insert the new add pronto.
2283 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2284 LHS->getName()), I);
2285 return BinaryOperator::CreateAnd(NewAdd, C2);
2290 // Try to fold constant add into select arguments.
2291 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2292 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2296 // add (cast *A to intptrtype) B ->
2297 // cast (GEP (cast *A to i8*) B) --> intptrtype
2299 CastInst *CI = dyn_cast<CastInst>(LHS);
2302 CI = dyn_cast<CastInst>(RHS);
2305 if (CI && CI->getType()->isSized() &&
2306 (CI->getType()->getScalarSizeInBits() ==
2307 TD->getIntPtrType()->getPrimitiveSizeInBits())
2308 && isa<PointerType>(CI->getOperand(0)->getType())) {
2310 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2311 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2312 Context->getPointerType(Type::Int8Ty, AS), I);
2313 GetElementPtrInst *GEP = GetElementPtrInst::Create(I2, Other, "ctg2");
2314 // A GEP formed from an arbitrary add may overflow.
2315 cast<GEPOperator>(GEP)->setHasNoPointerOverflow(false);
2316 I2 = InsertNewInstBefore(GEP, I);
2317 return new PtrToIntInst(I2, CI->getType());
2321 // add (select X 0 (sub n A)) A --> select X A n
2323 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2326 SI = dyn_cast<SelectInst>(RHS);
2329 if (SI && SI->hasOneUse()) {
2330 Value *TV = SI->getTrueValue();
2331 Value *FV = SI->getFalseValue();
2334 // Can we fold the add into the argument of the select?
2335 // We check both true and false select arguments for a matching subtract.
2336 if (match(FV, m_Zero(), *Context) &&
2337 match(TV, m_Sub(m_Value(N), m_Specific(A)), *Context))
2338 // Fold the add into the true select value.
2339 return SelectInst::Create(SI->getCondition(), N, A);
2340 if (match(TV, m_Zero(), *Context) &&
2341 match(FV, m_Sub(m_Value(N), m_Specific(A)), *Context))
2342 // Fold the add into the false select value.
2343 return SelectInst::Create(SI->getCondition(), A, N);
2347 // Check for (add (sext x), y), see if we can merge this into an
2348 // integer add followed by a sext.
2349 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2350 // (add (sext x), cst) --> (sext (add x, cst'))
2351 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2353 Context->getConstantExprTrunc(RHSC, LHSConv->getOperand(0)->getType());
2354 if (LHSConv->hasOneUse() &&
2355 Context->getConstantExprSExt(CI, I.getType()) == RHSC &&
2356 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2357 // Insert the new, smaller add.
2358 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2360 InsertNewInstBefore(NewAdd, I);
2361 return new SExtInst(NewAdd, I.getType());
2365 // (add (sext x), (sext y)) --> (sext (add int x, y))
2366 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2367 // Only do this if x/y have the same type, if at last one of them has a
2368 // single use (so we don't increase the number of sexts), and if the
2369 // integer add will not overflow.
2370 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2371 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2372 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2373 RHSConv->getOperand(0))) {
2374 // Insert the new integer add.
2375 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2376 RHSConv->getOperand(0),
2378 InsertNewInstBefore(NewAdd, I);
2379 return new SExtInst(NewAdd, I.getType());
2384 return Changed ? &I : 0;
2387 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2388 bool Changed = SimplifyCommutative(I);
2389 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2391 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2393 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2394 if (CFP->isExactlyValue(Context->getConstantFPNegativeZero
2395 (I.getType())->getValueAPF()))
2396 return ReplaceInstUsesWith(I, LHS);
2399 if (isa<PHINode>(LHS))
2400 if (Instruction *NV = FoldOpIntoPhi(I))
2405 // -A + -B --> -(A + B)
2406 if (Value *LHSV = dyn_castFNegVal(LHS, Context))
2407 return BinaryOperator::CreateFSub(RHS, LHSV);
2410 if (!isa<Constant>(RHS))
2411 if (Value *V = dyn_castFNegVal(RHS, Context))
2412 return BinaryOperator::CreateFSub(LHS, V);
2414 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2415 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2416 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2417 return ReplaceInstUsesWith(I, LHS);
2419 // Check for (add double (sitofp x), y), see if we can merge this into an
2420 // integer add followed by a promotion.
2421 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2422 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2423 // ... if the constant fits in the integer value. This is useful for things
2424 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2425 // requires a constant pool load, and generally allows the add to be better
2427 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2429 Context->getConstantExprFPToSI(CFP, LHSConv->getOperand(0)->getType());
2430 if (LHSConv->hasOneUse() &&
2431 Context->getConstantExprSIToFP(CI, I.getType()) == CFP &&
2432 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2433 // Insert the new integer add.
2434 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2436 InsertNewInstBefore(NewAdd, I);
2437 return new SIToFPInst(NewAdd, I.getType());
2441 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2442 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2443 // Only do this if x/y have the same type, if at last one of them has a
2444 // single use (so we don't increase the number of int->fp conversions),
2445 // and if the integer add will not overflow.
2446 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2447 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2448 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2449 RHSConv->getOperand(0))) {
2450 // Insert the new integer add.
2451 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2452 RHSConv->getOperand(0),
2454 InsertNewInstBefore(NewAdd, I);
2455 return new SIToFPInst(NewAdd, I.getType());
2460 return Changed ? &I : 0;
2463 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2464 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2466 if (Op0 == Op1) // sub X, X -> 0
2467 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2469 // If this is a 'B = x-(-A)', change to B = x+A...
2470 if (Value *V = dyn_castNegVal(Op1, Context))
2471 return BinaryOperator::CreateAdd(Op0, V);
2473 if (isa<UndefValue>(Op0))
2474 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2475 if (isa<UndefValue>(Op1))
2476 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2478 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2479 // Replace (-1 - A) with (~A)...
2480 if (C->isAllOnesValue())
2481 return BinaryOperator::CreateNot(*Context, Op1);
2483 // C - ~X == X + (1+C)
2485 if (match(Op1, m_Not(m_Value(X)), *Context))
2486 return BinaryOperator::CreateAdd(X, AddOne(C, Context));
2488 // -(X >>u 31) -> (X >>s 31)
2489 // -(X >>s 31) -> (X >>u 31)
2491 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2492 if (SI->getOpcode() == Instruction::LShr) {
2493 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2494 // Check to see if we are shifting out everything but the sign bit.
2495 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2496 SI->getType()->getPrimitiveSizeInBits()-1) {
2497 // Ok, the transformation is safe. Insert AShr.
2498 return BinaryOperator::Create(Instruction::AShr,
2499 SI->getOperand(0), CU, SI->getName());
2503 else if (SI->getOpcode() == Instruction::AShr) {
2504 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2505 // Check to see if we are shifting out everything but the sign bit.
2506 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2507 SI->getType()->getPrimitiveSizeInBits()-1) {
2508 // Ok, the transformation is safe. Insert LShr.
2509 return BinaryOperator::CreateLShr(
2510 SI->getOperand(0), CU, SI->getName());
2517 // Try to fold constant sub into select arguments.
2518 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2519 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2522 // C - zext(bool) -> bool ? C - 1 : C
2523 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2524 if (ZI->getSrcTy() == Type::Int1Ty)
2525 return SelectInst::Create(ZI->getOperand(0), SubOne(C, Context), C);
2528 if (I.getType() == Type::Int1Ty)
2529 return BinaryOperator::CreateXor(Op0, Op1);
2531 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2532 if (Op1I->getOpcode() == Instruction::Add) {
2533 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2534 return BinaryOperator::CreateNeg(*Context, Op1I->getOperand(1),
2536 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2537 return BinaryOperator::CreateNeg(*Context, Op1I->getOperand(0),
2539 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2540 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2541 // C1-(X+C2) --> (C1-C2)-X
2542 return BinaryOperator::CreateSub(
2543 Context->getConstantExprSub(CI1, CI2), Op1I->getOperand(0));
2547 if (Op1I->hasOneUse()) {
2548 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2549 // is not used by anyone else...
2551 if (Op1I->getOpcode() == Instruction::Sub) {
2552 // Swap the two operands of the subexpr...
2553 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2554 Op1I->setOperand(0, IIOp1);
2555 Op1I->setOperand(1, IIOp0);
2557 // Create the new top level add instruction...
2558 return BinaryOperator::CreateAdd(Op0, Op1);
2561 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2563 if (Op1I->getOpcode() == Instruction::And &&
2564 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2565 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2568 InsertNewInstBefore(BinaryOperator::CreateNot(*Context,
2569 OtherOp, "B.not"), I);
2570 return BinaryOperator::CreateAnd(Op0, NewNot);
2573 // 0 - (X sdiv C) -> (X sdiv -C)
2574 if (Op1I->getOpcode() == Instruction::SDiv)
2575 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2577 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2578 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2579 Context->getConstantExprNeg(DivRHS));
2581 // X - X*C --> X * (1-C)
2582 ConstantInt *C2 = 0;
2583 if (dyn_castFoldableMul(Op1I, C2, Context) == Op0) {
2585 Context->getConstantExprSub(Context->getConstantInt(I.getType(), 1),
2587 return BinaryOperator::CreateMul(Op0, CP1);
2592 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2593 if (Op0I->getOpcode() == Instruction::Add) {
2594 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2595 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2596 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2597 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2598 } else if (Op0I->getOpcode() == Instruction::Sub) {
2599 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2600 return BinaryOperator::CreateNeg(*Context, Op0I->getOperand(1),
2606 if (Value *X = dyn_castFoldableMul(Op0, C1, Context)) {
2607 if (X == Op1) // X*C - X --> X * (C-1)
2608 return BinaryOperator::CreateMul(Op1, SubOne(C1, Context));
2610 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2611 if (X == dyn_castFoldableMul(Op1, C2, Context))
2612 return BinaryOperator::CreateMul(X, Context->getConstantExprSub(C1, C2));
2617 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2618 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2620 // If this is a 'B = x-(-A)', change to B = x+A...
2621 if (Value *V = dyn_castFNegVal(Op1, Context))
2622 return BinaryOperator::CreateFAdd(Op0, V);
2624 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2625 if (Op1I->getOpcode() == Instruction::FAdd) {
2626 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2627 return BinaryOperator::CreateFNeg(*Context, Op1I->getOperand(1),
2629 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2630 return BinaryOperator::CreateFNeg(*Context, Op1I->getOperand(0),
2638 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2639 /// comparison only checks the sign bit. If it only checks the sign bit, set
2640 /// TrueIfSigned if the result of the comparison is true when the input value is
2642 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2643 bool &TrueIfSigned) {
2645 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2646 TrueIfSigned = true;
2647 return RHS->isZero();
2648 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2649 TrueIfSigned = true;
2650 return RHS->isAllOnesValue();
2651 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2652 TrueIfSigned = false;
2653 return RHS->isAllOnesValue();
2654 case ICmpInst::ICMP_UGT:
2655 // True if LHS u> RHS and RHS == high-bit-mask - 1
2656 TrueIfSigned = true;
2657 return RHS->getValue() ==
2658 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2659 case ICmpInst::ICMP_UGE:
2660 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2661 TrueIfSigned = true;
2662 return RHS->getValue().isSignBit();
2668 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2669 bool Changed = SimplifyCommutative(I);
2670 Value *Op0 = I.getOperand(0);
2672 // TODO: If Op1 is undef and Op0 is finite, return zero.
2673 if (!I.getType()->isFPOrFPVector() &&
2674 isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2675 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2677 // Simplify mul instructions with a constant RHS...
2678 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2679 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2681 // ((X << C1)*C2) == (X * (C2 << C1))
2682 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2683 if (SI->getOpcode() == Instruction::Shl)
2684 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2685 return BinaryOperator::CreateMul(SI->getOperand(0),
2686 Context->getConstantExprShl(CI, ShOp));
2689 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2690 if (CI->equalsInt(1)) // X * 1 == X
2691 return ReplaceInstUsesWith(I, Op0);
2692 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2693 return BinaryOperator::CreateNeg(*Context, Op0, I.getName());
2695 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2696 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2697 return BinaryOperator::CreateShl(Op0,
2698 Context->getConstantInt(Op0->getType(), Val.logBase2()));
2700 } else if (isa<VectorType>(Op1->getType())) {
2701 if (Op1->isNullValue())
2702 return ReplaceInstUsesWith(I, Op1);
2704 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2705 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2706 return BinaryOperator::CreateNeg(*Context, Op0, I.getName());
2708 // As above, vector X*splat(1.0) -> X in all defined cases.
2709 if (Constant *Splat = Op1V->getSplatValue()) {
2710 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2711 if (CI->equalsInt(1))
2712 return ReplaceInstUsesWith(I, Op0);
2717 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2718 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2719 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2720 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2721 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2723 InsertNewInstBefore(Add, I);
2724 Value *C1C2 = Context->getConstantExprMul(Op1,
2725 cast<Constant>(Op0I->getOperand(1)));
2726 return BinaryOperator::CreateAdd(Add, C1C2);
2730 // Try to fold constant mul into select arguments.
2731 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2732 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2735 if (isa<PHINode>(Op0))
2736 if (Instruction *NV = FoldOpIntoPhi(I))
2740 if (Value *Op0v = dyn_castNegVal(Op0, Context)) // -X * -Y = X*Y
2741 if (Value *Op1v = dyn_castNegVal(I.getOperand(1), Context))
2742 return BinaryOperator::CreateMul(Op0v, Op1v);
2744 // (X / Y) * Y = X - (X % Y)
2745 // (X / Y) * -Y = (X % Y) - X
2747 Value *Op1 = I.getOperand(1);
2748 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2750 (BO->getOpcode() != Instruction::UDiv &&
2751 BO->getOpcode() != Instruction::SDiv)) {
2753 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2755 Value *Neg = dyn_castNegVal(Op1, Context);
2756 if (BO && BO->hasOneUse() &&
2757 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2758 (BO->getOpcode() == Instruction::UDiv ||
2759 BO->getOpcode() == Instruction::SDiv)) {
2760 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2763 if (BO->getOpcode() == Instruction::UDiv)
2764 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2766 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2768 InsertNewInstBefore(Rem, I);
2772 return BinaryOperator::CreateSub(Op0BO, Rem);
2774 return BinaryOperator::CreateSub(Rem, Op0BO);
2778 if (I.getType() == Type::Int1Ty)
2779 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2781 // If one of the operands of the multiply is a cast from a boolean value, then
2782 // we know the bool is either zero or one, so this is a 'masking' multiply.
2783 // See if we can simplify things based on how the boolean was originally
2785 CastInst *BoolCast = 0;
2786 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2787 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2790 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2791 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2794 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2795 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2796 const Type *SCOpTy = SCIOp0->getType();
2799 // If the icmp is true iff the sign bit of X is set, then convert this
2800 // multiply into a shift/and combination.
2801 if (isa<ConstantInt>(SCIOp1) &&
2802 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2804 // Shift the X value right to turn it into "all signbits".
2805 Constant *Amt = Context->getConstantInt(SCIOp0->getType(),
2806 SCOpTy->getPrimitiveSizeInBits()-1);
2808 InsertNewInstBefore(
2809 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2810 BoolCast->getOperand(0)->getName()+
2813 // If the multiply type is not the same as the source type, sign extend
2814 // or truncate to the multiply type.
2815 if (I.getType() != V->getType()) {
2816 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2817 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2818 Instruction::CastOps opcode =
2819 (SrcBits == DstBits ? Instruction::BitCast :
2820 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2821 V = InsertCastBefore(opcode, V, I.getType(), I);
2824 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2825 return BinaryOperator::CreateAnd(V, OtherOp);
2830 return Changed ? &I : 0;
2833 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2834 bool Changed = SimplifyCommutative(I);
2835 Value *Op0 = I.getOperand(0);
2837 // Simplify mul instructions with a constant RHS...
2838 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2839 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2840 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2841 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2842 if (Op1F->isExactlyValue(1.0))
2843 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2844 } else if (isa<VectorType>(Op1->getType())) {
2845 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2846 // As above, vector X*splat(1.0) -> X in all defined cases.
2847 if (Constant *Splat = Op1V->getSplatValue()) {
2848 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2849 if (F->isExactlyValue(1.0))
2850 return ReplaceInstUsesWith(I, Op0);
2855 // Try to fold constant mul into select arguments.
2856 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2857 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2860 if (isa<PHINode>(Op0))
2861 if (Instruction *NV = FoldOpIntoPhi(I))
2865 if (Value *Op0v = dyn_castFNegVal(Op0, Context)) // -X * -Y = X*Y
2866 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1), Context))
2867 return BinaryOperator::CreateFMul(Op0v, Op1v);
2869 return Changed ? &I : 0;
2872 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2874 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2875 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2877 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2878 int NonNullOperand = -1;
2879 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2880 if (ST->isNullValue())
2882 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2883 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2884 if (ST->isNullValue())
2887 if (NonNullOperand == -1)
2890 Value *SelectCond = SI->getOperand(0);
2892 // Change the div/rem to use 'Y' instead of the select.
2893 I.setOperand(1, SI->getOperand(NonNullOperand));
2895 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2896 // problem. However, the select, or the condition of the select may have
2897 // multiple uses. Based on our knowledge that the operand must be non-zero,
2898 // propagate the known value for the select into other uses of it, and
2899 // propagate a known value of the condition into its other users.
2901 // If the select and condition only have a single use, don't bother with this,
2903 if (SI->use_empty() && SelectCond->hasOneUse())
2906 // Scan the current block backward, looking for other uses of SI.
2907 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2909 while (BBI != BBFront) {
2911 // If we found a call to a function, we can't assume it will return, so
2912 // information from below it cannot be propagated above it.
2913 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2916 // Replace uses of the select or its condition with the known values.
2917 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2920 *I = SI->getOperand(NonNullOperand);
2922 } else if (*I == SelectCond) {
2923 *I = NonNullOperand == 1 ? Context->getConstantIntTrue() :
2924 Context->getConstantIntFalse();
2929 // If we past the instruction, quit looking for it.
2932 if (&*BBI == SelectCond)
2935 // If we ran out of things to eliminate, break out of the loop.
2936 if (SelectCond == 0 && SI == 0)
2944 /// This function implements the transforms on div instructions that work
2945 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2946 /// used by the visitors to those instructions.
2947 /// @brief Transforms common to all three div instructions
2948 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2949 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2951 // undef / X -> 0 for integer.
2952 // undef / X -> undef for FP (the undef could be a snan).
2953 if (isa<UndefValue>(Op0)) {
2954 if (Op0->getType()->isFPOrFPVector())
2955 return ReplaceInstUsesWith(I, Op0);
2956 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2959 // X / undef -> undef
2960 if (isa<UndefValue>(Op1))
2961 return ReplaceInstUsesWith(I, Op1);
2966 /// This function implements the transforms common to both integer division
2967 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2968 /// division instructions.
2969 /// @brief Common integer divide transforms
2970 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2971 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2973 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2975 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2976 Constant *CI = Context->getConstantInt(Ty->getElementType(), 1);
2977 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2978 return ReplaceInstUsesWith(I, Context->getConstantVector(Elts));
2981 Constant *CI = Context->getConstantInt(I.getType(), 1);
2982 return ReplaceInstUsesWith(I, CI);
2985 if (Instruction *Common = commonDivTransforms(I))
2988 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2989 // This does not apply for fdiv.
2990 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2993 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2995 if (RHS->equalsInt(1))
2996 return ReplaceInstUsesWith(I, Op0);
2998 // (X / C1) / C2 -> X / (C1*C2)
2999 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3000 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3001 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3002 if (MultiplyOverflows(RHS, LHSRHS,
3003 I.getOpcode()==Instruction::SDiv, Context))
3004 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3006 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
3007 Context->getConstantExprMul(RHS, LHSRHS));
3010 if (!RHS->isZero()) { // avoid X udiv 0
3011 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3012 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3014 if (isa<PHINode>(Op0))
3015 if (Instruction *NV = FoldOpIntoPhi(I))
3020 // 0 / X == 0, we don't need to preserve faults!
3021 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3022 if (LHS->equalsInt(0))
3023 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3025 // It can't be division by zero, hence it must be division by one.
3026 if (I.getType() == Type::Int1Ty)
3027 return ReplaceInstUsesWith(I, Op0);
3029 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3030 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3033 return ReplaceInstUsesWith(I, Op0);
3039 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3040 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3042 // Handle the integer div common cases
3043 if (Instruction *Common = commonIDivTransforms(I))
3046 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3047 // X udiv C^2 -> X >> C
3048 // Check to see if this is an unsigned division with an exact power of 2,
3049 // if so, convert to a right shift.
3050 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3051 return BinaryOperator::CreateLShr(Op0,
3052 Context->getConstantInt(Op0->getType(), C->getValue().logBase2()));
3054 // X udiv C, where C >= signbit
3055 if (C->getValue().isNegative()) {
3056 Value *IC = InsertNewInstBefore(new ICmpInst(*Context,
3057 ICmpInst::ICMP_ULT, Op0, C),
3059 return SelectInst::Create(IC, Context->getNullValue(I.getType()),
3060 Context->getConstantInt(I.getType(), 1));
3064 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3065 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3066 if (RHSI->getOpcode() == Instruction::Shl &&
3067 isa<ConstantInt>(RHSI->getOperand(0))) {
3068 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3069 if (C1.isPowerOf2()) {
3070 Value *N = RHSI->getOperand(1);
3071 const Type *NTy = N->getType();
3072 if (uint32_t C2 = C1.logBase2()) {
3073 Constant *C2V = Context->getConstantInt(NTy, C2);
3074 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3076 return BinaryOperator::CreateLShr(Op0, N);
3081 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3082 // where C1&C2 are powers of two.
3083 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3084 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3085 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3086 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3087 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3088 // Compute the shift amounts
3089 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3090 // Construct the "on true" case of the select
3091 Constant *TC = Context->getConstantInt(Op0->getType(), TSA);
3092 Instruction *TSI = BinaryOperator::CreateLShr(
3093 Op0, TC, SI->getName()+".t");
3094 TSI = InsertNewInstBefore(TSI, I);
3096 // Construct the "on false" case of the select
3097 Constant *FC = Context->getConstantInt(Op0->getType(), FSA);
3098 Instruction *FSI = BinaryOperator::CreateLShr(
3099 Op0, FC, SI->getName()+".f");
3100 FSI = InsertNewInstBefore(FSI, I);
3102 // construct the select instruction and return it.
3103 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3109 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3110 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3112 // Handle the integer div common cases
3113 if (Instruction *Common = commonIDivTransforms(I))
3116 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3118 if (RHS->isAllOnesValue())
3119 return BinaryOperator::CreateNeg(*Context, Op0);
3122 // If the sign bits of both operands are zero (i.e. we can prove they are
3123 // unsigned inputs), turn this into a udiv.
3124 if (I.getType()->isInteger()) {
3125 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3126 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3127 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3128 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3135 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3136 return commonDivTransforms(I);
3139 /// This function implements the transforms on rem instructions that work
3140 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3141 /// is used by the visitors to those instructions.
3142 /// @brief Transforms common to all three rem instructions
3143 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3144 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3146 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3147 if (I.getType()->isFPOrFPVector())
3148 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3149 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3151 if (isa<UndefValue>(Op1))
3152 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3154 // Handle cases involving: rem X, (select Cond, Y, Z)
3155 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3161 /// This function implements the transforms common to both integer remainder
3162 /// instructions (urem and srem). It is called by the visitors to those integer
3163 /// remainder instructions.
3164 /// @brief Common integer remainder transforms
3165 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3166 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3168 if (Instruction *common = commonRemTransforms(I))
3171 // 0 % X == 0 for integer, we don't need to preserve faults!
3172 if (Constant *LHS = dyn_cast<Constant>(Op0))
3173 if (LHS->isNullValue())
3174 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3176 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3177 // X % 0 == undef, we don't need to preserve faults!
3178 if (RHS->equalsInt(0))
3179 return ReplaceInstUsesWith(I, Context->getUndef(I.getType()));
3181 if (RHS->equalsInt(1)) // X % 1 == 0
3182 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3184 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3185 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3186 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3188 } else if (isa<PHINode>(Op0I)) {
3189 if (Instruction *NV = FoldOpIntoPhi(I))
3193 // See if we can fold away this rem instruction.
3194 if (SimplifyDemandedInstructionBits(I))
3202 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3203 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3205 if (Instruction *common = commonIRemTransforms(I))
3208 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3209 // X urem C^2 -> X and C
3210 // Check to see if this is an unsigned remainder with an exact power of 2,
3211 // if so, convert to a bitwise and.
3212 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3213 if (C->getValue().isPowerOf2())
3214 return BinaryOperator::CreateAnd(Op0, SubOne(C, Context));
3217 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3218 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3219 if (RHSI->getOpcode() == Instruction::Shl &&
3220 isa<ConstantInt>(RHSI->getOperand(0))) {
3221 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3222 Constant *N1 = Context->getAllOnesValue(I.getType());
3223 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3225 return BinaryOperator::CreateAnd(Op0, Add);
3230 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3231 // where C1&C2 are powers of two.
3232 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3233 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3234 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3235 // STO == 0 and SFO == 0 handled above.
3236 if ((STO->getValue().isPowerOf2()) &&
3237 (SFO->getValue().isPowerOf2())) {
3238 Value *TrueAnd = InsertNewInstBefore(
3239 BinaryOperator::CreateAnd(Op0, SubOne(STO, Context),
3240 SI->getName()+".t"), I);
3241 Value *FalseAnd = InsertNewInstBefore(
3242 BinaryOperator::CreateAnd(Op0, SubOne(SFO, Context),
3243 SI->getName()+".f"), I);
3244 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3252 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3253 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3255 // Handle the integer rem common cases
3256 if (Instruction *common = commonIRemTransforms(I))
3259 if (Value *RHSNeg = dyn_castNegVal(Op1, Context))
3260 if (!isa<Constant>(RHSNeg) ||
3261 (isa<ConstantInt>(RHSNeg) &&
3262 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3264 AddUsesToWorkList(I);
3265 I.setOperand(1, RHSNeg);
3269 // If the sign bits of both operands are zero (i.e. we can prove they are
3270 // unsigned inputs), turn this into a urem.
3271 if (I.getType()->isInteger()) {
3272 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3273 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3274 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3275 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3279 // If it's a constant vector, flip any negative values positive.
3280 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3281 unsigned VWidth = RHSV->getNumOperands();
3283 bool hasNegative = false;
3284 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3285 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3286 if (RHS->getValue().isNegative())
3290 std::vector<Constant *> Elts(VWidth);
3291 for (unsigned i = 0; i != VWidth; ++i) {
3292 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3293 if (RHS->getValue().isNegative())
3294 Elts[i] = cast<ConstantInt>(Context->getConstantExprNeg(RHS));
3300 Constant *NewRHSV = Context->getConstantVector(Elts);
3301 if (NewRHSV != RHSV) {
3302 AddUsesToWorkList(I);
3303 I.setOperand(1, NewRHSV);
3312 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3313 return commonRemTransforms(I);
3316 // isOneBitSet - Return true if there is exactly one bit set in the specified
3318 static bool isOneBitSet(const ConstantInt *CI) {
3319 return CI->getValue().isPowerOf2();
3322 // isHighOnes - Return true if the constant is of the form 1+0+.
3323 // This is the same as lowones(~X).
3324 static bool isHighOnes(const ConstantInt *CI) {
3325 return (~CI->getValue() + 1).isPowerOf2();
3328 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3329 /// are carefully arranged to allow folding of expressions such as:
3331 /// (A < B) | (A > B) --> (A != B)
3333 /// Note that this is only valid if the first and second predicates have the
3334 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3336 /// Three bits are used to represent the condition, as follows:
3341 /// <=> Value Definition
3342 /// 000 0 Always false
3349 /// 111 7 Always true
3351 static unsigned getICmpCode(const ICmpInst *ICI) {
3352 switch (ICI->getPredicate()) {
3354 case ICmpInst::ICMP_UGT: return 1; // 001
3355 case ICmpInst::ICMP_SGT: return 1; // 001
3356 case ICmpInst::ICMP_EQ: return 2; // 010
3357 case ICmpInst::ICMP_UGE: return 3; // 011
3358 case ICmpInst::ICMP_SGE: return 3; // 011
3359 case ICmpInst::ICMP_ULT: return 4; // 100
3360 case ICmpInst::ICMP_SLT: return 4; // 100
3361 case ICmpInst::ICMP_NE: return 5; // 101
3362 case ICmpInst::ICMP_ULE: return 6; // 110
3363 case ICmpInst::ICMP_SLE: return 6; // 110
3366 llvm_unreachable("Invalid ICmp predicate!");
3371 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3372 /// predicate into a three bit mask. It also returns whether it is an ordered
3373 /// predicate by reference.
3374 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3377 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3378 case FCmpInst::FCMP_UNO: return 0; // 000
3379 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3380 case FCmpInst::FCMP_UGT: return 1; // 001
3381 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3382 case FCmpInst::FCMP_UEQ: return 2; // 010
3383 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3384 case FCmpInst::FCMP_UGE: return 3; // 011
3385 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3386 case FCmpInst::FCMP_ULT: return 4; // 100
3387 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3388 case FCmpInst::FCMP_UNE: return 5; // 101
3389 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3390 case FCmpInst::FCMP_ULE: return 6; // 110
3393 // Not expecting FCMP_FALSE and FCMP_TRUE;
3394 llvm_unreachable("Unexpected FCmp predicate!");
3399 /// getICmpValue - This is the complement of getICmpCode, which turns an
3400 /// opcode and two operands into either a constant true or false, or a brand
3401 /// new ICmp instruction. The sign is passed in to determine which kind
3402 /// of predicate to use in the new icmp instruction.
3403 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3404 LLVMContext *Context) {
3406 default: llvm_unreachable("Illegal ICmp code!");
3407 case 0: return Context->getConstantIntFalse();
3410 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, LHS, RHS);
3412 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, LHS, RHS);
3413 case 2: return new ICmpInst(*Context, ICmpInst::ICMP_EQ, LHS, RHS);
3416 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHS, RHS);
3418 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHS, RHS);
3421 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHS, RHS);
3423 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHS, RHS);
3424 case 5: return new ICmpInst(*Context, ICmpInst::ICMP_NE, LHS, RHS);
3427 return new ICmpInst(*Context, ICmpInst::ICMP_SLE, LHS, RHS);
3429 return new ICmpInst(*Context, ICmpInst::ICMP_ULE, LHS, RHS);
3430 case 7: return Context->getConstantIntTrue();
3434 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3435 /// opcode and two operands into either a FCmp instruction. isordered is passed
3436 /// in to determine which kind of predicate to use in the new fcmp instruction.
3437 static Value *getFCmpValue(bool isordered, unsigned code,
3438 Value *LHS, Value *RHS, LLVMContext *Context) {
3440 default: llvm_unreachable("Illegal FCmp code!");
3443 return new FCmpInst(*Context, FCmpInst::FCMP_ORD, LHS, RHS);
3445 return new FCmpInst(*Context, FCmpInst::FCMP_UNO, LHS, RHS);
3448 return new FCmpInst(*Context, FCmpInst::FCMP_OGT, LHS, RHS);
3450 return new FCmpInst(*Context, FCmpInst::FCMP_UGT, LHS, RHS);
3453 return new FCmpInst(*Context, FCmpInst::FCMP_OEQ, LHS, RHS);
3455 return new FCmpInst(*Context, FCmpInst::FCMP_UEQ, LHS, RHS);
3458 return new FCmpInst(*Context, FCmpInst::FCMP_OGE, LHS, RHS);
3460 return new FCmpInst(*Context, FCmpInst::FCMP_UGE, LHS, RHS);
3463 return new FCmpInst(*Context, FCmpInst::FCMP_OLT, LHS, RHS);
3465 return new FCmpInst(*Context, FCmpInst::FCMP_ULT, LHS, RHS);
3468 return new FCmpInst(*Context, FCmpInst::FCMP_ONE, LHS, RHS);
3470 return new FCmpInst(*Context, FCmpInst::FCMP_UNE, LHS, RHS);
3473 return new FCmpInst(*Context, FCmpInst::FCMP_OLE, LHS, RHS);
3475 return new FCmpInst(*Context, FCmpInst::FCMP_ULE, LHS, RHS);
3476 case 7: return Context->getConstantIntTrue();
3480 /// PredicatesFoldable - Return true if both predicates match sign or if at
3481 /// least one of them is an equality comparison (which is signless).
3482 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3483 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3484 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3485 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3489 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3490 struct FoldICmpLogical {
3493 ICmpInst::Predicate pred;
3494 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3495 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3496 pred(ICI->getPredicate()) {}
3497 bool shouldApply(Value *V) const {
3498 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3499 if (PredicatesFoldable(pred, ICI->getPredicate()))
3500 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3501 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3504 Instruction *apply(Instruction &Log) const {
3505 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3506 if (ICI->getOperand(0) != LHS) {
3507 assert(ICI->getOperand(1) == LHS);
3508 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3511 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3512 unsigned LHSCode = getICmpCode(ICI);
3513 unsigned RHSCode = getICmpCode(RHSICI);
3515 switch (Log.getOpcode()) {
3516 case Instruction::And: Code = LHSCode & RHSCode; break;
3517 case Instruction::Or: Code = LHSCode | RHSCode; break;
3518 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3519 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3522 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3523 ICmpInst::isSignedPredicate(ICI->getPredicate());
3525 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3526 if (Instruction *I = dyn_cast<Instruction>(RV))
3528 // Otherwise, it's a constant boolean value...
3529 return IC.ReplaceInstUsesWith(Log, RV);
3532 } // end anonymous namespace
3534 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3535 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3536 // guaranteed to be a binary operator.
3537 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3539 ConstantInt *AndRHS,
3540 BinaryOperator &TheAnd) {
3541 Value *X = Op->getOperand(0);
3542 Constant *Together = 0;
3544 Together = Context->getConstantExprAnd(AndRHS, OpRHS);
3546 switch (Op->getOpcode()) {
3547 case Instruction::Xor:
3548 if (Op->hasOneUse()) {
3549 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3550 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3551 InsertNewInstBefore(And, TheAnd);
3553 return BinaryOperator::CreateXor(And, Together);
3556 case Instruction::Or:
3557 if (Together == AndRHS) // (X | C) & C --> C
3558 return ReplaceInstUsesWith(TheAnd, AndRHS);
3560 if (Op->hasOneUse() && Together != OpRHS) {
3561 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3562 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3563 InsertNewInstBefore(Or, TheAnd);
3565 return BinaryOperator::CreateAnd(Or, AndRHS);
3568 case Instruction::Add:
3569 if (Op->hasOneUse()) {
3570 // Adding a one to a single bit bit-field should be turned into an XOR
3571 // of the bit. First thing to check is to see if this AND is with a
3572 // single bit constant.
3573 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3575 // If there is only one bit set...
3576 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3577 // Ok, at this point, we know that we are masking the result of the
3578 // ADD down to exactly one bit. If the constant we are adding has
3579 // no bits set below this bit, then we can eliminate the ADD.
3580 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3582 // Check to see if any bits below the one bit set in AndRHSV are set.
3583 if ((AddRHS & (AndRHSV-1)) == 0) {
3584 // If not, the only thing that can effect the output of the AND is
3585 // the bit specified by AndRHSV. If that bit is set, the effect of
3586 // the XOR is to toggle the bit. If it is clear, then the ADD has
3588 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3589 TheAnd.setOperand(0, X);
3592 // Pull the XOR out of the AND.
3593 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3594 InsertNewInstBefore(NewAnd, TheAnd);
3595 NewAnd->takeName(Op);
3596 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3603 case Instruction::Shl: {
3604 // We know that the AND will not produce any of the bits shifted in, so if
3605 // the anded constant includes them, clear them now!
3607 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3608 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3609 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3610 ConstantInt *CI = Context->getConstantInt(AndRHS->getValue() & ShlMask);
3612 if (CI->getValue() == ShlMask) {
3613 // Masking out bits that the shift already masks
3614 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3615 } else if (CI != AndRHS) { // Reducing bits set in and.
3616 TheAnd.setOperand(1, CI);
3621 case Instruction::LShr:
3623 // We know that the AND will not produce any of the bits shifted in, so if
3624 // the anded constant includes them, clear them now! This only applies to
3625 // unsigned shifts, because a signed shr may bring in set bits!
3627 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3628 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3629 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3630 ConstantInt *CI = Context->getConstantInt(AndRHS->getValue() & ShrMask);
3632 if (CI->getValue() == ShrMask) {
3633 // Masking out bits that the shift already masks.
3634 return ReplaceInstUsesWith(TheAnd, Op);
3635 } else if (CI != AndRHS) {
3636 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3641 case Instruction::AShr:
3643 // See if this is shifting in some sign extension, then masking it out
3645 if (Op->hasOneUse()) {
3646 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3647 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3648 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3649 Constant *C = Context->getConstantInt(AndRHS->getValue() & ShrMask);
3650 if (C == AndRHS) { // Masking out bits shifted in.
3651 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3652 // Make the argument unsigned.
3653 Value *ShVal = Op->getOperand(0);
3654 ShVal = InsertNewInstBefore(
3655 BinaryOperator::CreateLShr(ShVal, OpRHS,
3656 Op->getName()), TheAnd);
3657 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3666 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3667 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3668 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3669 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3670 /// insert new instructions.
3671 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3672 bool isSigned, bool Inside,
3674 assert(cast<ConstantInt>(Context->getConstantExprICmp((isSigned ?
3675 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3676 "Lo is not <= Hi in range emission code!");
3679 if (Lo == Hi) // Trivially false.
3680 return new ICmpInst(*Context, ICmpInst::ICMP_NE, V, V);
3682 // V >= Min && V < Hi --> V < Hi
3683 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3684 ICmpInst::Predicate pred = (isSigned ?
3685 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3686 return new ICmpInst(*Context, pred, V, Hi);
3689 // Emit V-Lo <u Hi-Lo
3690 Constant *NegLo = Context->getConstantExprNeg(Lo);
3691 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3692 InsertNewInstBefore(Add, IB);
3693 Constant *UpperBound = Context->getConstantExprAdd(NegLo, Hi);
3694 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, UpperBound);
3697 if (Lo == Hi) // Trivially true.
3698 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, V, V);
3700 // V < Min || V >= Hi -> V > Hi-1
3701 Hi = SubOne(cast<ConstantInt>(Hi), Context);
3702 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3703 ICmpInst::Predicate pred = (isSigned ?
3704 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3705 return new ICmpInst(*Context, pred, V, Hi);
3708 // Emit V-Lo >u Hi-1-Lo
3709 // Note that Hi has already had one subtracted from it, above.
3710 ConstantInt *NegLo = cast<ConstantInt>(Context->getConstantExprNeg(Lo));
3711 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3712 InsertNewInstBefore(Add, IB);
3713 Constant *LowerBound = Context->getConstantExprAdd(NegLo, Hi);
3714 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add, LowerBound);
3717 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3718 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3719 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3720 // not, since all 1s are not contiguous.
3721 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3722 const APInt& V = Val->getValue();
3723 uint32_t BitWidth = Val->getType()->getBitWidth();
3724 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3726 // look for the first zero bit after the run of ones
3727 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3728 // look for the first non-zero bit
3729 ME = V.getActiveBits();
3733 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3734 /// where isSub determines whether the operator is a sub. If we can fold one of
3735 /// the following xforms:
3737 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3738 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3739 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3741 /// return (A +/- B).
3743 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3744 ConstantInt *Mask, bool isSub,
3746 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3747 if (!LHSI || LHSI->getNumOperands() != 2 ||
3748 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3750 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3752 switch (LHSI->getOpcode()) {
3754 case Instruction::And:
3755 if (Context->getConstantExprAnd(N, Mask) == Mask) {
3756 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3757 if ((Mask->getValue().countLeadingZeros() +
3758 Mask->getValue().countPopulation()) ==
3759 Mask->getValue().getBitWidth())
3762 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3763 // part, we don't need any explicit masks to take them out of A. If that
3764 // is all N is, ignore it.
3765 uint32_t MB = 0, ME = 0;
3766 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3767 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3768 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3769 if (MaskedValueIsZero(RHS, Mask))
3774 case Instruction::Or:
3775 case Instruction::Xor:
3776 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3777 if ((Mask->getValue().countLeadingZeros() +
3778 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3779 && Context->getConstantExprAnd(N, Mask)->isNullValue())
3786 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3788 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3789 return InsertNewInstBefore(New, I);
3792 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3793 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3794 ICmpInst *LHS, ICmpInst *RHS) {
3796 ConstantInt *LHSCst, *RHSCst;
3797 ICmpInst::Predicate LHSCC, RHSCC;
3799 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3800 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3801 m_ConstantInt(LHSCst)), *Context) ||
3802 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3803 m_ConstantInt(RHSCst)), *Context))
3806 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3807 // where C is a power of 2
3808 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3809 LHSCst->getValue().isPowerOf2()) {
3810 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3811 InsertNewInstBefore(NewOr, I);
3812 return new ICmpInst(*Context, LHSCC, NewOr, LHSCst);
3815 // From here on, we only handle:
3816 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3817 if (Val != Val2) return 0;
3819 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3820 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3821 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3822 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3823 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3826 // We can't fold (ugt x, C) & (sgt x, C2).
3827 if (!PredicatesFoldable(LHSCC, RHSCC))
3830 // Ensure that the larger constant is on the RHS.
3832 if (ICmpInst::isSignedPredicate(LHSCC) ||
3833 (ICmpInst::isEquality(LHSCC) &&
3834 ICmpInst::isSignedPredicate(RHSCC)))
3835 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3837 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3840 std::swap(LHS, RHS);
3841 std::swap(LHSCst, RHSCst);
3842 std::swap(LHSCC, RHSCC);
3845 // At this point, we know we have have two icmp instructions
3846 // comparing a value against two constants and and'ing the result
3847 // together. Because of the above check, we know that we only have
3848 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3849 // (from the FoldICmpLogical check above), that the two constants
3850 // are not equal and that the larger constant is on the RHS
3851 assert(LHSCst != RHSCst && "Compares not folded above?");
3854 default: llvm_unreachable("Unknown integer condition code!");
3855 case ICmpInst::ICMP_EQ:
3857 default: llvm_unreachable("Unknown integer condition code!");
3858 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3859 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3860 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3861 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3862 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3863 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3864 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3865 return ReplaceInstUsesWith(I, LHS);
3867 case ICmpInst::ICMP_NE:
3869 default: llvm_unreachable("Unknown integer condition code!");
3870 case ICmpInst::ICMP_ULT:
3871 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X u< 14) -> X < 13
3872 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Val, LHSCst);
3873 break; // (X != 13 & X u< 15) -> no change
3874 case ICmpInst::ICMP_SLT:
3875 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X s< 14) -> X < 13
3876 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Val, LHSCst);
3877 break; // (X != 13 & X s< 15) -> no change
3878 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3879 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3880 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3881 return ReplaceInstUsesWith(I, RHS);
3882 case ICmpInst::ICMP_NE:
3883 if (LHSCst == SubOne(RHSCst, Context)){// (X != 13 & X != 14) -> X-13 >u 1
3884 Constant *AddCST = Context->getConstantExprNeg(LHSCst);
3885 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3886 Val->getName()+".off");
3887 InsertNewInstBefore(Add, I);
3888 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add,
3889 Context->getConstantInt(Add->getType(), 1));
3891 break; // (X != 13 & X != 15) -> no change
3894 case ICmpInst::ICMP_ULT:
3896 default: llvm_unreachable("Unknown integer condition code!");
3897 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3898 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3899 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3900 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3902 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3903 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3904 return ReplaceInstUsesWith(I, LHS);
3905 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3909 case ICmpInst::ICMP_SLT:
3911 default: llvm_unreachable("Unknown integer condition code!");
3912 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3913 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3914 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3915 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3917 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3918 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3919 return ReplaceInstUsesWith(I, LHS);
3920 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3924 case ICmpInst::ICMP_UGT:
3926 default: llvm_unreachable("Unknown integer condition code!");
3927 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3928 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3929 return ReplaceInstUsesWith(I, RHS);
3930 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3932 case ICmpInst::ICMP_NE:
3933 if (RHSCst == AddOne(LHSCst, Context)) // (X u> 13 & X != 14) -> X u> 14
3934 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3935 break; // (X u> 13 & X != 15) -> no change
3936 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3937 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3938 RHSCst, false, true, I);
3939 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3943 case ICmpInst::ICMP_SGT:
3945 default: llvm_unreachable("Unknown integer condition code!");
3946 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3947 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3948 return ReplaceInstUsesWith(I, RHS);
3949 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3951 case ICmpInst::ICMP_NE:
3952 if (RHSCst == AddOne(LHSCst, Context)) // (X s> 13 & X != 14) -> X s> 14
3953 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3954 break; // (X s> 13 & X != 15) -> no change
3955 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3956 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3957 RHSCst, true, true, I);
3958 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3968 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3969 bool Changed = SimplifyCommutative(I);
3970 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3972 if (isa<UndefValue>(Op1)) // X & undef -> 0
3973 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3977 return ReplaceInstUsesWith(I, Op1);
3979 // See if we can simplify any instructions used by the instruction whose sole
3980 // purpose is to compute bits we don't care about.
3981 if (SimplifyDemandedInstructionBits(I))
3983 if (isa<VectorType>(I.getType())) {
3984 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3985 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3986 return ReplaceInstUsesWith(I, I.getOperand(0));
3987 } else if (isa<ConstantAggregateZero>(Op1)) {
3988 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3992 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3993 const APInt& AndRHSMask = AndRHS->getValue();
3994 APInt NotAndRHS(~AndRHSMask);
3996 // Optimize a variety of ((val OP C1) & C2) combinations...
3997 if (isa<BinaryOperator>(Op0)) {
3998 Instruction *Op0I = cast<Instruction>(Op0);
3999 Value *Op0LHS = Op0I->getOperand(0);
4000 Value *Op0RHS = Op0I->getOperand(1);
4001 switch (Op0I->getOpcode()) {
4002 case Instruction::Xor:
4003 case Instruction::Or:
4004 // If the mask is only needed on one incoming arm, push it up.
4005 if (Op0I->hasOneUse()) {
4006 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
4007 // Not masking anything out for the LHS, move to RHS.
4008 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4009 Op0RHS->getName()+".masked");
4010 InsertNewInstBefore(NewRHS, I);
4011 return BinaryOperator::Create(
4012 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4014 if (!isa<Constant>(Op0RHS) &&
4015 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4016 // Not masking anything out for the RHS, move to LHS.
4017 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4018 Op0LHS->getName()+".masked");
4019 InsertNewInstBefore(NewLHS, I);
4020 return BinaryOperator::Create(
4021 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4026 case Instruction::Add:
4027 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4028 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4029 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4030 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4031 return BinaryOperator::CreateAnd(V, AndRHS);
4032 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4033 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4036 case Instruction::Sub:
4037 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4038 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4039 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4040 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4041 return BinaryOperator::CreateAnd(V, AndRHS);
4043 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4044 // has 1's for all bits that the subtraction with A might affect.
4045 if (Op0I->hasOneUse()) {
4046 uint32_t BitWidth = AndRHSMask.getBitWidth();
4047 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4048 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4050 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4051 if (!(A && A->isZero()) && // avoid infinite recursion.
4052 MaskedValueIsZero(Op0LHS, Mask)) {
4053 Instruction *NewNeg = BinaryOperator::CreateNeg(*Context, Op0RHS);
4054 InsertNewInstBefore(NewNeg, I);
4055 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4060 case Instruction::Shl:
4061 case Instruction::LShr:
4062 // (1 << x) & 1 --> zext(x == 0)
4063 // (1 >> x) & 1 --> zext(x == 0)
4064 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4065 Instruction *NewICmp = new ICmpInst(*Context, ICmpInst::ICMP_EQ,
4066 Op0RHS, Context->getNullValue(I.getType()));
4067 InsertNewInstBefore(NewICmp, I);
4068 return new ZExtInst(NewICmp, I.getType());
4073 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4074 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4076 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4077 // If this is an integer truncation or change from signed-to-unsigned, and
4078 // if the source is an and/or with immediate, transform it. This
4079 // frequently occurs for bitfield accesses.
4080 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4081 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4082 CastOp->getNumOperands() == 2)
4083 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4084 if (CastOp->getOpcode() == Instruction::And) {
4085 // Change: and (cast (and X, C1) to T), C2
4086 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4087 // This will fold the two constants together, which may allow
4088 // other simplifications.
4089 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4090 CastOp->getOperand(0), I.getType(),
4091 CastOp->getName()+".shrunk");
4092 NewCast = InsertNewInstBefore(NewCast, I);
4093 // trunc_or_bitcast(C1)&C2
4095 Context->getConstantExprTruncOrBitCast(AndCI,I.getType());
4096 C3 = Context->getConstantExprAnd(C3, AndRHS);
4097 return BinaryOperator::CreateAnd(NewCast, C3);
4098 } else if (CastOp->getOpcode() == Instruction::Or) {
4099 // Change: and (cast (or X, C1) to T), C2
4100 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4102 Context->getConstantExprTruncOrBitCast(AndCI,I.getType());
4103 if (Context->getConstantExprAnd(C3, AndRHS) == AndRHS)
4105 return ReplaceInstUsesWith(I, AndRHS);
4111 // Try to fold constant and into select arguments.
4112 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4113 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4115 if (isa<PHINode>(Op0))
4116 if (Instruction *NV = FoldOpIntoPhi(I))
4120 Value *Op0NotVal = dyn_castNotVal(Op0, Context);
4121 Value *Op1NotVal = dyn_castNotVal(Op1, Context);
4123 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4124 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
4126 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4127 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4128 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4129 I.getName()+".demorgan");
4130 InsertNewInstBefore(Or, I);
4131 return BinaryOperator::CreateNot(*Context, Or);
4135 Value *A = 0, *B = 0, *C = 0, *D = 0;
4136 if (match(Op0, m_Or(m_Value(A), m_Value(B)), *Context)) {
4137 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4138 return ReplaceInstUsesWith(I, Op1);
4140 // (A|B) & ~(A&B) -> A^B
4141 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4142 if ((A == C && B == D) || (A == D && B == C))
4143 return BinaryOperator::CreateXor(A, B);
4147 if (match(Op1, m_Or(m_Value(A), m_Value(B)), *Context)) {
4148 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4149 return ReplaceInstUsesWith(I, Op0);
4151 // ~(A&B) & (A|B) -> A^B
4152 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4153 if ((A == C && B == D) || (A == D && B == C))
4154 return BinaryOperator::CreateXor(A, B);
4158 if (Op0->hasOneUse() &&
4159 match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4160 if (A == Op1) { // (A^B)&A -> A&(A^B)
4161 I.swapOperands(); // Simplify below
4162 std::swap(Op0, Op1);
4163 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4164 cast<BinaryOperator>(Op0)->swapOperands();
4165 I.swapOperands(); // Simplify below
4166 std::swap(Op0, Op1);
4170 if (Op1->hasOneUse() &&
4171 match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4172 if (B == Op0) { // B&(A^B) -> B&(B^A)
4173 cast<BinaryOperator>(Op1)->swapOperands();
4176 if (A == Op0) { // A&(A^B) -> A & ~B
4177 Instruction *NotB = BinaryOperator::CreateNot(*Context, B, "tmp");
4178 InsertNewInstBefore(NotB, I);
4179 return BinaryOperator::CreateAnd(A, NotB);
4183 // (A&((~A)|B)) -> A&B
4184 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A)), *Context) ||
4185 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1))), *Context))
4186 return BinaryOperator::CreateAnd(A, Op1);
4187 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A)), *Context) ||
4188 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0))), *Context))
4189 return BinaryOperator::CreateAnd(A, Op0);
4192 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4193 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4194 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4197 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4198 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4202 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4203 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4204 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4205 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4206 const Type *SrcTy = Op0C->getOperand(0)->getType();
4207 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4208 // Only do this if the casts both really cause code to be generated.
4209 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4211 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4213 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4214 Op1C->getOperand(0),
4216 InsertNewInstBefore(NewOp, I);
4217 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4221 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4222 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4223 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4224 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4225 SI0->getOperand(1) == SI1->getOperand(1) &&
4226 (SI0->hasOneUse() || SI1->hasOneUse())) {
4227 Instruction *NewOp =
4228 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4230 SI0->getName()), I);
4231 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4232 SI1->getOperand(1));
4236 // If and'ing two fcmp, try combine them into one.
4237 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4238 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4239 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4240 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4241 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4242 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4243 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4244 // If either of the constants are nans, then the whole thing returns
4246 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4247 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4248 return new FCmpInst(*Context, FCmpInst::FCMP_ORD,
4249 LHS->getOperand(0), RHS->getOperand(0));
4252 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4253 FCmpInst::Predicate Op0CC, Op1CC;
4254 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS),
4255 m_Value(Op0RHS)), *Context) &&
4256 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS),
4257 m_Value(Op1RHS)), *Context)) {
4258 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4259 // Swap RHS operands to match LHS.
4260 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4261 std::swap(Op1LHS, Op1RHS);
4263 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4264 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4266 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4268 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4269 Op1CC == FCmpInst::FCMP_FALSE)
4270 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4271 else if (Op0CC == FCmpInst::FCMP_TRUE)
4272 return ReplaceInstUsesWith(I, Op1);
4273 else if (Op1CC == FCmpInst::FCMP_TRUE)
4274 return ReplaceInstUsesWith(I, Op0);
4277 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4278 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4280 std::swap(Op0, Op1);
4281 std::swap(Op0Pred, Op1Pred);
4282 std::swap(Op0Ordered, Op1Ordered);
4285 // uno && ueq -> uno && (uno || eq) -> ueq
4286 // ord && olt -> ord && (ord && lt) -> olt
4287 if (Op0Ordered == Op1Ordered)
4288 return ReplaceInstUsesWith(I, Op1);
4289 // uno && oeq -> uno && (ord && eq) -> false
4290 // uno && ord -> false
4292 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4293 // ord && ueq -> ord && (uno || eq) -> oeq
4294 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4295 Op0LHS, Op0RHS, Context));
4303 return Changed ? &I : 0;
4306 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4307 /// capable of providing pieces of a bswap. The subexpression provides pieces
4308 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4309 /// the expression came from the corresponding "byte swapped" byte in some other
4310 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4311 /// we know that the expression deposits the low byte of %X into the high byte
4312 /// of the bswap result and that all other bytes are zero. This expression is
4313 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4316 /// This function returns true if the match was unsuccessful and false if so.
4317 /// On entry to the function the "OverallLeftShift" is a signed integer value
4318 /// indicating the number of bytes that the subexpression is later shifted. For
4319 /// example, if the expression is later right shifted by 16 bits, the
4320 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4321 /// byte of ByteValues is actually being set.
4323 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4324 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4325 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4326 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4327 /// always in the local (OverallLeftShift) coordinate space.
4329 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4330 SmallVector<Value*, 8> &ByteValues) {
4331 if (Instruction *I = dyn_cast<Instruction>(V)) {
4332 // If this is an or instruction, it may be an inner node of the bswap.
4333 if (I->getOpcode() == Instruction::Or) {
4334 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4336 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4340 // If this is a logical shift by a constant multiple of 8, recurse with
4341 // OverallLeftShift and ByteMask adjusted.
4342 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4344 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4345 // Ensure the shift amount is defined and of a byte value.
4346 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4349 unsigned ByteShift = ShAmt >> 3;
4350 if (I->getOpcode() == Instruction::Shl) {
4351 // X << 2 -> collect(X, +2)
4352 OverallLeftShift += ByteShift;
4353 ByteMask >>= ByteShift;
4355 // X >>u 2 -> collect(X, -2)
4356 OverallLeftShift -= ByteShift;
4357 ByteMask <<= ByteShift;
4358 ByteMask &= (~0U >> (32-ByteValues.size()));
4361 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4362 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4364 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4368 // If this is a logical 'and' with a mask that clears bytes, clear the
4369 // corresponding bytes in ByteMask.
4370 if (I->getOpcode() == Instruction::And &&
4371 isa<ConstantInt>(I->getOperand(1))) {
4372 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4373 unsigned NumBytes = ByteValues.size();
4374 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4375 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4377 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4378 // If this byte is masked out by a later operation, we don't care what
4380 if ((ByteMask & (1 << i)) == 0)
4383 // If the AndMask is all zeros for this byte, clear the bit.
4384 APInt MaskB = AndMask & Byte;
4386 ByteMask &= ~(1U << i);
4390 // If the AndMask is not all ones for this byte, it's not a bytezap.
4394 // Otherwise, this byte is kept.
4397 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4402 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4403 // the input value to the bswap. Some observations: 1) if more than one byte
4404 // is demanded from this input, then it could not be successfully assembled
4405 // into a byteswap. At least one of the two bytes would not be aligned with
4406 // their ultimate destination.
4407 if (!isPowerOf2_32(ByteMask)) return true;
4408 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4410 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4411 // is demanded, it needs to go into byte 0 of the result. This means that the
4412 // byte needs to be shifted until it lands in the right byte bucket. The
4413 // shift amount depends on the position: if the byte is coming from the high
4414 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4415 // low part, it must be shifted left.
4416 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4417 if (InputByteNo < ByteValues.size()/2) {
4418 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4421 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4425 // If the destination byte value is already defined, the values are or'd
4426 // together, which isn't a bswap (unless it's an or of the same bits).
4427 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4429 ByteValues[DestByteNo] = V;
4433 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4434 /// If so, insert the new bswap intrinsic and return it.
4435 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4436 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4437 if (!ITy || ITy->getBitWidth() % 16 ||
4438 // ByteMask only allows up to 32-byte values.
4439 ITy->getBitWidth() > 32*8)
4440 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4442 /// ByteValues - For each byte of the result, we keep track of which value
4443 /// defines each byte.
4444 SmallVector<Value*, 8> ByteValues;
4445 ByteValues.resize(ITy->getBitWidth()/8);
4447 // Try to find all the pieces corresponding to the bswap.
4448 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4449 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4452 // Check to see if all of the bytes come from the same value.
4453 Value *V = ByteValues[0];
4454 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4456 // Check to make sure that all of the bytes come from the same value.
4457 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4458 if (ByteValues[i] != V)
4460 const Type *Tys[] = { ITy };
4461 Module *M = I.getParent()->getParent()->getParent();
4462 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4463 return CallInst::Create(F, V);
4466 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4467 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4468 /// we can simplify this expression to "cond ? C : D or B".
4469 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4471 LLVMContext *Context) {
4472 // If A is not a select of -1/0, this cannot match.
4474 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond)), *Context))
4477 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4478 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4479 return SelectInst::Create(Cond, C, B);
4480 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4481 return SelectInst::Create(Cond, C, B);
4482 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4483 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4484 return SelectInst::Create(Cond, C, D);
4485 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4486 return SelectInst::Create(Cond, C, D);
4490 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4491 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4492 ICmpInst *LHS, ICmpInst *RHS) {
4494 ConstantInt *LHSCst, *RHSCst;
4495 ICmpInst::Predicate LHSCC, RHSCC;
4497 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4498 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4499 m_ConstantInt(LHSCst)), *Context) ||
4500 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4501 m_ConstantInt(RHSCst)), *Context))
4504 // From here on, we only handle:
4505 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4506 if (Val != Val2) return 0;
4508 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4509 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4510 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4511 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4512 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4515 // We can't fold (ugt x, C) | (sgt x, C2).
4516 if (!PredicatesFoldable(LHSCC, RHSCC))
4519 // Ensure that the larger constant is on the RHS.
4521 if (ICmpInst::isSignedPredicate(LHSCC) ||
4522 (ICmpInst::isEquality(LHSCC) &&
4523 ICmpInst::isSignedPredicate(RHSCC)))
4524 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4526 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4529 std::swap(LHS, RHS);
4530 std::swap(LHSCst, RHSCst);
4531 std::swap(LHSCC, RHSCC);
4534 // At this point, we know we have have two icmp instructions
4535 // comparing a value against two constants and or'ing the result
4536 // together. Because of the above check, we know that we only have
4537 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4538 // FoldICmpLogical check above), that the two constants are not
4540 assert(LHSCst != RHSCst && "Compares not folded above?");
4543 default: llvm_unreachable("Unknown integer condition code!");
4544 case ICmpInst::ICMP_EQ:
4546 default: llvm_unreachable("Unknown integer condition code!");
4547 case ICmpInst::ICMP_EQ:
4548 if (LHSCst == SubOne(RHSCst, Context)) {
4549 // (X == 13 | X == 14) -> X-13 <u 2
4550 Constant *AddCST = Context->getConstantExprNeg(LHSCst);
4551 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4552 Val->getName()+".off");
4553 InsertNewInstBefore(Add, I);
4554 AddCST = Context->getConstantExprSub(AddOne(RHSCst, Context), LHSCst);
4555 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, AddCST);
4557 break; // (X == 13 | X == 15) -> no change
4558 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4559 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4561 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4562 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4563 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4564 return ReplaceInstUsesWith(I, RHS);
4567 case ICmpInst::ICMP_NE:
4569 default: llvm_unreachable("Unknown integer condition code!");
4570 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4571 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4572 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4573 return ReplaceInstUsesWith(I, LHS);
4574 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4575 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4576 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4577 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4580 case ICmpInst::ICMP_ULT:
4582 default: llvm_unreachable("Unknown integer condition code!");
4583 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4585 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4586 // If RHSCst is [us]MAXINT, it is always false. Not handling
4587 // this can cause overflow.
4588 if (RHSCst->isMaxValue(false))
4589 return ReplaceInstUsesWith(I, LHS);
4590 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4592 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4594 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4595 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4596 return ReplaceInstUsesWith(I, RHS);
4597 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4601 case ICmpInst::ICMP_SLT:
4603 default: llvm_unreachable("Unknown integer condition code!");
4604 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4606 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4607 // If RHSCst is [us]MAXINT, it is always false. Not handling
4608 // this can cause overflow.
4609 if (RHSCst->isMaxValue(true))
4610 return ReplaceInstUsesWith(I, LHS);
4611 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4613 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4615 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4616 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4617 return ReplaceInstUsesWith(I, RHS);
4618 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4622 case ICmpInst::ICMP_UGT:
4624 default: llvm_unreachable("Unknown integer condition code!");
4625 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4626 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4627 return ReplaceInstUsesWith(I, LHS);
4628 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4630 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4631 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4632 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4633 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4637 case ICmpInst::ICMP_SGT:
4639 default: llvm_unreachable("Unknown integer condition code!");
4640 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4641 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4642 return ReplaceInstUsesWith(I, LHS);
4643 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4645 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4646 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4647 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4648 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4656 /// FoldOrWithConstants - This helper function folds:
4658 /// ((A | B) & C1) | (B & C2)
4664 /// when the XOR of the two constants is "all ones" (-1).
4665 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4666 Value *A, Value *B, Value *C) {
4667 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4671 ConstantInt *CI2 = 0;
4672 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)), *Context)) return 0;
4674 APInt Xor = CI1->getValue() ^ CI2->getValue();
4675 if (!Xor.isAllOnesValue()) return 0;
4677 if (V1 == A || V1 == B) {
4678 Instruction *NewOp =
4679 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4680 return BinaryOperator::CreateOr(NewOp, V1);
4686 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4687 bool Changed = SimplifyCommutative(I);
4688 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4690 if (isa<UndefValue>(Op1)) // X | undef -> -1
4691 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4695 return ReplaceInstUsesWith(I, Op0);
4697 // See if we can simplify any instructions used by the instruction whose sole
4698 // purpose is to compute bits we don't care about.
4699 if (SimplifyDemandedInstructionBits(I))
4701 if (isa<VectorType>(I.getType())) {
4702 if (isa<ConstantAggregateZero>(Op1)) {
4703 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4704 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4705 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4706 return ReplaceInstUsesWith(I, I.getOperand(1));
4711 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4712 ConstantInt *C1 = 0; Value *X = 0;
4713 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4714 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1)), *Context) &&
4716 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4717 InsertNewInstBefore(Or, I);
4719 return BinaryOperator::CreateAnd(Or,
4720 Context->getConstantInt(RHS->getValue() | C1->getValue()));
4723 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4724 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1)), *Context) &&
4726 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4727 InsertNewInstBefore(Or, I);
4729 return BinaryOperator::CreateXor(Or,
4730 Context->getConstantInt(C1->getValue() & ~RHS->getValue()));
4733 // Try to fold constant and into select arguments.
4734 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4735 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4737 if (isa<PHINode>(Op0))
4738 if (Instruction *NV = FoldOpIntoPhi(I))
4742 Value *A = 0, *B = 0;
4743 ConstantInt *C1 = 0, *C2 = 0;
4745 if (match(Op0, m_And(m_Value(A), m_Value(B)), *Context))
4746 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4747 return ReplaceInstUsesWith(I, Op1);
4748 if (match(Op1, m_And(m_Value(A), m_Value(B)), *Context))
4749 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4750 return ReplaceInstUsesWith(I, Op0);
4752 // (A | B) | C and A | (B | C) -> bswap if possible.
4753 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4754 if (match(Op0, m_Or(m_Value(), m_Value()), *Context) ||
4755 match(Op1, m_Or(m_Value(), m_Value()), *Context) ||
4756 (match(Op0, m_Shift(m_Value(), m_Value()), *Context) &&
4757 match(Op1, m_Shift(m_Value(), m_Value()), *Context))) {
4758 if (Instruction *BSwap = MatchBSwap(I))
4762 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4763 if (Op0->hasOneUse() &&
4764 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4765 MaskedValueIsZero(Op1, C1->getValue())) {
4766 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4767 InsertNewInstBefore(NOr, I);
4769 return BinaryOperator::CreateXor(NOr, C1);
4772 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4773 if (Op1->hasOneUse() &&
4774 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4775 MaskedValueIsZero(Op0, C1->getValue())) {
4776 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4777 InsertNewInstBefore(NOr, I);
4779 return BinaryOperator::CreateXor(NOr, C1);
4783 Value *C = 0, *D = 0;
4784 if (match(Op0, m_And(m_Value(A), m_Value(C)), *Context) &&
4785 match(Op1, m_And(m_Value(B), m_Value(D)), *Context)) {
4786 Value *V1 = 0, *V2 = 0, *V3 = 0;
4787 C1 = dyn_cast<ConstantInt>(C);
4788 C2 = dyn_cast<ConstantInt>(D);
4789 if (C1 && C2) { // (A & C1)|(B & C2)
4790 // If we have: ((V + N) & C1) | (V & C2)
4791 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4792 // replace with V+N.
4793 if (C1->getValue() == ~C2->getValue()) {
4794 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4795 match(A, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4796 // Add commutes, try both ways.
4797 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4798 return ReplaceInstUsesWith(I, A);
4799 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4800 return ReplaceInstUsesWith(I, A);
4802 // Or commutes, try both ways.
4803 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4804 match(B, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4805 // Add commutes, try both ways.
4806 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4807 return ReplaceInstUsesWith(I, B);
4808 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4809 return ReplaceInstUsesWith(I, B);
4812 V1 = 0; V2 = 0; V3 = 0;
4815 // Check to see if we have any common things being and'ed. If so, find the
4816 // terms for V1 & (V2|V3).
4817 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4818 if (A == B) // (A & C)|(A & D) == A & (C|D)
4819 V1 = A, V2 = C, V3 = D;
4820 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4821 V1 = A, V2 = B, V3 = C;
4822 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4823 V1 = C, V2 = A, V3 = D;
4824 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4825 V1 = C, V2 = A, V3 = B;
4829 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4830 return BinaryOperator::CreateAnd(V1, Or);
4834 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4835 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4837 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4839 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4841 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4844 // ((A&~B)|(~A&B)) -> A^B
4845 if ((match(C, m_Not(m_Specific(D)), *Context) &&
4846 match(B, m_Not(m_Specific(A)), *Context)))
4847 return BinaryOperator::CreateXor(A, D);
4848 // ((~B&A)|(~A&B)) -> A^B
4849 if ((match(A, m_Not(m_Specific(D)), *Context) &&
4850 match(B, m_Not(m_Specific(C)), *Context)))
4851 return BinaryOperator::CreateXor(C, D);
4852 // ((A&~B)|(B&~A)) -> A^B
4853 if ((match(C, m_Not(m_Specific(B)), *Context) &&
4854 match(D, m_Not(m_Specific(A)), *Context)))
4855 return BinaryOperator::CreateXor(A, B);
4856 // ((~B&A)|(B&~A)) -> A^B
4857 if ((match(A, m_Not(m_Specific(B)), *Context) &&
4858 match(D, m_Not(m_Specific(C)), *Context)))
4859 return BinaryOperator::CreateXor(C, B);
4862 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4863 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4864 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4865 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4866 SI0->getOperand(1) == SI1->getOperand(1) &&
4867 (SI0->hasOneUse() || SI1->hasOneUse())) {
4868 Instruction *NewOp =
4869 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4871 SI0->getName()), I);
4872 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4873 SI1->getOperand(1));
4877 // ((A|B)&1)|(B&-2) -> (A&1) | B
4878 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4879 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4880 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4881 if (Ret) return Ret;
4883 // (B&-2)|((A|B)&1) -> (A&1) | B
4884 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4885 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4886 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4887 if (Ret) return Ret;
4890 if (match(Op0, m_Not(m_Value(A)), *Context)) { // ~A | Op1
4891 if (A == Op1) // ~A | A == -1
4892 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4896 // Note, A is still live here!
4897 if (match(Op1, m_Not(m_Value(B)), *Context)) { // Op0 | ~B
4899 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4901 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4902 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4903 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4904 I.getName()+".demorgan"), I);
4905 return BinaryOperator::CreateNot(*Context, And);
4909 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4910 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4911 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4914 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4915 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4919 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4920 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4921 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4922 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4923 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4924 !isa<ICmpInst>(Op1C->getOperand(0))) {
4925 const Type *SrcTy = Op0C->getOperand(0)->getType();
4926 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4927 // Only do this if the casts both really cause code to be
4929 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4931 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4933 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4934 Op1C->getOperand(0),
4936 InsertNewInstBefore(NewOp, I);
4937 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4944 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4945 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4946 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4947 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4948 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4949 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4950 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4951 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4952 // If either of the constants are nans, then the whole thing returns
4954 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4955 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4957 // Otherwise, no need to compare the two constants, compare the
4959 return new FCmpInst(*Context, FCmpInst::FCMP_UNO,
4960 LHS->getOperand(0), RHS->getOperand(0));
4963 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4964 FCmpInst::Predicate Op0CC, Op1CC;
4965 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS),
4966 m_Value(Op0RHS)), *Context) &&
4967 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS),
4968 m_Value(Op1RHS)), *Context)) {
4969 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4970 // Swap RHS operands to match LHS.
4971 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4972 std::swap(Op1LHS, Op1RHS);
4974 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4975 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4977 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4979 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4980 Op1CC == FCmpInst::FCMP_TRUE)
4981 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4982 else if (Op0CC == FCmpInst::FCMP_FALSE)
4983 return ReplaceInstUsesWith(I, Op1);
4984 else if (Op1CC == FCmpInst::FCMP_FALSE)
4985 return ReplaceInstUsesWith(I, Op0);
4988 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4989 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4990 if (Op0Ordered == Op1Ordered) {
4991 // If both are ordered or unordered, return a new fcmp with
4992 // or'ed predicates.
4993 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4994 Op0LHS, Op0RHS, Context);
4995 if (Instruction *I = dyn_cast<Instruction>(RV))
4997 // Otherwise, it's a constant boolean value...
4998 return ReplaceInstUsesWith(I, RV);
5006 return Changed ? &I : 0;
5011 // XorSelf - Implements: X ^ X --> 0
5014 XorSelf(Value *rhs) : RHS(rhs) {}
5015 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5016 Instruction *apply(BinaryOperator &Xor) const {
5023 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5024 bool Changed = SimplifyCommutative(I);
5025 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5027 if (isa<UndefValue>(Op1)) {
5028 if (isa<UndefValue>(Op0))
5029 // Handle undef ^ undef -> 0 special case. This is a common
5031 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
5032 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5035 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5036 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1), Context)) {
5037 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5038 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
5041 // See if we can simplify any instructions used by the instruction whose sole
5042 // purpose is to compute bits we don't care about.
5043 if (SimplifyDemandedInstructionBits(I))
5045 if (isa<VectorType>(I.getType()))
5046 if (isa<ConstantAggregateZero>(Op1))
5047 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5049 // Is this a ~ operation?
5050 if (Value *NotOp = dyn_castNotVal(&I, Context)) {
5051 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5052 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5053 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5054 if (Op0I->getOpcode() == Instruction::And ||
5055 Op0I->getOpcode() == Instruction::Or) {
5056 if (dyn_castNotVal(Op0I->getOperand(1), Context)) Op0I->swapOperands();
5057 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0), Context)) {
5059 BinaryOperator::CreateNot(*Context, Op0I->getOperand(1),
5060 Op0I->getOperand(1)->getName()+".not");
5061 InsertNewInstBefore(NotY, I);
5062 if (Op0I->getOpcode() == Instruction::And)
5063 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5065 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5072 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5073 if (RHS == Context->getConstantIntTrue() && Op0->hasOneUse()) {
5074 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5075 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5076 return new ICmpInst(*Context, ICI->getInversePredicate(),
5077 ICI->getOperand(0), ICI->getOperand(1));
5079 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5080 return new FCmpInst(*Context, FCI->getInversePredicate(),
5081 FCI->getOperand(0), FCI->getOperand(1));
5084 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5085 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5086 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5087 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5088 Instruction::CastOps Opcode = Op0C->getOpcode();
5089 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
5090 if (RHS == Context->getConstantExprCast(Opcode,
5091 Context->getConstantIntTrue(),
5092 Op0C->getDestTy())) {
5093 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
5095 CI->getOpcode(), CI->getInversePredicate(),
5096 CI->getOperand(0), CI->getOperand(1)), I);
5097 NewCI->takeName(CI);
5098 return CastInst::Create(Opcode, NewCI, Op0C->getType());
5105 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5106 // ~(c-X) == X-c-1 == X+(-c-1)
5107 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5108 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5109 Constant *NegOp0I0C = Context->getConstantExprNeg(Op0I0C);
5110 Constant *ConstantRHS = Context->getConstantExprSub(NegOp0I0C,
5111 Context->getConstantInt(I.getType(), 1));
5112 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5115 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5116 if (Op0I->getOpcode() == Instruction::Add) {
5117 // ~(X-c) --> (-c-1)-X
5118 if (RHS->isAllOnesValue()) {
5119 Constant *NegOp0CI = Context->getConstantExprNeg(Op0CI);
5120 return BinaryOperator::CreateSub(
5121 Context->getConstantExprSub(NegOp0CI,
5122 Context->getConstantInt(I.getType(), 1)),
5123 Op0I->getOperand(0));
5124 } else if (RHS->getValue().isSignBit()) {
5125 // (X + C) ^ signbit -> (X + C + signbit)
5127 Context->getConstantInt(RHS->getValue() + Op0CI->getValue());
5128 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5131 } else if (Op0I->getOpcode() == Instruction::Or) {
5132 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5133 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5134 Constant *NewRHS = Context->getConstantExprOr(Op0CI, RHS);
5135 // Anything in both C1 and C2 is known to be zero, remove it from
5137 Constant *CommonBits = Context->getConstantExprAnd(Op0CI, RHS);
5138 NewRHS = Context->getConstantExprAnd(NewRHS,
5139 Context->getConstantExprNot(CommonBits));
5140 AddToWorkList(Op0I);
5141 I.setOperand(0, Op0I->getOperand(0));
5142 I.setOperand(1, NewRHS);
5149 // Try to fold constant and into select arguments.
5150 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5151 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5153 if (isa<PHINode>(Op0))
5154 if (Instruction *NV = FoldOpIntoPhi(I))
5158 if (Value *X = dyn_castNotVal(Op0, Context)) // ~A ^ A == -1
5160 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
5162 if (Value *X = dyn_castNotVal(Op1, Context)) // A ^ ~A == -1
5164 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
5167 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5170 if (match(Op1I, m_Or(m_Value(A), m_Value(B)), *Context)) {
5171 if (A == Op0) { // B^(B|A) == (A|B)^B
5172 Op1I->swapOperands();
5174 std::swap(Op0, Op1);
5175 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5176 I.swapOperands(); // Simplified below.
5177 std::swap(Op0, Op1);
5179 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)), *Context)) {
5180 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5181 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)), *Context)) {
5182 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5183 } else if (match(Op1I, m_And(m_Value(A), m_Value(B)), *Context) &&
5185 if (A == Op0) { // A^(A&B) -> A^(B&A)
5186 Op1I->swapOperands();
5189 if (B == Op0) { // A^(B&A) -> (B&A)^A
5190 I.swapOperands(); // Simplified below.
5191 std::swap(Op0, Op1);
5196 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5199 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5200 Op0I->hasOneUse()) {
5201 if (A == Op1) // (B|A)^B == (A|B)^B
5203 if (B == Op1) { // (A|B)^B == A & ~B
5205 InsertNewInstBefore(BinaryOperator::CreateNot(*Context,
5207 return BinaryOperator::CreateAnd(A, NotB);
5209 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)), *Context)) {
5210 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5211 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)), *Context)) {
5212 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5213 } else if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5215 if (A == Op1) // (A&B)^A -> (B&A)^A
5217 if (B == Op1 && // (B&A)^A == ~B & A
5218 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5220 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, A, "tmp"), I);
5221 return BinaryOperator::CreateAnd(N, Op1);
5226 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5227 if (Op0I && Op1I && Op0I->isShift() &&
5228 Op0I->getOpcode() == Op1I->getOpcode() &&
5229 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5230 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5231 Instruction *NewOp =
5232 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5233 Op1I->getOperand(0),
5234 Op0I->getName()), I);
5235 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5236 Op1I->getOperand(1));
5240 Value *A, *B, *C, *D;
5241 // (A & B)^(A | B) -> A ^ B
5242 if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5243 match(Op1I, m_Or(m_Value(C), m_Value(D)), *Context)) {
5244 if ((A == C && B == D) || (A == D && B == C))
5245 return BinaryOperator::CreateXor(A, B);
5247 // (A | B)^(A & B) -> A ^ B
5248 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5249 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5250 if ((A == C && B == D) || (A == D && B == C))
5251 return BinaryOperator::CreateXor(A, B);
5255 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5256 match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5257 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5258 // (X & Y)^(X & Y) -> (Y^Z) & X
5259 Value *X = 0, *Y = 0, *Z = 0;
5261 X = A, Y = B, Z = D;
5263 X = A, Y = B, Z = C;
5265 X = B, Y = A, Z = D;
5267 X = B, Y = A, Z = C;
5270 Instruction *NewOp =
5271 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5272 return BinaryOperator::CreateAnd(NewOp, X);
5277 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5278 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5279 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
5282 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5283 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5284 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5285 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5286 const Type *SrcTy = Op0C->getOperand(0)->getType();
5287 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5288 // Only do this if the casts both really cause code to be generated.
5289 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5291 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5293 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5294 Op1C->getOperand(0),
5296 InsertNewInstBefore(NewOp, I);
5297 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5302 return Changed ? &I : 0;
5305 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5306 LLVMContext *Context) {
5307 return cast<ConstantInt>(Context->getConstantExprExtractElement(V, Idx));
5310 static bool HasAddOverflow(ConstantInt *Result,
5311 ConstantInt *In1, ConstantInt *In2,
5314 if (In2->getValue().isNegative())
5315 return Result->getValue().sgt(In1->getValue());
5317 return Result->getValue().slt(In1->getValue());
5319 return Result->getValue().ult(In1->getValue());
5322 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5323 /// overflowed for this type.
5324 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5325 Constant *In2, LLVMContext *Context,
5326 bool IsSigned = false) {
5327 Result = Context->getConstantExprAdd(In1, In2);
5329 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5330 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5331 Constant *Idx = Context->getConstantInt(Type::Int32Ty, i);
5332 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5333 ExtractElement(In1, Idx, Context),
5334 ExtractElement(In2, Idx, Context),
5341 return HasAddOverflow(cast<ConstantInt>(Result),
5342 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5346 static bool HasSubOverflow(ConstantInt *Result,
5347 ConstantInt *In1, ConstantInt *In2,
5350 if (In2->getValue().isNegative())
5351 return Result->getValue().slt(In1->getValue());
5353 return Result->getValue().sgt(In1->getValue());
5355 return Result->getValue().ugt(In1->getValue());
5358 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5359 /// overflowed for this type.
5360 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5361 Constant *In2, LLVMContext *Context,
5362 bool IsSigned = false) {
5363 Result = Context->getConstantExprSub(In1, In2);
5365 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5366 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5367 Constant *Idx = Context->getConstantInt(Type::Int32Ty, i);
5368 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5369 ExtractElement(In1, Idx, Context),
5370 ExtractElement(In2, Idx, Context),
5377 return HasSubOverflow(cast<ConstantInt>(Result),
5378 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5382 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5383 /// code necessary to compute the offset from the base pointer (without adding
5384 /// in the base pointer). Return the result as a signed integer of intptr size.
5385 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5386 TargetData &TD = IC.getTargetData();
5387 gep_type_iterator GTI = gep_type_begin(GEP);
5388 const Type *IntPtrTy = TD.getIntPtrType();
5389 LLVMContext *Context = IC.getContext();
5390 Value *Result = Context->getNullValue(IntPtrTy);
5392 // Build a mask for high order bits.
5393 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5394 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5396 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5399 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5400 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5401 if (OpC->isZero()) continue;
5403 // Handle a struct index, which adds its field offset to the pointer.
5404 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5405 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5407 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5409 Context->getConstantInt(RC->getValue() + APInt(IntPtrWidth, Size));
5411 Result = IC.InsertNewInstBefore(
5412 BinaryOperator::CreateAdd(Result,
5413 Context->getConstantInt(IntPtrTy, Size),
5414 GEP->getName()+".offs"), I);
5418 Constant *Scale = Context->getConstantInt(IntPtrTy, Size);
5420 Context->getConstantExprIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5421 Scale = Context->getConstantExprMul(OC, Scale);
5422 if (Constant *RC = dyn_cast<Constant>(Result))
5423 Result = Context->getConstantExprAdd(RC, Scale);
5425 // Emit an add instruction.
5426 Result = IC.InsertNewInstBefore(
5427 BinaryOperator::CreateAdd(Result, Scale,
5428 GEP->getName()+".offs"), I);
5432 // Convert to correct type.
5433 if (Op->getType() != IntPtrTy) {
5434 if (Constant *OpC = dyn_cast<Constant>(Op))
5435 Op = Context->getConstantExprIntegerCast(OpC, IntPtrTy, true);
5437 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5439 Op->getName()+".c"), I);
5442 Constant *Scale = Context->getConstantInt(IntPtrTy, Size);
5443 if (Constant *OpC = dyn_cast<Constant>(Op))
5444 Op = Context->getConstantExprMul(OpC, Scale);
5445 else // We'll let instcombine(mul) convert this to a shl if possible.
5446 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5447 GEP->getName()+".idx"), I);
5450 // Emit an add instruction.
5451 if (isa<Constant>(Op) && isa<Constant>(Result))
5452 Result = Context->getConstantExprAdd(cast<Constant>(Op),
5453 cast<Constant>(Result));
5455 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5456 GEP->getName()+".offs"), I);
5462 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5463 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5464 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5465 /// be complex, and scales are involved. The above expression would also be
5466 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5467 /// This later form is less amenable to optimization though, and we are allowed
5468 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5470 /// If we can't emit an optimized form for this expression, this returns null.
5472 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5474 TargetData &TD = IC.getTargetData();
5475 gep_type_iterator GTI = gep_type_begin(GEP);
5477 // Check to see if this gep only has a single variable index. If so, and if
5478 // any constant indices are a multiple of its scale, then we can compute this
5479 // in terms of the scale of the variable index. For example, if the GEP
5480 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5481 // because the expression will cross zero at the same point.
5482 unsigned i, e = GEP->getNumOperands();
5484 for (i = 1; i != e; ++i, ++GTI) {
5485 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5486 // Compute the aggregate offset of constant indices.
5487 if (CI->isZero()) continue;
5489 // Handle a struct index, which adds its field offset to the pointer.
5490 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5491 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5493 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5494 Offset += Size*CI->getSExtValue();
5497 // Found our variable index.
5502 // If there are no variable indices, we must have a constant offset, just
5503 // evaluate it the general way.
5504 if (i == e) return 0;
5506 Value *VariableIdx = GEP->getOperand(i);
5507 // Determine the scale factor of the variable element. For example, this is
5508 // 4 if the variable index is into an array of i32.
5509 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5511 // Verify that there are no other variable indices. If so, emit the hard way.
5512 for (++i, ++GTI; i != e; ++i, ++GTI) {
5513 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5516 // Compute the aggregate offset of constant indices.
5517 if (CI->isZero()) continue;
5519 // Handle a struct index, which adds its field offset to the pointer.
5520 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5521 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5523 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5524 Offset += Size*CI->getSExtValue();
5528 // Okay, we know we have a single variable index, which must be a
5529 // pointer/array/vector index. If there is no offset, life is simple, return
5531 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5533 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5534 // we don't need to bother extending: the extension won't affect where the
5535 // computation crosses zero.
5536 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5537 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5538 VariableIdx->getNameStart(), &I);
5542 // Otherwise, there is an index. The computation we will do will be modulo
5543 // the pointer size, so get it.
5544 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5546 Offset &= PtrSizeMask;
5547 VariableScale &= PtrSizeMask;
5549 // To do this transformation, any constant index must be a multiple of the
5550 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5551 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5552 // multiple of the variable scale.
5553 int64_t NewOffs = Offset / (int64_t)VariableScale;
5554 if (Offset != NewOffs*(int64_t)VariableScale)
5557 // Okay, we can do this evaluation. Start by converting the index to intptr.
5558 const Type *IntPtrTy = TD.getIntPtrType();
5559 if (VariableIdx->getType() != IntPtrTy)
5560 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5562 VariableIdx->getNameStart(), &I);
5563 Constant *OffsetVal = IC.getContext()->getConstantInt(IntPtrTy, NewOffs);
5564 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5568 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5569 /// else. At this point we know that the GEP is on the LHS of the comparison.
5570 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5571 ICmpInst::Predicate Cond,
5573 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5575 // Look through bitcasts.
5576 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5577 RHS = BCI->getOperand(0);
5579 Value *PtrBase = GEPLHS->getOperand(0);
5580 if (PtrBase == RHS) {
5581 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5582 // This transformation (ignoring the base and scales) is valid because we
5583 // know pointers can't overflow. See if we can output an optimized form.
5584 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5586 // If not, synthesize the offset the hard way.
5588 Offset = EmitGEPOffset(GEPLHS, I, *this);
5589 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), Offset,
5590 Context->getNullValue(Offset->getType()));
5591 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5592 // If the base pointers are different, but the indices are the same, just
5593 // compare the base pointer.
5594 if (PtrBase != GEPRHS->getOperand(0)) {
5595 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5596 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5597 GEPRHS->getOperand(0)->getType();
5599 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5600 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5601 IndicesTheSame = false;
5605 // If all indices are the same, just compare the base pointers.
5607 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond),
5608 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5610 // Otherwise, the base pointers are different and the indices are
5611 // different, bail out.
5615 // If one of the GEPs has all zero indices, recurse.
5616 bool AllZeros = true;
5617 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5618 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5619 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5624 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5625 ICmpInst::getSwappedPredicate(Cond), I);
5627 // If the other GEP has all zero indices, recurse.
5629 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5630 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5631 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5636 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5638 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5639 // If the GEPs only differ by one index, compare it.
5640 unsigned NumDifferences = 0; // Keep track of # differences.
5641 unsigned DiffOperand = 0; // The operand that differs.
5642 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5643 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5644 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5645 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5646 // Irreconcilable differences.
5650 if (NumDifferences++) break;
5655 if (NumDifferences == 0) // SAME GEP?
5656 return ReplaceInstUsesWith(I, // No comparison is needed here.
5657 Context->getConstantInt(Type::Int1Ty,
5658 ICmpInst::isTrueWhenEqual(Cond)));
5660 else if (NumDifferences == 1) {
5661 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5662 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5663 // Make sure we do a signed comparison here.
5664 return new ICmpInst(*Context,
5665 ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5669 // Only lower this if the icmp is the only user of the GEP or if we expect
5670 // the result to fold to a constant!
5671 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5672 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5673 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5674 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5675 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5676 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), L, R);
5682 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5684 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5687 if (!isa<ConstantFP>(RHSC)) return 0;
5688 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5690 // Get the width of the mantissa. We don't want to hack on conversions that
5691 // might lose information from the integer, e.g. "i64 -> float"
5692 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5693 if (MantissaWidth == -1) return 0; // Unknown.
5695 // Check to see that the input is converted from an integer type that is small
5696 // enough that preserves all bits. TODO: check here for "known" sign bits.
5697 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5698 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5700 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5701 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5705 // If the conversion would lose info, don't hack on this.
5706 if ((int)InputSize > MantissaWidth)
5709 // Otherwise, we can potentially simplify the comparison. We know that it
5710 // will always come through as an integer value and we know the constant is
5711 // not a NAN (it would have been previously simplified).
5712 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5714 ICmpInst::Predicate Pred;
5715 switch (I.getPredicate()) {
5716 default: llvm_unreachable("Unexpected predicate!");
5717 case FCmpInst::FCMP_UEQ:
5718 case FCmpInst::FCMP_OEQ:
5719 Pred = ICmpInst::ICMP_EQ;
5721 case FCmpInst::FCMP_UGT:
5722 case FCmpInst::FCMP_OGT:
5723 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5725 case FCmpInst::FCMP_UGE:
5726 case FCmpInst::FCMP_OGE:
5727 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5729 case FCmpInst::FCMP_ULT:
5730 case FCmpInst::FCMP_OLT:
5731 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5733 case FCmpInst::FCMP_ULE:
5734 case FCmpInst::FCMP_OLE:
5735 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5737 case FCmpInst::FCMP_UNE:
5738 case FCmpInst::FCMP_ONE:
5739 Pred = ICmpInst::ICMP_NE;
5741 case FCmpInst::FCMP_ORD:
5742 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5743 case FCmpInst::FCMP_UNO:
5744 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5747 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5749 // Now we know that the APFloat is a normal number, zero or inf.
5751 // See if the FP constant is too large for the integer. For example,
5752 // comparing an i8 to 300.0.
5753 unsigned IntWidth = IntTy->getScalarSizeInBits();
5756 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5757 // and large values.
5758 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5759 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5760 APFloat::rmNearestTiesToEven);
5761 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5762 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5763 Pred == ICmpInst::ICMP_SLE)
5764 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5765 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5768 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5769 // +INF and large values.
5770 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5771 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5772 APFloat::rmNearestTiesToEven);
5773 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5774 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5775 Pred == ICmpInst::ICMP_ULE)
5776 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5777 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5782 // See if the RHS value is < SignedMin.
5783 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5784 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5785 APFloat::rmNearestTiesToEven);
5786 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5787 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5788 Pred == ICmpInst::ICMP_SGE)
5789 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5790 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5794 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5795 // [0, UMAX], but it may still be fractional. See if it is fractional by
5796 // casting the FP value to the integer value and back, checking for equality.
5797 // Don't do this for zero, because -0.0 is not fractional.
5798 Constant *RHSInt = LHSUnsigned
5799 ? Context->getConstantExprFPToUI(RHSC, IntTy)
5800 : Context->getConstantExprFPToSI(RHSC, IntTy);
5801 if (!RHS.isZero()) {
5802 bool Equal = LHSUnsigned
5803 ? Context->getConstantExprUIToFP(RHSInt, RHSC->getType()) == RHSC
5804 : Context->getConstantExprSIToFP(RHSInt, RHSC->getType()) == RHSC;
5806 // If we had a comparison against a fractional value, we have to adjust
5807 // the compare predicate and sometimes the value. RHSC is rounded towards
5808 // zero at this point.
5810 default: llvm_unreachable("Unexpected integer comparison!");
5811 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5812 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5813 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5814 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5815 case ICmpInst::ICMP_ULE:
5816 // (float)int <= 4.4 --> int <= 4
5817 // (float)int <= -4.4 --> false
5818 if (RHS.isNegative())
5819 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5821 case ICmpInst::ICMP_SLE:
5822 // (float)int <= 4.4 --> int <= 4
5823 // (float)int <= -4.4 --> int < -4
5824 if (RHS.isNegative())
5825 Pred = ICmpInst::ICMP_SLT;
5827 case ICmpInst::ICMP_ULT:
5828 // (float)int < -4.4 --> false
5829 // (float)int < 4.4 --> int <= 4
5830 if (RHS.isNegative())
5831 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5832 Pred = ICmpInst::ICMP_ULE;
5834 case ICmpInst::ICMP_SLT:
5835 // (float)int < -4.4 --> int < -4
5836 // (float)int < 4.4 --> int <= 4
5837 if (!RHS.isNegative())
5838 Pred = ICmpInst::ICMP_SLE;
5840 case ICmpInst::ICMP_UGT:
5841 // (float)int > 4.4 --> int > 4
5842 // (float)int > -4.4 --> true
5843 if (RHS.isNegative())
5844 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5846 case ICmpInst::ICMP_SGT:
5847 // (float)int > 4.4 --> int > 4
5848 // (float)int > -4.4 --> int >= -4
5849 if (RHS.isNegative())
5850 Pred = ICmpInst::ICMP_SGE;
5852 case ICmpInst::ICMP_UGE:
5853 // (float)int >= -4.4 --> true
5854 // (float)int >= 4.4 --> int > 4
5855 if (!RHS.isNegative())
5856 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5857 Pred = ICmpInst::ICMP_UGT;
5859 case ICmpInst::ICMP_SGE:
5860 // (float)int >= -4.4 --> int >= -4
5861 // (float)int >= 4.4 --> int > 4
5862 if (!RHS.isNegative())
5863 Pred = ICmpInst::ICMP_SGT;
5869 // Lower this FP comparison into an appropriate integer version of the
5871 return new ICmpInst(*Context, Pred, LHSI->getOperand(0), RHSInt);
5874 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5875 bool Changed = SimplifyCompare(I);
5876 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5878 // Fold trivial predicates.
5879 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5880 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5881 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5882 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5884 // Simplify 'fcmp pred X, X'
5886 switch (I.getPredicate()) {
5887 default: llvm_unreachable("Unknown predicate!");
5888 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5889 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5890 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5891 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5892 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5893 case FCmpInst::FCMP_OLT: // True if ordered and less than
5894 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5895 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5897 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5898 case FCmpInst::FCMP_ULT: // True if unordered or less than
5899 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5900 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5901 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5902 I.setPredicate(FCmpInst::FCMP_UNO);
5903 I.setOperand(1, Context->getNullValue(Op0->getType()));
5906 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5907 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5908 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5909 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5910 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5911 I.setPredicate(FCmpInst::FCMP_ORD);
5912 I.setOperand(1, Context->getNullValue(Op0->getType()));
5917 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5918 return ReplaceInstUsesWith(I, Context->getUndef(Type::Int1Ty));
5920 // Handle fcmp with constant RHS
5921 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5922 // If the constant is a nan, see if we can fold the comparison based on it.
5923 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5924 if (CFP->getValueAPF().isNaN()) {
5925 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5926 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5927 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5928 "Comparison must be either ordered or unordered!");
5929 // True if unordered.
5930 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5934 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5935 switch (LHSI->getOpcode()) {
5936 case Instruction::PHI:
5937 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5938 // block. If in the same block, we're encouraging jump threading. If
5939 // not, we are just pessimizing the code by making an i1 phi.
5940 if (LHSI->getParent() == I.getParent())
5941 if (Instruction *NV = FoldOpIntoPhi(I))
5944 case Instruction::SIToFP:
5945 case Instruction::UIToFP:
5946 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5949 case Instruction::Select:
5950 // If either operand of the select is a constant, we can fold the
5951 // comparison into the select arms, which will cause one to be
5952 // constant folded and the select turned into a bitwise or.
5953 Value *Op1 = 0, *Op2 = 0;
5954 if (LHSI->hasOneUse()) {
5955 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5956 // Fold the known value into the constant operand.
5957 Op1 = Context->getConstantExprCompare(I.getPredicate(), C, RHSC);
5958 // Insert a new FCmp of the other select operand.
5959 Op2 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5960 LHSI->getOperand(2), RHSC,
5962 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5963 // Fold the known value into the constant operand.
5964 Op2 = Context->getConstantExprCompare(I.getPredicate(), C, RHSC);
5965 // Insert a new FCmp of the other select operand.
5966 Op1 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5967 LHSI->getOperand(1), RHSC,
5973 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5978 return Changed ? &I : 0;
5981 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5982 bool Changed = SimplifyCompare(I);
5983 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5984 const Type *Ty = Op0->getType();
5988 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
5989 I.isTrueWhenEqual()));
5991 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5992 return ReplaceInstUsesWith(I, Context->getUndef(Type::Int1Ty));
5994 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5995 // addresses never equal each other! We already know that Op0 != Op1.
5996 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5997 isa<ConstantPointerNull>(Op0)) &&
5998 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5999 isa<ConstantPointerNull>(Op1)))
6000 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
6001 !I.isTrueWhenEqual()));
6003 // icmp's with boolean values can always be turned into bitwise operations
6004 if (Ty == Type::Int1Ty) {
6005 switch (I.getPredicate()) {
6006 default: llvm_unreachable("Invalid icmp instruction!");
6007 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
6008 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
6009 InsertNewInstBefore(Xor, I);
6010 return BinaryOperator::CreateNot(*Context, Xor);
6012 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6013 return BinaryOperator::CreateXor(Op0, Op1);
6015 case ICmpInst::ICMP_UGT:
6016 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6018 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6019 Instruction *Not = BinaryOperator::CreateNot(*Context,
6020 Op0, I.getName()+"tmp");
6021 InsertNewInstBefore(Not, I);
6022 return BinaryOperator::CreateAnd(Not, Op1);
6024 case ICmpInst::ICMP_SGT:
6025 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6027 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6028 Instruction *Not = BinaryOperator::CreateNot(*Context,
6029 Op1, I.getName()+"tmp");
6030 InsertNewInstBefore(Not, I);
6031 return BinaryOperator::CreateAnd(Not, Op0);
6033 case ICmpInst::ICMP_UGE:
6034 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6036 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6037 Instruction *Not = BinaryOperator::CreateNot(*Context,
6038 Op0, I.getName()+"tmp");
6039 InsertNewInstBefore(Not, I);
6040 return BinaryOperator::CreateOr(Not, Op1);
6042 case ICmpInst::ICMP_SGE:
6043 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6045 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6046 Instruction *Not = BinaryOperator::CreateNot(*Context,
6047 Op1, I.getName()+"tmp");
6048 InsertNewInstBefore(Not, I);
6049 return BinaryOperator::CreateOr(Not, Op0);
6054 unsigned BitWidth = 0;
6056 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6057 else if (Ty->isIntOrIntVector())
6058 BitWidth = Ty->getScalarSizeInBits();
6060 bool isSignBit = false;
6062 // See if we are doing a comparison with a constant.
6063 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6064 Value *A = 0, *B = 0;
6066 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6067 if (I.isEquality() && CI->isNullValue() &&
6068 match(Op0, m_Sub(m_Value(A), m_Value(B)), *Context)) {
6069 // (icmp cond A B) if cond is equality
6070 return new ICmpInst(*Context, I.getPredicate(), A, B);
6073 // If we have an icmp le or icmp ge instruction, turn it into the
6074 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6075 // them being folded in the code below.
6076 switch (I.getPredicate()) {
6078 case ICmpInst::ICMP_ULE:
6079 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6080 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6081 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Op0,
6082 AddOne(CI, Context));
6083 case ICmpInst::ICMP_SLE:
6084 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6085 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6086 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6087 AddOne(CI, Context));
6088 case ICmpInst::ICMP_UGE:
6089 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6090 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6091 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Op0,
6092 SubOne(CI, Context));
6093 case ICmpInst::ICMP_SGE:
6094 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6095 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6096 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6097 SubOne(CI, Context));
6100 // If this comparison is a normal comparison, it demands all
6101 // bits, if it is a sign bit comparison, it only demands the sign bit.
6103 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6106 // See if we can fold the comparison based on range information we can get
6107 // by checking whether bits are known to be zero or one in the input.
6108 if (BitWidth != 0) {
6109 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6110 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6112 if (SimplifyDemandedBits(I.getOperandUse(0),
6113 isSignBit ? APInt::getSignBit(BitWidth)
6114 : APInt::getAllOnesValue(BitWidth),
6115 Op0KnownZero, Op0KnownOne, 0))
6117 if (SimplifyDemandedBits(I.getOperandUse(1),
6118 APInt::getAllOnesValue(BitWidth),
6119 Op1KnownZero, Op1KnownOne, 0))
6122 // Given the known and unknown bits, compute a range that the LHS could be
6123 // in. Compute the Min, Max and RHS values based on the known bits. For the
6124 // EQ and NE we use unsigned values.
6125 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6126 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6127 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6128 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6130 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6133 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6135 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6139 // If Min and Max are known to be the same, then SimplifyDemandedBits
6140 // figured out that the LHS is a constant. Just constant fold this now so
6141 // that code below can assume that Min != Max.
6142 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6143 return new ICmpInst(*Context, I.getPredicate(),
6144 Context->getConstantInt(Op0Min), Op1);
6145 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6146 return new ICmpInst(*Context, I.getPredicate(), Op0,
6147 Context->getConstantInt(Op1Min));
6149 // Based on the range information we know about the LHS, see if we can
6150 // simplify this comparison. For example, (x&4) < 8 is always true.
6151 switch (I.getPredicate()) {
6152 default: llvm_unreachable("Unknown icmp opcode!");
6153 case ICmpInst::ICMP_EQ:
6154 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6155 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6157 case ICmpInst::ICMP_NE:
6158 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6159 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6161 case ICmpInst::ICMP_ULT:
6162 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6163 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6164 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6165 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6166 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6167 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6168 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6169 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6170 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6171 SubOne(CI, Context));
6173 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6174 if (CI->isMinValue(true))
6175 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6176 Context->getAllOnesValue(Op0->getType()));
6179 case ICmpInst::ICMP_UGT:
6180 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6181 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6182 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6183 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6185 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6186 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6187 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6188 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6189 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6190 AddOne(CI, Context));
6192 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6193 if (CI->isMaxValue(true))
6194 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6195 Context->getNullValue(Op0->getType()));
6198 case ICmpInst::ICMP_SLT:
6199 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6200 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6201 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6202 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6203 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6204 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6205 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6206 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6207 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6208 SubOne(CI, Context));
6211 case ICmpInst::ICMP_SGT:
6212 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6213 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6214 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6215 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6217 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6218 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6219 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6220 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6221 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6222 AddOne(CI, Context));
6225 case ICmpInst::ICMP_SGE:
6226 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6227 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6228 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6229 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6230 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6232 case ICmpInst::ICMP_SLE:
6233 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6234 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6235 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6236 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6237 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6239 case ICmpInst::ICMP_UGE:
6240 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6241 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6242 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6243 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6244 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6246 case ICmpInst::ICMP_ULE:
6247 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6248 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6249 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6250 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6251 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6255 // Turn a signed comparison into an unsigned one if both operands
6256 // are known to have the same sign.
6257 if (I.isSignedPredicate() &&
6258 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6259 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6260 return new ICmpInst(*Context, I.getUnsignedPredicate(), Op0, Op1);
6263 // Test if the ICmpInst instruction is used exclusively by a select as
6264 // part of a minimum or maximum operation. If so, refrain from doing
6265 // any other folding. This helps out other analyses which understand
6266 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6267 // and CodeGen. And in this case, at least one of the comparison
6268 // operands has at least one user besides the compare (the select),
6269 // which would often largely negate the benefit of folding anyway.
6271 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6272 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6273 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6276 // See if we are doing a comparison between a constant and an instruction that
6277 // can be folded into the comparison.
6278 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6279 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6280 // instruction, see if that instruction also has constants so that the
6281 // instruction can be folded into the icmp
6282 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6283 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6287 // Handle icmp with constant (but not simple integer constant) RHS
6288 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6289 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6290 switch (LHSI->getOpcode()) {
6291 case Instruction::GetElementPtr:
6292 if (RHSC->isNullValue()) {
6293 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6294 bool isAllZeros = true;
6295 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6296 if (!isa<Constant>(LHSI->getOperand(i)) ||
6297 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6302 return new ICmpInst(*Context, I.getPredicate(), LHSI->getOperand(0),
6303 Context->getNullValue(LHSI->getOperand(0)->getType()));
6307 case Instruction::PHI:
6308 // Only fold icmp into the PHI if the phi and fcmp are in the same
6309 // block. If in the same block, we're encouraging jump threading. If
6310 // not, we are just pessimizing the code by making an i1 phi.
6311 if (LHSI->getParent() == I.getParent())
6312 if (Instruction *NV = FoldOpIntoPhi(I))
6315 case Instruction::Select: {
6316 // If either operand of the select is a constant, we can fold the
6317 // comparison into the select arms, which will cause one to be
6318 // constant folded and the select turned into a bitwise or.
6319 Value *Op1 = 0, *Op2 = 0;
6320 if (LHSI->hasOneUse()) {
6321 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6322 // Fold the known value into the constant operand.
6323 Op1 = Context->getConstantExprICmp(I.getPredicate(), C, RHSC);
6324 // Insert a new ICmp of the other select operand.
6325 Op2 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6326 LHSI->getOperand(2), RHSC,
6328 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6329 // Fold the known value into the constant operand.
6330 Op2 = Context->getConstantExprICmp(I.getPredicate(), C, RHSC);
6331 // Insert a new ICmp of the other select operand.
6332 Op1 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6333 LHSI->getOperand(1), RHSC,
6339 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6342 case Instruction::Malloc:
6343 // If we have (malloc != null), and if the malloc has a single use, we
6344 // can assume it is successful and remove the malloc.
6345 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6346 AddToWorkList(LHSI);
6347 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
6348 !I.isTrueWhenEqual()));
6354 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6355 if (User *GEP = dyn_castGetElementPtr(Op0))
6356 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6358 if (User *GEP = dyn_castGetElementPtr(Op1))
6359 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6360 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6363 // Test to see if the operands of the icmp are casted versions of other
6364 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6366 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6367 if (isa<PointerType>(Op0->getType()) &&
6368 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6369 // We keep moving the cast from the left operand over to the right
6370 // operand, where it can often be eliminated completely.
6371 Op0 = CI->getOperand(0);
6373 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6374 // so eliminate it as well.
6375 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6376 Op1 = CI2->getOperand(0);
6378 // If Op1 is a constant, we can fold the cast into the constant.
6379 if (Op0->getType() != Op1->getType()) {
6380 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6381 Op1 = Context->getConstantExprBitCast(Op1C, Op0->getType());
6383 // Otherwise, cast the RHS right before the icmp
6384 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6387 return new ICmpInst(*Context, I.getPredicate(), Op0, Op1);
6391 if (isa<CastInst>(Op0)) {
6392 // Handle the special case of: icmp (cast bool to X), <cst>
6393 // This comes up when you have code like
6396 // For generality, we handle any zero-extension of any operand comparison
6397 // with a constant or another cast from the same type.
6398 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6399 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6403 // See if it's the same type of instruction on the left and right.
6404 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6405 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6406 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6407 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6408 switch (Op0I->getOpcode()) {
6410 case Instruction::Add:
6411 case Instruction::Sub:
6412 case Instruction::Xor:
6413 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6414 return new ICmpInst(*Context, I.getPredicate(), Op0I->getOperand(0),
6415 Op1I->getOperand(0));
6416 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6417 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6418 if (CI->getValue().isSignBit()) {
6419 ICmpInst::Predicate Pred = I.isSignedPredicate()
6420 ? I.getUnsignedPredicate()
6421 : I.getSignedPredicate();
6422 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6423 Op1I->getOperand(0));
6426 if (CI->getValue().isMaxSignedValue()) {
6427 ICmpInst::Predicate Pred = I.isSignedPredicate()
6428 ? I.getUnsignedPredicate()
6429 : I.getSignedPredicate();
6430 Pred = I.getSwappedPredicate(Pred);
6431 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6432 Op1I->getOperand(0));
6436 case Instruction::Mul:
6437 if (!I.isEquality())
6440 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6441 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6442 // Mask = -1 >> count-trailing-zeros(Cst).
6443 if (!CI->isZero() && !CI->isOne()) {
6444 const APInt &AP = CI->getValue();
6445 ConstantInt *Mask = Context->getConstantInt(
6446 APInt::getLowBitsSet(AP.getBitWidth(),
6448 AP.countTrailingZeros()));
6449 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6451 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6453 InsertNewInstBefore(And1, I);
6454 InsertNewInstBefore(And2, I);
6455 return new ICmpInst(*Context, I.getPredicate(), And1, And2);
6464 // ~x < ~y --> y < x
6466 if (match(Op0, m_Not(m_Value(A)), *Context) &&
6467 match(Op1, m_Not(m_Value(B)), *Context))
6468 return new ICmpInst(*Context, I.getPredicate(), B, A);
6471 if (I.isEquality()) {
6472 Value *A, *B, *C, *D;
6474 // -x == -y --> x == y
6475 if (match(Op0, m_Neg(m_Value(A)), *Context) &&
6476 match(Op1, m_Neg(m_Value(B)), *Context))
6477 return new ICmpInst(*Context, I.getPredicate(), A, B);
6479 if (match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
6480 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6481 Value *OtherVal = A == Op1 ? B : A;
6482 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6483 Context->getNullValue(A->getType()));
6486 if (match(Op1, m_Xor(m_Value(C), m_Value(D)), *Context)) {
6487 // A^c1 == C^c2 --> A == C^(c1^c2)
6488 ConstantInt *C1, *C2;
6489 if (match(B, m_ConstantInt(C1), *Context) &&
6490 match(D, m_ConstantInt(C2), *Context) && Op1->hasOneUse()) {
6492 Context->getConstantInt(C1->getValue() ^ C2->getValue());
6493 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6494 return new ICmpInst(*Context, I.getPredicate(), A,
6495 InsertNewInstBefore(Xor, I));
6498 // A^B == A^D -> B == D
6499 if (A == C) return new ICmpInst(*Context, I.getPredicate(), B, D);
6500 if (A == D) return new ICmpInst(*Context, I.getPredicate(), B, C);
6501 if (B == C) return new ICmpInst(*Context, I.getPredicate(), A, D);
6502 if (B == D) return new ICmpInst(*Context, I.getPredicate(), A, C);
6506 if (match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context) &&
6507 (A == Op0 || B == Op0)) {
6508 // A == (A^B) -> B == 0
6509 Value *OtherVal = A == Op0 ? B : A;
6510 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6511 Context->getNullValue(A->getType()));
6514 // (A-B) == A -> B == 0
6515 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B)), *Context))
6516 return new ICmpInst(*Context, I.getPredicate(), B,
6517 Context->getNullValue(B->getType()));
6519 // A == (A-B) -> B == 0
6520 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B)), *Context))
6521 return new ICmpInst(*Context, I.getPredicate(), B,
6522 Context->getNullValue(B->getType()));
6524 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6525 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6526 match(Op0, m_And(m_Value(A), m_Value(B)), *Context) &&
6527 match(Op1, m_And(m_Value(C), m_Value(D)), *Context)) {
6528 Value *X = 0, *Y = 0, *Z = 0;
6531 X = B; Y = D; Z = A;
6532 } else if (A == D) {
6533 X = B; Y = C; Z = A;
6534 } else if (B == C) {
6535 X = A; Y = D; Z = B;
6536 } else if (B == D) {
6537 X = A; Y = C; Z = B;
6540 if (X) { // Build (X^Y) & Z
6541 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6542 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6543 I.setOperand(0, Op1);
6544 I.setOperand(1, Context->getNullValue(Op1->getType()));
6549 return Changed ? &I : 0;
6553 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6554 /// and CmpRHS are both known to be integer constants.
6555 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6556 ConstantInt *DivRHS) {
6557 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6558 const APInt &CmpRHSV = CmpRHS->getValue();
6560 // FIXME: If the operand types don't match the type of the divide
6561 // then don't attempt this transform. The code below doesn't have the
6562 // logic to deal with a signed divide and an unsigned compare (and
6563 // vice versa). This is because (x /s C1) <s C2 produces different
6564 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6565 // (x /u C1) <u C2. Simply casting the operands and result won't
6566 // work. :( The if statement below tests that condition and bails
6568 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6569 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6571 if (DivRHS->isZero())
6572 return 0; // The ProdOV computation fails on divide by zero.
6573 if (DivIsSigned && DivRHS->isAllOnesValue())
6574 return 0; // The overflow computation also screws up here
6575 if (DivRHS->isOne())
6576 return 0; // Not worth bothering, and eliminates some funny cases
6579 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6580 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6581 // C2 (CI). By solving for X we can turn this into a range check
6582 // instead of computing a divide.
6583 Constant *Prod = Context->getConstantExprMul(CmpRHS, DivRHS);
6585 // Determine if the product overflows by seeing if the product is
6586 // not equal to the divide. Make sure we do the same kind of divide
6587 // as in the LHS instruction that we're folding.
6588 bool ProdOV = (DivIsSigned ? Context->getConstantExprSDiv(Prod, DivRHS) :
6589 Context->getConstantExprUDiv(Prod, DivRHS)) != CmpRHS;
6591 // Get the ICmp opcode
6592 ICmpInst::Predicate Pred = ICI.getPredicate();
6594 // Figure out the interval that is being checked. For example, a comparison
6595 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6596 // Compute this interval based on the constants involved and the signedness of
6597 // the compare/divide. This computes a half-open interval, keeping track of
6598 // whether either value in the interval overflows. After analysis each
6599 // overflow variable is set to 0 if it's corresponding bound variable is valid
6600 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6601 int LoOverflow = 0, HiOverflow = 0;
6602 Constant *LoBound = 0, *HiBound = 0;
6604 if (!DivIsSigned) { // udiv
6605 // e.g. X/5 op 3 --> [15, 20)
6607 HiOverflow = LoOverflow = ProdOV;
6609 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6610 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6611 if (CmpRHSV == 0) { // (X / pos) op 0
6612 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6613 LoBound = cast<ConstantInt>(Context->getConstantExprNeg(SubOne(DivRHS,
6616 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6617 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6618 HiOverflow = LoOverflow = ProdOV;
6620 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6621 } else { // (X / pos) op neg
6622 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6623 HiBound = AddOne(Prod, Context);
6624 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6626 ConstantInt* DivNeg =
6627 cast<ConstantInt>(Context->getConstantExprNeg(DivRHS));
6628 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6632 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6633 if (CmpRHSV == 0) { // (X / neg) op 0
6634 // e.g. X/-5 op 0 --> [-4, 5)
6635 LoBound = AddOne(DivRHS, Context);
6636 HiBound = cast<ConstantInt>(Context->getConstantExprNeg(DivRHS));
6637 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6638 HiOverflow = 1; // [INTMIN+1, overflow)
6639 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6641 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6642 // e.g. X/-5 op 3 --> [-19, -14)
6643 HiBound = AddOne(Prod, Context);
6644 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6646 LoOverflow = AddWithOverflow(LoBound, HiBound,
6647 DivRHS, Context, true) ? -1 : 0;
6648 } else { // (X / neg) op neg
6649 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6650 LoOverflow = HiOverflow = ProdOV;
6652 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6655 // Dividing by a negative swaps the condition. LT <-> GT
6656 Pred = ICmpInst::getSwappedPredicate(Pred);
6659 Value *X = DivI->getOperand(0);
6661 default: llvm_unreachable("Unhandled icmp opcode!");
6662 case ICmpInst::ICMP_EQ:
6663 if (LoOverflow && HiOverflow)
6664 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6665 else if (HiOverflow)
6666 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6667 ICmpInst::ICMP_UGE, X, LoBound);
6668 else if (LoOverflow)
6669 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6670 ICmpInst::ICMP_ULT, X, HiBound);
6672 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6673 case ICmpInst::ICMP_NE:
6674 if (LoOverflow && HiOverflow)
6675 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6676 else if (HiOverflow)
6677 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6678 ICmpInst::ICMP_ULT, X, LoBound);
6679 else if (LoOverflow)
6680 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6681 ICmpInst::ICMP_UGE, X, HiBound);
6683 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6684 case ICmpInst::ICMP_ULT:
6685 case ICmpInst::ICMP_SLT:
6686 if (LoOverflow == +1) // Low bound is greater than input range.
6687 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6688 if (LoOverflow == -1) // Low bound is less than input range.
6689 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6690 return new ICmpInst(*Context, Pred, X, LoBound);
6691 case ICmpInst::ICMP_UGT:
6692 case ICmpInst::ICMP_SGT:
6693 if (HiOverflow == +1) // High bound greater than input range.
6694 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6695 else if (HiOverflow == -1) // High bound less than input range.
6696 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6697 if (Pred == ICmpInst::ICMP_UGT)
6698 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, X, HiBound);
6700 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, X, HiBound);
6705 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6707 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6710 const APInt &RHSV = RHS->getValue();
6712 switch (LHSI->getOpcode()) {
6713 case Instruction::Trunc:
6714 if (ICI.isEquality() && LHSI->hasOneUse()) {
6715 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6716 // of the high bits truncated out of x are known.
6717 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6718 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6719 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6720 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6721 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6723 // If all the high bits are known, we can do this xform.
6724 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6725 // Pull in the high bits from known-ones set.
6726 APInt NewRHS(RHS->getValue());
6727 NewRHS.zext(SrcBits);
6729 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6730 Context->getConstantInt(NewRHS));
6735 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6736 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6737 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6739 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6740 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6741 Value *CompareVal = LHSI->getOperand(0);
6743 // If the sign bit of the XorCST is not set, there is no change to
6744 // the operation, just stop using the Xor.
6745 if (!XorCST->getValue().isNegative()) {
6746 ICI.setOperand(0, CompareVal);
6747 AddToWorkList(LHSI);
6751 // Was the old condition true if the operand is positive?
6752 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6754 // If so, the new one isn't.
6755 isTrueIfPositive ^= true;
6757 if (isTrueIfPositive)
6758 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, CompareVal,
6759 SubOne(RHS, Context));
6761 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, CompareVal,
6762 AddOne(RHS, Context));
6765 if (LHSI->hasOneUse()) {
6766 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6767 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6768 const APInt &SignBit = XorCST->getValue();
6769 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6770 ? ICI.getUnsignedPredicate()
6771 : ICI.getSignedPredicate();
6772 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6773 Context->getConstantInt(RHSV ^ SignBit));
6776 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6777 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6778 const APInt &NotSignBit = XorCST->getValue();
6779 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6780 ? ICI.getUnsignedPredicate()
6781 : ICI.getSignedPredicate();
6782 Pred = ICI.getSwappedPredicate(Pred);
6783 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6784 Context->getConstantInt(RHSV ^ NotSignBit));
6789 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6790 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6791 LHSI->getOperand(0)->hasOneUse()) {
6792 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6794 // If the LHS is an AND of a truncating cast, we can widen the
6795 // and/compare to be the input width without changing the value
6796 // produced, eliminating a cast.
6797 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6798 // We can do this transformation if either the AND constant does not
6799 // have its sign bit set or if it is an equality comparison.
6800 // Extending a relational comparison when we're checking the sign
6801 // bit would not work.
6802 if (Cast->hasOneUse() &&
6803 (ICI.isEquality() ||
6804 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6806 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6807 APInt NewCST = AndCST->getValue();
6808 NewCST.zext(BitWidth);
6810 NewCI.zext(BitWidth);
6811 Instruction *NewAnd =
6812 BinaryOperator::CreateAnd(Cast->getOperand(0),
6813 Context->getConstantInt(NewCST),LHSI->getName());
6814 InsertNewInstBefore(NewAnd, ICI);
6815 return new ICmpInst(*Context, ICI.getPredicate(), NewAnd,
6816 Context->getConstantInt(NewCI));
6820 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6821 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6822 // happens a LOT in code produced by the C front-end, for bitfield
6824 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6825 if (Shift && !Shift->isShift())
6829 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6830 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6831 const Type *AndTy = AndCST->getType(); // Type of the and.
6833 // We can fold this as long as we can't shift unknown bits
6834 // into the mask. This can only happen with signed shift
6835 // rights, as they sign-extend.
6837 bool CanFold = Shift->isLogicalShift();
6839 // To test for the bad case of the signed shr, see if any
6840 // of the bits shifted in could be tested after the mask.
6841 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6842 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6844 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6845 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6846 AndCST->getValue()) == 0)
6852 if (Shift->getOpcode() == Instruction::Shl)
6853 NewCst = Context->getConstantExprLShr(RHS, ShAmt);
6855 NewCst = Context->getConstantExprShl(RHS, ShAmt);
6857 // Check to see if we are shifting out any of the bits being
6859 if (Context->getConstantExpr(Shift->getOpcode(),
6860 NewCst, ShAmt) != RHS) {
6861 // If we shifted bits out, the fold is not going to work out.
6862 // As a special case, check to see if this means that the
6863 // result is always true or false now.
6864 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6865 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6866 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6867 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6869 ICI.setOperand(1, NewCst);
6870 Constant *NewAndCST;
6871 if (Shift->getOpcode() == Instruction::Shl)
6872 NewAndCST = Context->getConstantExprLShr(AndCST, ShAmt);
6874 NewAndCST = Context->getConstantExprShl(AndCST, ShAmt);
6875 LHSI->setOperand(1, NewAndCST);
6876 LHSI->setOperand(0, Shift->getOperand(0));
6877 AddToWorkList(Shift); // Shift is dead.
6878 AddUsesToWorkList(ICI);
6884 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6885 // preferable because it allows the C<<Y expression to be hoisted out
6886 // of a loop if Y is invariant and X is not.
6887 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6888 ICI.isEquality() && !Shift->isArithmeticShift() &&
6889 !isa<Constant>(Shift->getOperand(0))) {
6892 if (Shift->getOpcode() == Instruction::LShr) {
6893 NS = BinaryOperator::CreateShl(AndCST,
6894 Shift->getOperand(1), "tmp");
6896 // Insert a logical shift.
6897 NS = BinaryOperator::CreateLShr(AndCST,
6898 Shift->getOperand(1), "tmp");
6900 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6902 // Compute X & (C << Y).
6903 Instruction *NewAnd =
6904 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6905 InsertNewInstBefore(NewAnd, ICI);
6907 ICI.setOperand(0, NewAnd);
6913 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6914 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6917 uint32_t TypeBits = RHSV.getBitWidth();
6919 // Check that the shift amount is in range. If not, don't perform
6920 // undefined shifts. When the shift is visited it will be
6922 if (ShAmt->uge(TypeBits))
6925 if (ICI.isEquality()) {
6926 // If we are comparing against bits always shifted out, the
6927 // comparison cannot succeed.
6929 Context->getConstantExprShl(Context->getConstantExprLShr(RHS, ShAmt),
6931 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6932 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6933 Constant *Cst = Context->getConstantInt(Type::Int1Ty, IsICMP_NE);
6934 return ReplaceInstUsesWith(ICI, Cst);
6937 if (LHSI->hasOneUse()) {
6938 // Otherwise strength reduce the shift into an and.
6939 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6941 Context->getConstantInt(APInt::getLowBitsSet(TypeBits,
6942 TypeBits-ShAmtVal));
6945 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6946 Mask, LHSI->getName()+".mask");
6947 Value *And = InsertNewInstBefore(AndI, ICI);
6948 return new ICmpInst(*Context, ICI.getPredicate(), And,
6949 Context->getConstantInt(RHSV.lshr(ShAmtVal)));
6953 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6954 bool TrueIfSigned = false;
6955 if (LHSI->hasOneUse() &&
6956 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6957 // (X << 31) <s 0 --> (X&1) != 0
6958 Constant *Mask = Context->getConstantInt(APInt(TypeBits, 1) <<
6959 (TypeBits-ShAmt->getZExtValue()-1));
6961 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6962 Mask, LHSI->getName()+".mask");
6963 Value *And = InsertNewInstBefore(AndI, ICI);
6965 return new ICmpInst(*Context,
6966 TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6967 And, Context->getNullValue(And->getType()));
6972 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6973 case Instruction::AShr: {
6974 // Only handle equality comparisons of shift-by-constant.
6975 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6976 if (!ShAmt || !ICI.isEquality()) break;
6978 // Check that the shift amount is in range. If not, don't perform
6979 // undefined shifts. When the shift is visited it will be
6981 uint32_t TypeBits = RHSV.getBitWidth();
6982 if (ShAmt->uge(TypeBits))
6985 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6987 // If we are comparing against bits always shifted out, the
6988 // comparison cannot succeed.
6989 APInt Comp = RHSV << ShAmtVal;
6990 if (LHSI->getOpcode() == Instruction::LShr)
6991 Comp = Comp.lshr(ShAmtVal);
6993 Comp = Comp.ashr(ShAmtVal);
6995 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6996 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6997 Constant *Cst = Context->getConstantInt(Type::Int1Ty, IsICMP_NE);
6998 return ReplaceInstUsesWith(ICI, Cst);
7001 // Otherwise, check to see if the bits shifted out are known to be zero.
7002 // If so, we can compare against the unshifted value:
7003 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
7004 if (LHSI->hasOneUse() &&
7005 MaskedValueIsZero(LHSI->getOperand(0),
7006 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
7007 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
7008 Context->getConstantExprShl(RHS, ShAmt));
7011 if (LHSI->hasOneUse()) {
7012 // Otherwise strength reduce the shift into an and.
7013 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7014 Constant *Mask = Context->getConstantInt(Val);
7017 BinaryOperator::CreateAnd(LHSI->getOperand(0),
7018 Mask, LHSI->getName()+".mask");
7019 Value *And = InsertNewInstBefore(AndI, ICI);
7020 return new ICmpInst(*Context, ICI.getPredicate(), And,
7021 Context->getConstantExprShl(RHS, ShAmt));
7026 case Instruction::SDiv:
7027 case Instruction::UDiv:
7028 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7029 // Fold this div into the comparison, producing a range check.
7030 // Determine, based on the divide type, what the range is being
7031 // checked. If there is an overflow on the low or high side, remember
7032 // it, otherwise compute the range [low, hi) bounding the new value.
7033 // See: InsertRangeTest above for the kinds of replacements possible.
7034 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7035 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7040 case Instruction::Add:
7041 // Fold: icmp pred (add, X, C1), C2
7043 if (!ICI.isEquality()) {
7044 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7046 const APInt &LHSV = LHSC->getValue();
7048 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7051 if (ICI.isSignedPredicate()) {
7052 if (CR.getLower().isSignBit()) {
7053 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7054 Context->getConstantInt(CR.getUpper()));
7055 } else if (CR.getUpper().isSignBit()) {
7056 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7057 Context->getConstantInt(CR.getLower()));
7060 if (CR.getLower().isMinValue()) {
7061 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7062 Context->getConstantInt(CR.getUpper()));
7063 } else if (CR.getUpper().isMinValue()) {
7064 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7065 Context->getConstantInt(CR.getLower()));
7072 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7073 if (ICI.isEquality()) {
7074 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7076 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7077 // the second operand is a constant, simplify a bit.
7078 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7079 switch (BO->getOpcode()) {
7080 case Instruction::SRem:
7081 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7082 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7083 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7084 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7085 Instruction *NewRem =
7086 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
7088 InsertNewInstBefore(NewRem, ICI);
7089 return new ICmpInst(*Context, ICI.getPredicate(), NewRem,
7090 Context->getNullValue(BO->getType()));
7094 case Instruction::Add:
7095 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7096 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7097 if (BO->hasOneUse())
7098 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7099 Context->getConstantExprSub(RHS, BOp1C));
7100 } else if (RHSV == 0) {
7101 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7102 // efficiently invertible, or if the add has just this one use.
7103 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7105 if (Value *NegVal = dyn_castNegVal(BOp1, Context))
7106 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, NegVal);
7107 else if (Value *NegVal = dyn_castNegVal(BOp0, Context))
7108 return new ICmpInst(*Context, ICI.getPredicate(), NegVal, BOp1);
7109 else if (BO->hasOneUse()) {
7110 Instruction *Neg = BinaryOperator::CreateNeg(*Context, BOp1);
7111 InsertNewInstBefore(Neg, ICI);
7113 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, Neg);
7117 case Instruction::Xor:
7118 // For the xor case, we can xor two constants together, eliminating
7119 // the explicit xor.
7120 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7121 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7122 Context->getConstantExprXor(RHS, BOC));
7125 case Instruction::Sub:
7126 // Replace (([sub|xor] A, B) != 0) with (A != B)
7128 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7132 case Instruction::Or:
7133 // If bits are being or'd in that are not present in the constant we
7134 // are comparing against, then the comparison could never succeed!
7135 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7136 Constant *NotCI = Context->getConstantExprNot(RHS);
7137 if (!Context->getConstantExprAnd(BOC, NotCI)->isNullValue())
7138 return ReplaceInstUsesWith(ICI,
7139 Context->getConstantInt(Type::Int1Ty,
7144 case Instruction::And:
7145 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7146 // If bits are being compared against that are and'd out, then the
7147 // comparison can never succeed!
7148 if ((RHSV & ~BOC->getValue()) != 0)
7149 return ReplaceInstUsesWith(ICI,
7150 Context->getConstantInt(Type::Int1Ty,
7153 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7154 if (RHS == BOC && RHSV.isPowerOf2())
7155 return new ICmpInst(*Context, isICMP_NE ? ICmpInst::ICMP_EQ :
7156 ICmpInst::ICMP_NE, LHSI,
7157 Context->getNullValue(RHS->getType()));
7159 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7160 if (BOC->getValue().isSignBit()) {
7161 Value *X = BO->getOperand(0);
7162 Constant *Zero = Context->getNullValue(X->getType());
7163 ICmpInst::Predicate pred = isICMP_NE ?
7164 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7165 return new ICmpInst(*Context, pred, X, Zero);
7168 // ((X & ~7) == 0) --> X < 8
7169 if (RHSV == 0 && isHighOnes(BOC)) {
7170 Value *X = BO->getOperand(0);
7171 Constant *NegX = Context->getConstantExprNeg(BOC);
7172 ICmpInst::Predicate pred = isICMP_NE ?
7173 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7174 return new ICmpInst(*Context, pred, X, NegX);
7179 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7180 // Handle icmp {eq|ne} <intrinsic>, intcst.
7181 if (II->getIntrinsicID() == Intrinsic::bswap) {
7183 ICI.setOperand(0, II->getOperand(1));
7184 ICI.setOperand(1, Context->getConstantInt(RHSV.byteSwap()));
7192 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7193 /// We only handle extending casts so far.
7195 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7196 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7197 Value *LHSCIOp = LHSCI->getOperand(0);
7198 const Type *SrcTy = LHSCIOp->getType();
7199 const Type *DestTy = LHSCI->getType();
7202 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7203 // integer type is the same size as the pointer type.
7204 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
7205 getTargetData().getPointerSizeInBits() ==
7206 cast<IntegerType>(DestTy)->getBitWidth()) {
7208 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7209 RHSOp = Context->getConstantExprIntToPtr(RHSC, SrcTy);
7210 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7211 RHSOp = RHSC->getOperand(0);
7212 // If the pointer types don't match, insert a bitcast.
7213 if (LHSCIOp->getType() != RHSOp->getType())
7214 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
7218 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSOp);
7221 // The code below only handles extension cast instructions, so far.
7223 if (LHSCI->getOpcode() != Instruction::ZExt &&
7224 LHSCI->getOpcode() != Instruction::SExt)
7227 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7228 bool isSignedCmp = ICI.isSignedPredicate();
7230 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7231 // Not an extension from the same type?
7232 RHSCIOp = CI->getOperand(0);
7233 if (RHSCIOp->getType() != LHSCIOp->getType())
7236 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7237 // and the other is a zext), then we can't handle this.
7238 if (CI->getOpcode() != LHSCI->getOpcode())
7241 // Deal with equality cases early.
7242 if (ICI.isEquality())
7243 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7245 // A signed comparison of sign extended values simplifies into a
7246 // signed comparison.
7247 if (isSignedCmp && isSignedExt)
7248 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7250 // The other three cases all fold into an unsigned comparison.
7251 return new ICmpInst(*Context, ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7254 // If we aren't dealing with a constant on the RHS, exit early
7255 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7259 // Compute the constant that would happen if we truncated to SrcTy then
7260 // reextended to DestTy.
7261 Constant *Res1 = Context->getConstantExprTrunc(CI, SrcTy);
7262 Constant *Res2 = Context->getConstantExprCast(LHSCI->getOpcode(),
7265 // If the re-extended constant didn't change...
7267 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7268 // For example, we might have:
7269 // %A = sext i16 %X to i32
7270 // %B = icmp ugt i32 %A, 1330
7271 // It is incorrect to transform this into
7272 // %B = icmp ugt i16 %X, 1330
7273 // because %A may have negative value.
7275 // However, we allow this when the compare is EQ/NE, because they are
7277 if (isSignedExt == isSignedCmp || ICI.isEquality())
7278 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, Res1);
7282 // The re-extended constant changed so the constant cannot be represented
7283 // in the shorter type. Consequently, we cannot emit a simple comparison.
7285 // First, handle some easy cases. We know the result cannot be equal at this
7286 // point so handle the ICI.isEquality() cases
7287 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7288 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
7289 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7290 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
7292 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7293 // should have been folded away previously and not enter in here.
7296 // We're performing a signed comparison.
7297 if (cast<ConstantInt>(CI)->getValue().isNegative())
7298 Result = Context->getConstantIntFalse(); // X < (small) --> false
7300 Result = Context->getConstantIntTrue(); // X < (large) --> true
7302 // We're performing an unsigned comparison.
7304 // We're performing an unsigned comp with a sign extended value.
7305 // This is true if the input is >= 0. [aka >s -1]
7306 Constant *NegOne = Context->getAllOnesValue(SrcTy);
7307 Result = InsertNewInstBefore(new ICmpInst(*Context, ICmpInst::ICMP_SGT,
7308 LHSCIOp, NegOne, ICI.getName()), ICI);
7310 // Unsigned extend & unsigned compare -> always true.
7311 Result = Context->getConstantIntTrue();
7315 // Finally, return the value computed.
7316 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7317 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7318 return ReplaceInstUsesWith(ICI, Result);
7320 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7321 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7322 "ICmp should be folded!");
7323 if (Constant *CI = dyn_cast<Constant>(Result))
7324 return ReplaceInstUsesWith(ICI, Context->getConstantExprNot(CI));
7325 return BinaryOperator::CreateNot(*Context, Result);
7328 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7329 return commonShiftTransforms(I);
7332 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7333 return commonShiftTransforms(I);
7336 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7337 if (Instruction *R = commonShiftTransforms(I))
7340 Value *Op0 = I.getOperand(0);
7342 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7343 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7344 if (CSI->isAllOnesValue())
7345 return ReplaceInstUsesWith(I, CSI);
7347 // See if we can turn a signed shr into an unsigned shr.
7348 if (MaskedValueIsZero(Op0,
7349 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7350 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7352 // Arithmetic shifting an all-sign-bit value is a no-op.
7353 unsigned NumSignBits = ComputeNumSignBits(Op0);
7354 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7355 return ReplaceInstUsesWith(I, Op0);
7360 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7361 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7362 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7364 // shl X, 0 == X and shr X, 0 == X
7365 // shl 0, X == 0 and shr 0, X == 0
7366 if (Op1 == Context->getNullValue(Op1->getType()) ||
7367 Op0 == Context->getNullValue(Op0->getType()))
7368 return ReplaceInstUsesWith(I, Op0);
7370 if (isa<UndefValue>(Op0)) {
7371 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7372 return ReplaceInstUsesWith(I, Op0);
7373 else // undef << X -> 0, undef >>u X -> 0
7374 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7376 if (isa<UndefValue>(Op1)) {
7377 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7378 return ReplaceInstUsesWith(I, Op0);
7379 else // X << undef, X >>u undef -> 0
7380 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7383 // See if we can fold away this shift.
7384 if (SimplifyDemandedInstructionBits(I))
7387 // Try to fold constant and into select arguments.
7388 if (isa<Constant>(Op0))
7389 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7390 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7393 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7394 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7399 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7400 BinaryOperator &I) {
7401 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7403 // See if we can simplify any instructions used by the instruction whose sole
7404 // purpose is to compute bits we don't care about.
7405 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7407 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7410 if (Op1->uge(TypeBits)) {
7411 if (I.getOpcode() != Instruction::AShr)
7412 return ReplaceInstUsesWith(I, Context->getNullValue(Op0->getType()));
7414 I.setOperand(1, Context->getConstantInt(I.getType(), TypeBits-1));
7419 // ((X*C1) << C2) == (X * (C1 << C2))
7420 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7421 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7422 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7423 return BinaryOperator::CreateMul(BO->getOperand(0),
7424 Context->getConstantExprShl(BOOp, Op1));
7426 // Try to fold constant and into select arguments.
7427 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7428 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7430 if (isa<PHINode>(Op0))
7431 if (Instruction *NV = FoldOpIntoPhi(I))
7434 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7435 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7436 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7437 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7438 // place. Don't try to do this transformation in this case. Also, we
7439 // require that the input operand is a shift-by-constant so that we have
7440 // confidence that the shifts will get folded together. We could do this
7441 // xform in more cases, but it is unlikely to be profitable.
7442 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7443 isa<ConstantInt>(TrOp->getOperand(1))) {
7444 // Okay, we'll do this xform. Make the shift of shift.
7445 Constant *ShAmt = Context->getConstantExprZExt(Op1, TrOp->getType());
7446 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7448 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7450 // For logical shifts, the truncation has the effect of making the high
7451 // part of the register be zeros. Emulate this by inserting an AND to
7452 // clear the top bits as needed. This 'and' will usually be zapped by
7453 // other xforms later if dead.
7454 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7455 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7456 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7458 // The mask we constructed says what the trunc would do if occurring
7459 // between the shifts. We want to know the effect *after* the second
7460 // shift. We know that it is a logical shift by a constant, so adjust the
7461 // mask as appropriate.
7462 if (I.getOpcode() == Instruction::Shl)
7463 MaskV <<= Op1->getZExtValue();
7465 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7466 MaskV = MaskV.lshr(Op1->getZExtValue());
7470 BinaryOperator::CreateAnd(NSh, Context->getConstantInt(MaskV),
7472 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7474 // Return the value truncated to the interesting size.
7475 return new TruncInst(And, I.getType());
7479 if (Op0->hasOneUse()) {
7480 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7481 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7484 switch (Op0BO->getOpcode()) {
7486 case Instruction::Add:
7487 case Instruction::And:
7488 case Instruction::Or:
7489 case Instruction::Xor: {
7490 // These operators commute.
7491 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7492 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7493 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7494 m_Specific(Op1)), *Context)){
7495 Instruction *YS = BinaryOperator::CreateShl(
7496 Op0BO->getOperand(0), Op1,
7498 InsertNewInstBefore(YS, I); // (Y << C)
7500 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7501 Op0BO->getOperand(1)->getName());
7502 InsertNewInstBefore(X, I); // (X + (Y << C))
7503 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7504 return BinaryOperator::CreateAnd(X, Context->getConstantInt(
7505 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7508 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7509 Value *Op0BOOp1 = Op0BO->getOperand(1);
7510 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7512 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7513 m_ConstantInt(CC)), *Context) &&
7514 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7515 Instruction *YS = BinaryOperator::CreateShl(
7516 Op0BO->getOperand(0), Op1,
7518 InsertNewInstBefore(YS, I); // (Y << C)
7520 BinaryOperator::CreateAnd(V1,
7521 Context->getConstantExprShl(CC, Op1),
7522 V1->getName()+".mask");
7523 InsertNewInstBefore(XM, I); // X & (CC << C)
7525 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7530 case Instruction::Sub: {
7531 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7532 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7533 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7534 m_Specific(Op1)), *Context)){
7535 Instruction *YS = BinaryOperator::CreateShl(
7536 Op0BO->getOperand(1), Op1,
7538 InsertNewInstBefore(YS, I); // (Y << C)
7540 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7541 Op0BO->getOperand(0)->getName());
7542 InsertNewInstBefore(X, I); // (X + (Y << C))
7543 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7544 return BinaryOperator::CreateAnd(X, Context->getConstantInt(
7545 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7548 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7549 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7550 match(Op0BO->getOperand(0),
7551 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7552 m_ConstantInt(CC)), *Context) && V2 == Op1 &&
7553 cast<BinaryOperator>(Op0BO->getOperand(0))
7554 ->getOperand(0)->hasOneUse()) {
7555 Instruction *YS = BinaryOperator::CreateShl(
7556 Op0BO->getOperand(1), Op1,
7558 InsertNewInstBefore(YS, I); // (Y << C)
7560 BinaryOperator::CreateAnd(V1,
7561 Context->getConstantExprShl(CC, Op1),
7562 V1->getName()+".mask");
7563 InsertNewInstBefore(XM, I); // X & (CC << C)
7565 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7573 // If the operand is an bitwise operator with a constant RHS, and the
7574 // shift is the only use, we can pull it out of the shift.
7575 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7576 bool isValid = true; // Valid only for And, Or, Xor
7577 bool highBitSet = false; // Transform if high bit of constant set?
7579 switch (Op0BO->getOpcode()) {
7580 default: isValid = false; break; // Do not perform transform!
7581 case Instruction::Add:
7582 isValid = isLeftShift;
7584 case Instruction::Or:
7585 case Instruction::Xor:
7588 case Instruction::And:
7593 // If this is a signed shift right, and the high bit is modified
7594 // by the logical operation, do not perform the transformation.
7595 // The highBitSet boolean indicates the value of the high bit of
7596 // the constant which would cause it to be modified for this
7599 if (isValid && I.getOpcode() == Instruction::AShr)
7600 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7603 Constant *NewRHS = Context->getConstantExpr(I.getOpcode(), Op0C, Op1);
7605 Instruction *NewShift =
7606 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7607 InsertNewInstBefore(NewShift, I);
7608 NewShift->takeName(Op0BO);
7610 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7617 // Find out if this is a shift of a shift by a constant.
7618 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7619 if (ShiftOp && !ShiftOp->isShift())
7622 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7623 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7624 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7625 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7626 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7627 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7628 Value *X = ShiftOp->getOperand(0);
7630 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7632 const IntegerType *Ty = cast<IntegerType>(I.getType());
7634 // Check for (X << c1) << c2 and (X >> c1) >> c2
7635 if (I.getOpcode() == ShiftOp->getOpcode()) {
7636 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7638 if (AmtSum >= TypeBits) {
7639 if (I.getOpcode() != Instruction::AShr)
7640 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7641 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7644 return BinaryOperator::Create(I.getOpcode(), X,
7645 Context->getConstantInt(Ty, AmtSum));
7646 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7647 I.getOpcode() == Instruction::AShr) {
7648 if (AmtSum >= TypeBits)
7649 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7651 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7652 return BinaryOperator::CreateLShr(X, Context->getConstantInt(Ty, AmtSum));
7653 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7654 I.getOpcode() == Instruction::LShr) {
7655 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7656 if (AmtSum >= TypeBits)
7657 AmtSum = TypeBits-1;
7659 Instruction *Shift =
7660 BinaryOperator::CreateAShr(X, Context->getConstantInt(Ty, AmtSum));
7661 InsertNewInstBefore(Shift, I);
7663 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7664 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7667 // Okay, if we get here, one shift must be left, and the other shift must be
7668 // right. See if the amounts are equal.
7669 if (ShiftAmt1 == ShiftAmt2) {
7670 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7671 if (I.getOpcode() == Instruction::Shl) {
7672 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7673 return BinaryOperator::CreateAnd(X, Context->getConstantInt(Mask));
7675 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7676 if (I.getOpcode() == Instruction::LShr) {
7677 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7678 return BinaryOperator::CreateAnd(X, Context->getConstantInt(Mask));
7680 // We can simplify ((X << C) >>s C) into a trunc + sext.
7681 // NOTE: we could do this for any C, but that would make 'unusual' integer
7682 // types. For now, just stick to ones well-supported by the code
7684 const Type *SExtType = 0;
7685 switch (Ty->getBitWidth() - ShiftAmt1) {
7692 SExtType = Context->getIntegerType(Ty->getBitWidth() - ShiftAmt1);
7697 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7698 InsertNewInstBefore(NewTrunc, I);
7699 return new SExtInst(NewTrunc, Ty);
7701 // Otherwise, we can't handle it yet.
7702 } else if (ShiftAmt1 < ShiftAmt2) {
7703 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7705 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7706 if (I.getOpcode() == Instruction::Shl) {
7707 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7708 ShiftOp->getOpcode() == Instruction::AShr);
7709 Instruction *Shift =
7710 BinaryOperator::CreateShl(X, Context->getConstantInt(Ty, ShiftDiff));
7711 InsertNewInstBefore(Shift, I);
7713 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7714 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7717 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7718 if (I.getOpcode() == Instruction::LShr) {
7719 assert(ShiftOp->getOpcode() == Instruction::Shl);
7720 Instruction *Shift =
7721 BinaryOperator::CreateLShr(X, Context->getConstantInt(Ty, ShiftDiff));
7722 InsertNewInstBefore(Shift, I);
7724 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7725 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7728 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7730 assert(ShiftAmt2 < ShiftAmt1);
7731 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7733 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7734 if (I.getOpcode() == Instruction::Shl) {
7735 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7736 ShiftOp->getOpcode() == Instruction::AShr);
7737 Instruction *Shift =
7738 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7739 Context->getConstantInt(Ty, ShiftDiff));
7740 InsertNewInstBefore(Shift, I);
7742 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7743 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(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, Context->getConstantInt(Ty, ShiftDiff));
7751 InsertNewInstBefore(Shift, I);
7753 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7754 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7757 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7764 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7765 /// expression. If so, decompose it, returning some value X, such that Val is
7768 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7769 int &Offset, LLVMContext *Context) {
7770 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7771 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7772 Offset = CI->getZExtValue();
7774 return Context->getConstantInt(Type::Int32Ty, 0);
7775 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7776 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7777 if (I->getOpcode() == Instruction::Shl) {
7778 // This is a value scaled by '1 << the shift amt'.
7779 Scale = 1U << RHS->getZExtValue();
7781 return I->getOperand(0);
7782 } else if (I->getOpcode() == Instruction::Mul) {
7783 // This value is scaled by 'RHS'.
7784 Scale = RHS->getZExtValue();
7786 return I->getOperand(0);
7787 } else if (I->getOpcode() == Instruction::Add) {
7788 // We have X+C. Check to see if we really have (X*C2)+C1,
7789 // where C1 is divisible by C2.
7792 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7794 Offset += RHS->getZExtValue();
7801 // Otherwise, we can't look past this.
7808 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7809 /// try to eliminate the cast by moving the type information into the alloc.
7810 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7811 AllocationInst &AI) {
7812 const PointerType *PTy = cast<PointerType>(CI.getType());
7814 // Remove any uses of AI that are dead.
7815 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7817 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7818 Instruction *User = cast<Instruction>(*UI++);
7819 if (isInstructionTriviallyDead(User)) {
7820 while (UI != E && *UI == User)
7821 ++UI; // If this instruction uses AI more than once, don't break UI.
7824 DOUT << "IC: DCE: " << *User;
7825 EraseInstFromFunction(*User);
7829 // Get the type really allocated and the type casted to.
7830 const Type *AllocElTy = AI.getAllocatedType();
7831 const Type *CastElTy = PTy->getElementType();
7832 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7834 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7835 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7836 if (CastElTyAlign < AllocElTyAlign) return 0;
7838 // If the allocation has multiple uses, only promote it if we are strictly
7839 // increasing the alignment of the resultant allocation. If we keep it the
7840 // same, we open the door to infinite loops of various kinds. (A reference
7841 // from a dbg.declare doesn't count as a use for this purpose.)
7842 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7843 CastElTyAlign == AllocElTyAlign) return 0;
7845 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7846 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7847 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7849 // See if we can satisfy the modulus by pulling a scale out of the array
7851 unsigned ArraySizeScale;
7853 Value *NumElements = // See if the array size is a decomposable linear expr.
7854 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7855 ArrayOffset, Context);
7857 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7859 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7860 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7862 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7867 // If the allocation size is constant, form a constant mul expression
7868 Amt = Context->getConstantInt(Type::Int32Ty, Scale);
7869 if (isa<ConstantInt>(NumElements))
7870 Amt = Context->getConstantExprMul(cast<ConstantInt>(NumElements),
7871 cast<ConstantInt>(Amt));
7872 // otherwise multiply the amount and the number of elements
7874 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7875 Amt = InsertNewInstBefore(Tmp, AI);
7879 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7880 Value *Off = Context->getConstantInt(Type::Int32Ty, Offset, true);
7881 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7882 Amt = InsertNewInstBefore(Tmp, AI);
7885 AllocationInst *New;
7886 if (isa<MallocInst>(AI))
7887 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7889 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7890 InsertNewInstBefore(New, AI);
7893 // If the allocation has one real use plus a dbg.declare, just remove the
7895 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7896 EraseInstFromFunction(*DI);
7898 // If the allocation has multiple real uses, insert a cast and change all
7899 // things that used it to use the new cast. This will also hack on CI, but it
7901 else if (!AI.hasOneUse()) {
7902 AddUsesToWorkList(AI);
7903 // New is the allocation instruction, pointer typed. AI is the original
7904 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7905 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7906 InsertNewInstBefore(NewCast, AI);
7907 AI.replaceAllUsesWith(NewCast);
7909 return ReplaceInstUsesWith(CI, New);
7912 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7913 /// and return it as type Ty without inserting any new casts and without
7914 /// changing the computed value. This is used by code that tries to decide
7915 /// whether promoting or shrinking integer operations to wider or smaller types
7916 /// will allow us to eliminate a truncate or extend.
7918 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7919 /// extension operation if Ty is larger.
7921 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7922 /// should return true if trunc(V) can be computed by computing V in the smaller
7923 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7924 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7925 /// efficiently truncated.
7927 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7928 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7929 /// the final result.
7930 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7932 int &NumCastsRemoved){
7933 // We can always evaluate constants in another type.
7934 if (isa<Constant>(V))
7937 Instruction *I = dyn_cast<Instruction>(V);
7938 if (!I) return false;
7940 const Type *OrigTy = V->getType();
7942 // If this is an extension or truncate, we can often eliminate it.
7943 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7944 // If this is a cast from the destination type, we can trivially eliminate
7945 // it, and this will remove a cast overall.
7946 if (I->getOperand(0)->getType() == Ty) {
7947 // If the first operand is itself a cast, and is eliminable, do not count
7948 // this as an eliminable cast. We would prefer to eliminate those two
7950 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7956 // We can't extend or shrink something that has multiple uses: doing so would
7957 // require duplicating the instruction in general, which isn't profitable.
7958 if (!I->hasOneUse()) return false;
7960 unsigned Opc = I->getOpcode();
7962 case Instruction::Add:
7963 case Instruction::Sub:
7964 case Instruction::Mul:
7965 case Instruction::And:
7966 case Instruction::Or:
7967 case Instruction::Xor:
7968 // These operators can all arbitrarily be extended or truncated.
7969 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7971 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7974 case Instruction::UDiv:
7975 case Instruction::URem: {
7976 // UDiv and URem can be truncated if all the truncated bits are zero.
7977 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7978 uint32_t BitWidth = Ty->getScalarSizeInBits();
7979 if (BitWidth < OrigBitWidth) {
7980 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7981 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7982 MaskedValueIsZero(I->getOperand(1), Mask)) {
7983 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7985 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7991 case Instruction::Shl:
7992 // If we are truncating the result of this SHL, and if it's a shift of a
7993 // constant amount, we can always perform a SHL in a smaller type.
7994 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7995 uint32_t BitWidth = Ty->getScalarSizeInBits();
7996 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7997 CI->getLimitedValue(BitWidth) < BitWidth)
7998 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8002 case Instruction::LShr:
8003 // If this is a truncate of a logical shr, we can truncate it to a smaller
8004 // lshr iff we know that the bits we would otherwise be shifting in are
8006 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
8007 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8008 uint32_t BitWidth = Ty->getScalarSizeInBits();
8009 if (BitWidth < OrigBitWidth &&
8010 MaskedValueIsZero(I->getOperand(0),
8011 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
8012 CI->getLimitedValue(BitWidth) < BitWidth) {
8013 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8018 case Instruction::ZExt:
8019 case Instruction::SExt:
8020 case Instruction::Trunc:
8021 // If this is the same kind of case as our original (e.g. zext+zext), we
8022 // can safely replace it. Note that replacing it does not reduce the number
8023 // of casts in the input.
8027 // sext (zext ty1), ty2 -> zext ty2
8028 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8031 case Instruction::Select: {
8032 SelectInst *SI = cast<SelectInst>(I);
8033 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8035 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8038 case Instruction::PHI: {
8039 // We can change a phi if we can change all operands.
8040 PHINode *PN = cast<PHINode>(I);
8041 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8042 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8048 // TODO: Can handle more cases here.
8055 /// EvaluateInDifferentType - Given an expression that
8056 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8057 /// evaluate the expression.
8058 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8060 if (Constant *C = dyn_cast<Constant>(V))
8061 return Context->getConstantExprIntegerCast(C, Ty,
8062 isSigned /*Sext or ZExt*/);
8064 // Otherwise, it must be an instruction.
8065 Instruction *I = cast<Instruction>(V);
8066 Instruction *Res = 0;
8067 unsigned Opc = I->getOpcode();
8069 case Instruction::Add:
8070 case Instruction::Sub:
8071 case Instruction::Mul:
8072 case Instruction::And:
8073 case Instruction::Or:
8074 case Instruction::Xor:
8075 case Instruction::AShr:
8076 case Instruction::LShr:
8077 case Instruction::Shl:
8078 case Instruction::UDiv:
8079 case Instruction::URem: {
8080 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8081 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8082 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8085 case Instruction::Trunc:
8086 case Instruction::ZExt:
8087 case Instruction::SExt:
8088 // If the source type of the cast is the type we're trying for then we can
8089 // just return the source. There's no need to insert it because it is not
8091 if (I->getOperand(0)->getType() == Ty)
8092 return I->getOperand(0);
8094 // Otherwise, must be the same type of cast, so just reinsert a new one.
8095 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8098 case Instruction::Select: {
8099 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8100 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8101 Res = SelectInst::Create(I->getOperand(0), True, False);
8104 case Instruction::PHI: {
8105 PHINode *OPN = cast<PHINode>(I);
8106 PHINode *NPN = PHINode::Create(Ty);
8107 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8108 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8109 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8115 // TODO: Can handle more cases here.
8116 llvm_unreachable("Unreachable!");
8121 return InsertNewInstBefore(Res, *I);
8124 /// @brief Implement the transforms common to all CastInst visitors.
8125 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8126 Value *Src = CI.getOperand(0);
8128 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8129 // eliminate it now.
8130 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8131 if (Instruction::CastOps opc =
8132 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8133 // The first cast (CSrc) is eliminable so we need to fix up or replace
8134 // the second cast (CI). CSrc will then have a good chance of being dead.
8135 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8139 // If we are casting a select then fold the cast into the select
8140 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8141 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8144 // If we are casting a PHI then fold the cast into the PHI
8145 if (isa<PHINode>(Src))
8146 if (Instruction *NV = FoldOpIntoPhi(CI))
8152 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8153 /// or not there is a sequence of GEP indices into the type that will land us at
8154 /// the specified offset. If so, fill them into NewIndices and return the
8155 /// resultant element type, otherwise return null.
8156 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8157 SmallVectorImpl<Value*> &NewIndices,
8158 const TargetData *TD,
8159 LLVMContext *Context) {
8160 if (!Ty->isSized()) return 0;
8162 // Start with the index over the outer type. Note that the type size
8163 // might be zero (even if the offset isn't zero) if the indexed type
8164 // is something like [0 x {int, int}]
8165 const Type *IntPtrTy = TD->getIntPtrType();
8166 int64_t FirstIdx = 0;
8167 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8168 FirstIdx = Offset/TySize;
8169 Offset -= FirstIdx*TySize;
8171 // Handle hosts where % returns negative instead of values [0..TySize).
8175 assert(Offset >= 0);
8177 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8180 NewIndices.push_back(Context->getConstantInt(IntPtrTy, FirstIdx));
8182 // Index into the types. If we fail, set OrigBase to null.
8184 // Indexing into tail padding between struct/array elements.
8185 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8188 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8189 const StructLayout *SL = TD->getStructLayout(STy);
8190 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8191 "Offset must stay within the indexed type");
8193 unsigned Elt = SL->getElementContainingOffset(Offset);
8194 NewIndices.push_back(Context->getConstantInt(Type::Int32Ty, Elt));
8196 Offset -= SL->getElementOffset(Elt);
8197 Ty = STy->getElementType(Elt);
8198 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8199 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8200 assert(EltSize && "Cannot index into a zero-sized array");
8201 NewIndices.push_back(Context->getConstantInt(IntPtrTy,Offset/EltSize));
8203 Ty = AT->getElementType();
8205 // Otherwise, we can't index into the middle of this atomic type, bail.
8213 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8214 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8215 Value *Src = CI.getOperand(0);
8217 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8218 // If casting the result of a getelementptr instruction with no offset, turn
8219 // this into a cast of the original pointer!
8220 if (GEP->hasAllZeroIndices()) {
8221 // Changing the cast operand is usually not a good idea but it is safe
8222 // here because the pointer operand is being replaced with another
8223 // pointer operand so the opcode doesn't need to change.
8225 CI.setOperand(0, GEP->getOperand(0));
8229 // If the GEP has a single use, and the base pointer is a bitcast, and the
8230 // GEP computes a constant offset, see if we can convert these three
8231 // instructions into fewer. This typically happens with unions and other
8232 // non-type-safe code.
8233 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8234 if (GEP->hasAllConstantIndices()) {
8235 // We are guaranteed to get a constant from EmitGEPOffset.
8236 ConstantInt *OffsetV =
8237 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8238 int64_t Offset = OffsetV->getSExtValue();
8240 // Get the base pointer input of the bitcast, and the type it points to.
8241 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8242 const Type *GEPIdxTy =
8243 cast<PointerType>(OrigBase->getType())->getElementType();
8244 SmallVector<Value*, 8> NewIndices;
8245 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8246 // If we were able to index down into an element, create the GEP
8247 // and bitcast the result. This eliminates one bitcast, potentially
8249 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
8251 NewIndices.end(), "");
8252 InsertNewInstBefore(NGEP, CI);
8253 NGEP->takeName(GEP);
8255 if (isa<BitCastInst>(CI))
8256 return new BitCastInst(NGEP, CI.getType());
8257 assert(isa<PtrToIntInst>(CI));
8258 return new PtrToIntInst(NGEP, CI.getType());
8264 return commonCastTransforms(CI);
8267 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8268 /// type like i42. We don't want to introduce operations on random non-legal
8269 /// integer types where they don't already exist in the code. In the future,
8270 /// we should consider making this based off target-data, so that 32-bit targets
8271 /// won't get i64 operations etc.
8272 static bool isSafeIntegerType(const Type *Ty) {
8273 switch (Ty->getPrimitiveSizeInBits()) {
8284 /// commonIntCastTransforms - This function implements the common transforms
8285 /// for trunc, zext, and sext.
8286 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8287 if (Instruction *Result = commonCastTransforms(CI))
8290 Value *Src = CI.getOperand(0);
8291 const Type *SrcTy = Src->getType();
8292 const Type *DestTy = CI.getType();
8293 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8294 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8296 // See if we can simplify any instructions used by the LHS whose sole
8297 // purpose is to compute bits we don't care about.
8298 if (SimplifyDemandedInstructionBits(CI))
8301 // If the source isn't an instruction or has more than one use then we
8302 // can't do anything more.
8303 Instruction *SrcI = dyn_cast<Instruction>(Src);
8304 if (!SrcI || !Src->hasOneUse())
8307 // Attempt to propagate the cast into the instruction for int->int casts.
8308 int NumCastsRemoved = 0;
8309 // Only do this if the dest type is a simple type, don't convert the
8310 // expression tree to something weird like i93 unless the source is also
8312 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8313 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8314 CanEvaluateInDifferentType(SrcI, DestTy,
8315 CI.getOpcode(), NumCastsRemoved)) {
8316 // If this cast is a truncate, evaluting in a different type always
8317 // eliminates the cast, so it is always a win. If this is a zero-extension,
8318 // we need to do an AND to maintain the clear top-part of the computation,
8319 // so we require that the input have eliminated at least one cast. If this
8320 // is a sign extension, we insert two new casts (to do the extension) so we
8321 // require that two casts have been eliminated.
8322 bool DoXForm = false;
8323 bool JustReplace = false;
8324 switch (CI.getOpcode()) {
8326 // All the others use floating point so we shouldn't actually
8327 // get here because of the check above.
8328 llvm_unreachable("Unknown cast type");
8329 case Instruction::Trunc:
8332 case Instruction::ZExt: {
8333 DoXForm = NumCastsRemoved >= 1;
8334 if (!DoXForm && 0) {
8335 // If it's unnecessary to issue an AND to clear the high bits, it's
8336 // always profitable to do this xform.
8337 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8338 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8339 if (MaskedValueIsZero(TryRes, Mask))
8340 return ReplaceInstUsesWith(CI, TryRes);
8342 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8343 if (TryI->use_empty())
8344 EraseInstFromFunction(*TryI);
8348 case Instruction::SExt: {
8349 DoXForm = NumCastsRemoved >= 2;
8350 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8351 // If we do not have to emit the truncate + sext pair, then it's always
8352 // profitable to do this xform.
8354 // It's not safe to eliminate the trunc + sext pair if one of the
8355 // eliminated cast is a truncate. e.g.
8356 // t2 = trunc i32 t1 to i16
8357 // t3 = sext i16 t2 to i32
8360 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8361 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8362 if (NumSignBits > (DestBitSize - SrcBitSize))
8363 return ReplaceInstUsesWith(CI, TryRes);
8365 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8366 if (TryI->use_empty())
8367 EraseInstFromFunction(*TryI);
8374 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
8376 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8377 CI.getOpcode() == Instruction::SExt);
8379 // Just replace this cast with the result.
8380 return ReplaceInstUsesWith(CI, Res);
8382 assert(Res->getType() == DestTy);
8383 switch (CI.getOpcode()) {
8384 default: llvm_unreachable("Unknown cast type!");
8385 case Instruction::Trunc:
8386 // Just replace this cast with the result.
8387 return ReplaceInstUsesWith(CI, Res);
8388 case Instruction::ZExt: {
8389 assert(SrcBitSize < DestBitSize && "Not a zext?");
8391 // If the high bits are already zero, just replace this cast with the
8393 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8394 if (MaskedValueIsZero(Res, Mask))
8395 return ReplaceInstUsesWith(CI, Res);
8397 // We need to emit an AND to clear the high bits.
8398 Constant *C = Context->getConstantInt(APInt::getLowBitsSet(DestBitSize,
8400 return BinaryOperator::CreateAnd(Res, C);
8402 case Instruction::SExt: {
8403 // If the high bits are already filled with sign bit, just replace this
8404 // cast with the result.
8405 unsigned NumSignBits = ComputeNumSignBits(Res);
8406 if (NumSignBits > (DestBitSize - SrcBitSize))
8407 return ReplaceInstUsesWith(CI, Res);
8409 // We need to emit a cast to truncate, then a cast to sext.
8410 return CastInst::Create(Instruction::SExt,
8411 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8418 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8419 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8421 switch (SrcI->getOpcode()) {
8422 case Instruction::Add:
8423 case Instruction::Mul:
8424 case Instruction::And:
8425 case Instruction::Or:
8426 case Instruction::Xor:
8427 // If we are discarding information, rewrite.
8428 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8429 // Don't insert two casts unless at least one can be eliminated.
8430 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8431 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8432 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8433 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8434 return BinaryOperator::Create(
8435 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8439 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8440 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8441 SrcI->getOpcode() == Instruction::Xor &&
8442 Op1 == Context->getConstantIntTrue() &&
8443 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8444 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8445 return BinaryOperator::CreateXor(New,
8446 Context->getConstantInt(CI.getType(), 1));
8450 case Instruction::Shl: {
8451 // Canonicalize trunc inside shl, if we can.
8452 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8453 if (CI && DestBitSize < SrcBitSize &&
8454 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8455 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8456 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8457 return BinaryOperator::CreateShl(Op0c, Op1c);
8465 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8466 if (Instruction *Result = commonIntCastTransforms(CI))
8469 Value *Src = CI.getOperand(0);
8470 const Type *Ty = CI.getType();
8471 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8472 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8474 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8475 if (DestBitWidth == 1 &&
8476 isa<VectorType>(Ty) == isa<VectorType>(Src->getType())) {
8477 Constant *One = Context->getConstantInt(Src->getType(), 1);
8478 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8479 Value *Zero = Context->getNullValue(Src->getType());
8480 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Src, Zero);
8483 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8484 ConstantInt *ShAmtV = 0;
8486 if (Src->hasOneUse() &&
8487 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)), *Context)) {
8488 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8490 // Get a mask for the bits shifting in.
8491 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8492 if (MaskedValueIsZero(ShiftOp, Mask)) {
8493 if (ShAmt >= DestBitWidth) // All zeros.
8494 return ReplaceInstUsesWith(CI, Context->getNullValue(Ty));
8496 // Okay, we can shrink this. Truncate the input, then return a new
8498 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8499 Value *V2 = Context->getConstantExprTrunc(ShAmtV, Ty);
8500 return BinaryOperator::CreateLShr(V1, V2);
8507 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8508 /// in order to eliminate the icmp.
8509 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8511 // If we are just checking for a icmp eq of a single bit and zext'ing it
8512 // to an integer, then shift the bit to the appropriate place and then
8513 // cast to integer to avoid the comparison.
8514 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8515 const APInt &Op1CV = Op1C->getValue();
8517 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8518 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8519 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8520 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8521 if (!DoXform) return ICI;
8523 Value *In = ICI->getOperand(0);
8524 Value *Sh = Context->getConstantInt(In->getType(),
8525 In->getType()->getScalarSizeInBits()-1);
8526 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8527 In->getName()+".lobit"),
8529 if (In->getType() != CI.getType())
8530 In = CastInst::CreateIntegerCast(In, CI.getType(),
8531 false/*ZExt*/, "tmp", &CI);
8533 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8534 Constant *One = Context->getConstantInt(In->getType(), 1);
8535 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8536 In->getName()+".not"),
8540 return ReplaceInstUsesWith(CI, In);
8545 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8546 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8547 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8548 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8549 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8550 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8551 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8552 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8553 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8554 // This only works for EQ and NE
8555 ICI->isEquality()) {
8556 // If Op1C some other power of two, convert:
8557 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8558 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8559 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8560 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8562 APInt KnownZeroMask(~KnownZero);
8563 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8564 if (!DoXform) return ICI;
8566 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8567 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8568 // (X&4) == 2 --> false
8569 // (X&4) != 2 --> true
8570 Constant *Res = Context->getConstantInt(Type::Int1Ty, isNE);
8571 Res = Context->getConstantExprZExt(Res, CI.getType());
8572 return ReplaceInstUsesWith(CI, Res);
8575 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8576 Value *In = ICI->getOperand(0);
8578 // Perform a logical shr by shiftamt.
8579 // Insert the shift to put the result in the low bit.
8580 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8581 Context->getConstantInt(In->getType(), ShiftAmt),
8582 In->getName()+".lobit"), CI);
8585 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8586 Constant *One = Context->getConstantInt(In->getType(), 1);
8587 In = BinaryOperator::CreateXor(In, One, "tmp");
8588 InsertNewInstBefore(cast<Instruction>(In), CI);
8591 if (CI.getType() == In->getType())
8592 return ReplaceInstUsesWith(CI, In);
8594 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8602 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8603 // If one of the common conversion will work ..
8604 if (Instruction *Result = commonIntCastTransforms(CI))
8607 Value *Src = CI.getOperand(0);
8609 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8610 // types and if the sizes are just right we can convert this into a logical
8611 // 'and' which will be much cheaper than the pair of casts.
8612 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8613 // Get the sizes of the types involved. We know that the intermediate type
8614 // will be smaller than A or C, but don't know the relation between A and C.
8615 Value *A = CSrc->getOperand(0);
8616 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8617 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8618 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8619 // If we're actually extending zero bits, then if
8620 // SrcSize < DstSize: zext(a & mask)
8621 // SrcSize == DstSize: a & mask
8622 // SrcSize > DstSize: trunc(a) & mask
8623 if (SrcSize < DstSize) {
8624 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8625 Constant *AndConst = Context->getConstantInt(A->getType(), AndValue);
8627 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8628 InsertNewInstBefore(And, CI);
8629 return new ZExtInst(And, CI.getType());
8630 } else if (SrcSize == DstSize) {
8631 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8632 return BinaryOperator::CreateAnd(A, Context->getConstantInt(A->getType(),
8634 } else if (SrcSize > DstSize) {
8635 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8636 InsertNewInstBefore(Trunc, CI);
8637 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8638 return BinaryOperator::CreateAnd(Trunc,
8639 Context->getConstantInt(Trunc->getType(),
8644 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8645 return transformZExtICmp(ICI, CI);
8647 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8648 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8649 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8650 // of the (zext icmp) will be transformed.
8651 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8652 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8653 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8654 (transformZExtICmp(LHS, CI, false) ||
8655 transformZExtICmp(RHS, CI, false))) {
8656 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8657 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8658 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8662 // zext(trunc(t) & C) -> (t & zext(C)).
8663 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8664 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8665 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8666 Value *TI0 = TI->getOperand(0);
8667 if (TI0->getType() == CI.getType())
8669 BinaryOperator::CreateAnd(TI0,
8670 Context->getConstantExprZExt(C, CI.getType()));
8673 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8674 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8675 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8676 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8677 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8678 And->getOperand(1) == C)
8679 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8680 Value *TI0 = TI->getOperand(0);
8681 if (TI0->getType() == CI.getType()) {
8682 Constant *ZC = Context->getConstantExprZExt(C, CI.getType());
8683 Instruction *NewAnd = BinaryOperator::CreateAnd(TI0, ZC, "tmp");
8684 InsertNewInstBefore(NewAnd, *And);
8685 return BinaryOperator::CreateXor(NewAnd, ZC);
8692 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8693 if (Instruction *I = commonIntCastTransforms(CI))
8696 Value *Src = CI.getOperand(0);
8698 // Canonicalize sign-extend from i1 to a select.
8699 if (Src->getType() == Type::Int1Ty)
8700 return SelectInst::Create(Src,
8701 Context->getAllOnesValue(CI.getType()),
8702 Context->getNullValue(CI.getType()));
8704 // See if the value being truncated is already sign extended. If so, just
8705 // eliminate the trunc/sext pair.
8706 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8707 Value *Op = cast<User>(Src)->getOperand(0);
8708 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8709 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8710 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8711 unsigned NumSignBits = ComputeNumSignBits(Op);
8713 if (OpBits == DestBits) {
8714 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8715 // bits, it is already ready.
8716 if (NumSignBits > DestBits-MidBits)
8717 return ReplaceInstUsesWith(CI, Op);
8718 } else if (OpBits < DestBits) {
8719 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8720 // bits, just sext from i32.
8721 if (NumSignBits > OpBits-MidBits)
8722 return new SExtInst(Op, CI.getType(), "tmp");
8724 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8725 // bits, just truncate to i32.
8726 if (NumSignBits > OpBits-MidBits)
8727 return new TruncInst(Op, CI.getType(), "tmp");
8731 // If the input is a shl/ashr pair of a same constant, then this is a sign
8732 // extension from a smaller value. If we could trust arbitrary bitwidth
8733 // integers, we could turn this into a truncate to the smaller bit and then
8734 // use a sext for the whole extension. Since we don't, look deeper and check
8735 // for a truncate. If the source and dest are the same type, eliminate the
8736 // trunc and extend and just do shifts. For example, turn:
8737 // %a = trunc i32 %i to i8
8738 // %b = shl i8 %a, 6
8739 // %c = ashr i8 %b, 6
8740 // %d = sext i8 %c to i32
8742 // %a = shl i32 %i, 30
8743 // %d = ashr i32 %a, 30
8745 ConstantInt *BA = 0, *CA = 0;
8746 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8747 m_ConstantInt(CA)), *Context) &&
8748 BA == CA && isa<TruncInst>(A)) {
8749 Value *I = cast<TruncInst>(A)->getOperand(0);
8750 if (I->getType() == CI.getType()) {
8751 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8752 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8753 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8754 Constant *ShAmtV = Context->getConstantInt(CI.getType(), ShAmt);
8755 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8757 return BinaryOperator::CreateAShr(I, ShAmtV);
8764 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8765 /// in the specified FP type without changing its value.
8766 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8767 LLVMContext *Context) {
8769 APFloat F = CFP->getValueAPF();
8770 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8772 return Context->getConstantFP(F);
8776 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8777 /// through it until we get the source value.
8778 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8779 if (Instruction *I = dyn_cast<Instruction>(V))
8780 if (I->getOpcode() == Instruction::FPExt)
8781 return LookThroughFPExtensions(I->getOperand(0), Context);
8783 // If this value is a constant, return the constant in the smallest FP type
8784 // that can accurately represent it. This allows us to turn
8785 // (float)((double)X+2.0) into x+2.0f.
8786 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8787 if (CFP->getType() == Type::PPC_FP128Ty)
8788 return V; // No constant folding of this.
8789 // See if the value can be truncated to float and then reextended.
8790 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8792 if (CFP->getType() == Type::DoubleTy)
8793 return V; // Won't shrink.
8794 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8796 // Don't try to shrink to various long double types.
8802 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8803 if (Instruction *I = commonCastTransforms(CI))
8806 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8807 // smaller than the destination type, we can eliminate the truncate by doing
8808 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8809 // many builtins (sqrt, etc).
8810 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8811 if (OpI && OpI->hasOneUse()) {
8812 switch (OpI->getOpcode()) {
8814 case Instruction::FAdd:
8815 case Instruction::FSub:
8816 case Instruction::FMul:
8817 case Instruction::FDiv:
8818 case Instruction::FRem:
8819 const Type *SrcTy = OpI->getType();
8820 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8821 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8822 if (LHSTrunc->getType() != SrcTy &&
8823 RHSTrunc->getType() != SrcTy) {
8824 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8825 // If the source types were both smaller than the destination type of
8826 // the cast, do this xform.
8827 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8828 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8829 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8831 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8833 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8842 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8843 return commonCastTransforms(CI);
8846 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8847 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8849 return commonCastTransforms(FI);
8851 // fptoui(uitofp(X)) --> X
8852 // fptoui(sitofp(X)) --> X
8853 // This is safe if the intermediate type has enough bits in its mantissa to
8854 // accurately represent all values of X. For example, do not do this with
8855 // i64->float->i64. This is also safe for sitofp case, because any negative
8856 // 'X' value would cause an undefined result for the fptoui.
8857 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8858 OpI->getOperand(0)->getType() == FI.getType() &&
8859 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8860 OpI->getType()->getFPMantissaWidth())
8861 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8863 return commonCastTransforms(FI);
8866 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8867 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8869 return commonCastTransforms(FI);
8871 // fptosi(sitofp(X)) --> X
8872 // fptosi(uitofp(X)) --> X
8873 // This is safe if the intermediate type has enough bits in its mantissa to
8874 // accurately represent all values of X. For example, do not do this with
8875 // i64->float->i64. This is also safe for sitofp case, because any negative
8876 // 'X' value would cause an undefined result for the fptoui.
8877 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8878 OpI->getOperand(0)->getType() == FI.getType() &&
8879 (int)FI.getType()->getScalarSizeInBits() <=
8880 OpI->getType()->getFPMantissaWidth())
8881 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8883 return commonCastTransforms(FI);
8886 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8887 return commonCastTransforms(CI);
8890 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8891 return commonCastTransforms(CI);
8894 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8895 // If the destination integer type is smaller than the intptr_t type for
8896 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8897 // trunc to be exposed to other transforms. Don't do this for extending
8898 // ptrtoint's, because we don't know if the target sign or zero extends its
8900 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8901 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8902 TD->getIntPtrType(),
8904 return new TruncInst(P, CI.getType());
8907 return commonPointerCastTransforms(CI);
8910 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8911 // If the source integer type is larger than the intptr_t type for
8912 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8913 // allows the trunc to be exposed to other transforms. Don't do this for
8914 // extending inttoptr's, because we don't know if the target sign or zero
8915 // extends to pointers.
8916 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
8917 TD->getPointerSizeInBits()) {
8918 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8919 TD->getIntPtrType(),
8921 return new IntToPtrInst(P, CI.getType());
8924 if (Instruction *I = commonCastTransforms(CI))
8927 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8928 if (!DestPointee->isSized()) return 0;
8930 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8933 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8934 m_ConstantInt(Cst)), *Context)) {
8935 // If the source and destination operands have the same type, see if this
8936 // is a single-index GEP.
8937 if (X->getType() == CI.getType()) {
8938 // Get the size of the pointee type.
8939 uint64_t Size = TD->getTypeAllocSize(DestPointee);
8941 // Convert the constant to intptr type.
8942 APInt Offset = Cst->getValue();
8943 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8945 // If Offset is evenly divisible by Size, we can do this xform.
8946 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8947 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8948 GetElementPtrInst *GEP =
8949 GetElementPtrInst::Create(X, Context->getConstantInt(Offset));
8950 // A gep synthesized from inttoptr+add+ptrtoint must be assumed to
8951 // potentially overflow, in the absense of further analysis.
8952 cast<GEPOperator>(GEP)->setHasNoPointerOverflow(false);
8956 // TODO: Could handle other cases, e.g. where add is indexing into field of
8958 } else if (CI.getOperand(0)->hasOneUse() &&
8959 match(CI.getOperand(0), m_Add(m_Value(X),
8960 m_ConstantInt(Cst)), *Context)) {
8961 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8962 // "inttoptr+GEP" instead of "add+intptr".
8964 // Get the size of the pointee type.
8965 uint64_t Size = TD->getTypeAllocSize(DestPointee);
8967 // Convert the constant to intptr type.
8968 APInt Offset = Cst->getValue();
8969 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8971 // If Offset is evenly divisible by Size, we can do this xform.
8972 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8973 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8975 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8977 GetElementPtrInst *GEP =
8978 GetElementPtrInst::Create(P, Context->getConstantInt(Offset), "tmp");
8979 // A gep synthesized from inttoptr+add+ptrtoint must be assumed to
8980 // potentially overflow, in the absense of further analysis.
8981 cast<GEPOperator>(GEP)->setHasNoPointerOverflow(false);
8988 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8989 // If the operands are integer typed then apply the integer transforms,
8990 // otherwise just apply the common ones.
8991 Value *Src = CI.getOperand(0);
8992 const Type *SrcTy = Src->getType();
8993 const Type *DestTy = CI.getType();
8995 if (isa<PointerType>(SrcTy)) {
8996 if (Instruction *I = commonPointerCastTransforms(CI))
8999 if (Instruction *Result = commonCastTransforms(CI))
9004 // Get rid of casts from one type to the same type. These are useless and can
9005 // be replaced by the operand.
9006 if (DestTy == Src->getType())
9007 return ReplaceInstUsesWith(CI, Src);
9009 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
9010 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
9011 const Type *DstElTy = DstPTy->getElementType();
9012 const Type *SrcElTy = SrcPTy->getElementType();
9014 // If the address spaces don't match, don't eliminate the bitcast, which is
9015 // required for changing types.
9016 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
9019 // If we are casting a malloc or alloca to a pointer to a type of the same
9020 // size, rewrite the allocation instruction to allocate the "right" type.
9021 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
9022 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
9025 // If the source and destination are pointers, and this cast is equivalent
9026 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
9027 // This can enhance SROA and other transforms that want type-safe pointers.
9028 Constant *ZeroUInt = Context->getNullValue(Type::Int32Ty);
9029 unsigned NumZeros = 0;
9030 while (SrcElTy != DstElTy &&
9031 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
9032 SrcElTy->getNumContainedTypes() /* not "{}" */) {
9033 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
9037 // If we found a path from the src to dest, create the getelementptr now.
9038 if (SrcElTy == DstElTy) {
9039 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
9040 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
9041 ((Instruction*) NULL));
9045 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9046 if (SVI->hasOneUse()) {
9047 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9048 // a bitconvert to a vector with the same # elts.
9049 if (isa<VectorType>(DestTy) &&
9050 cast<VectorType>(DestTy)->getNumElements() ==
9051 SVI->getType()->getNumElements() &&
9052 SVI->getType()->getNumElements() ==
9053 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9055 // If either of the operands is a cast from CI.getType(), then
9056 // evaluating the shuffle in the casted destination's type will allow
9057 // us to eliminate at least one cast.
9058 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9059 Tmp->getOperand(0)->getType() == DestTy) ||
9060 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9061 Tmp->getOperand(0)->getType() == DestTy)) {
9062 Value *LHS = InsertCastBefore(Instruction::BitCast,
9063 SVI->getOperand(0), DestTy, CI);
9064 Value *RHS = InsertCastBefore(Instruction::BitCast,
9065 SVI->getOperand(1), DestTy, CI);
9066 // Return a new shuffle vector. Use the same element ID's, as we
9067 // know the vector types match #elts.
9068 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9076 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9078 /// %D = select %cond, %C, %A
9080 /// %C = select %cond, %B, 0
9083 /// Assuming that the specified instruction is an operand to the select, return
9084 /// a bitmask indicating which operands of this instruction are foldable if they
9085 /// equal the other incoming value of the select.
9087 static unsigned GetSelectFoldableOperands(Instruction *I) {
9088 switch (I->getOpcode()) {
9089 case Instruction::Add:
9090 case Instruction::Mul:
9091 case Instruction::And:
9092 case Instruction::Or:
9093 case Instruction::Xor:
9094 return 3; // Can fold through either operand.
9095 case Instruction::Sub: // Can only fold on the amount subtracted.
9096 case Instruction::Shl: // Can only fold on the shift amount.
9097 case Instruction::LShr:
9098 case Instruction::AShr:
9101 return 0; // Cannot fold
9105 /// GetSelectFoldableConstant - For the same transformation as the previous
9106 /// function, return the identity constant that goes into the select.
9107 static Constant *GetSelectFoldableConstant(Instruction *I,
9108 LLVMContext *Context) {
9109 switch (I->getOpcode()) {
9110 default: llvm_unreachable("This cannot happen!");
9111 case Instruction::Add:
9112 case Instruction::Sub:
9113 case Instruction::Or:
9114 case Instruction::Xor:
9115 case Instruction::Shl:
9116 case Instruction::LShr:
9117 case Instruction::AShr:
9118 return Context->getNullValue(I->getType());
9119 case Instruction::And:
9120 return Context->getAllOnesValue(I->getType());
9121 case Instruction::Mul:
9122 return Context->getConstantInt(I->getType(), 1);
9126 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9127 /// have the same opcode and only one use each. Try to simplify this.
9128 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9130 if (TI->getNumOperands() == 1) {
9131 // If this is a non-volatile load or a cast from the same type,
9134 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9137 return 0; // unknown unary op.
9140 // Fold this by inserting a select from the input values.
9141 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9142 FI->getOperand(0), SI.getName()+".v");
9143 InsertNewInstBefore(NewSI, SI);
9144 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9148 // Only handle binary operators here.
9149 if (!isa<BinaryOperator>(TI))
9152 // Figure out if the operations have any operands in common.
9153 Value *MatchOp, *OtherOpT, *OtherOpF;
9155 if (TI->getOperand(0) == FI->getOperand(0)) {
9156 MatchOp = TI->getOperand(0);
9157 OtherOpT = TI->getOperand(1);
9158 OtherOpF = FI->getOperand(1);
9159 MatchIsOpZero = true;
9160 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9161 MatchOp = TI->getOperand(1);
9162 OtherOpT = TI->getOperand(0);
9163 OtherOpF = FI->getOperand(0);
9164 MatchIsOpZero = false;
9165 } else if (!TI->isCommutative()) {
9167 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9168 MatchOp = TI->getOperand(0);
9169 OtherOpT = TI->getOperand(1);
9170 OtherOpF = FI->getOperand(0);
9171 MatchIsOpZero = true;
9172 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9173 MatchOp = TI->getOperand(1);
9174 OtherOpT = TI->getOperand(0);
9175 OtherOpF = FI->getOperand(1);
9176 MatchIsOpZero = true;
9181 // If we reach here, they do have operations in common.
9182 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9183 OtherOpF, SI.getName()+".v");
9184 InsertNewInstBefore(NewSI, SI);
9186 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9188 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9190 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9192 llvm_unreachable("Shouldn't get here");
9196 static bool isSelect01(Constant *C1, Constant *C2) {
9197 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9200 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9203 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9206 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9207 /// facilitate further optimization.
9208 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9210 // See the comment above GetSelectFoldableOperands for a description of the
9211 // transformation we are doing here.
9212 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9213 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9214 !isa<Constant>(FalseVal)) {
9215 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9216 unsigned OpToFold = 0;
9217 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9219 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9224 Constant *C = GetSelectFoldableConstant(TVI, Context);
9225 Value *OOp = TVI->getOperand(2-OpToFold);
9226 // Avoid creating select between 2 constants unless it's selecting
9228 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9229 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9230 InsertNewInstBefore(NewSel, SI);
9231 NewSel->takeName(TVI);
9232 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9233 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9234 llvm_unreachable("Unknown instruction!!");
9241 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9242 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9243 !isa<Constant>(TrueVal)) {
9244 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9245 unsigned OpToFold = 0;
9246 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9248 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9253 Constant *C = GetSelectFoldableConstant(FVI, Context);
9254 Value *OOp = FVI->getOperand(2-OpToFold);
9255 // Avoid creating select between 2 constants unless it's selecting
9257 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9258 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9259 InsertNewInstBefore(NewSel, SI);
9260 NewSel->takeName(FVI);
9261 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9262 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9263 llvm_unreachable("Unknown instruction!!");
9273 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9274 /// ICmpInst as its first operand.
9276 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9278 bool Changed = false;
9279 ICmpInst::Predicate Pred = ICI->getPredicate();
9280 Value *CmpLHS = ICI->getOperand(0);
9281 Value *CmpRHS = ICI->getOperand(1);
9282 Value *TrueVal = SI.getTrueValue();
9283 Value *FalseVal = SI.getFalseValue();
9285 // Check cases where the comparison is with a constant that
9286 // can be adjusted to fit the min/max idiom. We may edit ICI in
9287 // place here, so make sure the select is the only user.
9288 if (ICI->hasOneUse())
9289 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9292 case ICmpInst::ICMP_ULT:
9293 case ICmpInst::ICMP_SLT: {
9294 // X < MIN ? T : F --> F
9295 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9296 return ReplaceInstUsesWith(SI, FalseVal);
9297 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9298 Constant *AdjustedRHS = SubOne(CI, Context);
9299 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9300 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9301 Pred = ICmpInst::getSwappedPredicate(Pred);
9302 CmpRHS = AdjustedRHS;
9303 std::swap(FalseVal, TrueVal);
9304 ICI->setPredicate(Pred);
9305 ICI->setOperand(1, CmpRHS);
9306 SI.setOperand(1, TrueVal);
9307 SI.setOperand(2, FalseVal);
9312 case ICmpInst::ICMP_UGT:
9313 case ICmpInst::ICMP_SGT: {
9314 // X > MAX ? T : F --> F
9315 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9316 return ReplaceInstUsesWith(SI, FalseVal);
9317 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9318 Constant *AdjustedRHS = AddOne(CI, Context);
9319 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9320 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9321 Pred = ICmpInst::getSwappedPredicate(Pred);
9322 CmpRHS = AdjustedRHS;
9323 std::swap(FalseVal, TrueVal);
9324 ICI->setPredicate(Pred);
9325 ICI->setOperand(1, CmpRHS);
9326 SI.setOperand(1, TrueVal);
9327 SI.setOperand(2, FalseVal);
9334 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9335 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9336 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9337 if (match(TrueVal, m_ConstantInt<-1>(), *Context) &&
9338 match(FalseVal, m_ConstantInt<0>(), *Context))
9339 Pred = ICI->getPredicate();
9340 else if (match(TrueVal, m_ConstantInt<0>(), *Context) &&
9341 match(FalseVal, m_ConstantInt<-1>(), *Context))
9342 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9344 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9345 // If we are just checking for a icmp eq of a single bit and zext'ing it
9346 // to an integer, then shift the bit to the appropriate place and then
9347 // cast to integer to avoid the comparison.
9348 const APInt &Op1CV = CI->getValue();
9350 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9351 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9352 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9353 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9354 Value *In = ICI->getOperand(0);
9355 Value *Sh = Context->getConstantInt(In->getType(),
9356 In->getType()->getScalarSizeInBits()-1);
9357 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9358 In->getName()+".lobit"),
9360 if (In->getType() != SI.getType())
9361 In = CastInst::CreateIntegerCast(In, SI.getType(),
9362 true/*SExt*/, "tmp", ICI);
9364 if (Pred == ICmpInst::ICMP_SGT)
9365 In = InsertNewInstBefore(BinaryOperator::CreateNot(*Context, In,
9366 In->getName()+".not"), *ICI);
9368 return ReplaceInstUsesWith(SI, In);
9373 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9374 // Transform (X == Y) ? X : Y -> Y
9375 if (Pred == ICmpInst::ICMP_EQ)
9376 return ReplaceInstUsesWith(SI, FalseVal);
9377 // Transform (X != Y) ? X : Y -> X
9378 if (Pred == ICmpInst::ICMP_NE)
9379 return ReplaceInstUsesWith(SI, TrueVal);
9380 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9382 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9383 // Transform (X == Y) ? Y : X -> X
9384 if (Pred == ICmpInst::ICMP_EQ)
9385 return ReplaceInstUsesWith(SI, FalseVal);
9386 // Transform (X != Y) ? Y : X -> Y
9387 if (Pred == ICmpInst::ICMP_NE)
9388 return ReplaceInstUsesWith(SI, TrueVal);
9389 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9392 /// NOTE: if we wanted to, this is where to detect integer ABS
9394 return Changed ? &SI : 0;
9397 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9398 Value *CondVal = SI.getCondition();
9399 Value *TrueVal = SI.getTrueValue();
9400 Value *FalseVal = SI.getFalseValue();
9402 // select true, X, Y -> X
9403 // select false, X, Y -> Y
9404 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9405 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9407 // select C, X, X -> X
9408 if (TrueVal == FalseVal)
9409 return ReplaceInstUsesWith(SI, TrueVal);
9411 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9412 return ReplaceInstUsesWith(SI, FalseVal);
9413 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9414 return ReplaceInstUsesWith(SI, TrueVal);
9415 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9416 if (isa<Constant>(TrueVal))
9417 return ReplaceInstUsesWith(SI, TrueVal);
9419 return ReplaceInstUsesWith(SI, FalseVal);
9422 if (SI.getType() == Type::Int1Ty) {
9423 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9424 if (C->getZExtValue()) {
9425 // Change: A = select B, true, C --> A = or B, C
9426 return BinaryOperator::CreateOr(CondVal, FalseVal);
9428 // Change: A = select B, false, C --> A = and !B, C
9430 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9431 "not."+CondVal->getName()), SI);
9432 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9434 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9435 if (C->getZExtValue() == false) {
9436 // Change: A = select B, C, false --> A = and B, C
9437 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9439 // Change: A = select B, C, true --> A = or !B, C
9441 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9442 "not."+CondVal->getName()), SI);
9443 return BinaryOperator::CreateOr(NotCond, TrueVal);
9447 // select a, b, a -> a&b
9448 // select a, a, b -> a|b
9449 if (CondVal == TrueVal)
9450 return BinaryOperator::CreateOr(CondVal, FalseVal);
9451 else if (CondVal == FalseVal)
9452 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9455 // Selecting between two integer constants?
9456 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9457 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9458 // select C, 1, 0 -> zext C to int
9459 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9460 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9461 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9462 // select C, 0, 1 -> zext !C to int
9464 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9465 "not."+CondVal->getName()), SI);
9466 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9469 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9470 // If one of the constants is zero (we know they can't both be) and we
9471 // have an icmp instruction with zero, and we have an 'and' with the
9472 // non-constant value, eliminate this whole mess. This corresponds to
9473 // cases like this: ((X & 27) ? 27 : 0)
9474 if (TrueValC->isZero() || FalseValC->isZero())
9475 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9476 cast<Constant>(IC->getOperand(1))->isNullValue())
9477 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9478 if (ICA->getOpcode() == Instruction::And &&
9479 isa<ConstantInt>(ICA->getOperand(1)) &&
9480 (ICA->getOperand(1) == TrueValC ||
9481 ICA->getOperand(1) == FalseValC) &&
9482 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9483 // Okay, now we know that everything is set up, we just don't
9484 // know whether we have a icmp_ne or icmp_eq and whether the
9485 // true or false val is the zero.
9486 bool ShouldNotVal = !TrueValC->isZero();
9487 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9490 V = InsertNewInstBefore(BinaryOperator::Create(
9491 Instruction::Xor, V, ICA->getOperand(1)), SI);
9492 return ReplaceInstUsesWith(SI, V);
9497 // See if we are selecting two values based on a comparison of the two values.
9498 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9499 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9500 // Transform (X == Y) ? X : Y -> Y
9501 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9502 // This is not safe in general for floating point:
9503 // consider X== -0, Y== +0.
9504 // It becomes safe if either operand is a nonzero constant.
9505 ConstantFP *CFPt, *CFPf;
9506 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9507 !CFPt->getValueAPF().isZero()) ||
9508 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9509 !CFPf->getValueAPF().isZero()))
9510 return ReplaceInstUsesWith(SI, FalseVal);
9512 // Transform (X != Y) ? X : Y -> X
9513 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9514 return ReplaceInstUsesWith(SI, TrueVal);
9515 // NOTE: if we wanted to, this is where to detect MIN/MAX
9517 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9518 // Transform (X == Y) ? Y : X -> X
9519 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9520 // This is not safe in general for floating point:
9521 // consider X== -0, Y== +0.
9522 // It becomes safe if either operand is a nonzero constant.
9523 ConstantFP *CFPt, *CFPf;
9524 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9525 !CFPt->getValueAPF().isZero()) ||
9526 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9527 !CFPf->getValueAPF().isZero()))
9528 return ReplaceInstUsesWith(SI, FalseVal);
9530 // Transform (X != Y) ? Y : X -> Y
9531 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9532 return ReplaceInstUsesWith(SI, TrueVal);
9533 // NOTE: if we wanted to, this is where to detect MIN/MAX
9535 // NOTE: if we wanted to, this is where to detect ABS
9538 // See if we are selecting two values based on a comparison of the two values.
9539 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9540 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9543 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9544 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9545 if (TI->hasOneUse() && FI->hasOneUse()) {
9546 Instruction *AddOp = 0, *SubOp = 0;
9548 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9549 if (TI->getOpcode() == FI->getOpcode())
9550 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9553 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9554 // even legal for FP.
9555 if ((TI->getOpcode() == Instruction::Sub &&
9556 FI->getOpcode() == Instruction::Add) ||
9557 (TI->getOpcode() == Instruction::FSub &&
9558 FI->getOpcode() == Instruction::FAdd)) {
9559 AddOp = FI; SubOp = TI;
9560 } else if ((FI->getOpcode() == Instruction::Sub &&
9561 TI->getOpcode() == Instruction::Add) ||
9562 (FI->getOpcode() == Instruction::FSub &&
9563 TI->getOpcode() == Instruction::FAdd)) {
9564 AddOp = TI; SubOp = FI;
9568 Value *OtherAddOp = 0;
9569 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9570 OtherAddOp = AddOp->getOperand(1);
9571 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9572 OtherAddOp = AddOp->getOperand(0);
9576 // So at this point we know we have (Y -> OtherAddOp):
9577 // select C, (add X, Y), (sub X, Z)
9578 Value *NegVal; // Compute -Z
9579 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9580 NegVal = Context->getConstantExprNeg(C);
9582 NegVal = InsertNewInstBefore(
9583 BinaryOperator::CreateNeg(*Context, SubOp->getOperand(1),
9587 Value *NewTrueOp = OtherAddOp;
9588 Value *NewFalseOp = NegVal;
9590 std::swap(NewTrueOp, NewFalseOp);
9591 Instruction *NewSel =
9592 SelectInst::Create(CondVal, NewTrueOp,
9593 NewFalseOp, SI.getName() + ".p");
9595 NewSel = InsertNewInstBefore(NewSel, SI);
9596 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9601 // See if we can fold the select into one of our operands.
9602 if (SI.getType()->isInteger()) {
9603 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9608 if (BinaryOperator::isNot(CondVal)) {
9609 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9610 SI.setOperand(1, FalseVal);
9611 SI.setOperand(2, TrueVal);
9618 /// EnforceKnownAlignment - If the specified pointer points to an object that
9619 /// we control, modify the object's alignment to PrefAlign. This isn't
9620 /// often possible though. If alignment is important, a more reliable approach
9621 /// is to simply align all global variables and allocation instructions to
9622 /// their preferred alignment from the beginning.
9624 static unsigned EnforceKnownAlignment(Value *V,
9625 unsigned Align, unsigned PrefAlign) {
9627 User *U = dyn_cast<User>(V);
9628 if (!U) return Align;
9630 switch (Operator::getOpcode(U)) {
9632 case Instruction::BitCast:
9633 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9634 case Instruction::GetElementPtr: {
9635 // If all indexes are zero, it is just the alignment of the base pointer.
9636 bool AllZeroOperands = true;
9637 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9638 if (!isa<Constant>(*i) ||
9639 !cast<Constant>(*i)->isNullValue()) {
9640 AllZeroOperands = false;
9644 if (AllZeroOperands) {
9645 // Treat this like a bitcast.
9646 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9652 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9653 // If there is a large requested alignment and we can, bump up the alignment
9655 if (!GV->isDeclaration()) {
9656 if (GV->getAlignment() >= PrefAlign)
9657 Align = GV->getAlignment();
9659 GV->setAlignment(PrefAlign);
9663 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9664 // If there is a requested alignment and if this is an alloca, round up. We
9665 // don't do this for malloc, because some systems can't respect the request.
9666 if (isa<AllocaInst>(AI)) {
9667 if (AI->getAlignment() >= PrefAlign)
9668 Align = AI->getAlignment();
9670 AI->setAlignment(PrefAlign);
9679 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9680 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9681 /// and it is more than the alignment of the ultimate object, see if we can
9682 /// increase the alignment of the ultimate object, making this check succeed.
9683 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9684 unsigned PrefAlign) {
9685 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9686 sizeof(PrefAlign) * CHAR_BIT;
9687 APInt Mask = APInt::getAllOnesValue(BitWidth);
9688 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9689 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9690 unsigned TrailZ = KnownZero.countTrailingOnes();
9691 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9693 if (PrefAlign > Align)
9694 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9696 // We don't need to make any adjustment.
9700 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9701 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9702 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9703 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9704 unsigned CopyAlign = MI->getAlignment();
9706 if (CopyAlign < MinAlign) {
9707 MI->setAlignment(Context->getConstantInt(MI->getAlignmentType(),
9712 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9714 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9715 if (MemOpLength == 0) return 0;
9717 // Source and destination pointer types are always "i8*" for intrinsic. See
9718 // if the size is something we can handle with a single primitive load/store.
9719 // A single load+store correctly handles overlapping memory in the memmove
9721 unsigned Size = MemOpLength->getZExtValue();
9722 if (Size == 0) return MI; // Delete this mem transfer.
9724 if (Size > 8 || (Size&(Size-1)))
9725 return 0; // If not 1/2/4/8 bytes, exit.
9727 // Use an integer load+store unless we can find something better.
9729 Context->getPointerTypeUnqual(Context->getIntegerType(Size<<3));
9731 // Memcpy forces the use of i8* for the source and destination. That means
9732 // that if you're using memcpy to move one double around, you'll get a cast
9733 // from double* to i8*. We'd much rather use a double load+store rather than
9734 // an i64 load+store, here because this improves the odds that the source or
9735 // dest address will be promotable. See if we can find a better type than the
9736 // integer datatype.
9737 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9738 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9739 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9740 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9741 // down through these levels if so.
9742 while (!SrcETy->isSingleValueType()) {
9743 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9744 if (STy->getNumElements() == 1)
9745 SrcETy = STy->getElementType(0);
9748 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9749 if (ATy->getNumElements() == 1)
9750 SrcETy = ATy->getElementType();
9757 if (SrcETy->isSingleValueType())
9758 NewPtrTy = Context->getPointerTypeUnqual(SrcETy);
9763 // If the memcpy/memmove provides better alignment info than we can
9765 SrcAlign = std::max(SrcAlign, CopyAlign);
9766 DstAlign = std::max(DstAlign, CopyAlign);
9768 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9769 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9770 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9771 InsertNewInstBefore(L, *MI);
9772 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9774 // Set the size of the copy to 0, it will be deleted on the next iteration.
9775 MI->setOperand(3, Context->getNullValue(MemOpLength->getType()));
9779 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9780 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9781 if (MI->getAlignment() < Alignment) {
9782 MI->setAlignment(Context->getConstantInt(MI->getAlignmentType(),
9787 // Extract the length and alignment and fill if they are constant.
9788 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9789 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9790 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9792 uint64_t Len = LenC->getZExtValue();
9793 Alignment = MI->getAlignment();
9795 // If the length is zero, this is a no-op
9796 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9798 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9799 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9800 const Type *ITy = Context->getIntegerType(Len*8); // n=1 -> i8.
9802 Value *Dest = MI->getDest();
9803 Dest = InsertBitCastBefore(Dest, Context->getPointerTypeUnqual(ITy), *MI);
9805 // Alignment 0 is identity for alignment 1 for memset, but not store.
9806 if (Alignment == 0) Alignment = 1;
9808 // Extract the fill value and store.
9809 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9810 InsertNewInstBefore(new StoreInst(Context->getConstantInt(ITy, Fill),
9811 Dest, false, Alignment), *MI);
9813 // Set the size of the copy to 0, it will be deleted on the next iteration.
9814 MI->setLength(Context->getNullValue(LenC->getType()));
9822 /// visitCallInst - CallInst simplification. This mostly only handles folding
9823 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9824 /// the heavy lifting.
9826 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9827 // If the caller function is nounwind, mark the call as nounwind, even if the
9829 if (CI.getParent()->getParent()->doesNotThrow() &&
9830 !CI.doesNotThrow()) {
9831 CI.setDoesNotThrow();
9837 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9838 if (!II) return visitCallSite(&CI);
9840 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9842 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9843 bool Changed = false;
9845 // memmove/cpy/set of zero bytes is a noop.
9846 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9847 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9849 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9850 if (CI->getZExtValue() == 1) {
9851 // Replace the instruction with just byte operations. We would
9852 // transform other cases to loads/stores, but we don't know if
9853 // alignment is sufficient.
9857 // If we have a memmove and the source operation is a constant global,
9858 // then the source and dest pointers can't alias, so we can change this
9859 // into a call to memcpy.
9860 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9861 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9862 if (GVSrc->isConstant()) {
9863 Module *M = CI.getParent()->getParent()->getParent();
9864 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9866 Tys[0] = CI.getOperand(3)->getType();
9868 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9872 // memmove(x,x,size) -> noop.
9873 if (MMI->getSource() == MMI->getDest())
9874 return EraseInstFromFunction(CI);
9877 // If we can determine a pointer alignment that is bigger than currently
9878 // set, update the alignment.
9879 if (isa<MemTransferInst>(MI)) {
9880 if (Instruction *I = SimplifyMemTransfer(MI))
9882 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9883 if (Instruction *I = SimplifyMemSet(MSI))
9887 if (Changed) return II;
9890 switch (II->getIntrinsicID()) {
9892 case Intrinsic::bswap:
9893 // bswap(bswap(x)) -> x
9894 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9895 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9896 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9898 case Intrinsic::ppc_altivec_lvx:
9899 case Intrinsic::ppc_altivec_lvxl:
9900 case Intrinsic::x86_sse_loadu_ps:
9901 case Intrinsic::x86_sse2_loadu_pd:
9902 case Intrinsic::x86_sse2_loadu_dq:
9903 // Turn PPC lvx -> load if the pointer is known aligned.
9904 // Turn X86 loadups -> load if the pointer is known aligned.
9905 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9906 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9907 Context->getPointerTypeUnqual(II->getType()),
9909 return new LoadInst(Ptr);
9912 case Intrinsic::ppc_altivec_stvx:
9913 case Intrinsic::ppc_altivec_stvxl:
9914 // Turn stvx -> store if the pointer is known aligned.
9915 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9916 const Type *OpPtrTy =
9917 Context->getPointerTypeUnqual(II->getOperand(1)->getType());
9918 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9919 return new StoreInst(II->getOperand(1), Ptr);
9922 case Intrinsic::x86_sse_storeu_ps:
9923 case Intrinsic::x86_sse2_storeu_pd:
9924 case Intrinsic::x86_sse2_storeu_dq:
9925 // Turn X86 storeu -> store if the pointer is known aligned.
9926 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9927 const Type *OpPtrTy =
9928 Context->getPointerTypeUnqual(II->getOperand(2)->getType());
9929 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9930 return new StoreInst(II->getOperand(2), Ptr);
9934 case Intrinsic::x86_sse_cvttss2si: {
9935 // These intrinsics only demands the 0th element of its input vector. If
9936 // we can simplify the input based on that, do so now.
9938 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9939 APInt DemandedElts(VWidth, 1);
9940 APInt UndefElts(VWidth, 0);
9941 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9943 II->setOperand(1, V);
9949 case Intrinsic::ppc_altivec_vperm:
9950 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9951 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9952 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9954 // Check that all of the elements are integer constants or undefs.
9955 bool AllEltsOk = true;
9956 for (unsigned i = 0; i != 16; ++i) {
9957 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9958 !isa<UndefValue>(Mask->getOperand(i))) {
9965 // Cast the input vectors to byte vectors.
9966 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9967 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9968 Value *Result = Context->getUndef(Op0->getType());
9970 // Only extract each element once.
9971 Value *ExtractedElts[32];
9972 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9974 for (unsigned i = 0; i != 16; ++i) {
9975 if (isa<UndefValue>(Mask->getOperand(i)))
9977 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9978 Idx &= 31; // Match the hardware behavior.
9980 if (ExtractedElts[Idx] == 0) {
9982 new ExtractElementInst(Idx < 16 ? Op0 : Op1,
9983 Context->getConstantInt(Type::Int32Ty, Idx&15, false), "tmp");
9984 InsertNewInstBefore(Elt, CI);
9985 ExtractedElts[Idx] = Elt;
9988 // Insert this value into the result vector.
9989 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9990 Context->getConstantInt(Type::Int32Ty, i, false),
9992 InsertNewInstBefore(cast<Instruction>(Result), CI);
9994 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9999 case Intrinsic::stackrestore: {
10000 // If the save is right next to the restore, remove the restore. This can
10001 // happen when variable allocas are DCE'd.
10002 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
10003 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
10004 BasicBlock::iterator BI = SS;
10006 return EraseInstFromFunction(CI);
10010 // Scan down this block to see if there is another stack restore in the
10011 // same block without an intervening call/alloca.
10012 BasicBlock::iterator BI = II;
10013 TerminatorInst *TI = II->getParent()->getTerminator();
10014 bool CannotRemove = false;
10015 for (++BI; &*BI != TI; ++BI) {
10016 if (isa<AllocaInst>(BI)) {
10017 CannotRemove = true;
10020 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
10021 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
10022 // If there is a stackrestore below this one, remove this one.
10023 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10024 return EraseInstFromFunction(CI);
10025 // Otherwise, ignore the intrinsic.
10027 // If we found a non-intrinsic call, we can't remove the stack
10029 CannotRemove = true;
10035 // If the stack restore is in a return/unwind block and if there are no
10036 // allocas or calls between the restore and the return, nuke the restore.
10037 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10038 return EraseInstFromFunction(CI);
10043 return visitCallSite(II);
10046 // InvokeInst simplification
10048 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10049 return visitCallSite(&II);
10052 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10053 /// passed through the varargs area, we can eliminate the use of the cast.
10054 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10055 const CastInst * const CI,
10056 const TargetData * const TD,
10058 if (!CI->isLosslessCast())
10061 // The size of ByVal arguments is derived from the type, so we
10062 // can't change to a type with a different size. If the size were
10063 // passed explicitly we could avoid this check.
10064 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10067 const Type* SrcTy =
10068 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10069 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10070 if (!SrcTy->isSized() || !DstTy->isSized())
10072 if (TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10077 // visitCallSite - Improvements for call and invoke instructions.
10079 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10080 bool Changed = false;
10082 // If the callee is a constexpr cast of a function, attempt to move the cast
10083 // to the arguments of the call/invoke.
10084 if (transformConstExprCastCall(CS)) return 0;
10086 Value *Callee = CS.getCalledValue();
10088 if (Function *CalleeF = dyn_cast<Function>(Callee))
10089 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10090 Instruction *OldCall = CS.getInstruction();
10091 // If the call and callee calling conventions don't match, this call must
10092 // be unreachable, as the call is undefined.
10093 new StoreInst(Context->getConstantIntTrue(),
10094 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)),
10096 if (!OldCall->use_empty())
10097 OldCall->replaceAllUsesWith(Context->getUndef(OldCall->getType()));
10098 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10099 return EraseInstFromFunction(*OldCall);
10103 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10104 // This instruction is not reachable, just remove it. We insert a store to
10105 // undef so that we know that this code is not reachable, despite the fact
10106 // that we can't modify the CFG here.
10107 new StoreInst(Context->getConstantIntTrue(),
10108 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)),
10109 CS.getInstruction());
10111 if (!CS.getInstruction()->use_empty())
10112 CS.getInstruction()->
10113 replaceAllUsesWith(Context->getUndef(CS.getInstruction()->getType()));
10115 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10116 // Don't break the CFG, insert a dummy cond branch.
10117 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10118 Context->getConstantIntTrue(), II);
10120 return EraseInstFromFunction(*CS.getInstruction());
10123 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10124 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10125 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10126 return transformCallThroughTrampoline(CS);
10128 const PointerType *PTy = cast<PointerType>(Callee->getType());
10129 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10130 if (FTy->isVarArg()) {
10131 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10132 // See if we can optimize any arguments passed through the varargs area of
10134 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10135 E = CS.arg_end(); I != E; ++I, ++ix) {
10136 CastInst *CI = dyn_cast<CastInst>(*I);
10137 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10138 *I = CI->getOperand(0);
10144 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10145 // Inline asm calls cannot throw - mark them 'nounwind'.
10146 CS.setDoesNotThrow();
10150 return Changed ? CS.getInstruction() : 0;
10153 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10154 // attempt to move the cast to the arguments of the call/invoke.
10156 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10157 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10158 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10159 if (CE->getOpcode() != Instruction::BitCast ||
10160 !isa<Function>(CE->getOperand(0)))
10162 Function *Callee = cast<Function>(CE->getOperand(0));
10163 Instruction *Caller = CS.getInstruction();
10164 const AttrListPtr &CallerPAL = CS.getAttributes();
10166 // Okay, this is a cast from a function to a different type. Unless doing so
10167 // would cause a type conversion of one of our arguments, change this call to
10168 // be a direct call with arguments casted to the appropriate types.
10170 const FunctionType *FT = Callee->getFunctionType();
10171 const Type *OldRetTy = Caller->getType();
10172 const Type *NewRetTy = FT->getReturnType();
10174 if (isa<StructType>(NewRetTy))
10175 return false; // TODO: Handle multiple return values.
10177 // Check to see if we are changing the return type...
10178 if (OldRetTy != NewRetTy) {
10179 if (Callee->isDeclaration() &&
10180 // Conversion is ok if changing from one pointer type to another or from
10181 // a pointer to an integer of the same size.
10182 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
10183 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
10184 return false; // Cannot transform this return value.
10186 if (!Caller->use_empty() &&
10187 // void -> non-void is handled specially
10188 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
10189 return false; // Cannot transform this return value.
10191 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10192 Attributes RAttrs = CallerPAL.getRetAttributes();
10193 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10194 return false; // Attribute not compatible with transformed value.
10197 // If the callsite is an invoke instruction, and the return value is used by
10198 // a PHI node in a successor, we cannot change the return type of the call
10199 // because there is no place to put the cast instruction (without breaking
10200 // the critical edge). Bail out in this case.
10201 if (!Caller->use_empty())
10202 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10203 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10205 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10206 if (PN->getParent() == II->getNormalDest() ||
10207 PN->getParent() == II->getUnwindDest())
10211 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10212 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10214 CallSite::arg_iterator AI = CS.arg_begin();
10215 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10216 const Type *ParamTy = FT->getParamType(i);
10217 const Type *ActTy = (*AI)->getType();
10219 if (!CastInst::isCastable(ActTy, ParamTy))
10220 return false; // Cannot transform this parameter value.
10222 if (CallerPAL.getParamAttributes(i + 1)
10223 & Attribute::typeIncompatible(ParamTy))
10224 return false; // Attribute not compatible with transformed value.
10226 // Converting from one pointer type to another or between a pointer and an
10227 // integer of the same size is safe even if we do not have a body.
10228 bool isConvertible = ActTy == ParamTy ||
10229 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
10230 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
10231 if (Callee->isDeclaration() && !isConvertible) return false;
10234 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10235 Callee->isDeclaration())
10236 return false; // Do not delete arguments unless we have a function body.
10238 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10239 !CallerPAL.isEmpty())
10240 // In this case we have more arguments than the new function type, but we
10241 // won't be dropping them. Check that these extra arguments have attributes
10242 // that are compatible with being a vararg call argument.
10243 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10244 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10246 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10247 if (PAttrs & Attribute::VarArgsIncompatible)
10251 // Okay, we decided that this is a safe thing to do: go ahead and start
10252 // inserting cast instructions as necessary...
10253 std::vector<Value*> Args;
10254 Args.reserve(NumActualArgs);
10255 SmallVector<AttributeWithIndex, 8> attrVec;
10256 attrVec.reserve(NumCommonArgs);
10258 // Get any return attributes.
10259 Attributes RAttrs = CallerPAL.getRetAttributes();
10261 // If the return value is not being used, the type may not be compatible
10262 // with the existing attributes. Wipe out any problematic attributes.
10263 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10265 // Add the new return attributes.
10267 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10269 AI = CS.arg_begin();
10270 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10271 const Type *ParamTy = FT->getParamType(i);
10272 if ((*AI)->getType() == ParamTy) {
10273 Args.push_back(*AI);
10275 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10276 false, ParamTy, false);
10277 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
10278 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
10281 // Add any parameter attributes.
10282 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10283 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10286 // If the function takes more arguments than the call was taking, add them
10288 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10289 Args.push_back(Context->getNullValue(FT->getParamType(i)));
10291 // If we are removing arguments to the function, emit an obnoxious warning...
10292 if (FT->getNumParams() < NumActualArgs) {
10293 if (!FT->isVarArg()) {
10294 cerr << "WARNING: While resolving call to function '"
10295 << Callee->getName() << "' arguments were dropped!\n";
10297 // Add all of the arguments in their promoted form to the arg list...
10298 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10299 const Type *PTy = getPromotedType((*AI)->getType());
10300 if (PTy != (*AI)->getType()) {
10301 // Must promote to pass through va_arg area!
10302 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
10304 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
10305 InsertNewInstBefore(Cast, *Caller);
10306 Args.push_back(Cast);
10308 Args.push_back(*AI);
10311 // Add any parameter attributes.
10312 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10313 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10318 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10319 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10321 if (NewRetTy == Type::VoidTy)
10322 Caller->setName(""); // Void type should not have a name.
10324 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
10327 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10328 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10329 Args.begin(), Args.end(),
10330 Caller->getName(), Caller);
10331 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10332 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10334 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10335 Caller->getName(), Caller);
10336 CallInst *CI = cast<CallInst>(Caller);
10337 if (CI->isTailCall())
10338 cast<CallInst>(NC)->setTailCall();
10339 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10340 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10343 // Insert a cast of the return type as necessary.
10345 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10346 if (NV->getType() != Type::VoidTy) {
10347 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10349 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10351 // If this is an invoke instruction, we should insert it after the first
10352 // non-phi, instruction in the normal successor block.
10353 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10354 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10355 InsertNewInstBefore(NC, *I);
10357 // Otherwise, it's a call, just insert cast right after the call instr
10358 InsertNewInstBefore(NC, *Caller);
10360 AddUsersToWorkList(*Caller);
10362 NV = Context->getUndef(Caller->getType());
10366 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10367 Caller->replaceAllUsesWith(NV);
10368 Caller->eraseFromParent();
10369 RemoveFromWorkList(Caller);
10373 // transformCallThroughTrampoline - Turn a call to a function created by the
10374 // init_trampoline intrinsic into a direct call to the underlying function.
10376 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10377 Value *Callee = CS.getCalledValue();
10378 const PointerType *PTy = cast<PointerType>(Callee->getType());
10379 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10380 const AttrListPtr &Attrs = CS.getAttributes();
10382 // If the call already has the 'nest' attribute somewhere then give up -
10383 // otherwise 'nest' would occur twice after splicing in the chain.
10384 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10387 IntrinsicInst *Tramp =
10388 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10390 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10391 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10392 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10394 const AttrListPtr &NestAttrs = NestF->getAttributes();
10395 if (!NestAttrs.isEmpty()) {
10396 unsigned NestIdx = 1;
10397 const Type *NestTy = 0;
10398 Attributes NestAttr = Attribute::None;
10400 // Look for a parameter marked with the 'nest' attribute.
10401 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10402 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10403 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10404 // Record the parameter type and any other attributes.
10406 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10411 Instruction *Caller = CS.getInstruction();
10412 std::vector<Value*> NewArgs;
10413 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10415 SmallVector<AttributeWithIndex, 8> NewAttrs;
10416 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10418 // Insert the nest argument into the call argument list, which may
10419 // mean appending it. Likewise for attributes.
10421 // Add any result attributes.
10422 if (Attributes Attr = Attrs.getRetAttributes())
10423 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10427 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10429 if (Idx == NestIdx) {
10430 // Add the chain argument and attributes.
10431 Value *NestVal = Tramp->getOperand(3);
10432 if (NestVal->getType() != NestTy)
10433 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10434 NewArgs.push_back(NestVal);
10435 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10441 // Add the original argument and attributes.
10442 NewArgs.push_back(*I);
10443 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10445 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10451 // Add any function attributes.
10452 if (Attributes Attr = Attrs.getFnAttributes())
10453 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10455 // The trampoline may have been bitcast to a bogus type (FTy).
10456 // Handle this by synthesizing a new function type, equal to FTy
10457 // with the chain parameter inserted.
10459 std::vector<const Type*> NewTypes;
10460 NewTypes.reserve(FTy->getNumParams()+1);
10462 // Insert the chain's type into the list of parameter types, which may
10463 // mean appending it.
10466 FunctionType::param_iterator I = FTy->param_begin(),
10467 E = FTy->param_end();
10470 if (Idx == NestIdx)
10471 // Add the chain's type.
10472 NewTypes.push_back(NestTy);
10477 // Add the original type.
10478 NewTypes.push_back(*I);
10484 // Replace the trampoline call with a direct call. Let the generic
10485 // code sort out any function type mismatches.
10486 FunctionType *NewFTy =
10487 Context->getFunctionType(FTy->getReturnType(), NewTypes,
10489 Constant *NewCallee =
10490 NestF->getType() == Context->getPointerTypeUnqual(NewFTy) ?
10491 NestF : Context->getConstantExprBitCast(NestF,
10492 Context->getPointerTypeUnqual(NewFTy));
10493 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
10495 Instruction *NewCaller;
10496 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10497 NewCaller = InvokeInst::Create(NewCallee,
10498 II->getNormalDest(), II->getUnwindDest(),
10499 NewArgs.begin(), NewArgs.end(),
10500 Caller->getName(), Caller);
10501 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10502 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10504 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10505 Caller->getName(), Caller);
10506 if (cast<CallInst>(Caller)->isTailCall())
10507 cast<CallInst>(NewCaller)->setTailCall();
10508 cast<CallInst>(NewCaller)->
10509 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10510 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10512 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10513 Caller->replaceAllUsesWith(NewCaller);
10514 Caller->eraseFromParent();
10515 RemoveFromWorkList(Caller);
10520 // Replace the trampoline call with a direct call. Since there is no 'nest'
10521 // parameter, there is no need to adjust the argument list. Let the generic
10522 // code sort out any function type mismatches.
10523 Constant *NewCallee =
10524 NestF->getType() == PTy ? NestF :
10525 Context->getConstantExprBitCast(NestF, PTy);
10526 CS.setCalledFunction(NewCallee);
10527 return CS.getInstruction();
10530 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10531 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10532 /// and a single binop.
10533 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10534 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10535 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10536 unsigned Opc = FirstInst->getOpcode();
10537 Value *LHSVal = FirstInst->getOperand(0);
10538 Value *RHSVal = FirstInst->getOperand(1);
10540 const Type *LHSType = LHSVal->getType();
10541 const Type *RHSType = RHSVal->getType();
10543 // Scan to see if all operands are the same opcode, all have one use, and all
10544 // kill their operands (i.e. the operands have one use).
10545 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10546 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10547 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10548 // Verify type of the LHS matches so we don't fold cmp's of different
10549 // types or GEP's with different index types.
10550 I->getOperand(0)->getType() != LHSType ||
10551 I->getOperand(1)->getType() != RHSType)
10554 // If they are CmpInst instructions, check their predicates
10555 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10556 if (cast<CmpInst>(I)->getPredicate() !=
10557 cast<CmpInst>(FirstInst)->getPredicate())
10560 // Keep track of which operand needs a phi node.
10561 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10562 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10565 // Otherwise, this is safe to transform!
10567 Value *InLHS = FirstInst->getOperand(0);
10568 Value *InRHS = FirstInst->getOperand(1);
10569 PHINode *NewLHS = 0, *NewRHS = 0;
10571 NewLHS = PHINode::Create(LHSType,
10572 FirstInst->getOperand(0)->getName() + ".pn");
10573 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10574 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10575 InsertNewInstBefore(NewLHS, PN);
10580 NewRHS = PHINode::Create(RHSType,
10581 FirstInst->getOperand(1)->getName() + ".pn");
10582 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10583 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10584 InsertNewInstBefore(NewRHS, PN);
10588 // Add all operands to the new PHIs.
10589 if (NewLHS || NewRHS) {
10590 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10591 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10593 Value *NewInLHS = InInst->getOperand(0);
10594 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10597 Value *NewInRHS = InInst->getOperand(1);
10598 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10603 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10604 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10605 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10606 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10610 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10611 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10613 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10614 FirstInst->op_end());
10615 // This is true if all GEP bases are allocas and if all indices into them are
10617 bool AllBasePointersAreAllocas = true;
10619 // Scan to see if all operands are the same opcode, all have one use, and all
10620 // kill their operands (i.e. the operands have one use).
10621 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10622 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10623 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10624 GEP->getNumOperands() != FirstInst->getNumOperands())
10627 // Keep track of whether or not all GEPs are of alloca pointers.
10628 if (AllBasePointersAreAllocas &&
10629 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10630 !GEP->hasAllConstantIndices()))
10631 AllBasePointersAreAllocas = false;
10633 // Compare the operand lists.
10634 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10635 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10638 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10639 // if one of the PHIs has a constant for the index. The index may be
10640 // substantially cheaper to compute for the constants, so making it a
10641 // variable index could pessimize the path. This also handles the case
10642 // for struct indices, which must always be constant.
10643 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10644 isa<ConstantInt>(GEP->getOperand(op)))
10647 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10649 FixedOperands[op] = 0; // Needs a PHI.
10653 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10654 // bother doing this transformation. At best, this will just save a bit of
10655 // offset calculation, but all the predecessors will have to materialize the
10656 // stack address into a register anyway. We'd actually rather *clone* the
10657 // load up into the predecessors so that we have a load of a gep of an alloca,
10658 // which can usually all be folded into the load.
10659 if (AllBasePointersAreAllocas)
10662 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10663 // that is variable.
10664 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10666 bool HasAnyPHIs = false;
10667 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10668 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10669 Value *FirstOp = FirstInst->getOperand(i);
10670 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10671 FirstOp->getName()+".pn");
10672 InsertNewInstBefore(NewPN, PN);
10674 NewPN->reserveOperandSpace(e);
10675 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10676 OperandPhis[i] = NewPN;
10677 FixedOperands[i] = NewPN;
10682 // Add all operands to the new PHIs.
10684 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10685 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10686 BasicBlock *InBB = PN.getIncomingBlock(i);
10688 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10689 if (PHINode *OpPhi = OperandPhis[op])
10690 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10694 Value *Base = FixedOperands[0];
10695 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10696 FixedOperands.end());
10700 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10701 /// sink the load out of the block that defines it. This means that it must be
10702 /// obvious the value of the load is not changed from the point of the load to
10703 /// the end of the block it is in.
10705 /// Finally, it is safe, but not profitable, to sink a load targetting a
10706 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10708 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10709 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10711 for (++BBI; BBI != E; ++BBI)
10712 if (BBI->mayWriteToMemory())
10715 // Check for non-address taken alloca. If not address-taken already, it isn't
10716 // profitable to do this xform.
10717 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10718 bool isAddressTaken = false;
10719 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10721 if (isa<LoadInst>(UI)) continue;
10722 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10723 // If storing TO the alloca, then the address isn't taken.
10724 if (SI->getOperand(1) == AI) continue;
10726 isAddressTaken = true;
10730 if (!isAddressTaken && AI->isStaticAlloca())
10734 // If this load is a load from a GEP with a constant offset from an alloca,
10735 // then we don't want to sink it. In its present form, it will be
10736 // load [constant stack offset]. Sinking it will cause us to have to
10737 // materialize the stack addresses in each predecessor in a register only to
10738 // do a shared load from register in the successor.
10739 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10740 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10741 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10748 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10749 // operator and they all are only used by the PHI, PHI together their
10750 // inputs, and do the operation once, to the result of the PHI.
10751 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10752 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10754 // Scan the instruction, looking for input operations that can be folded away.
10755 // If all input operands to the phi are the same instruction (e.g. a cast from
10756 // the same type or "+42") we can pull the operation through the PHI, reducing
10757 // code size and simplifying code.
10758 Constant *ConstantOp = 0;
10759 const Type *CastSrcTy = 0;
10760 bool isVolatile = false;
10761 if (isa<CastInst>(FirstInst)) {
10762 CastSrcTy = FirstInst->getOperand(0)->getType();
10763 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10764 // Can fold binop, compare or shift here if the RHS is a constant,
10765 // otherwise call FoldPHIArgBinOpIntoPHI.
10766 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10767 if (ConstantOp == 0)
10768 return FoldPHIArgBinOpIntoPHI(PN);
10769 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10770 isVolatile = LI->isVolatile();
10771 // We can't sink the load if the loaded value could be modified between the
10772 // load and the PHI.
10773 if (LI->getParent() != PN.getIncomingBlock(0) ||
10774 !isSafeAndProfitableToSinkLoad(LI))
10777 // If the PHI is of volatile loads and the load block has multiple
10778 // successors, sinking it would remove a load of the volatile value from
10779 // the path through the other successor.
10781 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10784 } else if (isa<GetElementPtrInst>(FirstInst)) {
10785 return FoldPHIArgGEPIntoPHI(PN);
10787 return 0; // Cannot fold this operation.
10790 // Check to see if all arguments are the same operation.
10791 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10792 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10793 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10794 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10797 if (I->getOperand(0)->getType() != CastSrcTy)
10798 return 0; // Cast operation must match.
10799 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10800 // We can't sink the load if the loaded value could be modified between
10801 // the load and the PHI.
10802 if (LI->isVolatile() != isVolatile ||
10803 LI->getParent() != PN.getIncomingBlock(i) ||
10804 !isSafeAndProfitableToSinkLoad(LI))
10807 // If the PHI is of volatile loads and the load block has multiple
10808 // successors, sinking it would remove a load of the volatile value from
10809 // the path through the other successor.
10811 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10814 } else if (I->getOperand(1) != ConstantOp) {
10819 // Okay, they are all the same operation. Create a new PHI node of the
10820 // correct type, and PHI together all of the LHS's of the instructions.
10821 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10822 PN.getName()+".in");
10823 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10825 Value *InVal = FirstInst->getOperand(0);
10826 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10828 // Add all operands to the new PHI.
10829 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10830 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10831 if (NewInVal != InVal)
10833 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10838 // The new PHI unions all of the same values together. This is really
10839 // common, so we handle it intelligently here for compile-time speed.
10843 InsertNewInstBefore(NewPN, PN);
10847 // Insert and return the new operation.
10848 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10849 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10850 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10851 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10852 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10853 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10854 PhiVal, ConstantOp);
10855 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10857 // If this was a volatile load that we are merging, make sure to loop through
10858 // and mark all the input loads as non-volatile. If we don't do this, we will
10859 // insert a new volatile load and the old ones will not be deletable.
10861 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10862 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10864 return new LoadInst(PhiVal, "", isVolatile);
10867 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10869 static bool DeadPHICycle(PHINode *PN,
10870 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10871 if (PN->use_empty()) return true;
10872 if (!PN->hasOneUse()) return false;
10874 // Remember this node, and if we find the cycle, return.
10875 if (!PotentiallyDeadPHIs.insert(PN))
10878 // Don't scan crazily complex things.
10879 if (PotentiallyDeadPHIs.size() == 16)
10882 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10883 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10888 /// PHIsEqualValue - Return true if this phi node is always equal to
10889 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10890 /// z = some value; x = phi (y, z); y = phi (x, z)
10891 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10892 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10893 // See if we already saw this PHI node.
10894 if (!ValueEqualPHIs.insert(PN))
10897 // Don't scan crazily complex things.
10898 if (ValueEqualPHIs.size() == 16)
10901 // Scan the operands to see if they are either phi nodes or are equal to
10903 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10904 Value *Op = PN->getIncomingValue(i);
10905 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10906 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10908 } else if (Op != NonPhiInVal)
10916 // PHINode simplification
10918 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10919 // If LCSSA is around, don't mess with Phi nodes
10920 if (MustPreserveLCSSA) return 0;
10922 if (Value *V = PN.hasConstantValue())
10923 return ReplaceInstUsesWith(PN, V);
10925 // If all PHI operands are the same operation, pull them through the PHI,
10926 // reducing code size.
10927 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10928 isa<Instruction>(PN.getIncomingValue(1)) &&
10929 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10930 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10931 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10932 // than themselves more than once.
10933 PN.getIncomingValue(0)->hasOneUse())
10934 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10937 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10938 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10939 // PHI)... break the cycle.
10940 if (PN.hasOneUse()) {
10941 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10942 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10943 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10944 PotentiallyDeadPHIs.insert(&PN);
10945 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10946 return ReplaceInstUsesWith(PN, Context->getUndef(PN.getType()));
10949 // If this phi has a single use, and if that use just computes a value for
10950 // the next iteration of a loop, delete the phi. This occurs with unused
10951 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10952 // common case here is good because the only other things that catch this
10953 // are induction variable analysis (sometimes) and ADCE, which is only run
10955 if (PHIUser->hasOneUse() &&
10956 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10957 PHIUser->use_back() == &PN) {
10958 return ReplaceInstUsesWith(PN, Context->getUndef(PN.getType()));
10962 // We sometimes end up with phi cycles that non-obviously end up being the
10963 // same value, for example:
10964 // z = some value; x = phi (y, z); y = phi (x, z)
10965 // where the phi nodes don't necessarily need to be in the same block. Do a
10966 // quick check to see if the PHI node only contains a single non-phi value, if
10967 // so, scan to see if the phi cycle is actually equal to that value.
10969 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10970 // Scan for the first non-phi operand.
10971 while (InValNo != NumOperandVals &&
10972 isa<PHINode>(PN.getIncomingValue(InValNo)))
10975 if (InValNo != NumOperandVals) {
10976 Value *NonPhiInVal = PN.getOperand(InValNo);
10978 // Scan the rest of the operands to see if there are any conflicts, if so
10979 // there is no need to recursively scan other phis.
10980 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10981 Value *OpVal = PN.getIncomingValue(InValNo);
10982 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10986 // If we scanned over all operands, then we have one unique value plus
10987 // phi values. Scan PHI nodes to see if they all merge in each other or
10989 if (InValNo == NumOperandVals) {
10990 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10991 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10992 return ReplaceInstUsesWith(PN, NonPhiInVal);
10999 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
11000 Instruction *InsertPoint,
11001 InstCombiner *IC) {
11002 unsigned PtrSize = DTy->getScalarSizeInBits();
11003 unsigned VTySize = V->getType()->getScalarSizeInBits();
11004 // We must cast correctly to the pointer type. Ensure that we
11005 // sign extend the integer value if it is smaller as this is
11006 // used for address computation.
11007 Instruction::CastOps opcode =
11008 (VTySize < PtrSize ? Instruction::SExt :
11009 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
11010 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
11014 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
11015 Value *PtrOp = GEP.getOperand(0);
11016 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
11017 // If so, eliminate the noop.
11018 if (GEP.getNumOperands() == 1)
11019 return ReplaceInstUsesWith(GEP, PtrOp);
11021 if (isa<UndefValue>(GEP.getOperand(0)))
11022 return ReplaceInstUsesWith(GEP, Context->getUndef(GEP.getType()));
11024 bool HasZeroPointerIndex = false;
11025 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
11026 HasZeroPointerIndex = C->isNullValue();
11028 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11029 return ReplaceInstUsesWith(GEP, PtrOp);
11031 // Eliminate unneeded casts for indices.
11032 bool MadeChange = false;
11034 gep_type_iterator GTI = gep_type_begin(GEP);
11035 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
11036 i != e; ++i, ++GTI) {
11037 if (isa<SequentialType>(*GTI)) {
11038 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
11039 if (CI->getOpcode() == Instruction::ZExt ||
11040 CI->getOpcode() == Instruction::SExt) {
11041 const Type *SrcTy = CI->getOperand(0)->getType();
11042 // We can eliminate a cast from i32 to i64 iff the target
11043 // is a 32-bit pointer target.
11044 if (SrcTy->getScalarSizeInBits() >= TD->getPointerSizeInBits()) {
11046 *i = CI->getOperand(0);
11050 // If we are using a wider index than needed for this platform, shrink it
11051 // to what we need. If narrower, sign-extend it to what we need.
11052 // If the incoming value needs a cast instruction,
11053 // insert it. This explicit cast can make subsequent optimizations more
11056 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
11057 if (Constant *C = dyn_cast<Constant>(Op)) {
11058 *i = Context->getConstantExprTrunc(C, TD->getIntPtrType());
11061 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
11066 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
11067 if (Constant *C = dyn_cast<Constant>(Op)) {
11068 *i = Context->getConstantExprSExt(C, TD->getIntPtrType());
11071 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
11079 if (MadeChange) return &GEP;
11081 // Combine Indices - If the source pointer to this getelementptr instruction
11082 // is a getelementptr instruction, combine the indices of the two
11083 // getelementptr instructions into a single instruction.
11085 SmallVector<Value*, 8> SrcGEPOperands;
11086 if (User *Src = dyn_castGetElementPtr(PtrOp))
11087 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
11089 if (!SrcGEPOperands.empty()) {
11090 // Note that if our source is a gep chain itself that we wait for that
11091 // chain to be resolved before we perform this transformation. This
11092 // avoids us creating a TON of code in some cases.
11094 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
11095 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
11096 return 0; // Wait until our source is folded to completion.
11098 SmallVector<Value*, 8> Indices;
11100 // Find out whether the last index in the source GEP is a sequential idx.
11101 bool EndsWithSequential = false;
11102 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
11103 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
11104 EndsWithSequential = !isa<StructType>(*I);
11106 // Can we combine the two pointer arithmetics offsets?
11107 if (EndsWithSequential) {
11108 // Replace: gep (gep %P, long B), long A, ...
11109 // With: T = long A+B; gep %P, T, ...
11111 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
11112 if (SO1 == Context->getNullValue(SO1->getType())) {
11114 } else if (GO1 == Context->getNullValue(GO1->getType())) {
11117 // If they aren't the same type, convert both to an integer of the
11118 // target's pointer size.
11119 if (SO1->getType() != GO1->getType()) {
11120 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
11122 Context->getConstantExprIntegerCast(SO1C, GO1->getType(), true);
11123 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
11125 Context->getConstantExprIntegerCast(GO1C, SO1->getType(), true);
11127 unsigned PS = TD->getPointerSizeInBits();
11128 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
11129 // Convert GO1 to SO1's type.
11130 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
11132 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
11133 // Convert SO1 to GO1's type.
11134 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
11136 const Type *PT = TD->getIntPtrType();
11137 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
11138 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
11142 if (isa<Constant>(SO1) && isa<Constant>(GO1))
11143 Sum = Context->getConstantExprAdd(cast<Constant>(SO1),
11144 cast<Constant>(GO1));
11146 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11147 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
11151 // Recycle the GEP we already have if possible.
11152 if (SrcGEPOperands.size() == 2) {
11153 GEP.setOperand(0, SrcGEPOperands[0]);
11154 GEP.setOperand(1, Sum);
11157 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11158 SrcGEPOperands.end()-1);
11159 Indices.push_back(Sum);
11160 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
11162 } else if (isa<Constant>(*GEP.idx_begin()) &&
11163 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11164 SrcGEPOperands.size() != 1) {
11165 // Otherwise we can do the fold if the first index of the GEP is a zero
11166 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11167 SrcGEPOperands.end());
11168 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
11171 if (!Indices.empty())
11172 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
11173 Indices.end(), GEP.getName());
11175 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
11176 // GEP of global variable. If all of the indices for this GEP are
11177 // constants, we can promote this to a constexpr instead of an instruction.
11179 // Scan for nonconstants...
11180 SmallVector<Constant*, 8> Indices;
11181 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
11182 for (; I != E && isa<Constant>(*I); ++I)
11183 Indices.push_back(cast<Constant>(*I));
11185 if (I == E) { // If they are all constants...
11186 Constant *CE = Context->getConstantExprGetElementPtr(GV,
11187 &Indices[0],Indices.size());
11189 // Replace all uses of the GEP with the new constexpr...
11190 return ReplaceInstUsesWith(GEP, CE);
11192 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
11193 if (!isa<PointerType>(X->getType())) {
11194 // Not interesting. Source pointer must be a cast from pointer.
11195 } else if (HasZeroPointerIndex) {
11196 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11197 // into : GEP [10 x i8]* X, i32 0, ...
11199 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11200 // into : GEP i8* X, ...
11202 // This occurs when the program declares an array extern like "int X[];"
11203 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11204 const PointerType *XTy = cast<PointerType>(X->getType());
11205 if (const ArrayType *CATy =
11206 dyn_cast<ArrayType>(CPTy->getElementType())) {
11207 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11208 if (CATy->getElementType() == XTy->getElementType()) {
11209 // -> GEP i8* X, ...
11210 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11211 return GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11213 } else if (const ArrayType *XATy =
11214 dyn_cast<ArrayType>(XTy->getElementType())) {
11215 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11216 if (CATy->getElementType() == XATy->getElementType()) {
11217 // -> GEP [10 x i8]* X, i32 0, ...
11218 // At this point, we know that the cast source type is a pointer
11219 // to an array of the same type as the destination pointer
11220 // array. Because the array type is never stepped over (there
11221 // is a leading zero) we can fold the cast into this GEP.
11222 GEP.setOperand(0, X);
11227 } else if (GEP.getNumOperands() == 2) {
11228 // Transform things like:
11229 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11230 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11231 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11232 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11233 if (isa<ArrayType>(SrcElTy) &&
11234 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11235 TD->getTypeAllocSize(ResElTy)) {
11237 Idx[0] = Context->getNullValue(Type::Int32Ty);
11238 Idx[1] = GEP.getOperand(1);
11239 Value *V = InsertNewInstBefore(
11240 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
11241 // V and GEP are both pointer types --> BitCast
11242 return new BitCastInst(V, GEP.getType());
11245 // Transform things like:
11246 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11247 // (where tmp = 8*tmp2) into:
11248 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11250 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
11251 uint64_t ArrayEltSize =
11252 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11254 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11255 // allow either a mul, shift, or constant here.
11257 ConstantInt *Scale = 0;
11258 if (ArrayEltSize == 1) {
11259 NewIdx = GEP.getOperand(1);
11261 Context->getConstantInt(cast<IntegerType>(NewIdx->getType()), 1);
11262 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11263 NewIdx = Context->getConstantInt(CI->getType(), 1);
11265 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11266 if (Inst->getOpcode() == Instruction::Shl &&
11267 isa<ConstantInt>(Inst->getOperand(1))) {
11268 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11269 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11270 Scale = Context->getConstantInt(cast<IntegerType>(Inst->getType()),
11272 NewIdx = Inst->getOperand(0);
11273 } else if (Inst->getOpcode() == Instruction::Mul &&
11274 isa<ConstantInt>(Inst->getOperand(1))) {
11275 Scale = cast<ConstantInt>(Inst->getOperand(1));
11276 NewIdx = Inst->getOperand(0);
11280 // If the index will be to exactly the right offset with the scale taken
11281 // out, perform the transformation. Note, we don't know whether Scale is
11282 // signed or not. We'll use unsigned version of division/modulo
11283 // operation after making sure Scale doesn't have the sign bit set.
11284 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11285 Scale->getZExtValue() % ArrayEltSize == 0) {
11286 Scale = Context->getConstantInt(Scale->getType(),
11287 Scale->getZExtValue() / ArrayEltSize);
11288 if (Scale->getZExtValue() != 1) {
11290 Context->getConstantExprIntegerCast(Scale, NewIdx->getType(),
11292 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
11293 NewIdx = InsertNewInstBefore(Sc, GEP);
11296 // Insert the new GEP instruction.
11298 Idx[0] = Context->getNullValue(Type::Int32Ty);
11300 Instruction *NewGEP =
11301 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11302 NewGEP = InsertNewInstBefore(NewGEP, GEP);
11303 // The NewGEP must be pointer typed, so must the old one -> BitCast
11304 return new BitCastInst(NewGEP, GEP.getType());
11310 /// See if we can simplify:
11311 /// X = bitcast A to B*
11312 /// Y = gep X, <...constant indices...>
11313 /// into a gep of the original struct. This is important for SROA and alias
11314 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11315 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11316 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11317 // Determine how much the GEP moves the pointer. We are guaranteed to get
11318 // a constant back from EmitGEPOffset.
11319 ConstantInt *OffsetV =
11320 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11321 int64_t Offset = OffsetV->getSExtValue();
11323 // If this GEP instruction doesn't move the pointer, just replace the GEP
11324 // with a bitcast of the real input to the dest type.
11326 // If the bitcast is of an allocation, and the allocation will be
11327 // converted to match the type of the cast, don't touch this.
11328 if (isa<AllocationInst>(BCI->getOperand(0))) {
11329 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11330 if (Instruction *I = visitBitCast(*BCI)) {
11333 BCI->getParent()->getInstList().insert(BCI, I);
11334 ReplaceInstUsesWith(*BCI, I);
11339 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11342 // Otherwise, if the offset is non-zero, we need to find out if there is a
11343 // field at Offset in 'A's type. If so, we can pull the cast through the
11345 SmallVector<Value*, 8> NewIndices;
11347 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11348 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11349 Instruction *NGEP =
11350 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11352 if (NGEP->getType() == GEP.getType()) return NGEP;
11353 InsertNewInstBefore(NGEP, GEP);
11354 NGEP->takeName(&GEP);
11355 return new BitCastInst(NGEP, GEP.getType());
11363 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11364 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11365 if (AI.isArrayAllocation()) { // Check C != 1
11366 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11367 const Type *NewTy =
11368 Context->getArrayType(AI.getAllocatedType(), C->getZExtValue());
11369 AllocationInst *New = 0;
11371 // Create and insert the replacement instruction...
11372 if (isa<MallocInst>(AI))
11373 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11375 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11376 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11379 InsertNewInstBefore(New, AI);
11381 // Scan to the end of the allocation instructions, to skip over a block of
11382 // allocas if possible...also skip interleaved debug info
11384 BasicBlock::iterator It = New;
11385 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11387 // Now that I is pointing to the first non-allocation-inst in the block,
11388 // insert our getelementptr instruction...
11390 Value *NullIdx = Context->getNullValue(Type::Int32Ty);
11394 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11395 New->getName()+".sub", It);
11397 // Now make everything use the getelementptr instead of the original
11399 return ReplaceInstUsesWith(AI, V);
11400 } else if (isa<UndefValue>(AI.getArraySize())) {
11401 return ReplaceInstUsesWith(AI, Context->getNullValue(AI.getType()));
11405 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11406 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11407 // Note that we only do this for alloca's, because malloc should allocate
11408 // and return a unique pointer, even for a zero byte allocation.
11409 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11410 return ReplaceInstUsesWith(AI, Context->getNullValue(AI.getType()));
11412 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11413 if (AI.getAlignment() == 0)
11414 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11420 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11421 Value *Op = FI.getOperand(0);
11423 // free undef -> unreachable.
11424 if (isa<UndefValue>(Op)) {
11425 // Insert a new store to null because we cannot modify the CFG here.
11426 new StoreInst(Context->getConstantIntTrue(),
11427 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)), &FI);
11428 return EraseInstFromFunction(FI);
11431 // If we have 'free null' delete the instruction. This can happen in stl code
11432 // when lots of inlining happens.
11433 if (isa<ConstantPointerNull>(Op))
11434 return EraseInstFromFunction(FI);
11436 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11437 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11438 FI.setOperand(0, CI->getOperand(0));
11442 // Change free (gep X, 0,0,0,0) into free(X)
11443 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11444 if (GEPI->hasAllZeroIndices()) {
11445 AddToWorkList(GEPI);
11446 FI.setOperand(0, GEPI->getOperand(0));
11451 // Change free(malloc) into nothing, if the malloc has a single use.
11452 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11453 if (MI->hasOneUse()) {
11454 EraseInstFromFunction(FI);
11455 return EraseInstFromFunction(*MI);
11462 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11463 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11464 const TargetData *TD) {
11465 User *CI = cast<User>(LI.getOperand(0));
11466 Value *CastOp = CI->getOperand(0);
11467 LLVMContext *Context = IC.getContext();
11470 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11471 // Instead of loading constant c string, use corresponding integer value
11472 // directly if string length is small enough.
11474 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11475 unsigned len = Str.length();
11476 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11477 unsigned numBits = Ty->getPrimitiveSizeInBits();
11478 // Replace LI with immediate integer store.
11479 if ((numBits >> 3) == len + 1) {
11480 APInt StrVal(numBits, 0);
11481 APInt SingleChar(numBits, 0);
11482 if (TD->isLittleEndian()) {
11483 for (signed i = len-1; i >= 0; i--) {
11484 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11485 StrVal = (StrVal << 8) | SingleChar;
11488 for (unsigned i = 0; i < len; i++) {
11489 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11490 StrVal = (StrVal << 8) | SingleChar;
11492 // Append NULL at the end.
11494 StrVal = (StrVal << 8) | SingleChar;
11496 Value *NL = Context->getConstantInt(StrVal);
11497 return IC.ReplaceInstUsesWith(LI, NL);
11503 const PointerType *DestTy = cast<PointerType>(CI->getType());
11504 const Type *DestPTy = DestTy->getElementType();
11505 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11507 // If the address spaces don't match, don't eliminate the cast.
11508 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11511 const Type *SrcPTy = SrcTy->getElementType();
11513 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11514 isa<VectorType>(DestPTy)) {
11515 // If the source is an array, the code below will not succeed. Check to
11516 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11518 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11519 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11520 if (ASrcTy->getNumElements() != 0) {
11522 Idxs[0] = Idxs[1] = Context->getNullValue(Type::Int32Ty);
11523 CastOp = Context->getConstantExprGetElementPtr(CSrc, Idxs, 2);
11524 SrcTy = cast<PointerType>(CastOp->getType());
11525 SrcPTy = SrcTy->getElementType();
11528 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11529 isa<VectorType>(SrcPTy)) &&
11530 // Do not allow turning this into a load of an integer, which is then
11531 // casted to a pointer, this pessimizes pointer analysis a lot.
11532 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11533 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11534 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11536 // Okay, we are casting from one integer or pointer type to another of
11537 // the same size. Instead of casting the pointer before the load, cast
11538 // the result of the loaded value.
11539 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11541 LI.isVolatile()),LI);
11542 // Now cast the result of the load.
11543 return new BitCastInst(NewLoad, LI.getType());
11550 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11551 Value *Op = LI.getOperand(0);
11553 // Attempt to improve the alignment.
11554 unsigned KnownAlign =
11555 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11557 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11558 LI.getAlignment()))
11559 LI.setAlignment(KnownAlign);
11561 // load (cast X) --> cast (load X) iff safe
11562 if (isa<CastInst>(Op))
11563 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11566 // None of the following transforms are legal for volatile loads.
11567 if (LI.isVolatile()) return 0;
11569 // Do really simple store-to-load forwarding and load CSE, to catch cases
11570 // where there are several consequtive memory accesses to the same location,
11571 // separated by a few arithmetic operations.
11572 BasicBlock::iterator BBI = &LI;
11573 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11574 return ReplaceInstUsesWith(LI, AvailableVal);
11576 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11577 const Value *GEPI0 = GEPI->getOperand(0);
11578 // TODO: Consider a target hook for valid address spaces for this xform.
11579 if (isa<ConstantPointerNull>(GEPI0) &&
11580 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11581 // Insert a new store to null instruction before the load to indicate
11582 // that this code is not reachable. We do this instead of inserting
11583 // an unreachable instruction directly because we cannot modify the
11585 new StoreInst(Context->getUndef(LI.getType()),
11586 Context->getNullValue(Op->getType()), &LI);
11587 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11591 if (Constant *C = dyn_cast<Constant>(Op)) {
11592 // load null/undef -> undef
11593 // TODO: Consider a target hook for valid address spaces for this xform.
11594 if (isa<UndefValue>(C) || (C->isNullValue() &&
11595 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11596 // Insert a new store to null instruction before the load to indicate that
11597 // this code is not reachable. We do this instead of inserting an
11598 // unreachable instruction directly because we cannot modify the CFG.
11599 new StoreInst(Context->getUndef(LI.getType()),
11600 Context->getNullValue(Op->getType()), &LI);
11601 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11604 // Instcombine load (constant global) into the value loaded.
11605 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11606 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11607 return ReplaceInstUsesWith(LI, GV->getInitializer());
11609 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11610 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11611 if (CE->getOpcode() == Instruction::GetElementPtr) {
11612 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11613 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11615 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11617 return ReplaceInstUsesWith(LI, V);
11618 if (CE->getOperand(0)->isNullValue()) {
11619 // Insert a new store to null instruction before the load to indicate
11620 // that this code is not reachable. We do this instead of inserting
11621 // an unreachable instruction directly because we cannot modify the
11623 new StoreInst(Context->getUndef(LI.getType()),
11624 Context->getNullValue(Op->getType()), &LI);
11625 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11628 } else if (CE->isCast()) {
11629 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11635 // If this load comes from anywhere in a constant global, and if the global
11636 // is all undef or zero, we know what it loads.
11637 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11638 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11639 if (GV->getInitializer()->isNullValue())
11640 return ReplaceInstUsesWith(LI, Context->getNullValue(LI.getType()));
11641 else if (isa<UndefValue>(GV->getInitializer()))
11642 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11646 if (Op->hasOneUse()) {
11647 // Change select and PHI nodes to select values instead of addresses: this
11648 // helps alias analysis out a lot, allows many others simplifications, and
11649 // exposes redundancy in the code.
11651 // Note that we cannot do the transformation unless we know that the
11652 // introduced loads cannot trap! Something like this is valid as long as
11653 // the condition is always false: load (select bool %C, int* null, int* %G),
11654 // but it would not be valid if we transformed it to load from null
11655 // unconditionally.
11657 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11658 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11659 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11660 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11661 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11662 SI->getOperand(1)->getName()+".val"), LI);
11663 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11664 SI->getOperand(2)->getName()+".val"), LI);
11665 return SelectInst::Create(SI->getCondition(), V1, V2);
11668 // load (select (cond, null, P)) -> load P
11669 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11670 if (C->isNullValue()) {
11671 LI.setOperand(0, SI->getOperand(2));
11675 // load (select (cond, P, null)) -> load P
11676 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11677 if (C->isNullValue()) {
11678 LI.setOperand(0, SI->getOperand(1));
11686 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11687 /// when possible. This makes it generally easy to do alias analysis and/or
11688 /// SROA/mem2reg of the memory object.
11689 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11690 User *CI = cast<User>(SI.getOperand(1));
11691 Value *CastOp = CI->getOperand(0);
11692 LLVMContext *Context = IC.getContext();
11694 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11695 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11696 if (SrcTy == 0) return 0;
11698 const Type *SrcPTy = SrcTy->getElementType();
11700 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11703 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11704 /// to its first element. This allows us to handle things like:
11705 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11706 /// on 32-bit hosts.
11707 SmallVector<Value*, 4> NewGEPIndices;
11709 // If the source is an array, the code below will not succeed. Check to
11710 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11712 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11713 // Index through pointer.
11714 Constant *Zero = Context->getNullValue(Type::Int32Ty);
11715 NewGEPIndices.push_back(Zero);
11718 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11719 if (!STy->getNumElements()) /* Struct can be empty {} */
11721 NewGEPIndices.push_back(Zero);
11722 SrcPTy = STy->getElementType(0);
11723 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11724 NewGEPIndices.push_back(Zero);
11725 SrcPTy = ATy->getElementType();
11731 SrcTy = Context->getPointerType(SrcPTy, SrcTy->getAddressSpace());
11734 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11737 // If the pointers point into different address spaces or if they point to
11738 // values with different sizes, we can't do the transformation.
11739 if (SrcTy->getAddressSpace() !=
11740 cast<PointerType>(CI->getType())->getAddressSpace() ||
11741 IC.getTargetData().getTypeSizeInBits(SrcPTy) !=
11742 IC.getTargetData().getTypeSizeInBits(DestPTy))
11745 // Okay, we are casting from one integer or pointer type to another of
11746 // the same size. Instead of casting the pointer before
11747 // the store, cast the value to be stored.
11749 Value *SIOp0 = SI.getOperand(0);
11750 Instruction::CastOps opcode = Instruction::BitCast;
11751 const Type* CastSrcTy = SIOp0->getType();
11752 const Type* CastDstTy = SrcPTy;
11753 if (isa<PointerType>(CastDstTy)) {
11754 if (CastSrcTy->isInteger())
11755 opcode = Instruction::IntToPtr;
11756 } else if (isa<IntegerType>(CastDstTy)) {
11757 if (isa<PointerType>(SIOp0->getType()))
11758 opcode = Instruction::PtrToInt;
11761 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11762 // emit a GEP to index into its first field.
11763 if (!NewGEPIndices.empty()) {
11764 if (Constant *C = dyn_cast<Constant>(CastOp))
11765 CastOp = Context->getConstantExprGetElementPtr(C, &NewGEPIndices[0],
11766 NewGEPIndices.size());
11768 CastOp = IC.InsertNewInstBefore(
11769 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11770 NewGEPIndices.end()), SI);
11773 if (Constant *C = dyn_cast<Constant>(SIOp0))
11774 NewCast = Context->getConstantExprCast(opcode, C, CastDstTy);
11776 NewCast = IC.InsertNewInstBefore(
11777 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11779 return new StoreInst(NewCast, CastOp);
11782 /// equivalentAddressValues - Test if A and B will obviously have the same
11783 /// value. This includes recognizing that %t0 and %t1 will have the same
11784 /// value in code like this:
11785 /// %t0 = getelementptr \@a, 0, 3
11786 /// store i32 0, i32* %t0
11787 /// %t1 = getelementptr \@a, 0, 3
11788 /// %t2 = load i32* %t1
11790 static bool equivalentAddressValues(Value *A, Value *B) {
11791 // Test if the values are trivially equivalent.
11792 if (A == B) return true;
11794 // Test if the values come form identical arithmetic instructions.
11795 if (isa<BinaryOperator>(A) ||
11796 isa<CastInst>(A) ||
11798 isa<GetElementPtrInst>(A))
11799 if (Instruction *BI = dyn_cast<Instruction>(B))
11800 if (cast<Instruction>(A)->isIdenticalTo(BI))
11803 // Otherwise they may not be equivalent.
11807 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11808 // return the llvm.dbg.declare.
11809 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11810 if (!V->hasNUses(2))
11812 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11814 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11816 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11817 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11824 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11825 Value *Val = SI.getOperand(0);
11826 Value *Ptr = SI.getOperand(1);
11828 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11829 EraseInstFromFunction(SI);
11834 // If the RHS is an alloca with a single use, zapify the store, making the
11836 // If the RHS is an alloca with a two uses, the other one being a
11837 // llvm.dbg.declare, zapify the store and the declare, making the
11838 // alloca dead. We must do this to prevent declare's from affecting
11840 if (!SI.isVolatile()) {
11841 if (Ptr->hasOneUse()) {
11842 if (isa<AllocaInst>(Ptr)) {
11843 EraseInstFromFunction(SI);
11847 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11848 if (isa<AllocaInst>(GEP->getOperand(0))) {
11849 if (GEP->getOperand(0)->hasOneUse()) {
11850 EraseInstFromFunction(SI);
11854 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11855 EraseInstFromFunction(*DI);
11856 EraseInstFromFunction(SI);
11863 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11864 EraseInstFromFunction(*DI);
11865 EraseInstFromFunction(SI);
11871 // Attempt to improve the alignment.
11872 unsigned KnownAlign =
11873 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11875 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11876 SI.getAlignment()))
11877 SI.setAlignment(KnownAlign);
11879 // Do really simple DSE, to catch cases where there are several consecutive
11880 // stores to the same location, separated by a few arithmetic operations. This
11881 // situation often occurs with bitfield accesses.
11882 BasicBlock::iterator BBI = &SI;
11883 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11886 // Don't count debug info directives, lest they affect codegen,
11887 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11888 // It is necessary for correctness to skip those that feed into a
11889 // llvm.dbg.declare, as these are not present when debugging is off.
11890 if (isa<DbgInfoIntrinsic>(BBI) ||
11891 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11896 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11897 // Prev store isn't volatile, and stores to the same location?
11898 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11899 SI.getOperand(1))) {
11902 EraseInstFromFunction(*PrevSI);
11908 // If this is a load, we have to stop. However, if the loaded value is from
11909 // the pointer we're loading and is producing the pointer we're storing,
11910 // then *this* store is dead (X = load P; store X -> P).
11911 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11912 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11913 !SI.isVolatile()) {
11914 EraseInstFromFunction(SI);
11918 // Otherwise, this is a load from some other location. Stores before it
11919 // may not be dead.
11923 // Don't skip over loads or things that can modify memory.
11924 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11929 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11931 // store X, null -> turns into 'unreachable' in SimplifyCFG
11932 if (isa<ConstantPointerNull>(Ptr) &&
11933 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11934 if (!isa<UndefValue>(Val)) {
11935 SI.setOperand(0, Context->getUndef(Val->getType()));
11936 if (Instruction *U = dyn_cast<Instruction>(Val))
11937 AddToWorkList(U); // Dropped a use.
11940 return 0; // Do not modify these!
11943 // store undef, Ptr -> noop
11944 if (isa<UndefValue>(Val)) {
11945 EraseInstFromFunction(SI);
11950 // If the pointer destination is a cast, see if we can fold the cast into the
11952 if (isa<CastInst>(Ptr))
11953 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11955 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11957 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11961 // If this store is the last instruction in the basic block (possibly
11962 // excepting debug info instructions and the pointer bitcasts that feed
11963 // into them), and if the block ends with an unconditional branch, try
11964 // to move it to the successor block.
11968 } while (isa<DbgInfoIntrinsic>(BBI) ||
11969 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11970 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11971 if (BI->isUnconditional())
11972 if (SimplifyStoreAtEndOfBlock(SI))
11973 return 0; // xform done!
11978 /// SimplifyStoreAtEndOfBlock - Turn things like:
11979 /// if () { *P = v1; } else { *P = v2 }
11980 /// into a phi node with a store in the successor.
11982 /// Simplify things like:
11983 /// *P = v1; if () { *P = v2; }
11984 /// into a phi node with a store in the successor.
11986 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11987 BasicBlock *StoreBB = SI.getParent();
11989 // Check to see if the successor block has exactly two incoming edges. If
11990 // so, see if the other predecessor contains a store to the same location.
11991 // if so, insert a PHI node (if needed) and move the stores down.
11992 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11994 // Determine whether Dest has exactly two predecessors and, if so, compute
11995 // the other predecessor.
11996 pred_iterator PI = pred_begin(DestBB);
11997 BasicBlock *OtherBB = 0;
11998 if (*PI != StoreBB)
12001 if (PI == pred_end(DestBB))
12004 if (*PI != StoreBB) {
12009 if (++PI != pred_end(DestBB))
12012 // Bail out if all the relevant blocks aren't distinct (this can happen,
12013 // for example, if SI is in an infinite loop)
12014 if (StoreBB == DestBB || OtherBB == DestBB)
12017 // Verify that the other block ends in a branch and is not otherwise empty.
12018 BasicBlock::iterator BBI = OtherBB->getTerminator();
12019 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
12020 if (!OtherBr || BBI == OtherBB->begin())
12023 // If the other block ends in an unconditional branch, check for the 'if then
12024 // else' case. there is an instruction before the branch.
12025 StoreInst *OtherStore = 0;
12026 if (OtherBr->isUnconditional()) {
12028 // Skip over debugging info.
12029 while (isa<DbgInfoIntrinsic>(BBI) ||
12030 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12031 if (BBI==OtherBB->begin())
12035 // If this isn't a store, or isn't a store to the same location, bail out.
12036 OtherStore = dyn_cast<StoreInst>(BBI);
12037 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
12040 // Otherwise, the other block ended with a conditional branch. If one of the
12041 // destinations is StoreBB, then we have the if/then case.
12042 if (OtherBr->getSuccessor(0) != StoreBB &&
12043 OtherBr->getSuccessor(1) != StoreBB)
12046 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12047 // if/then triangle. See if there is a store to the same ptr as SI that
12048 // lives in OtherBB.
12050 // Check to see if we find the matching store.
12051 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12052 if (OtherStore->getOperand(1) != SI.getOperand(1))
12056 // If we find something that may be using or overwriting the stored
12057 // value, or if we run out of instructions, we can't do the xform.
12058 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12059 BBI == OtherBB->begin())
12063 // In order to eliminate the store in OtherBr, we have to
12064 // make sure nothing reads or overwrites the stored value in
12066 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12067 // FIXME: This should really be AA driven.
12068 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12073 // Insert a PHI node now if we need it.
12074 Value *MergedVal = OtherStore->getOperand(0);
12075 if (MergedVal != SI.getOperand(0)) {
12076 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12077 PN->reserveOperandSpace(2);
12078 PN->addIncoming(SI.getOperand(0), SI.getParent());
12079 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12080 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12083 // Advance to a place where it is safe to insert the new store and
12085 BBI = DestBB->getFirstNonPHI();
12086 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12087 OtherStore->isVolatile()), *BBI);
12089 // Nuke the old stores.
12090 EraseInstFromFunction(SI);
12091 EraseInstFromFunction(*OtherStore);
12097 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12098 // Change br (not X), label True, label False to: br X, label False, True
12100 BasicBlock *TrueDest;
12101 BasicBlock *FalseDest;
12102 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest), *Context) &&
12103 !isa<Constant>(X)) {
12104 // Swap Destinations and condition...
12105 BI.setCondition(X);
12106 BI.setSuccessor(0, FalseDest);
12107 BI.setSuccessor(1, TrueDest);
12111 // Cannonicalize fcmp_one -> fcmp_oeq
12112 FCmpInst::Predicate FPred; Value *Y;
12113 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12114 TrueDest, FalseDest), *Context))
12115 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12116 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
12117 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
12118 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
12119 Instruction *NewSCC = new FCmpInst(I, NewPred, X, Y, "");
12120 NewSCC->takeName(I);
12121 // Swap Destinations and condition...
12122 BI.setCondition(NewSCC);
12123 BI.setSuccessor(0, FalseDest);
12124 BI.setSuccessor(1, TrueDest);
12125 RemoveFromWorkList(I);
12126 I->eraseFromParent();
12127 AddToWorkList(NewSCC);
12131 // Cannonicalize icmp_ne -> icmp_eq
12132 ICmpInst::Predicate IPred;
12133 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12134 TrueDest, FalseDest), *Context))
12135 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12136 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12137 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
12138 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
12139 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
12140 Instruction *NewSCC = new ICmpInst(I, NewPred, X, Y, "");
12141 NewSCC->takeName(I);
12142 // Swap Destinations and condition...
12143 BI.setCondition(NewSCC);
12144 BI.setSuccessor(0, FalseDest);
12145 BI.setSuccessor(1, TrueDest);
12146 RemoveFromWorkList(I);
12147 I->eraseFromParent();;
12148 AddToWorkList(NewSCC);
12155 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12156 Value *Cond = SI.getCondition();
12157 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12158 if (I->getOpcode() == Instruction::Add)
12159 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12160 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12161 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12163 Context->getConstantExprSub(cast<Constant>(SI.getOperand(i)),
12165 SI.setOperand(0, I->getOperand(0));
12173 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12174 Value *Agg = EV.getAggregateOperand();
12176 if (!EV.hasIndices())
12177 return ReplaceInstUsesWith(EV, Agg);
12179 if (Constant *C = dyn_cast<Constant>(Agg)) {
12180 if (isa<UndefValue>(C))
12181 return ReplaceInstUsesWith(EV, Context->getUndef(EV.getType()));
12183 if (isa<ConstantAggregateZero>(C))
12184 return ReplaceInstUsesWith(EV, Context->getNullValue(EV.getType()));
12186 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12187 // Extract the element indexed by the first index out of the constant
12188 Value *V = C->getOperand(*EV.idx_begin());
12189 if (EV.getNumIndices() > 1)
12190 // Extract the remaining indices out of the constant indexed by the
12192 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12194 return ReplaceInstUsesWith(EV, V);
12196 return 0; // Can't handle other constants
12198 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12199 // We're extracting from an insertvalue instruction, compare the indices
12200 const unsigned *exti, *exte, *insi, *inse;
12201 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12202 exte = EV.idx_end(), inse = IV->idx_end();
12203 exti != exte && insi != inse;
12205 if (*insi != *exti)
12206 // The insert and extract both reference distinctly different elements.
12207 // This means the extract is not influenced by the insert, and we can
12208 // replace the aggregate operand of the extract with the aggregate
12209 // operand of the insert. i.e., replace
12210 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12211 // %E = extractvalue { i32, { i32 } } %I, 0
12213 // %E = extractvalue { i32, { i32 } } %A, 0
12214 return ExtractValueInst::Create(IV->getAggregateOperand(),
12215 EV.idx_begin(), EV.idx_end());
12217 if (exti == exte && insi == inse)
12218 // Both iterators are at the end: Index lists are identical. Replace
12219 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12220 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12222 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12223 if (exti == exte) {
12224 // The extract list is a prefix of the insert list. i.e. replace
12225 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12226 // %E = extractvalue { i32, { i32 } } %I, 1
12228 // %X = extractvalue { i32, { i32 } } %A, 1
12229 // %E = insertvalue { i32 } %X, i32 42, 0
12230 // by switching the order of the insert and extract (though the
12231 // insertvalue should be left in, since it may have other uses).
12232 Value *NewEV = InsertNewInstBefore(
12233 ExtractValueInst::Create(IV->getAggregateOperand(),
12234 EV.idx_begin(), EV.idx_end()),
12236 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12240 // The insert list is a prefix of the extract list
12241 // We can simply remove the common indices from the extract and make it
12242 // operate on the inserted value instead of the insertvalue result.
12244 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12245 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12247 // %E extractvalue { i32 } { i32 42 }, 0
12248 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12251 // Can't simplify extracts from other values. Note that nested extracts are
12252 // already simplified implicitely by the above (extract ( extract (insert) )
12253 // will be translated into extract ( insert ( extract ) ) first and then just
12254 // the value inserted, if appropriate).
12258 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12259 /// is to leave as a vector operation.
12260 static bool CheapToScalarize(Value *V, bool isConstant) {
12261 if (isa<ConstantAggregateZero>(V))
12263 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12264 if (isConstant) return true;
12265 // If all elts are the same, we can extract.
12266 Constant *Op0 = C->getOperand(0);
12267 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12268 if (C->getOperand(i) != Op0)
12272 Instruction *I = dyn_cast<Instruction>(V);
12273 if (!I) return false;
12275 // Insert element gets simplified to the inserted element or is deleted if
12276 // this is constant idx extract element and its a constant idx insertelt.
12277 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12278 isa<ConstantInt>(I->getOperand(2)))
12280 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12282 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12283 if (BO->hasOneUse() &&
12284 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12285 CheapToScalarize(BO->getOperand(1), isConstant)))
12287 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12288 if (CI->hasOneUse() &&
12289 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12290 CheapToScalarize(CI->getOperand(1), isConstant)))
12296 /// Read and decode a shufflevector mask.
12298 /// It turns undef elements into values that are larger than the number of
12299 /// elements in the input.
12300 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12301 unsigned NElts = SVI->getType()->getNumElements();
12302 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12303 return std::vector<unsigned>(NElts, 0);
12304 if (isa<UndefValue>(SVI->getOperand(2)))
12305 return std::vector<unsigned>(NElts, 2*NElts);
12307 std::vector<unsigned> Result;
12308 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12309 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12310 if (isa<UndefValue>(*i))
12311 Result.push_back(NElts*2); // undef -> 8
12313 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12317 /// FindScalarElement - Given a vector and an element number, see if the scalar
12318 /// value is already around as a register, for example if it were inserted then
12319 /// extracted from the vector.
12320 static Value *FindScalarElement(Value *V, unsigned EltNo,
12321 LLVMContext *Context) {
12322 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12323 const VectorType *PTy = cast<VectorType>(V->getType());
12324 unsigned Width = PTy->getNumElements();
12325 if (EltNo >= Width) // Out of range access.
12326 return Context->getUndef(PTy->getElementType());
12328 if (isa<UndefValue>(V))
12329 return Context->getUndef(PTy->getElementType());
12330 else if (isa<ConstantAggregateZero>(V))
12331 return Context->getNullValue(PTy->getElementType());
12332 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12333 return CP->getOperand(EltNo);
12334 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12335 // If this is an insert to a variable element, we don't know what it is.
12336 if (!isa<ConstantInt>(III->getOperand(2)))
12338 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12340 // If this is an insert to the element we are looking for, return the
12342 if (EltNo == IIElt)
12343 return III->getOperand(1);
12345 // Otherwise, the insertelement doesn't modify the value, recurse on its
12347 return FindScalarElement(III->getOperand(0), EltNo, Context);
12348 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12349 unsigned LHSWidth =
12350 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12351 unsigned InEl = getShuffleMask(SVI)[EltNo];
12352 if (InEl < LHSWidth)
12353 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12354 else if (InEl < LHSWidth*2)
12355 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12357 return Context->getUndef(PTy->getElementType());
12360 // Otherwise, we don't know.
12364 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12365 // If vector val is undef, replace extract with scalar undef.
12366 if (isa<UndefValue>(EI.getOperand(0)))
12367 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12369 // If vector val is constant 0, replace extract with scalar 0.
12370 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12371 return ReplaceInstUsesWith(EI, Context->getNullValue(EI.getType()));
12373 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12374 // If vector val is constant with all elements the same, replace EI with
12375 // that element. When the elements are not identical, we cannot replace yet
12376 // (we do that below, but only when the index is constant).
12377 Constant *op0 = C->getOperand(0);
12378 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12379 if (C->getOperand(i) != op0) {
12384 return ReplaceInstUsesWith(EI, op0);
12387 // If extracting a specified index from the vector, see if we can recursively
12388 // find a previously computed scalar that was inserted into the vector.
12389 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12390 unsigned IndexVal = IdxC->getZExtValue();
12391 unsigned VectorWidth =
12392 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12394 // If this is extracting an invalid index, turn this into undef, to avoid
12395 // crashing the code below.
12396 if (IndexVal >= VectorWidth)
12397 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12399 // This instruction only demands the single element from the input vector.
12400 // If the input vector has a single use, simplify it based on this use
12402 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12403 APInt UndefElts(VectorWidth, 0);
12404 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12405 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12406 DemandedMask, UndefElts)) {
12407 EI.setOperand(0, V);
12412 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12413 return ReplaceInstUsesWith(EI, Elt);
12415 // If the this extractelement is directly using a bitcast from a vector of
12416 // the same number of elements, see if we can find the source element from
12417 // it. In this case, we will end up needing to bitcast the scalars.
12418 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12419 if (const VectorType *VT =
12420 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12421 if (VT->getNumElements() == VectorWidth)
12422 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12423 IndexVal, Context))
12424 return new BitCastInst(Elt, EI.getType());
12428 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12429 if (I->hasOneUse()) {
12430 // Push extractelement into predecessor operation if legal and
12431 // profitable to do so
12432 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12433 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12434 if (CheapToScalarize(BO, isConstantElt)) {
12435 ExtractElementInst *newEI0 =
12436 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
12437 EI.getName()+".lhs");
12438 ExtractElementInst *newEI1 =
12439 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
12440 EI.getName()+".rhs");
12441 InsertNewInstBefore(newEI0, EI);
12442 InsertNewInstBefore(newEI1, EI);
12443 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12445 } else if (isa<LoadInst>(I)) {
12447 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12448 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12449 Context->getPointerType(EI.getType(), AS),EI);
12450 GetElementPtrInst *GEP =
12451 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12452 InsertNewInstBefore(GEP, EI);
12453 return new LoadInst(GEP);
12456 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12457 // Extracting the inserted element?
12458 if (IE->getOperand(2) == EI.getOperand(1))
12459 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12460 // If the inserted and extracted elements are constants, they must not
12461 // be the same value, extract from the pre-inserted value instead.
12462 if (isa<Constant>(IE->getOperand(2)) &&
12463 isa<Constant>(EI.getOperand(1))) {
12464 AddUsesToWorkList(EI);
12465 EI.setOperand(0, IE->getOperand(0));
12468 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12469 // If this is extracting an element from a shufflevector, figure out where
12470 // it came from and extract from the appropriate input element instead.
12471 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12472 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12474 unsigned LHSWidth =
12475 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12477 if (SrcIdx < LHSWidth)
12478 Src = SVI->getOperand(0);
12479 else if (SrcIdx < LHSWidth*2) {
12480 SrcIdx -= LHSWidth;
12481 Src = SVI->getOperand(1);
12483 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12485 return new ExtractElementInst(Src,
12486 Context->getConstantInt(Type::Int32Ty, SrcIdx, false));
12493 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12494 /// elements from either LHS or RHS, return the shuffle mask and true.
12495 /// Otherwise, return false.
12496 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12497 std::vector<Constant*> &Mask,
12498 LLVMContext *Context) {
12499 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12500 "Invalid CollectSingleShuffleElements");
12501 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12503 if (isa<UndefValue>(V)) {
12504 Mask.assign(NumElts, Context->getUndef(Type::Int32Ty));
12506 } else if (V == LHS) {
12507 for (unsigned i = 0; i != NumElts; ++i)
12508 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i));
12510 } else if (V == RHS) {
12511 for (unsigned i = 0; i != NumElts; ++i)
12512 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i+NumElts));
12514 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12515 // If this is an insert of an extract from some other vector, include it.
12516 Value *VecOp = IEI->getOperand(0);
12517 Value *ScalarOp = IEI->getOperand(1);
12518 Value *IdxOp = IEI->getOperand(2);
12520 if (!isa<ConstantInt>(IdxOp))
12522 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12524 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12525 // Okay, we can handle this if the vector we are insertinting into is
12526 // transitively ok.
12527 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12528 // If so, update the mask to reflect the inserted undef.
12529 Mask[InsertedIdx] = Context->getUndef(Type::Int32Ty);
12532 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12533 if (isa<ConstantInt>(EI->getOperand(1)) &&
12534 EI->getOperand(0)->getType() == V->getType()) {
12535 unsigned ExtractedIdx =
12536 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12538 // This must be extracting from either LHS or RHS.
12539 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
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 value.
12544 if (EI->getOperand(0) == LHS) {
12545 Mask[InsertedIdx % NumElts] =
12546 Context->getConstantInt(Type::Int32Ty, ExtractedIdx);
12548 assert(EI->getOperand(0) == RHS);
12549 Mask[InsertedIdx % NumElts] =
12550 Context->getConstantInt(Type::Int32Ty, ExtractedIdx+NumElts);
12559 // TODO: Handle shufflevector here!
12564 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12565 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12566 /// that computes V and the LHS value of the shuffle.
12567 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12568 Value *&RHS, LLVMContext *Context) {
12569 assert(isa<VectorType>(V->getType()) &&
12570 (RHS == 0 || V->getType() == RHS->getType()) &&
12571 "Invalid shuffle!");
12572 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12574 if (isa<UndefValue>(V)) {
12575 Mask.assign(NumElts, Context->getUndef(Type::Int32Ty));
12577 } else if (isa<ConstantAggregateZero>(V)) {
12578 Mask.assign(NumElts, Context->getConstantInt(Type::Int32Ty, 0));
12580 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12581 // If this is an insert of an extract from some other vector, include it.
12582 Value *VecOp = IEI->getOperand(0);
12583 Value *ScalarOp = IEI->getOperand(1);
12584 Value *IdxOp = IEI->getOperand(2);
12586 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12587 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12588 EI->getOperand(0)->getType() == V->getType()) {
12589 unsigned ExtractedIdx =
12590 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12591 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12593 // Either the extracted from or inserted into vector must be RHSVec,
12594 // otherwise we'd end up with a shuffle of three inputs.
12595 if (EI->getOperand(0) == RHS || RHS == 0) {
12596 RHS = EI->getOperand(0);
12597 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12598 Mask[InsertedIdx % NumElts] =
12599 Context->getConstantInt(Type::Int32Ty, NumElts+ExtractedIdx);
12603 if (VecOp == RHS) {
12604 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12606 // Everything but the extracted element is replaced with the RHS.
12607 for (unsigned i = 0; i != NumElts; ++i) {
12608 if (i != InsertedIdx)
12609 Mask[i] = Context->getConstantInt(Type::Int32Ty, NumElts+i);
12614 // If this insertelement is a chain that comes from exactly these two
12615 // vectors, return the vector and the effective shuffle.
12616 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12618 return EI->getOperand(0);
12623 // TODO: Handle shufflevector here!
12625 // Otherwise, can't do anything fancy. Return an identity vector.
12626 for (unsigned i = 0; i != NumElts; ++i)
12627 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i));
12631 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12632 Value *VecOp = IE.getOperand(0);
12633 Value *ScalarOp = IE.getOperand(1);
12634 Value *IdxOp = IE.getOperand(2);
12636 // Inserting an undef or into an undefined place, remove this.
12637 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12638 ReplaceInstUsesWith(IE, VecOp);
12640 // If the inserted element was extracted from some other vector, and if the
12641 // indexes are constant, try to turn this into a shufflevector operation.
12642 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12643 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12644 EI->getOperand(0)->getType() == IE.getType()) {
12645 unsigned NumVectorElts = IE.getType()->getNumElements();
12646 unsigned ExtractedIdx =
12647 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12648 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12650 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12651 return ReplaceInstUsesWith(IE, VecOp);
12653 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12654 return ReplaceInstUsesWith(IE, Context->getUndef(IE.getType()));
12656 // If we are extracting a value from a vector, then inserting it right
12657 // back into the same place, just use the input vector.
12658 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12659 return ReplaceInstUsesWith(IE, VecOp);
12661 // We could theoretically do this for ANY input. However, doing so could
12662 // turn chains of insertelement instructions into a chain of shufflevector
12663 // instructions, and right now we do not merge shufflevectors. As such,
12664 // only do this in a situation where it is clear that there is benefit.
12665 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12666 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12667 // the values of VecOp, except then one read from EIOp0.
12668 // Build a new shuffle mask.
12669 std::vector<Constant*> Mask;
12670 if (isa<UndefValue>(VecOp))
12671 Mask.assign(NumVectorElts, Context->getUndef(Type::Int32Ty));
12673 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12674 Mask.assign(NumVectorElts, Context->getConstantInt(Type::Int32Ty,
12677 Mask[InsertedIdx] =
12678 Context->getConstantInt(Type::Int32Ty, ExtractedIdx);
12679 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12680 Context->getConstantVector(Mask));
12683 // If this insertelement isn't used by some other insertelement, turn it
12684 // (and any insertelements it points to), into one big shuffle.
12685 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12686 std::vector<Constant*> Mask;
12688 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12689 if (RHS == 0) RHS = Context->getUndef(LHS->getType());
12690 // We now have a shuffle of LHS, RHS, Mask.
12691 return new ShuffleVectorInst(LHS, RHS,
12692 Context->getConstantVector(Mask));
12697 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12698 APInt UndefElts(VWidth, 0);
12699 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12700 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12707 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12708 Value *LHS = SVI.getOperand(0);
12709 Value *RHS = SVI.getOperand(1);
12710 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12712 bool MadeChange = false;
12714 // Undefined shuffle mask -> undefined value.
12715 if (isa<UndefValue>(SVI.getOperand(2)))
12716 return ReplaceInstUsesWith(SVI, Context->getUndef(SVI.getType()));
12718 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12720 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12723 APInt UndefElts(VWidth, 0);
12724 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12725 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12726 LHS = SVI.getOperand(0);
12727 RHS = SVI.getOperand(1);
12731 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12732 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12733 if (LHS == RHS || isa<UndefValue>(LHS)) {
12734 if (isa<UndefValue>(LHS) && LHS == RHS) {
12735 // shuffle(undef,undef,mask) -> undef.
12736 return ReplaceInstUsesWith(SVI, LHS);
12739 // Remap any references to RHS to use LHS.
12740 std::vector<Constant*> Elts;
12741 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12742 if (Mask[i] >= 2*e)
12743 Elts.push_back(Context->getUndef(Type::Int32Ty));
12745 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12746 (Mask[i] < e && isa<UndefValue>(LHS))) {
12747 Mask[i] = 2*e; // Turn into undef.
12748 Elts.push_back(Context->getUndef(Type::Int32Ty));
12750 Mask[i] = Mask[i] % e; // Force to LHS.
12751 Elts.push_back(Context->getConstantInt(Type::Int32Ty, Mask[i]));
12755 SVI.setOperand(0, SVI.getOperand(1));
12756 SVI.setOperand(1, Context->getUndef(RHS->getType()));
12757 SVI.setOperand(2, Context->getConstantVector(Elts));
12758 LHS = SVI.getOperand(0);
12759 RHS = SVI.getOperand(1);
12763 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12764 bool isLHSID = true, isRHSID = true;
12766 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12767 if (Mask[i] >= e*2) continue; // Ignore undef values.
12768 // Is this an identity shuffle of the LHS value?
12769 isLHSID &= (Mask[i] == i);
12771 // Is this an identity shuffle of the RHS value?
12772 isRHSID &= (Mask[i]-e == i);
12775 // Eliminate identity shuffles.
12776 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12777 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12779 // If the LHS is a shufflevector itself, see if we can combine it with this
12780 // one without producing an unusual shuffle. Here we are really conservative:
12781 // we are absolutely afraid of producing a shuffle mask not in the input
12782 // program, because the code gen may not be smart enough to turn a merged
12783 // shuffle into two specific shuffles: it may produce worse code. As such,
12784 // we only merge two shuffles if the result is one of the two input shuffle
12785 // masks. In this case, merging the shuffles just removes one instruction,
12786 // which we know is safe. This is good for things like turning:
12787 // (splat(splat)) -> splat.
12788 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12789 if (isa<UndefValue>(RHS)) {
12790 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12792 std::vector<unsigned> NewMask;
12793 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12794 if (Mask[i] >= 2*e)
12795 NewMask.push_back(2*e);
12797 NewMask.push_back(LHSMask[Mask[i]]);
12799 // If the result mask is equal to the src shuffle or this shuffle mask, do
12800 // the replacement.
12801 if (NewMask == LHSMask || NewMask == Mask) {
12802 unsigned LHSInNElts =
12803 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12804 std::vector<Constant*> Elts;
12805 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12806 if (NewMask[i] >= LHSInNElts*2) {
12807 Elts.push_back(Context->getUndef(Type::Int32Ty));
12809 Elts.push_back(Context->getConstantInt(Type::Int32Ty, NewMask[i]));
12812 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12813 LHSSVI->getOperand(1),
12814 Context->getConstantVector(Elts));
12819 return MadeChange ? &SVI : 0;
12825 /// TryToSinkInstruction - Try to move the specified instruction from its
12826 /// current block into the beginning of DestBlock, which can only happen if it's
12827 /// safe to move the instruction past all of the instructions between it and the
12828 /// end of its block.
12829 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12830 assert(I->hasOneUse() && "Invariants didn't hold!");
12832 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12833 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12836 // Do not sink alloca instructions out of the entry block.
12837 if (isa<AllocaInst>(I) && I->getParent() ==
12838 &DestBlock->getParent()->getEntryBlock())
12841 // We can only sink load instructions if there is nothing between the load and
12842 // the end of block that could change the value.
12843 if (I->mayReadFromMemory()) {
12844 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12846 if (Scan->mayWriteToMemory())
12850 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12852 CopyPrecedingStopPoint(I, InsertPos);
12853 I->moveBefore(InsertPos);
12859 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12860 /// all reachable code to the worklist.
12862 /// This has a couple of tricks to make the code faster and more powerful. In
12863 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12864 /// them to the worklist (this significantly speeds up instcombine on code where
12865 /// many instructions are dead or constant). Additionally, if we find a branch
12866 /// whose condition is a known constant, we only visit the reachable successors.
12868 static void AddReachableCodeToWorklist(BasicBlock *BB,
12869 SmallPtrSet<BasicBlock*, 64> &Visited,
12871 const TargetData *TD) {
12872 SmallVector<BasicBlock*, 256> Worklist;
12873 Worklist.push_back(BB);
12875 while (!Worklist.empty()) {
12876 BB = Worklist.back();
12877 Worklist.pop_back();
12879 // We have now visited this block! If we've already been here, ignore it.
12880 if (!Visited.insert(BB)) continue;
12882 DbgInfoIntrinsic *DBI_Prev = NULL;
12883 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12884 Instruction *Inst = BBI++;
12886 // DCE instruction if trivially dead.
12887 if (isInstructionTriviallyDead(Inst)) {
12889 DOUT << "IC: DCE: " << *Inst;
12890 Inst->eraseFromParent();
12894 // ConstantProp instruction if trivially constant.
12895 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12896 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12897 Inst->replaceAllUsesWith(C);
12899 Inst->eraseFromParent();
12903 // If there are two consecutive llvm.dbg.stoppoint calls then
12904 // it is likely that the optimizer deleted code in between these
12906 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12909 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12910 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12911 IC.RemoveFromWorkList(DBI_Prev);
12912 DBI_Prev->eraseFromParent();
12914 DBI_Prev = DBI_Next;
12919 IC.AddToWorkList(Inst);
12922 // Recursively visit successors. If this is a branch or switch on a
12923 // constant, only visit the reachable successor.
12924 TerminatorInst *TI = BB->getTerminator();
12925 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12926 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12927 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12928 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12929 Worklist.push_back(ReachableBB);
12932 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12933 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12934 // See if this is an explicit destination.
12935 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12936 if (SI->getCaseValue(i) == Cond) {
12937 BasicBlock *ReachableBB = SI->getSuccessor(i);
12938 Worklist.push_back(ReachableBB);
12942 // Otherwise it is the default destination.
12943 Worklist.push_back(SI->getSuccessor(0));
12948 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12949 Worklist.push_back(TI->getSuccessor(i));
12953 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12954 bool Changed = false;
12955 TD = &getAnalysis<TargetData>();
12957 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12958 << F.getNameStr() << "\n");
12961 // Do a depth-first traversal of the function, populate the worklist with
12962 // the reachable instructions. Ignore blocks that are not reachable. Keep
12963 // track of which blocks we visit.
12964 SmallPtrSet<BasicBlock*, 64> Visited;
12965 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12967 // Do a quick scan over the function. If we find any blocks that are
12968 // unreachable, remove any instructions inside of them. This prevents
12969 // the instcombine code from having to deal with some bad special cases.
12970 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12971 if (!Visited.count(BB)) {
12972 Instruction *Term = BB->getTerminator();
12973 while (Term != BB->begin()) { // Remove instrs bottom-up
12974 BasicBlock::iterator I = Term; --I;
12976 DOUT << "IC: DCE: " << *I;
12977 // A debug intrinsic shouldn't force another iteration if we weren't
12978 // going to do one without it.
12979 if (!isa<DbgInfoIntrinsic>(I)) {
12983 if (!I->use_empty())
12984 I->replaceAllUsesWith(Context->getUndef(I->getType()));
12985 I->eraseFromParent();
12990 while (!Worklist.empty()) {
12991 Instruction *I = RemoveOneFromWorkList();
12992 if (I == 0) continue; // skip null values.
12994 // Check to see if we can DCE the instruction.
12995 if (isInstructionTriviallyDead(I)) {
12996 // Add operands to the worklist.
12997 if (I->getNumOperands() < 4)
12998 AddUsesToWorkList(*I);
13001 DOUT << "IC: DCE: " << *I;
13003 I->eraseFromParent();
13004 RemoveFromWorkList(I);
13009 // Instruction isn't dead, see if we can constant propagate it.
13010 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
13011 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
13013 // Add operands to the worklist.
13014 AddUsesToWorkList(*I);
13015 ReplaceInstUsesWith(*I, C);
13018 I->eraseFromParent();
13019 RemoveFromWorkList(I);
13025 // See if we can constant fold its operands.
13026 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
13027 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
13028 if (Constant *NewC = ConstantFoldConstantExpression(CE,
13029 F.getContext(), TD))
13036 // See if we can trivially sink this instruction to a successor basic block.
13037 if (I->hasOneUse()) {
13038 BasicBlock *BB = I->getParent();
13039 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
13040 if (UserParent != BB) {
13041 bool UserIsSuccessor = false;
13042 // See if the user is one of our successors.
13043 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13044 if (*SI == UserParent) {
13045 UserIsSuccessor = true;
13049 // If the user is one of our immediate successors, and if that successor
13050 // only has us as a predecessors (we'd have to split the critical edge
13051 // otherwise), we can keep going.
13052 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
13053 next(pred_begin(UserParent)) == pred_end(UserParent))
13054 // Okay, the CFG is simple enough, try to sink this instruction.
13055 Changed |= TryToSinkInstruction(I, UserParent);
13059 // Now that we have an instruction, try combining it to simplify it...
13063 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
13064 if (Instruction *Result = visit(*I)) {
13066 // Should we replace the old instruction with a new one?
13068 DOUT << "IC: Old = " << *I
13069 << " New = " << *Result;
13071 // Everything uses the new instruction now.
13072 I->replaceAllUsesWith(Result);
13074 // Push the new instruction and any users onto the worklist.
13075 AddToWorkList(Result);
13076 AddUsersToWorkList(*Result);
13078 // Move the name to the new instruction first.
13079 Result->takeName(I);
13081 // Insert the new instruction into the basic block...
13082 BasicBlock *InstParent = I->getParent();
13083 BasicBlock::iterator InsertPos = I;
13085 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13086 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13089 InstParent->getInstList().insert(InsertPos, Result);
13091 // Make sure that we reprocess all operands now that we reduced their
13093 AddUsesToWorkList(*I);
13095 // Instructions can end up on the worklist more than once. Make sure
13096 // we do not process an instruction that has been deleted.
13097 RemoveFromWorkList(I);
13099 // Erase the old instruction.
13100 InstParent->getInstList().erase(I);
13103 DOUT << "IC: Mod = " << OrigI
13104 << " New = " << *I;
13107 // If the instruction was modified, it's possible that it is now dead.
13108 // if so, remove it.
13109 if (isInstructionTriviallyDead(I)) {
13110 // Make sure we process all operands now that we are reducing their
13112 AddUsesToWorkList(*I);
13114 // Instructions may end up in the worklist more than once. Erase all
13115 // occurrences of this instruction.
13116 RemoveFromWorkList(I);
13117 I->eraseFromParent();
13120 AddUsersToWorkList(*I);
13127 assert(WorklistMap.empty() && "Worklist empty, but map not?");
13129 // Do an explicit clear, this shrinks the map if needed.
13130 WorklistMap.clear();
13135 bool InstCombiner::runOnFunction(Function &F) {
13136 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13138 bool EverMadeChange = false;
13140 // Iterate while there is work to do.
13141 unsigned Iteration = 0;
13142 while (DoOneIteration(F, Iteration++))
13143 EverMadeChange = true;
13144 return EverMadeChange;
13147 FunctionPass *llvm::createInstructionCombiningPass() {
13148 return new InstCombiner();