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 (Operator *O = dyn_cast<Operator>(V)) {
445 if (O->getOpcode() == Instruction::BitCast)
446 return O->getOperand(0);
447 if (GEPOperator *GEP = dyn_cast<GEPOperator>(V))
448 if (GEP->hasAllZeroIndices())
449 return GEP->getPointerOperand();
454 /// This function is a wrapper around CastInst::isEliminableCastPair. It
455 /// simply extracts arguments and returns what that function returns.
456 static Instruction::CastOps
457 isEliminableCastPair(
458 const CastInst *CI, ///< The first cast instruction
459 unsigned opcode, ///< The opcode of the second cast instruction
460 const Type *DstTy, ///< The target type for the second cast instruction
461 TargetData *TD ///< The target data for pointer size
464 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
465 const Type *MidTy = CI->getType(); // B from above
467 // Get the opcodes of the two Cast instructions
468 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
469 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
471 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
472 DstTy, TD->getIntPtrType());
474 // We don't want to form an inttoptr or ptrtoint that converts to an integer
475 // type that differs from the pointer size.
476 if ((Res == Instruction::IntToPtr && SrcTy != TD->getIntPtrType()) ||
477 (Res == Instruction::PtrToInt && DstTy != TD->getIntPtrType()))
480 return Instruction::CastOps(Res);
483 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
484 /// in any code being generated. It does not require codegen if V is simple
485 /// enough or if the cast can be folded into other casts.
486 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
487 const Type *Ty, TargetData *TD) {
488 if (V->getType() == Ty || isa<Constant>(V)) return false;
490 // If this is another cast that can be eliminated, it isn't codegen either.
491 if (const CastInst *CI = dyn_cast<CastInst>(V))
492 if (isEliminableCastPair(CI, opcode, Ty, TD))
497 // SimplifyCommutative - This performs a few simplifications for commutative
500 // 1. Order operands such that they are listed from right (least complex) to
501 // left (most complex). This puts constants before unary operators before
504 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
505 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
507 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
508 bool Changed = false;
509 if (getComplexity(Context, I.getOperand(0)) <
510 getComplexity(Context, I.getOperand(1)))
511 Changed = !I.swapOperands();
513 if (!I.isAssociative()) return Changed;
514 Instruction::BinaryOps Opcode = I.getOpcode();
515 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
516 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
517 if (isa<Constant>(I.getOperand(1))) {
518 Constant *Folded = Context->getConstantExpr(I.getOpcode(),
519 cast<Constant>(I.getOperand(1)),
520 cast<Constant>(Op->getOperand(1)));
521 I.setOperand(0, Op->getOperand(0));
522 I.setOperand(1, Folded);
524 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
525 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
526 isOnlyUse(Op) && isOnlyUse(Op1)) {
527 Constant *C1 = cast<Constant>(Op->getOperand(1));
528 Constant *C2 = cast<Constant>(Op1->getOperand(1));
530 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
531 Constant *Folded = Context->getConstantExpr(I.getOpcode(), C1, C2);
532 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
536 I.setOperand(0, New);
537 I.setOperand(1, Folded);
544 /// SimplifyCompare - For a CmpInst this function just orders the operands
545 /// so that theyare listed from right (least complex) to left (most complex).
546 /// This puts constants before unary operators before binary operators.
547 bool InstCombiner::SimplifyCompare(CmpInst &I) {
548 if (getComplexity(Context, I.getOperand(0)) >=
549 getComplexity(Context, I.getOperand(1)))
552 // Compare instructions are not associative so there's nothing else we can do.
556 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
557 // if the LHS is a constant zero (which is the 'negate' form).
559 static inline Value *dyn_castNegVal(Value *V, LLVMContext *Context) {
560 if (BinaryOperator::isNeg(V))
561 return BinaryOperator::getNegArgument(V);
563 // Constants can be considered to be negated values if they can be folded.
564 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
565 return Context->getConstantExprNeg(C);
567 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
568 if (C->getType()->getElementType()->isInteger())
569 return Context->getConstantExprNeg(C);
574 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
575 // instruction if the LHS is a constant negative zero (which is the 'negate'
578 static inline Value *dyn_castFNegVal(Value *V, LLVMContext *Context) {
579 if (BinaryOperator::isFNeg(V))
580 return BinaryOperator::getFNegArgument(V);
582 // Constants can be considered to be negated values if they can be folded.
583 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
584 return Context->getConstantExprFNeg(C);
586 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
587 if (C->getType()->getElementType()->isFloatingPoint())
588 return Context->getConstantExprFNeg(C);
593 static inline Value *dyn_castNotVal(Value *V, LLVMContext *Context) {
594 if (BinaryOperator::isNot(V))
595 return BinaryOperator::getNotArgument(V);
597 // Constants can be considered to be not'ed values...
598 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
599 return Context->getConstantInt(~C->getValue());
603 // dyn_castFoldableMul - If this value is a multiply that can be folded into
604 // other computations (because it has a constant operand), return the
605 // non-constant operand of the multiply, and set CST to point to the multiplier.
606 // Otherwise, return null.
608 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST,
609 LLVMContext *Context) {
610 if (V->hasOneUse() && V->getType()->isInteger())
611 if (Instruction *I = dyn_cast<Instruction>(V)) {
612 if (I->getOpcode() == Instruction::Mul)
613 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
614 return I->getOperand(0);
615 if (I->getOpcode() == Instruction::Shl)
616 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
617 // The multiplier is really 1 << CST.
618 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
619 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
620 CST = Context->getConstantInt(APInt(BitWidth, 1).shl(CSTVal));
621 return I->getOperand(0);
627 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
628 /// expression, return it.
629 static User *dyn_castGetElementPtr(Value *V) {
630 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
631 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
632 if (CE->getOpcode() == Instruction::GetElementPtr)
633 return cast<User>(V);
637 /// AddOne - Add one to a ConstantInt
638 static Constant *AddOne(Constant *C, LLVMContext *Context) {
639 return Context->getConstantExprAdd(C,
640 Context->getConstantInt(C->getType(), 1));
642 /// SubOne - Subtract one from a ConstantInt
643 static Constant *SubOne(ConstantInt *C, LLVMContext *Context) {
644 return Context->getConstantExprSub(C,
645 Context->getConstantInt(C->getType(), 1));
647 /// MultiplyOverflows - True if the multiply can not be expressed in an int
649 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign,
650 LLVMContext *Context) {
651 uint32_t W = C1->getBitWidth();
652 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
661 APInt MulExt = LHSExt * RHSExt;
664 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
665 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
666 return MulExt.slt(Min) || MulExt.sgt(Max);
668 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
672 /// ShrinkDemandedConstant - Check to see if the specified operand of the
673 /// specified instruction is a constant integer. If so, check to see if there
674 /// are any bits set in the constant that are not demanded. If so, shrink the
675 /// constant and return true.
676 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
677 APInt Demanded, LLVMContext *Context) {
678 assert(I && "No instruction?");
679 assert(OpNo < I->getNumOperands() && "Operand index too large");
681 // If the operand is not a constant integer, nothing to do.
682 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
683 if (!OpC) return false;
685 // If there are no bits set that aren't demanded, nothing to do.
686 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
687 if ((~Demanded & OpC->getValue()) == 0)
690 // This instruction is producing bits that are not demanded. Shrink the RHS.
691 Demanded &= OpC->getValue();
692 I->setOperand(OpNo, Context->getConstantInt(Demanded));
696 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
697 // set of known zero and one bits, compute the maximum and minimum values that
698 // could have the specified known zero and known one bits, returning them in
700 static void ComputeSignedMinMaxValuesFromKnownBits(const APInt& KnownZero,
701 const APInt& KnownOne,
702 APInt& Min, APInt& Max) {
703 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
704 KnownZero.getBitWidth() == Min.getBitWidth() &&
705 KnownZero.getBitWidth() == Max.getBitWidth() &&
706 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
707 APInt UnknownBits = ~(KnownZero|KnownOne);
709 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
710 // bit if it is unknown.
712 Max = KnownOne|UnknownBits;
714 if (UnknownBits.isNegative()) { // Sign bit is unknown
715 Min.set(Min.getBitWidth()-1);
716 Max.clear(Max.getBitWidth()-1);
720 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
721 // a set of known zero and one bits, compute the maximum and minimum values that
722 // could have the specified known zero and known one bits, returning them in
724 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
725 const APInt &KnownOne,
726 APInt &Min, APInt &Max) {
727 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
728 KnownZero.getBitWidth() == Min.getBitWidth() &&
729 KnownZero.getBitWidth() == Max.getBitWidth() &&
730 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
731 APInt UnknownBits = ~(KnownZero|KnownOne);
733 // The minimum value is when the unknown bits are all zeros.
735 // The maximum value is when the unknown bits are all ones.
736 Max = KnownOne|UnknownBits;
739 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
740 /// SimplifyDemandedBits knows about. See if the instruction has any
741 /// properties that allow us to simplify its operands.
742 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
743 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
744 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
745 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
747 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
748 KnownZero, KnownOne, 0);
749 if (V == 0) return false;
750 if (V == &Inst) return true;
751 ReplaceInstUsesWith(Inst, V);
755 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
756 /// specified instruction operand if possible, updating it in place. It returns
757 /// true if it made any change and false otherwise.
758 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
759 APInt &KnownZero, APInt &KnownOne,
761 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
762 KnownZero, KnownOne, Depth);
763 if (NewVal == 0) return false;
769 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
770 /// value based on the demanded bits. When this function is called, it is known
771 /// that only the bits set in DemandedMask of the result of V are ever used
772 /// downstream. Consequently, depending on the mask and V, it may be possible
773 /// to replace V with a constant or one of its operands. In such cases, this
774 /// function does the replacement and returns true. In all other cases, it
775 /// returns false after analyzing the expression and setting KnownOne and known
776 /// to be one in the expression. KnownZero contains all the bits that are known
777 /// to be zero in the expression. These are provided to potentially allow the
778 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
779 /// the expression. KnownOne and KnownZero always follow the invariant that
780 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
781 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
782 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
783 /// and KnownOne must all be the same.
785 /// This returns null if it did not change anything and it permits no
786 /// simplification. This returns V itself if it did some simplification of V's
787 /// operands based on the information about what bits are demanded. This returns
788 /// some other non-null value if it found out that V is equal to another value
789 /// in the context where the specified bits are demanded, but not for all users.
790 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
791 APInt &KnownZero, APInt &KnownOne,
793 assert(V != 0 && "Null pointer of Value???");
794 assert(Depth <= 6 && "Limit Search Depth");
795 uint32_t BitWidth = DemandedMask.getBitWidth();
796 const Type *VTy = V->getType();
797 assert((TD || !isa<PointerType>(VTy)) &&
798 "SimplifyDemandedBits needs to know bit widths!");
799 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
800 (!VTy->isIntOrIntVector() ||
801 VTy->getScalarSizeInBits() == BitWidth) &&
802 KnownZero.getBitWidth() == BitWidth &&
803 KnownOne.getBitWidth() == BitWidth &&
804 "Value *V, DemandedMask, KnownZero and KnownOne "
805 "must have same BitWidth");
806 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
807 // We know all of the bits for a constant!
808 KnownOne = CI->getValue() & DemandedMask;
809 KnownZero = ~KnownOne & DemandedMask;
812 if (isa<ConstantPointerNull>(V)) {
813 // We know all of the bits for a constant!
815 KnownZero = DemandedMask;
821 if (DemandedMask == 0) { // Not demanding any bits from V.
822 if (isa<UndefValue>(V))
824 return Context->getUndef(VTy);
827 if (Depth == 6) // Limit search depth.
830 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
831 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
833 Instruction *I = dyn_cast<Instruction>(V);
835 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
836 return 0; // Only analyze instructions.
839 // If there are multiple uses of this value and we aren't at the root, then
840 // we can't do any simplifications of the operands, because DemandedMask
841 // only reflects the bits demanded by *one* of the users.
842 if (Depth != 0 && !I->hasOneUse()) {
843 // Despite the fact that we can't simplify this instruction in all User's
844 // context, we can at least compute the knownzero/knownone bits, and we can
845 // do simplifications that apply to *just* the one user if we know that
846 // this instruction has a simpler value in that context.
847 if (I->getOpcode() == Instruction::And) {
848 // If either the LHS or the RHS are Zero, the result is zero.
849 ComputeMaskedBits(I->getOperand(1), DemandedMask,
850 RHSKnownZero, RHSKnownOne, Depth+1);
851 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
852 LHSKnownZero, LHSKnownOne, Depth+1);
854 // If all of the demanded bits are known 1 on one side, return the other.
855 // These bits cannot contribute to the result of the 'and' in this
857 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
858 (DemandedMask & ~LHSKnownZero))
859 return I->getOperand(0);
860 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
861 (DemandedMask & ~RHSKnownZero))
862 return I->getOperand(1);
864 // If all of the demanded bits in the inputs are known zeros, return zero.
865 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
866 return Context->getNullValue(VTy);
868 } else if (I->getOpcode() == Instruction::Or) {
869 // We can simplify (X|Y) -> X or Y in the user's context if we know that
870 // only bits from X or Y are demanded.
872 // If either the LHS or the RHS are One, the result is One.
873 ComputeMaskedBits(I->getOperand(1), DemandedMask,
874 RHSKnownZero, RHSKnownOne, Depth+1);
875 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
876 LHSKnownZero, LHSKnownOne, Depth+1);
878 // If all of the demanded bits are known zero on one side, return the
879 // other. These bits cannot contribute to the result of the 'or' in this
881 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
882 (DemandedMask & ~LHSKnownOne))
883 return I->getOperand(0);
884 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
885 (DemandedMask & ~RHSKnownOne))
886 return I->getOperand(1);
888 // If all of the potentially set bits on one side are known to be set on
889 // the other side, just use the 'other' side.
890 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
891 (DemandedMask & (~RHSKnownZero)))
892 return I->getOperand(0);
893 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
894 (DemandedMask & (~LHSKnownZero)))
895 return I->getOperand(1);
898 // Compute the KnownZero/KnownOne bits to simplify things downstream.
899 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
903 // If this is the root being simplified, allow it to have multiple uses,
904 // just set the DemandedMask to all bits so that we can try to simplify the
905 // operands. This allows visitTruncInst (for example) to simplify the
906 // operand of a trunc without duplicating all the logic below.
907 if (Depth == 0 && !V->hasOneUse())
908 DemandedMask = APInt::getAllOnesValue(BitWidth);
910 switch (I->getOpcode()) {
912 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
914 case Instruction::And:
915 // If either the LHS or the RHS are Zero, the result is zero.
916 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
917 RHSKnownZero, RHSKnownOne, Depth+1) ||
918 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
919 LHSKnownZero, LHSKnownOne, Depth+1))
921 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
922 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
924 // If all of the demanded bits are known 1 on one side, return the other.
925 // These bits cannot contribute to the result of the 'and'.
926 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
927 (DemandedMask & ~LHSKnownZero))
928 return I->getOperand(0);
929 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
930 (DemandedMask & ~RHSKnownZero))
931 return I->getOperand(1);
933 // If all of the demanded bits in the inputs are known zeros, return zero.
934 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
935 return Context->getNullValue(VTy);
937 // If the RHS is a constant, see if we can simplify it.
938 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero, Context))
941 // Output known-1 bits are only known if set in both the LHS & RHS.
942 RHSKnownOne &= LHSKnownOne;
943 // Output known-0 are known to be clear if zero in either the LHS | RHS.
944 RHSKnownZero |= LHSKnownZero;
946 case Instruction::Or:
947 // If either the LHS or the RHS are One, the result is One.
948 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
949 RHSKnownZero, RHSKnownOne, Depth+1) ||
950 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
951 LHSKnownZero, LHSKnownOne, Depth+1))
953 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
954 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
956 // If all of the demanded bits are known zero on one side, return the other.
957 // These bits cannot contribute to the result of the 'or'.
958 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
959 (DemandedMask & ~LHSKnownOne))
960 return I->getOperand(0);
961 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
962 (DemandedMask & ~RHSKnownOne))
963 return I->getOperand(1);
965 // If all of the potentially set bits on one side are known to be set on
966 // the other side, just use the 'other' side.
967 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
968 (DemandedMask & (~RHSKnownZero)))
969 return I->getOperand(0);
970 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
971 (DemandedMask & (~LHSKnownZero)))
972 return I->getOperand(1);
974 // If the RHS is a constant, see if we can simplify it.
975 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context))
978 // Output known-0 bits are only known if clear in both the LHS & RHS.
979 RHSKnownZero &= LHSKnownZero;
980 // Output known-1 are known to be set if set in either the LHS | RHS.
981 RHSKnownOne |= LHSKnownOne;
983 case Instruction::Xor: {
984 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
985 RHSKnownZero, RHSKnownOne, Depth+1) ||
986 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
987 LHSKnownZero, LHSKnownOne, Depth+1))
989 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
990 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
992 // If all of the demanded bits are known zero on one side, return the other.
993 // These bits cannot contribute to the result of the 'xor'.
994 if ((DemandedMask & RHSKnownZero) == DemandedMask)
995 return I->getOperand(0);
996 if ((DemandedMask & LHSKnownZero) == DemandedMask)
997 return I->getOperand(1);
999 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1000 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1001 (RHSKnownOne & LHSKnownOne);
1002 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1003 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1004 (RHSKnownOne & LHSKnownZero);
1006 // If all of the demanded bits are known to be zero on one side or the
1007 // other, turn this into an *inclusive* or.
1008 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1009 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1011 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1013 return InsertNewInstBefore(Or, *I);
1016 // If all of the demanded bits on one side are known, and all of the set
1017 // bits on that side are also known to be set on the other side, turn this
1018 // into an AND, as we know the bits will be cleared.
1019 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1020 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1022 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1023 Constant *AndC = Context->getConstantInt(~RHSKnownOne & DemandedMask);
1025 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1026 return InsertNewInstBefore(And, *I);
1030 // If the RHS is a constant, see if we can simplify it.
1031 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1032 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context))
1035 RHSKnownZero = KnownZeroOut;
1036 RHSKnownOne = KnownOneOut;
1039 case Instruction::Select:
1040 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1041 RHSKnownZero, RHSKnownOne, Depth+1) ||
1042 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1043 LHSKnownZero, LHSKnownOne, Depth+1))
1045 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1046 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1048 // If the operands are constants, see if we can simplify them.
1049 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context) ||
1050 ShrinkDemandedConstant(I, 2, DemandedMask, Context))
1053 // Only known if known in both the LHS and RHS.
1054 RHSKnownOne &= LHSKnownOne;
1055 RHSKnownZero &= LHSKnownZero;
1057 case Instruction::Trunc: {
1058 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1059 DemandedMask.zext(truncBf);
1060 RHSKnownZero.zext(truncBf);
1061 RHSKnownOne.zext(truncBf);
1062 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1063 RHSKnownZero, RHSKnownOne, Depth+1))
1065 DemandedMask.trunc(BitWidth);
1066 RHSKnownZero.trunc(BitWidth);
1067 RHSKnownOne.trunc(BitWidth);
1068 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1071 case Instruction::BitCast:
1072 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1073 return false; // vector->int or fp->int?
1075 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1076 if (const VectorType *SrcVTy =
1077 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1078 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1079 // Don't touch a bitcast between vectors of different element counts.
1082 // Don't touch a scalar-to-vector bitcast.
1084 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1085 // Don't touch a vector-to-scalar bitcast.
1088 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1089 RHSKnownZero, RHSKnownOne, Depth+1))
1091 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1093 case Instruction::ZExt: {
1094 // Compute the bits in the result that are not present in the input.
1095 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1097 DemandedMask.trunc(SrcBitWidth);
1098 RHSKnownZero.trunc(SrcBitWidth);
1099 RHSKnownOne.trunc(SrcBitWidth);
1100 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1101 RHSKnownZero, RHSKnownOne, Depth+1))
1103 DemandedMask.zext(BitWidth);
1104 RHSKnownZero.zext(BitWidth);
1105 RHSKnownOne.zext(BitWidth);
1106 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1107 // The top bits are known to be zero.
1108 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1111 case Instruction::SExt: {
1112 // Compute the bits in the result that are not present in the input.
1113 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1115 APInt InputDemandedBits = DemandedMask &
1116 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1118 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1119 // If any of the sign extended bits are demanded, we know that the sign
1121 if ((NewBits & DemandedMask) != 0)
1122 InputDemandedBits.set(SrcBitWidth-1);
1124 InputDemandedBits.trunc(SrcBitWidth);
1125 RHSKnownZero.trunc(SrcBitWidth);
1126 RHSKnownOne.trunc(SrcBitWidth);
1127 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1128 RHSKnownZero, RHSKnownOne, Depth+1))
1130 InputDemandedBits.zext(BitWidth);
1131 RHSKnownZero.zext(BitWidth);
1132 RHSKnownOne.zext(BitWidth);
1133 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1135 // If the sign bit of the input is known set or clear, then we know the
1136 // top bits of the result.
1138 // If the input sign bit is known zero, or if the NewBits are not demanded
1139 // convert this into a zero extension.
1140 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1141 // Convert to ZExt cast
1142 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1143 return InsertNewInstBefore(NewCast, *I);
1144 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1145 RHSKnownOne |= NewBits;
1149 case Instruction::Add: {
1150 // Figure out what the input bits are. If the top bits of the and result
1151 // are not demanded, then the add doesn't demand them from its input
1153 unsigned NLZ = DemandedMask.countLeadingZeros();
1155 // If there is a constant on the RHS, there are a variety of xformations
1157 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1158 // If null, this should be simplified elsewhere. Some of the xforms here
1159 // won't work if the RHS is zero.
1163 // If the top bit of the output is demanded, demand everything from the
1164 // input. Otherwise, we demand all the input bits except NLZ top bits.
1165 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1167 // Find information about known zero/one bits in the input.
1168 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1169 LHSKnownZero, LHSKnownOne, Depth+1))
1172 // If the RHS of the add has bits set that can't affect the input, reduce
1174 if (ShrinkDemandedConstant(I, 1, InDemandedBits, Context))
1177 // Avoid excess work.
1178 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1181 // Turn it into OR if input bits are zero.
1182 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1184 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1186 return InsertNewInstBefore(Or, *I);
1189 // We can say something about the output known-zero and known-one bits,
1190 // depending on potential carries from the input constant and the
1191 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1192 // bits set and the RHS constant is 0x01001, then we know we have a known
1193 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1195 // To compute this, we first compute the potential carry bits. These are
1196 // the bits which may be modified. I'm not aware of a better way to do
1198 const APInt &RHSVal = RHS->getValue();
1199 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1201 // Now that we know which bits have carries, compute the known-1/0 sets.
1203 // Bits are known one if they are known zero in one operand and one in the
1204 // other, and there is no input carry.
1205 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1206 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1208 // Bits are known zero if they are known zero in both operands and there
1209 // is no input carry.
1210 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1212 // If the high-bits of this ADD are not demanded, then it does not demand
1213 // the high bits of its LHS or RHS.
1214 if (DemandedMask[BitWidth-1] == 0) {
1215 // Right fill the mask of bits for this ADD to demand the most
1216 // significant bit and all those below it.
1217 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1218 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1219 LHSKnownZero, LHSKnownOne, Depth+1) ||
1220 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1221 LHSKnownZero, LHSKnownOne, Depth+1))
1227 case Instruction::Sub:
1228 // If the high-bits of this SUB are not demanded, then it does not demand
1229 // the high bits of its LHS or RHS.
1230 if (DemandedMask[BitWidth-1] == 0) {
1231 // Right fill the mask of bits for this SUB to demand the most
1232 // significant bit and all those below it.
1233 uint32_t NLZ = DemandedMask.countLeadingZeros();
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))
1241 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1242 // the known zeros and ones.
1243 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1245 case Instruction::Shl:
1246 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1247 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1248 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1249 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1250 RHSKnownZero, RHSKnownOne, Depth+1))
1252 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1253 RHSKnownZero <<= ShiftAmt;
1254 RHSKnownOne <<= ShiftAmt;
1255 // low bits known zero.
1257 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1260 case Instruction::LShr:
1261 // For a logical shift right
1262 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1263 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1265 // Unsigned shift right.
1266 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1267 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1268 RHSKnownZero, RHSKnownOne, Depth+1))
1270 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1271 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1272 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1274 // Compute the new bits that are at the top now.
1275 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1276 RHSKnownZero |= HighBits; // high bits known zero.
1280 case Instruction::AShr:
1281 // If this is an arithmetic shift right and only the low-bit is set, we can
1282 // always convert this into a logical shr, even if the shift amount is
1283 // variable. The low bit of the shift cannot be an input sign bit unless
1284 // the shift amount is >= the size of the datatype, which is undefined.
1285 if (DemandedMask == 1) {
1286 // Perform the logical shift right.
1287 Instruction *NewVal = BinaryOperator::CreateLShr(
1288 I->getOperand(0), I->getOperand(1), I->getName());
1289 return InsertNewInstBefore(NewVal, *I);
1292 // If the sign bit is the only bit demanded by this ashr, then there is no
1293 // need to do it, the shift doesn't change the high bit.
1294 if (DemandedMask.isSignBit())
1295 return I->getOperand(0);
1297 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1298 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1300 // Signed shift right.
1301 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1302 // If any of the "high bits" are demanded, we should set the sign bit as
1304 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1305 DemandedMaskIn.set(BitWidth-1);
1306 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1307 RHSKnownZero, RHSKnownOne, Depth+1))
1309 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1310 // Compute the new bits that are at the top now.
1311 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1312 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1313 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1315 // Handle the sign bits.
1316 APInt SignBit(APInt::getSignBit(BitWidth));
1317 // Adjust to where it is now in the mask.
1318 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1320 // If the input sign bit is known to be zero, or if none of the top bits
1321 // are demanded, turn this into an unsigned shift right.
1322 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1323 (HighBits & ~DemandedMask) == HighBits) {
1324 // Perform the logical shift right.
1325 Instruction *NewVal = BinaryOperator::CreateLShr(
1326 I->getOperand(0), SA, I->getName());
1327 return InsertNewInstBefore(NewVal, *I);
1328 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1329 RHSKnownOne |= HighBits;
1333 case Instruction::SRem:
1334 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1335 APInt RA = Rem->getValue().abs();
1336 if (RA.isPowerOf2()) {
1337 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1338 return I->getOperand(0);
1340 APInt LowBits = RA - 1;
1341 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1342 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1343 LHSKnownZero, LHSKnownOne, Depth+1))
1346 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1347 LHSKnownZero |= ~LowBits;
1349 KnownZero |= LHSKnownZero & DemandedMask;
1351 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1355 case Instruction::URem: {
1356 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1357 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1358 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1359 KnownZero2, KnownOne2, Depth+1) ||
1360 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1361 KnownZero2, KnownOne2, Depth+1))
1364 unsigned Leaders = KnownZero2.countLeadingOnes();
1365 Leaders = std::max(Leaders,
1366 KnownZero2.countLeadingOnes());
1367 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1370 case Instruction::Call:
1371 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1372 switch (II->getIntrinsicID()) {
1374 case Intrinsic::bswap: {
1375 // If the only bits demanded come from one byte of the bswap result,
1376 // just shift the input byte into position to eliminate the bswap.
1377 unsigned NLZ = DemandedMask.countLeadingZeros();
1378 unsigned NTZ = DemandedMask.countTrailingZeros();
1380 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1381 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1382 // have 14 leading zeros, round to 8.
1385 // If we need exactly one byte, we can do this transformation.
1386 if (BitWidth-NLZ-NTZ == 8) {
1387 unsigned ResultBit = NTZ;
1388 unsigned InputBit = BitWidth-NTZ-8;
1390 // Replace this with either a left or right shift to get the byte into
1392 Instruction *NewVal;
1393 if (InputBit > ResultBit)
1394 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1395 Context->getConstantInt(I->getType(), InputBit-ResultBit));
1397 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1398 Context->getConstantInt(I->getType(), ResultBit-InputBit));
1399 NewVal->takeName(I);
1400 return InsertNewInstBefore(NewVal, *I);
1403 // TODO: Could compute known zero/one bits based on the input.
1408 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1412 // If the client is only demanding bits that we know, return the known
1414 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1415 Constant *C = Context->getConstantInt(RHSKnownOne);
1416 if (isa<PointerType>(V->getType()))
1417 C = Context->getConstantExprIntToPtr(C, V->getType());
1424 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1425 /// any number of elements. DemandedElts contains the set of elements that are
1426 /// actually used by the caller. This method analyzes which elements of the
1427 /// operand are undef and returns that information in UndefElts.
1429 /// If the information about demanded elements can be used to simplify the
1430 /// operation, the operation is simplified, then the resultant value is
1431 /// returned. This returns null if no change was made.
1432 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1435 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1436 APInt EltMask(APInt::getAllOnesValue(VWidth));
1437 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1439 if (isa<UndefValue>(V)) {
1440 // If the entire vector is undefined, just return this info.
1441 UndefElts = EltMask;
1443 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1444 UndefElts = EltMask;
1445 return Context->getUndef(V->getType());
1449 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1450 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1451 Constant *Undef = Context->getUndef(EltTy);
1453 std::vector<Constant*> Elts;
1454 for (unsigned i = 0; i != VWidth; ++i)
1455 if (!DemandedElts[i]) { // If not demanded, set to undef.
1456 Elts.push_back(Undef);
1458 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1459 Elts.push_back(Undef);
1461 } else { // Otherwise, defined.
1462 Elts.push_back(CP->getOperand(i));
1465 // If we changed the constant, return it.
1466 Constant *NewCP = Context->getConstantVector(Elts);
1467 return NewCP != CP ? NewCP : 0;
1468 } else if (isa<ConstantAggregateZero>(V)) {
1469 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1472 // Check if this is identity. If so, return 0 since we are not simplifying
1474 if (DemandedElts == ((1ULL << VWidth) -1))
1477 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1478 Constant *Zero = Context->getNullValue(EltTy);
1479 Constant *Undef = Context->getUndef(EltTy);
1480 std::vector<Constant*> Elts;
1481 for (unsigned i = 0; i != VWidth; ++i) {
1482 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1483 Elts.push_back(Elt);
1485 UndefElts = DemandedElts ^ EltMask;
1486 return Context->getConstantVector(Elts);
1489 // Limit search depth.
1493 // If multiple users are using the root value, procede with
1494 // simplification conservatively assuming that all elements
1496 if (!V->hasOneUse()) {
1497 // Quit if we find multiple users of a non-root value though.
1498 // They'll be handled when it's their turn to be visited by
1499 // the main instcombine process.
1501 // TODO: Just compute the UndefElts information recursively.
1504 // Conservatively assume that all elements are needed.
1505 DemandedElts = EltMask;
1508 Instruction *I = dyn_cast<Instruction>(V);
1509 if (!I) return 0; // Only analyze instructions.
1511 bool MadeChange = false;
1512 APInt UndefElts2(VWidth, 0);
1514 switch (I->getOpcode()) {
1517 case Instruction::InsertElement: {
1518 // If this is a variable index, we don't know which element it overwrites.
1519 // demand exactly the same input as we produce.
1520 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1522 // Note that we can't propagate undef elt info, because we don't know
1523 // which elt is getting updated.
1524 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1525 UndefElts2, Depth+1);
1526 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1530 // If this is inserting an element that isn't demanded, remove this
1532 unsigned IdxNo = Idx->getZExtValue();
1533 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1534 return AddSoonDeadInstToWorklist(*I, 0);
1536 // Otherwise, the element inserted overwrites whatever was there, so the
1537 // input demanded set is simpler than the output set.
1538 APInt DemandedElts2 = DemandedElts;
1539 DemandedElts2.clear(IdxNo);
1540 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1541 UndefElts, Depth+1);
1542 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1544 // The inserted element is defined.
1545 UndefElts.clear(IdxNo);
1548 case Instruction::ShuffleVector: {
1549 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1550 uint64_t LHSVWidth =
1551 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1552 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1553 for (unsigned i = 0; i < VWidth; i++) {
1554 if (DemandedElts[i]) {
1555 unsigned MaskVal = Shuffle->getMaskValue(i);
1556 if (MaskVal != -1u) {
1557 assert(MaskVal < LHSVWidth * 2 &&
1558 "shufflevector mask index out of range!");
1559 if (MaskVal < LHSVWidth)
1560 LeftDemanded.set(MaskVal);
1562 RightDemanded.set(MaskVal - LHSVWidth);
1567 APInt UndefElts4(LHSVWidth, 0);
1568 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1569 UndefElts4, Depth+1);
1570 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1572 APInt UndefElts3(LHSVWidth, 0);
1573 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1574 UndefElts3, Depth+1);
1575 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1577 bool NewUndefElts = false;
1578 for (unsigned i = 0; i < VWidth; i++) {
1579 unsigned MaskVal = Shuffle->getMaskValue(i);
1580 if (MaskVal == -1u) {
1582 } else if (MaskVal < LHSVWidth) {
1583 if (UndefElts4[MaskVal]) {
1584 NewUndefElts = true;
1588 if (UndefElts3[MaskVal - LHSVWidth]) {
1589 NewUndefElts = true;
1596 // Add additional discovered undefs.
1597 std::vector<Constant*> Elts;
1598 for (unsigned i = 0; i < VWidth; ++i) {
1600 Elts.push_back(Context->getUndef(Type::Int32Ty));
1602 Elts.push_back(Context->getConstantInt(Type::Int32Ty,
1603 Shuffle->getMaskValue(i)));
1605 I->setOperand(2, Context->getConstantVector(Elts));
1610 case Instruction::BitCast: {
1611 // Vector->vector casts only.
1612 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1614 unsigned InVWidth = VTy->getNumElements();
1615 APInt InputDemandedElts(InVWidth, 0);
1618 if (VWidth == InVWidth) {
1619 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1620 // elements as are demanded of us.
1622 InputDemandedElts = DemandedElts;
1623 } else if (VWidth > InVWidth) {
1627 // If there are more elements in the result than there are in the source,
1628 // then an input element is live if any of the corresponding output
1629 // elements are live.
1630 Ratio = VWidth/InVWidth;
1631 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1632 if (DemandedElts[OutIdx])
1633 InputDemandedElts.set(OutIdx/Ratio);
1639 // If there are more elements in the source than there are in the result,
1640 // then an input element is live if the corresponding output element is
1642 Ratio = InVWidth/VWidth;
1643 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1644 if (DemandedElts[InIdx/Ratio])
1645 InputDemandedElts.set(InIdx);
1648 // div/rem demand all inputs, because they don't want divide by zero.
1649 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1650 UndefElts2, Depth+1);
1652 I->setOperand(0, TmpV);
1656 UndefElts = UndefElts2;
1657 if (VWidth > InVWidth) {
1658 llvm_unreachable("Unimp");
1659 // If there are more elements in the result than there are in the source,
1660 // then an output element is undef if the corresponding input element is
1662 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1663 if (UndefElts2[OutIdx/Ratio])
1664 UndefElts.set(OutIdx);
1665 } else if (VWidth < InVWidth) {
1666 llvm_unreachable("Unimp");
1667 // If there are more elements in the source than there are in the result,
1668 // then a result element is undef if all of the corresponding input
1669 // elements are undef.
1670 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1671 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1672 if (!UndefElts2[InIdx]) // Not undef?
1673 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1677 case Instruction::And:
1678 case Instruction::Or:
1679 case Instruction::Xor:
1680 case Instruction::Add:
1681 case Instruction::Sub:
1682 case Instruction::Mul:
1683 // div/rem demand all inputs, because they don't want divide by zero.
1684 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1685 UndefElts, Depth+1);
1686 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1687 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1688 UndefElts2, Depth+1);
1689 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1691 // Output elements are undefined if both are undefined. Consider things
1692 // like undef&0. The result is known zero, not undef.
1693 UndefElts &= UndefElts2;
1696 case Instruction::Call: {
1697 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1699 switch (II->getIntrinsicID()) {
1702 // Binary vector operations that work column-wise. A dest element is a
1703 // function of the corresponding input elements from the two inputs.
1704 case Intrinsic::x86_sse_sub_ss:
1705 case Intrinsic::x86_sse_mul_ss:
1706 case Intrinsic::x86_sse_min_ss:
1707 case Intrinsic::x86_sse_max_ss:
1708 case Intrinsic::x86_sse2_sub_sd:
1709 case Intrinsic::x86_sse2_mul_sd:
1710 case Intrinsic::x86_sse2_min_sd:
1711 case Intrinsic::x86_sse2_max_sd:
1712 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1713 UndefElts, Depth+1);
1714 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1715 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1716 UndefElts2, Depth+1);
1717 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1719 // If only the low elt is demanded and this is a scalarizable intrinsic,
1720 // scalarize it now.
1721 if (DemandedElts == 1) {
1722 switch (II->getIntrinsicID()) {
1724 case Intrinsic::x86_sse_sub_ss:
1725 case Intrinsic::x86_sse_mul_ss:
1726 case Intrinsic::x86_sse2_sub_sd:
1727 case Intrinsic::x86_sse2_mul_sd:
1728 // TODO: Lower MIN/MAX/ABS/etc
1729 Value *LHS = II->getOperand(1);
1730 Value *RHS = II->getOperand(2);
1731 // Extract the element as scalars.
1732 LHS = InsertNewInstBefore(new ExtractElementInst(LHS,
1733 Context->getConstantInt(Type::Int32Ty, 0U, false), "tmp"), *II);
1734 RHS = InsertNewInstBefore(new ExtractElementInst(RHS,
1735 Context->getConstantInt(Type::Int32Ty, 0U, false), "tmp"), *II);
1737 switch (II->getIntrinsicID()) {
1738 default: llvm_unreachable("Case stmts out of sync!");
1739 case Intrinsic::x86_sse_sub_ss:
1740 case Intrinsic::x86_sse2_sub_sd:
1741 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1742 II->getName()), *II);
1744 case Intrinsic::x86_sse_mul_ss:
1745 case Intrinsic::x86_sse2_mul_sd:
1746 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1747 II->getName()), *II);
1752 InsertElementInst::Create(
1753 Context->getUndef(II->getType()), TmpV,
1754 Context->getConstantInt(Type::Int32Ty, 0U, false), II->getName());
1755 InsertNewInstBefore(New, *II);
1756 AddSoonDeadInstToWorklist(*II, 0);
1761 // Output elements are undefined if both are undefined. Consider things
1762 // like undef&0. The result is known zero, not undef.
1763 UndefElts &= UndefElts2;
1769 return MadeChange ? I : 0;
1773 /// AssociativeOpt - Perform an optimization on an associative operator. This
1774 /// function is designed to check a chain of associative operators for a
1775 /// potential to apply a certain optimization. Since the optimization may be
1776 /// applicable if the expression was reassociated, this checks the chain, then
1777 /// reassociates the expression as necessary to expose the optimization
1778 /// opportunity. This makes use of a special Functor, which must define
1779 /// 'shouldApply' and 'apply' methods.
1781 template<typename Functor>
1782 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F,
1783 LLVMContext *Context) {
1784 unsigned Opcode = Root.getOpcode();
1785 Value *LHS = Root.getOperand(0);
1787 // Quick check, see if the immediate LHS matches...
1788 if (F.shouldApply(LHS))
1789 return F.apply(Root);
1791 // Otherwise, if the LHS is not of the same opcode as the root, return.
1792 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1793 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1794 // Should we apply this transform to the RHS?
1795 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1797 // If not to the RHS, check to see if we should apply to the LHS...
1798 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1799 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1803 // If the functor wants to apply the optimization to the RHS of LHSI,
1804 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1806 // Now all of the instructions are in the current basic block, go ahead
1807 // and perform the reassociation.
1808 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1810 // First move the selected RHS to the LHS of the root...
1811 Root.setOperand(0, LHSI->getOperand(1));
1813 // Make what used to be the LHS of the root be the user of the root...
1814 Value *ExtraOperand = TmpLHSI->getOperand(1);
1815 if (&Root == TmpLHSI) {
1816 Root.replaceAllUsesWith(Context->getNullValue(TmpLHSI->getType()));
1819 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1820 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1821 BasicBlock::iterator ARI = &Root; ++ARI;
1822 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1825 // Now propagate the ExtraOperand down the chain of instructions until we
1827 while (TmpLHSI != LHSI) {
1828 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1829 // Move the instruction to immediately before the chain we are
1830 // constructing to avoid breaking dominance properties.
1831 NextLHSI->moveBefore(ARI);
1834 Value *NextOp = NextLHSI->getOperand(1);
1835 NextLHSI->setOperand(1, ExtraOperand);
1837 ExtraOperand = NextOp;
1840 // Now that the instructions are reassociated, have the functor perform
1841 // the transformation...
1842 return F.apply(Root);
1845 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1852 // AddRHS - Implements: X + X --> X << 1
1855 LLVMContext *Context;
1856 AddRHS(Value *rhs, LLVMContext *C) : RHS(rhs), Context(C) {}
1857 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1858 Instruction *apply(BinaryOperator &Add) const {
1859 return BinaryOperator::CreateShl(Add.getOperand(0),
1860 Context->getConstantInt(Add.getType(), 1));
1864 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1866 struct AddMaskingAnd {
1868 LLVMContext *Context;
1869 AddMaskingAnd(Constant *c, LLVMContext *C) : C2(c), Context(C) {}
1870 bool shouldApply(Value *LHS) const {
1872 return match(LHS, m_And(m_Value(), m_ConstantInt(C1)), *Context) &&
1873 Context->getConstantExprAnd(C1, C2)->isNullValue();
1875 Instruction *apply(BinaryOperator &Add) const {
1876 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1882 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1884 LLVMContext *Context = IC->getContext();
1886 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1887 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1890 // Figure out if the constant is the left or the right argument.
1891 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1892 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1894 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1896 return Context->getConstantExpr(I.getOpcode(), SOC, ConstOperand);
1897 return Context->getConstantExpr(I.getOpcode(), ConstOperand, SOC);
1900 Value *Op0 = SO, *Op1 = ConstOperand;
1902 std::swap(Op0, Op1);
1904 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1905 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1906 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1907 New = CmpInst::Create(*Context, CI->getOpcode(), CI->getPredicate(),
1908 Op0, Op1, SO->getName()+".cmp");
1910 llvm_unreachable("Unknown binary instruction type!");
1912 return IC->InsertNewInstBefore(New, I);
1915 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1916 // constant as the other operand, try to fold the binary operator into the
1917 // select arguments. This also works for Cast instructions, which obviously do
1918 // not have a second operand.
1919 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1921 // Don't modify shared select instructions
1922 if (!SI->hasOneUse()) return 0;
1923 Value *TV = SI->getOperand(1);
1924 Value *FV = SI->getOperand(2);
1926 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1927 // Bool selects with constant operands can be folded to logical ops.
1928 if (SI->getType() == Type::Int1Ty) return 0;
1930 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1931 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1933 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1940 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1941 /// node as operand #0, see if we can fold the instruction into the PHI (which
1942 /// is only possible if all operands to the PHI are constants).
1943 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1944 PHINode *PN = cast<PHINode>(I.getOperand(0));
1945 unsigned NumPHIValues = PN->getNumIncomingValues();
1946 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1948 // Check to see if all of the operands of the PHI are constants. If there is
1949 // one non-constant value, remember the BB it is. If there is more than one
1950 // or if *it* is a PHI, bail out.
1951 BasicBlock *NonConstBB = 0;
1952 for (unsigned i = 0; i != NumPHIValues; ++i)
1953 if (!isa<Constant>(PN->getIncomingValue(i))) {
1954 if (NonConstBB) return 0; // More than one non-const value.
1955 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1956 NonConstBB = PN->getIncomingBlock(i);
1958 // If the incoming non-constant value is in I's block, we have an infinite
1960 if (NonConstBB == I.getParent())
1964 // If there is exactly one non-constant value, we can insert a copy of the
1965 // operation in that block. However, if this is a critical edge, we would be
1966 // inserting the computation one some other paths (e.g. inside a loop). Only
1967 // do this if the pred block is unconditionally branching into the phi block.
1969 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1970 if (!BI || !BI->isUnconditional()) return 0;
1973 // Okay, we can do the transformation: create the new PHI node.
1974 PHINode *NewPN = PHINode::Create(I.getType(), "");
1975 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1976 InsertNewInstBefore(NewPN, *PN);
1977 NewPN->takeName(PN);
1979 // Next, add all of the operands to the PHI.
1980 if (I.getNumOperands() == 2) {
1981 Constant *C = cast<Constant>(I.getOperand(1));
1982 for (unsigned i = 0; i != NumPHIValues; ++i) {
1984 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1985 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1986 InV = Context->getConstantExprCompare(CI->getPredicate(), InC, C);
1988 InV = Context->getConstantExpr(I.getOpcode(), InC, C);
1990 assert(PN->getIncomingBlock(i) == NonConstBB);
1991 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1992 InV = BinaryOperator::Create(BO->getOpcode(),
1993 PN->getIncomingValue(i), C, "phitmp",
1994 NonConstBB->getTerminator());
1995 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1996 InV = CmpInst::Create(*Context, CI->getOpcode(),
1998 PN->getIncomingValue(i), C, "phitmp",
1999 NonConstBB->getTerminator());
2001 llvm_unreachable("Unknown binop!");
2003 AddToWorkList(cast<Instruction>(InV));
2005 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2008 CastInst *CI = cast<CastInst>(&I);
2009 const Type *RetTy = CI->getType();
2010 for (unsigned i = 0; i != NumPHIValues; ++i) {
2012 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2013 InV = Context->getConstantExprCast(CI->getOpcode(), InC, RetTy);
2015 assert(PN->getIncomingBlock(i) == NonConstBB);
2016 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2017 I.getType(), "phitmp",
2018 NonConstBB->getTerminator());
2019 AddToWorkList(cast<Instruction>(InV));
2021 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2024 return ReplaceInstUsesWith(I, NewPN);
2028 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2029 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2030 /// This basically requires proving that the add in the original type would not
2031 /// overflow to change the sign bit or have a carry out.
2032 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2033 // There are different heuristics we can use for this. Here are some simple
2036 // Add has the property that adding any two 2's complement numbers can only
2037 // have one carry bit which can change a sign. As such, if LHS and RHS each
2038 // have at least two sign bits, we know that the addition of the two values will
2039 // sign extend fine.
2040 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2044 // If one of the operands only has one non-zero bit, and if the other operand
2045 // has a known-zero bit in a more significant place than it (not including the
2046 // sign bit) the ripple may go up to and fill the zero, but won't change the
2047 // sign. For example, (X & ~4) + 1.
2055 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2056 bool Changed = SimplifyCommutative(I);
2057 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2059 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2060 // X + undef -> undef
2061 if (isa<UndefValue>(RHS))
2062 return ReplaceInstUsesWith(I, RHS);
2065 if (RHSC->isNullValue())
2066 return ReplaceInstUsesWith(I, LHS);
2068 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2069 // X + (signbit) --> X ^ signbit
2070 const APInt& Val = CI->getValue();
2071 uint32_t BitWidth = Val.getBitWidth();
2072 if (Val == APInt::getSignBit(BitWidth))
2073 return BinaryOperator::CreateXor(LHS, RHS);
2075 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2076 // (X & 254)+1 -> (X&254)|1
2077 if (SimplifyDemandedInstructionBits(I))
2080 // zext(bool) + C -> bool ? C + 1 : C
2081 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2082 if (ZI->getSrcTy() == Type::Int1Ty)
2083 return SelectInst::Create(ZI->getOperand(0), AddOne(CI, Context), CI);
2086 if (isa<PHINode>(LHS))
2087 if (Instruction *NV = FoldOpIntoPhi(I))
2090 ConstantInt *XorRHS = 0;
2092 if (isa<ConstantInt>(RHSC) &&
2093 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)), *Context)) {
2094 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2095 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2097 uint32_t Size = TySizeBits / 2;
2098 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2099 APInt CFF80Val(-C0080Val);
2101 if (TySizeBits > Size) {
2102 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2103 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2104 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2105 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2106 // This is a sign extend if the top bits are known zero.
2107 if (!MaskedValueIsZero(XorLHS,
2108 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2109 Size = 0; // Not a sign ext, but can't be any others either.
2114 C0080Val = APIntOps::lshr(C0080Val, Size);
2115 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2116 } while (Size >= 1);
2118 // FIXME: This shouldn't be necessary. When the backends can handle types
2119 // with funny bit widths then this switch statement should be removed. It
2120 // is just here to get the size of the "middle" type back up to something
2121 // that the back ends can handle.
2122 const Type *MiddleType = 0;
2125 case 32: MiddleType = Type::Int32Ty; break;
2126 case 16: MiddleType = Type::Int16Ty; break;
2127 case 8: MiddleType = Type::Int8Ty; break;
2130 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2131 InsertNewInstBefore(NewTrunc, I);
2132 return new SExtInst(NewTrunc, I.getType(), I.getName());
2137 if (I.getType() == Type::Int1Ty)
2138 return BinaryOperator::CreateXor(LHS, RHS);
2141 if (I.getType()->isInteger()) {
2142 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS, Context), Context))
2145 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2146 if (RHSI->getOpcode() == Instruction::Sub)
2147 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2148 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2150 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2151 if (LHSI->getOpcode() == Instruction::Sub)
2152 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2153 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2158 // -A + -B --> -(A + B)
2159 if (Value *LHSV = dyn_castNegVal(LHS, Context)) {
2160 if (LHS->getType()->isIntOrIntVector()) {
2161 if (Value *RHSV = dyn_castNegVal(RHS, Context)) {
2162 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2163 InsertNewInstBefore(NewAdd, I);
2164 return BinaryOperator::CreateNeg(*Context, NewAdd);
2168 return BinaryOperator::CreateSub(RHS, LHSV);
2172 if (!isa<Constant>(RHS))
2173 if (Value *V = dyn_castNegVal(RHS, Context))
2174 return BinaryOperator::CreateSub(LHS, V);
2178 if (Value *X = dyn_castFoldableMul(LHS, C2, Context)) {
2179 if (X == RHS) // X*C + X --> X * (C+1)
2180 return BinaryOperator::CreateMul(RHS, AddOne(C2, Context));
2182 // X*C1 + X*C2 --> X * (C1+C2)
2184 if (X == dyn_castFoldableMul(RHS, C1, Context))
2185 return BinaryOperator::CreateMul(X, Context->getConstantExprAdd(C1, C2));
2188 // X + X*C --> X * (C+1)
2189 if (dyn_castFoldableMul(RHS, C2, Context) == LHS)
2190 return BinaryOperator::CreateMul(LHS, AddOne(C2, Context));
2192 // X + ~X --> -1 since ~X = -X-1
2193 if (dyn_castNotVal(LHS, Context) == RHS ||
2194 dyn_castNotVal(RHS, Context) == LHS)
2195 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
2198 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2199 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2)), *Context))
2200 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2, Context), Context))
2203 // A+B --> A|B iff A and B have no bits set in common.
2204 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2205 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2206 APInt LHSKnownOne(IT->getBitWidth(), 0);
2207 APInt LHSKnownZero(IT->getBitWidth(), 0);
2208 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2209 if (LHSKnownZero != 0) {
2210 APInt RHSKnownOne(IT->getBitWidth(), 0);
2211 APInt RHSKnownZero(IT->getBitWidth(), 0);
2212 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2214 // No bits in common -> bitwise or.
2215 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2216 return BinaryOperator::CreateOr(LHS, RHS);
2220 // W*X + Y*Z --> W * (X+Z) iff W == Y
2221 if (I.getType()->isIntOrIntVector()) {
2222 Value *W, *X, *Y, *Z;
2223 if (match(LHS, m_Mul(m_Value(W), m_Value(X)), *Context) &&
2224 match(RHS, m_Mul(m_Value(Y), m_Value(Z)), *Context)) {
2228 } else if (Y == X) {
2230 } else if (X == Z) {
2237 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2238 LHS->getName()), I);
2239 return BinaryOperator::CreateMul(W, NewAdd);
2244 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2246 if (match(LHS, m_Not(m_Value(X)), *Context)) // ~X + C --> (C-1) - X
2247 return BinaryOperator::CreateSub(SubOne(CRHS, Context), X);
2249 // (X & FF00) + xx00 -> (X+xx00) & FF00
2250 if (LHS->hasOneUse() &&
2251 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)), *Context)) {
2252 Constant *Anded = Context->getConstantExprAnd(CRHS, C2);
2253 if (Anded == CRHS) {
2254 // See if all bits from the first bit set in the Add RHS up are included
2255 // in the mask. First, get the rightmost bit.
2256 const APInt& AddRHSV = CRHS->getValue();
2258 // Form a mask of all bits from the lowest bit added through the top.
2259 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2261 // See if the and mask includes all of these bits.
2262 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2264 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2265 // Okay, the xform is safe. Insert the new add pronto.
2266 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2267 LHS->getName()), I);
2268 return BinaryOperator::CreateAnd(NewAdd, C2);
2273 // Try to fold constant add into select arguments.
2274 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2275 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2279 // add (cast *A to intptrtype) B ->
2280 // cast (GEP (cast *A to i8*) B) --> intptrtype
2282 CastInst *CI = dyn_cast<CastInst>(LHS);
2285 CI = dyn_cast<CastInst>(RHS);
2288 if (CI && CI->getType()->isSized() &&
2289 (CI->getType()->getScalarSizeInBits() ==
2290 TD->getIntPtrType()->getPrimitiveSizeInBits())
2291 && isa<PointerType>(CI->getOperand(0)->getType())) {
2293 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2294 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2295 Context->getPointerType(Type::Int8Ty, AS), I);
2296 GetElementPtrInst *GEP = GetElementPtrInst::Create(I2, Other, "ctg2");
2297 // A GEP formed from an arbitrary add may overflow.
2298 cast<GEPOperator>(GEP)->setHasNoPointerOverflow(false);
2299 I2 = InsertNewInstBefore(GEP, I);
2300 return new PtrToIntInst(I2, CI->getType());
2304 // add (select X 0 (sub n A)) A --> select X A n
2306 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2309 SI = dyn_cast<SelectInst>(RHS);
2312 if (SI && SI->hasOneUse()) {
2313 Value *TV = SI->getTrueValue();
2314 Value *FV = SI->getFalseValue();
2317 // Can we fold the add into the argument of the select?
2318 // We check both true and false select arguments for a matching subtract.
2319 if (match(FV, m_Zero(), *Context) &&
2320 match(TV, m_Sub(m_Value(N), m_Specific(A)), *Context))
2321 // Fold the add into the true select value.
2322 return SelectInst::Create(SI->getCondition(), N, A);
2323 if (match(TV, m_Zero(), *Context) &&
2324 match(FV, m_Sub(m_Value(N), m_Specific(A)), *Context))
2325 // Fold the add into the false select value.
2326 return SelectInst::Create(SI->getCondition(), A, N);
2330 // Check for (add (sext x), y), see if we can merge this into an
2331 // integer add followed by a sext.
2332 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2333 // (add (sext x), cst) --> (sext (add x, cst'))
2334 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2336 Context->getConstantExprTrunc(RHSC, LHSConv->getOperand(0)->getType());
2337 if (LHSConv->hasOneUse() &&
2338 Context->getConstantExprSExt(CI, I.getType()) == RHSC &&
2339 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2340 // Insert the new, smaller add.
2341 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2343 InsertNewInstBefore(NewAdd, I);
2344 return new SExtInst(NewAdd, I.getType());
2348 // (add (sext x), (sext y)) --> (sext (add int x, y))
2349 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2350 // Only do this if x/y have the same type, if at last one of them has a
2351 // single use (so we don't increase the number of sexts), and if the
2352 // integer add will not overflow.
2353 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2354 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2355 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2356 RHSConv->getOperand(0))) {
2357 // Insert the new integer add.
2358 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2359 RHSConv->getOperand(0),
2361 InsertNewInstBefore(NewAdd, I);
2362 return new SExtInst(NewAdd, I.getType());
2367 return Changed ? &I : 0;
2370 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2371 bool Changed = SimplifyCommutative(I);
2372 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2374 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2376 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2377 if (CFP->isExactlyValue(Context->getConstantFPNegativeZero
2378 (I.getType())->getValueAPF()))
2379 return ReplaceInstUsesWith(I, LHS);
2382 if (isa<PHINode>(LHS))
2383 if (Instruction *NV = FoldOpIntoPhi(I))
2388 // -A + -B --> -(A + B)
2389 if (Value *LHSV = dyn_castFNegVal(LHS, Context))
2390 return BinaryOperator::CreateFSub(RHS, LHSV);
2393 if (!isa<Constant>(RHS))
2394 if (Value *V = dyn_castFNegVal(RHS, Context))
2395 return BinaryOperator::CreateFSub(LHS, V);
2397 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2398 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2399 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2400 return ReplaceInstUsesWith(I, LHS);
2402 // Check for (add double (sitofp x), y), see if we can merge this into an
2403 // integer add followed by a promotion.
2404 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2405 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2406 // ... if the constant fits in the integer value. This is useful for things
2407 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2408 // requires a constant pool load, and generally allows the add to be better
2410 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2412 Context->getConstantExprFPToSI(CFP, LHSConv->getOperand(0)->getType());
2413 if (LHSConv->hasOneUse() &&
2414 Context->getConstantExprSIToFP(CI, I.getType()) == CFP &&
2415 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2416 // Insert the new integer add.
2417 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2419 InsertNewInstBefore(NewAdd, I);
2420 return new SIToFPInst(NewAdd, I.getType());
2424 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2425 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2426 // Only do this if x/y have the same type, if at last one of them has a
2427 // single use (so we don't increase the number of int->fp conversions),
2428 // and if the integer add will not overflow.
2429 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2430 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2431 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2432 RHSConv->getOperand(0))) {
2433 // Insert the new integer add.
2434 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2435 RHSConv->getOperand(0),
2437 InsertNewInstBefore(NewAdd, I);
2438 return new SIToFPInst(NewAdd, I.getType());
2443 return Changed ? &I : 0;
2446 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2447 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2449 if (Op0 == Op1) // sub X, X -> 0
2450 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2452 // If this is a 'B = x-(-A)', change to B = x+A...
2453 if (Value *V = dyn_castNegVal(Op1, Context))
2454 return BinaryOperator::CreateAdd(Op0, V);
2456 if (isa<UndefValue>(Op0))
2457 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2458 if (isa<UndefValue>(Op1))
2459 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2461 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2462 // Replace (-1 - A) with (~A)...
2463 if (C->isAllOnesValue())
2464 return BinaryOperator::CreateNot(*Context, Op1);
2466 // C - ~X == X + (1+C)
2468 if (match(Op1, m_Not(m_Value(X)), *Context))
2469 return BinaryOperator::CreateAdd(X, AddOne(C, Context));
2471 // -(X >>u 31) -> (X >>s 31)
2472 // -(X >>s 31) -> (X >>u 31)
2474 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2475 if (SI->getOpcode() == Instruction::LShr) {
2476 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2477 // Check to see if we are shifting out everything but the sign bit.
2478 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2479 SI->getType()->getPrimitiveSizeInBits()-1) {
2480 // Ok, the transformation is safe. Insert AShr.
2481 return BinaryOperator::Create(Instruction::AShr,
2482 SI->getOperand(0), CU, SI->getName());
2486 else if (SI->getOpcode() == Instruction::AShr) {
2487 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2488 // Check to see if we are shifting out everything but the sign bit.
2489 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2490 SI->getType()->getPrimitiveSizeInBits()-1) {
2491 // Ok, the transformation is safe. Insert LShr.
2492 return BinaryOperator::CreateLShr(
2493 SI->getOperand(0), CU, SI->getName());
2500 // Try to fold constant sub into select arguments.
2501 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2502 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2505 // C - zext(bool) -> bool ? C - 1 : C
2506 if (ZExtInst *ZI = dyn_cast<ZExtInst>(Op1))
2507 if (ZI->getSrcTy() == Type::Int1Ty)
2508 return SelectInst::Create(ZI->getOperand(0), SubOne(C, Context), C);
2511 if (I.getType() == Type::Int1Ty)
2512 return BinaryOperator::CreateXor(Op0, Op1);
2514 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2515 if (Op1I->getOpcode() == Instruction::Add) {
2516 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2517 return BinaryOperator::CreateNeg(*Context, Op1I->getOperand(1),
2519 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2520 return BinaryOperator::CreateNeg(*Context, Op1I->getOperand(0),
2522 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2523 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2524 // C1-(X+C2) --> (C1-C2)-X
2525 return BinaryOperator::CreateSub(
2526 Context->getConstantExprSub(CI1, CI2), Op1I->getOperand(0));
2530 if (Op1I->hasOneUse()) {
2531 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2532 // is not used by anyone else...
2534 if (Op1I->getOpcode() == Instruction::Sub) {
2535 // Swap the two operands of the subexpr...
2536 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2537 Op1I->setOperand(0, IIOp1);
2538 Op1I->setOperand(1, IIOp0);
2540 // Create the new top level add instruction...
2541 return BinaryOperator::CreateAdd(Op0, Op1);
2544 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2546 if (Op1I->getOpcode() == Instruction::And &&
2547 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2548 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2551 InsertNewInstBefore(BinaryOperator::CreateNot(*Context,
2552 OtherOp, "B.not"), I);
2553 return BinaryOperator::CreateAnd(Op0, NewNot);
2556 // 0 - (X sdiv C) -> (X sdiv -C)
2557 if (Op1I->getOpcode() == Instruction::SDiv)
2558 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2560 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2561 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2562 Context->getConstantExprNeg(DivRHS));
2564 // X - X*C --> X * (1-C)
2565 ConstantInt *C2 = 0;
2566 if (dyn_castFoldableMul(Op1I, C2, Context) == Op0) {
2568 Context->getConstantExprSub(Context->getConstantInt(I.getType(), 1),
2570 return BinaryOperator::CreateMul(Op0, CP1);
2575 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2576 if (Op0I->getOpcode() == Instruction::Add) {
2577 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2578 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2579 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2580 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2581 } else if (Op0I->getOpcode() == Instruction::Sub) {
2582 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2583 return BinaryOperator::CreateNeg(*Context, Op0I->getOperand(1),
2589 if (Value *X = dyn_castFoldableMul(Op0, C1, Context)) {
2590 if (X == Op1) // X*C - X --> X * (C-1)
2591 return BinaryOperator::CreateMul(Op1, SubOne(C1, Context));
2593 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2594 if (X == dyn_castFoldableMul(Op1, C2, Context))
2595 return BinaryOperator::CreateMul(X, Context->getConstantExprSub(C1, C2));
2600 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2601 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2603 // If this is a 'B = x-(-A)', change to B = x+A...
2604 if (Value *V = dyn_castFNegVal(Op1, Context))
2605 return BinaryOperator::CreateFAdd(Op0, V);
2607 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2608 if (Op1I->getOpcode() == Instruction::FAdd) {
2609 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2610 return BinaryOperator::CreateFNeg(*Context, Op1I->getOperand(1),
2612 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2613 return BinaryOperator::CreateFNeg(*Context, Op1I->getOperand(0),
2621 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2622 /// comparison only checks the sign bit. If it only checks the sign bit, set
2623 /// TrueIfSigned if the result of the comparison is true when the input value is
2625 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2626 bool &TrueIfSigned) {
2628 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2629 TrueIfSigned = true;
2630 return RHS->isZero();
2631 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2632 TrueIfSigned = true;
2633 return RHS->isAllOnesValue();
2634 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2635 TrueIfSigned = false;
2636 return RHS->isAllOnesValue();
2637 case ICmpInst::ICMP_UGT:
2638 // True if LHS u> RHS and RHS == high-bit-mask - 1
2639 TrueIfSigned = true;
2640 return RHS->getValue() ==
2641 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2642 case ICmpInst::ICMP_UGE:
2643 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2644 TrueIfSigned = true;
2645 return RHS->getValue().isSignBit();
2651 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2652 bool Changed = SimplifyCommutative(I);
2653 Value *Op0 = I.getOperand(0);
2655 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2656 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2658 // Simplify mul instructions with a constant RHS...
2659 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2660 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2662 // ((X << C1)*C2) == (X * (C2 << C1))
2663 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2664 if (SI->getOpcode() == Instruction::Shl)
2665 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2666 return BinaryOperator::CreateMul(SI->getOperand(0),
2667 Context->getConstantExprShl(CI, ShOp));
2670 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2671 if (CI->equalsInt(1)) // X * 1 == X
2672 return ReplaceInstUsesWith(I, Op0);
2673 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2674 return BinaryOperator::CreateNeg(*Context, Op0, I.getName());
2676 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2677 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2678 return BinaryOperator::CreateShl(Op0,
2679 Context->getConstantInt(Op0->getType(), Val.logBase2()));
2681 } else if (isa<VectorType>(Op1->getType())) {
2682 if (Op1->isNullValue())
2683 return ReplaceInstUsesWith(I, Op1);
2685 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2686 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2687 return BinaryOperator::CreateNeg(*Context, Op0, I.getName());
2689 // As above, vector X*splat(1.0) -> X in all defined cases.
2690 if (Constant *Splat = Op1V->getSplatValue()) {
2691 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2692 if (CI->equalsInt(1))
2693 return ReplaceInstUsesWith(I, Op0);
2698 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2699 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2700 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2701 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2702 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2704 InsertNewInstBefore(Add, I);
2705 Value *C1C2 = Context->getConstantExprMul(Op1,
2706 cast<Constant>(Op0I->getOperand(1)));
2707 return BinaryOperator::CreateAdd(Add, C1C2);
2711 // Try to fold constant mul into select arguments.
2712 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2713 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2716 if (isa<PHINode>(Op0))
2717 if (Instruction *NV = FoldOpIntoPhi(I))
2721 if (Value *Op0v = dyn_castNegVal(Op0, Context)) // -X * -Y = X*Y
2722 if (Value *Op1v = dyn_castNegVal(I.getOperand(1), Context))
2723 return BinaryOperator::CreateMul(Op0v, Op1v);
2725 // (X / Y) * Y = X - (X % Y)
2726 // (X / Y) * -Y = (X % Y) - X
2728 Value *Op1 = I.getOperand(1);
2729 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2731 (BO->getOpcode() != Instruction::UDiv &&
2732 BO->getOpcode() != Instruction::SDiv)) {
2734 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2736 Value *Neg = dyn_castNegVal(Op1, Context);
2737 if (BO && BO->hasOneUse() &&
2738 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2739 (BO->getOpcode() == Instruction::UDiv ||
2740 BO->getOpcode() == Instruction::SDiv)) {
2741 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2744 if (BO->getOpcode() == Instruction::UDiv)
2745 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2747 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2749 InsertNewInstBefore(Rem, I);
2753 return BinaryOperator::CreateSub(Op0BO, Rem);
2755 return BinaryOperator::CreateSub(Rem, Op0BO);
2759 if (I.getType() == Type::Int1Ty)
2760 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2762 // If one of the operands of the multiply is a cast from a boolean value, then
2763 // we know the bool is either zero or one, so this is a 'masking' multiply.
2764 // See if we can simplify things based on how the boolean was originally
2766 CastInst *BoolCast = 0;
2767 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2768 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2771 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2772 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2775 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2776 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2777 const Type *SCOpTy = SCIOp0->getType();
2780 // If the icmp is true iff the sign bit of X is set, then convert this
2781 // multiply into a shift/and combination.
2782 if (isa<ConstantInt>(SCIOp1) &&
2783 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2785 // Shift the X value right to turn it into "all signbits".
2786 Constant *Amt = Context->getConstantInt(SCIOp0->getType(),
2787 SCOpTy->getPrimitiveSizeInBits()-1);
2789 InsertNewInstBefore(
2790 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2791 BoolCast->getOperand(0)->getName()+
2794 // If the multiply type is not the same as the source type, sign extend
2795 // or truncate to the multiply type.
2796 if (I.getType() != V->getType()) {
2797 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2798 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2799 Instruction::CastOps opcode =
2800 (SrcBits == DstBits ? Instruction::BitCast :
2801 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2802 V = InsertCastBefore(opcode, V, I.getType(), I);
2805 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2806 return BinaryOperator::CreateAnd(V, OtherOp);
2811 return Changed ? &I : 0;
2814 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2815 bool Changed = SimplifyCommutative(I);
2816 Value *Op0 = I.getOperand(0);
2818 // Simplify mul instructions with a constant RHS...
2819 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2820 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2821 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2822 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2823 if (Op1F->isExactlyValue(1.0))
2824 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2825 } else if (isa<VectorType>(Op1->getType())) {
2826 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2827 // As above, vector X*splat(1.0) -> X in all defined cases.
2828 if (Constant *Splat = Op1V->getSplatValue()) {
2829 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2830 if (F->isExactlyValue(1.0))
2831 return ReplaceInstUsesWith(I, Op0);
2836 // Try to fold constant mul into select arguments.
2837 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2838 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2841 if (isa<PHINode>(Op0))
2842 if (Instruction *NV = FoldOpIntoPhi(I))
2846 if (Value *Op0v = dyn_castFNegVal(Op0, Context)) // -X * -Y = X*Y
2847 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1), Context))
2848 return BinaryOperator::CreateFMul(Op0v, Op1v);
2850 return Changed ? &I : 0;
2853 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2855 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2856 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2858 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2859 int NonNullOperand = -1;
2860 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2861 if (ST->isNullValue())
2863 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2864 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2865 if (ST->isNullValue())
2868 if (NonNullOperand == -1)
2871 Value *SelectCond = SI->getOperand(0);
2873 // Change the div/rem to use 'Y' instead of the select.
2874 I.setOperand(1, SI->getOperand(NonNullOperand));
2876 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2877 // problem. However, the select, or the condition of the select may have
2878 // multiple uses. Based on our knowledge that the operand must be non-zero,
2879 // propagate the known value for the select into other uses of it, and
2880 // propagate a known value of the condition into its other users.
2882 // If the select and condition only have a single use, don't bother with this,
2884 if (SI->use_empty() && SelectCond->hasOneUse())
2887 // Scan the current block backward, looking for other uses of SI.
2888 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2890 while (BBI != BBFront) {
2892 // If we found a call to a function, we can't assume it will return, so
2893 // information from below it cannot be propagated above it.
2894 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2897 // Replace uses of the select or its condition with the known values.
2898 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2901 *I = SI->getOperand(NonNullOperand);
2903 } else if (*I == SelectCond) {
2904 *I = NonNullOperand == 1 ? Context->getConstantIntTrue() :
2905 Context->getConstantIntFalse();
2910 // If we past the instruction, quit looking for it.
2913 if (&*BBI == SelectCond)
2916 // If we ran out of things to eliminate, break out of the loop.
2917 if (SelectCond == 0 && SI == 0)
2925 /// This function implements the transforms on div instructions that work
2926 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2927 /// used by the visitors to those instructions.
2928 /// @brief Transforms common to all three div instructions
2929 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2930 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2932 // undef / X -> 0 for integer.
2933 // undef / X -> undef for FP (the undef could be a snan).
2934 if (isa<UndefValue>(Op0)) {
2935 if (Op0->getType()->isFPOrFPVector())
2936 return ReplaceInstUsesWith(I, Op0);
2937 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2940 // X / undef -> undef
2941 if (isa<UndefValue>(Op1))
2942 return ReplaceInstUsesWith(I, Op1);
2947 /// This function implements the transforms common to both integer division
2948 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2949 /// division instructions.
2950 /// @brief Common integer divide transforms
2951 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2952 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2954 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2956 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2957 Constant *CI = Context->getConstantInt(Ty->getElementType(), 1);
2958 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2959 return ReplaceInstUsesWith(I, Context->getConstantVector(Elts));
2962 Constant *CI = Context->getConstantInt(I.getType(), 1);
2963 return ReplaceInstUsesWith(I, CI);
2966 if (Instruction *Common = commonDivTransforms(I))
2969 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2970 // This does not apply for fdiv.
2971 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2974 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2976 if (RHS->equalsInt(1))
2977 return ReplaceInstUsesWith(I, Op0);
2979 // (X / C1) / C2 -> X / (C1*C2)
2980 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2981 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2982 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2983 if (MultiplyOverflows(RHS, LHSRHS,
2984 I.getOpcode()==Instruction::SDiv, Context))
2985 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2987 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2988 Context->getConstantExprMul(RHS, LHSRHS));
2991 if (!RHS->isZero()) { // avoid X udiv 0
2992 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2993 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2995 if (isa<PHINode>(Op0))
2996 if (Instruction *NV = FoldOpIntoPhi(I))
3001 // 0 / X == 0, we don't need to preserve faults!
3002 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3003 if (LHS->equalsInt(0))
3004 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3006 // It can't be division by zero, hence it must be division by one.
3007 if (I.getType() == Type::Int1Ty)
3008 return ReplaceInstUsesWith(I, Op0);
3010 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3011 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3014 return ReplaceInstUsesWith(I, Op0);
3020 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3021 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3023 // Handle the integer div common cases
3024 if (Instruction *Common = commonIDivTransforms(I))
3027 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3028 // X udiv C^2 -> X >> C
3029 // Check to see if this is an unsigned division with an exact power of 2,
3030 // if so, convert to a right shift.
3031 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3032 return BinaryOperator::CreateLShr(Op0,
3033 Context->getConstantInt(Op0->getType(), C->getValue().logBase2()));
3035 // X udiv C, where C >= signbit
3036 if (C->getValue().isNegative()) {
3037 Value *IC = InsertNewInstBefore(new ICmpInst(*Context,
3038 ICmpInst::ICMP_ULT, Op0, C),
3040 return SelectInst::Create(IC, Context->getNullValue(I.getType()),
3041 Context->getConstantInt(I.getType(), 1));
3045 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3046 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3047 if (RHSI->getOpcode() == Instruction::Shl &&
3048 isa<ConstantInt>(RHSI->getOperand(0))) {
3049 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3050 if (C1.isPowerOf2()) {
3051 Value *N = RHSI->getOperand(1);
3052 const Type *NTy = N->getType();
3053 if (uint32_t C2 = C1.logBase2()) {
3054 Constant *C2V = Context->getConstantInt(NTy, C2);
3055 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3057 return BinaryOperator::CreateLShr(Op0, N);
3062 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3063 // where C1&C2 are powers of two.
3064 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3065 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3066 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3067 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3068 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3069 // Compute the shift amounts
3070 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3071 // Construct the "on true" case of the select
3072 Constant *TC = Context->getConstantInt(Op0->getType(), TSA);
3073 Instruction *TSI = BinaryOperator::CreateLShr(
3074 Op0, TC, SI->getName()+".t");
3075 TSI = InsertNewInstBefore(TSI, I);
3077 // Construct the "on false" case of the select
3078 Constant *FC = Context->getConstantInt(Op0->getType(), FSA);
3079 Instruction *FSI = BinaryOperator::CreateLShr(
3080 Op0, FC, SI->getName()+".f");
3081 FSI = InsertNewInstBefore(FSI, I);
3083 // construct the select instruction and return it.
3084 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3090 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3091 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3093 // Handle the integer div common cases
3094 if (Instruction *Common = commonIDivTransforms(I))
3097 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3099 if (RHS->isAllOnesValue())
3100 return BinaryOperator::CreateNeg(*Context, Op0);
3103 // If the sign bits of both operands are zero (i.e. we can prove they are
3104 // unsigned inputs), turn this into a udiv.
3105 if (I.getType()->isInteger()) {
3106 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3107 if (MaskedValueIsZero(Op0, Mask)) {
3108 if (MaskedValueIsZero(Op1, Mask)) {
3109 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3110 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3112 ConstantInt *ShiftedInt;
3113 if (match(Op1, m_Shl(m_ConstantInt(ShiftedInt), m_Value()), *Context) &&
3114 ShiftedInt->getValue().isPowerOf2()) {
3115 // X sdiv (1 << Y) -> X udiv (1 << Y) ( -> X u>> Y)
3116 // Safe because the only negative value (1 << Y) can take on is
3117 // INT_MIN, and X sdiv INT_MIN == X udiv INT_MIN == 0 if X doesn't have
3118 // the sign bit set.
3119 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3127 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3128 return commonDivTransforms(I);
3131 /// This function implements the transforms on rem instructions that work
3132 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3133 /// is used by the visitors to those instructions.
3134 /// @brief Transforms common to all three rem instructions
3135 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3136 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3138 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3139 if (I.getType()->isFPOrFPVector())
3140 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3141 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3143 if (isa<UndefValue>(Op1))
3144 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3146 // Handle cases involving: rem X, (select Cond, Y, Z)
3147 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3153 /// This function implements the transforms common to both integer remainder
3154 /// instructions (urem and srem). It is called by the visitors to those integer
3155 /// remainder instructions.
3156 /// @brief Common integer remainder transforms
3157 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3158 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3160 if (Instruction *common = commonRemTransforms(I))
3163 // 0 % X == 0 for integer, we don't need to preserve faults!
3164 if (Constant *LHS = dyn_cast<Constant>(Op0))
3165 if (LHS->isNullValue())
3166 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3168 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3169 // X % 0 == undef, we don't need to preserve faults!
3170 if (RHS->equalsInt(0))
3171 return ReplaceInstUsesWith(I, Context->getUndef(I.getType()));
3173 if (RHS->equalsInt(1)) // X % 1 == 0
3174 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3176 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3177 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3178 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3180 } else if (isa<PHINode>(Op0I)) {
3181 if (Instruction *NV = FoldOpIntoPhi(I))
3185 // See if we can fold away this rem instruction.
3186 if (SimplifyDemandedInstructionBits(I))
3194 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3195 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3197 if (Instruction *common = commonIRemTransforms(I))
3200 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3201 // X urem C^2 -> X and C
3202 // Check to see if this is an unsigned remainder with an exact power of 2,
3203 // if so, convert to a bitwise and.
3204 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3205 if (C->getValue().isPowerOf2())
3206 return BinaryOperator::CreateAnd(Op0, SubOne(C, Context));
3209 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3210 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3211 if (RHSI->getOpcode() == Instruction::Shl &&
3212 isa<ConstantInt>(RHSI->getOperand(0))) {
3213 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3214 Constant *N1 = Context->getAllOnesValue(I.getType());
3215 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3217 return BinaryOperator::CreateAnd(Op0, Add);
3222 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3223 // where C1&C2 are powers of two.
3224 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3225 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3226 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3227 // STO == 0 and SFO == 0 handled above.
3228 if ((STO->getValue().isPowerOf2()) &&
3229 (SFO->getValue().isPowerOf2())) {
3230 Value *TrueAnd = InsertNewInstBefore(
3231 BinaryOperator::CreateAnd(Op0, SubOne(STO, Context),
3232 SI->getName()+".t"), I);
3233 Value *FalseAnd = InsertNewInstBefore(
3234 BinaryOperator::CreateAnd(Op0, SubOne(SFO, Context),
3235 SI->getName()+".f"), I);
3236 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3244 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3245 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3247 // Handle the integer rem common cases
3248 if (Instruction *common = commonIRemTransforms(I))
3251 if (Value *RHSNeg = dyn_castNegVal(Op1, Context))
3252 if (!isa<Constant>(RHSNeg) ||
3253 (isa<ConstantInt>(RHSNeg) &&
3254 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3256 AddUsesToWorkList(I);
3257 I.setOperand(1, RHSNeg);
3261 // If the sign bits of both operands are zero (i.e. we can prove they are
3262 // unsigned inputs), turn this into a urem.
3263 if (I.getType()->isInteger()) {
3264 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3265 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3266 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3267 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3271 // If it's a constant vector, flip any negative values positive.
3272 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3273 unsigned VWidth = RHSV->getNumOperands();
3275 bool hasNegative = false;
3276 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3277 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3278 if (RHS->getValue().isNegative())
3282 std::vector<Constant *> Elts(VWidth);
3283 for (unsigned i = 0; i != VWidth; ++i) {
3284 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3285 if (RHS->getValue().isNegative())
3286 Elts[i] = cast<ConstantInt>(Context->getConstantExprNeg(RHS));
3292 Constant *NewRHSV = Context->getConstantVector(Elts);
3293 if (NewRHSV != RHSV) {
3294 AddUsesToWorkList(I);
3295 I.setOperand(1, NewRHSV);
3304 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3305 return commonRemTransforms(I);
3308 // isOneBitSet - Return true if there is exactly one bit set in the specified
3310 static bool isOneBitSet(const ConstantInt *CI) {
3311 return CI->getValue().isPowerOf2();
3314 // isHighOnes - Return true if the constant is of the form 1+0+.
3315 // This is the same as lowones(~X).
3316 static bool isHighOnes(const ConstantInt *CI) {
3317 return (~CI->getValue() + 1).isPowerOf2();
3320 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3321 /// are carefully arranged to allow folding of expressions such as:
3323 /// (A < B) | (A > B) --> (A != B)
3325 /// Note that this is only valid if the first and second predicates have the
3326 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3328 /// Three bits are used to represent the condition, as follows:
3333 /// <=> Value Definition
3334 /// 000 0 Always false
3341 /// 111 7 Always true
3343 static unsigned getICmpCode(const ICmpInst *ICI) {
3344 switch (ICI->getPredicate()) {
3346 case ICmpInst::ICMP_UGT: return 1; // 001
3347 case ICmpInst::ICMP_SGT: return 1; // 001
3348 case ICmpInst::ICMP_EQ: return 2; // 010
3349 case ICmpInst::ICMP_UGE: return 3; // 011
3350 case ICmpInst::ICMP_SGE: return 3; // 011
3351 case ICmpInst::ICMP_ULT: return 4; // 100
3352 case ICmpInst::ICMP_SLT: return 4; // 100
3353 case ICmpInst::ICMP_NE: return 5; // 101
3354 case ICmpInst::ICMP_ULE: return 6; // 110
3355 case ICmpInst::ICMP_SLE: return 6; // 110
3358 llvm_unreachable("Invalid ICmp predicate!");
3363 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3364 /// predicate into a three bit mask. It also returns whether it is an ordered
3365 /// predicate by reference.
3366 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3369 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3370 case FCmpInst::FCMP_UNO: return 0; // 000
3371 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3372 case FCmpInst::FCMP_UGT: return 1; // 001
3373 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3374 case FCmpInst::FCMP_UEQ: return 2; // 010
3375 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3376 case FCmpInst::FCMP_UGE: return 3; // 011
3377 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3378 case FCmpInst::FCMP_ULT: return 4; // 100
3379 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3380 case FCmpInst::FCMP_UNE: return 5; // 101
3381 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3382 case FCmpInst::FCMP_ULE: return 6; // 110
3385 // Not expecting FCMP_FALSE and FCMP_TRUE;
3386 llvm_unreachable("Unexpected FCmp predicate!");
3391 /// getICmpValue - This is the complement of getICmpCode, which turns an
3392 /// opcode and two operands into either a constant true or false, or a brand
3393 /// new ICmp instruction. The sign is passed in to determine which kind
3394 /// of predicate to use in the new icmp instruction.
3395 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3396 LLVMContext *Context) {
3398 default: llvm_unreachable("Illegal ICmp code!");
3399 case 0: return Context->getConstantIntFalse();
3402 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, LHS, RHS);
3404 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, LHS, RHS);
3405 case 2: return new ICmpInst(*Context, ICmpInst::ICMP_EQ, LHS, RHS);
3408 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHS, RHS);
3410 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHS, RHS);
3413 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHS, RHS);
3415 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHS, RHS);
3416 case 5: return new ICmpInst(*Context, ICmpInst::ICMP_NE, LHS, RHS);
3419 return new ICmpInst(*Context, ICmpInst::ICMP_SLE, LHS, RHS);
3421 return new ICmpInst(*Context, ICmpInst::ICMP_ULE, LHS, RHS);
3422 case 7: return Context->getConstantIntTrue();
3426 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3427 /// opcode and two operands into either a FCmp instruction. isordered is passed
3428 /// in to determine which kind of predicate to use in the new fcmp instruction.
3429 static Value *getFCmpValue(bool isordered, unsigned code,
3430 Value *LHS, Value *RHS, LLVMContext *Context) {
3432 default: llvm_unreachable("Illegal FCmp code!");
3435 return new FCmpInst(*Context, FCmpInst::FCMP_ORD, LHS, RHS);
3437 return new FCmpInst(*Context, FCmpInst::FCMP_UNO, LHS, RHS);
3440 return new FCmpInst(*Context, FCmpInst::FCMP_OGT, LHS, RHS);
3442 return new FCmpInst(*Context, FCmpInst::FCMP_UGT, LHS, RHS);
3445 return new FCmpInst(*Context, FCmpInst::FCMP_OEQ, LHS, RHS);
3447 return new FCmpInst(*Context, FCmpInst::FCMP_UEQ, LHS, RHS);
3450 return new FCmpInst(*Context, FCmpInst::FCMP_OGE, LHS, RHS);
3452 return new FCmpInst(*Context, FCmpInst::FCMP_UGE, LHS, RHS);
3455 return new FCmpInst(*Context, FCmpInst::FCMP_OLT, LHS, RHS);
3457 return new FCmpInst(*Context, FCmpInst::FCMP_ULT, LHS, RHS);
3460 return new FCmpInst(*Context, FCmpInst::FCMP_ONE, LHS, RHS);
3462 return new FCmpInst(*Context, FCmpInst::FCMP_UNE, LHS, RHS);
3465 return new FCmpInst(*Context, FCmpInst::FCMP_OLE, LHS, RHS);
3467 return new FCmpInst(*Context, FCmpInst::FCMP_ULE, LHS, RHS);
3468 case 7: return Context->getConstantIntTrue();
3472 /// PredicatesFoldable - Return true if both predicates match sign or if at
3473 /// least one of them is an equality comparison (which is signless).
3474 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3475 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3476 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3477 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3481 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3482 struct FoldICmpLogical {
3485 ICmpInst::Predicate pred;
3486 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3487 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3488 pred(ICI->getPredicate()) {}
3489 bool shouldApply(Value *V) const {
3490 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3491 if (PredicatesFoldable(pred, ICI->getPredicate()))
3492 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3493 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3496 Instruction *apply(Instruction &Log) const {
3497 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3498 if (ICI->getOperand(0) != LHS) {
3499 assert(ICI->getOperand(1) == LHS);
3500 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3503 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3504 unsigned LHSCode = getICmpCode(ICI);
3505 unsigned RHSCode = getICmpCode(RHSICI);
3507 switch (Log.getOpcode()) {
3508 case Instruction::And: Code = LHSCode & RHSCode; break;
3509 case Instruction::Or: Code = LHSCode | RHSCode; break;
3510 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3511 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3514 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3515 ICmpInst::isSignedPredicate(ICI->getPredicate());
3517 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3518 if (Instruction *I = dyn_cast<Instruction>(RV))
3520 // Otherwise, it's a constant boolean value...
3521 return IC.ReplaceInstUsesWith(Log, RV);
3524 } // end anonymous namespace
3526 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3527 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3528 // guaranteed to be a binary operator.
3529 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3531 ConstantInt *AndRHS,
3532 BinaryOperator &TheAnd) {
3533 Value *X = Op->getOperand(0);
3534 Constant *Together = 0;
3536 Together = Context->getConstantExprAnd(AndRHS, OpRHS);
3538 switch (Op->getOpcode()) {
3539 case Instruction::Xor:
3540 if (Op->hasOneUse()) {
3541 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3542 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3543 InsertNewInstBefore(And, TheAnd);
3545 return BinaryOperator::CreateXor(And, Together);
3548 case Instruction::Or:
3549 if (Together == AndRHS) // (X | C) & C --> C
3550 return ReplaceInstUsesWith(TheAnd, AndRHS);
3552 if (Op->hasOneUse() && Together != OpRHS) {
3553 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3554 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3555 InsertNewInstBefore(Or, TheAnd);
3557 return BinaryOperator::CreateAnd(Or, AndRHS);
3560 case Instruction::Add:
3561 if (Op->hasOneUse()) {
3562 // Adding a one to a single bit bit-field should be turned into an XOR
3563 // of the bit. First thing to check is to see if this AND is with a
3564 // single bit constant.
3565 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3567 // If there is only one bit set...
3568 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3569 // Ok, at this point, we know that we are masking the result of the
3570 // ADD down to exactly one bit. If the constant we are adding has
3571 // no bits set below this bit, then we can eliminate the ADD.
3572 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3574 // Check to see if any bits below the one bit set in AndRHSV are set.
3575 if ((AddRHS & (AndRHSV-1)) == 0) {
3576 // If not, the only thing that can effect the output of the AND is
3577 // the bit specified by AndRHSV. If that bit is set, the effect of
3578 // the XOR is to toggle the bit. If it is clear, then the ADD has
3580 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3581 TheAnd.setOperand(0, X);
3584 // Pull the XOR out of the AND.
3585 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3586 InsertNewInstBefore(NewAnd, TheAnd);
3587 NewAnd->takeName(Op);
3588 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3595 case Instruction::Shl: {
3596 // We know that the AND will not produce any of the bits shifted in, so if
3597 // the anded constant includes them, clear them now!
3599 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3600 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3601 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3602 ConstantInt *CI = Context->getConstantInt(AndRHS->getValue() & ShlMask);
3604 if (CI->getValue() == ShlMask) {
3605 // Masking out bits that the shift already masks
3606 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3607 } else if (CI != AndRHS) { // Reducing bits set in and.
3608 TheAnd.setOperand(1, CI);
3613 case Instruction::LShr:
3615 // We know that the AND will not produce any of the bits shifted in, so if
3616 // the anded constant includes them, clear them now! This only applies to
3617 // unsigned shifts, because a signed shr may bring in set bits!
3619 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3620 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3621 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3622 ConstantInt *CI = Context->getConstantInt(AndRHS->getValue() & ShrMask);
3624 if (CI->getValue() == ShrMask) {
3625 // Masking out bits that the shift already masks.
3626 return ReplaceInstUsesWith(TheAnd, Op);
3627 } else if (CI != AndRHS) {
3628 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3633 case Instruction::AShr:
3635 // See if this is shifting in some sign extension, then masking it out
3637 if (Op->hasOneUse()) {
3638 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3639 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3640 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3641 Constant *C = Context->getConstantInt(AndRHS->getValue() & ShrMask);
3642 if (C == AndRHS) { // Masking out bits shifted in.
3643 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3644 // Make the argument unsigned.
3645 Value *ShVal = Op->getOperand(0);
3646 ShVal = InsertNewInstBefore(
3647 BinaryOperator::CreateLShr(ShVal, OpRHS,
3648 Op->getName()), TheAnd);
3649 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3658 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3659 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3660 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3661 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3662 /// insert new instructions.
3663 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3664 bool isSigned, bool Inside,
3666 assert(cast<ConstantInt>(Context->getConstantExprICmp((isSigned ?
3667 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3668 "Lo is not <= Hi in range emission code!");
3671 if (Lo == Hi) // Trivially false.
3672 return new ICmpInst(*Context, ICmpInst::ICMP_NE, V, V);
3674 // V >= Min && V < Hi --> V < Hi
3675 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3676 ICmpInst::Predicate pred = (isSigned ?
3677 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3678 return new ICmpInst(*Context, pred, V, Hi);
3681 // Emit V-Lo <u Hi-Lo
3682 Constant *NegLo = Context->getConstantExprNeg(Lo);
3683 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3684 InsertNewInstBefore(Add, IB);
3685 Constant *UpperBound = Context->getConstantExprAdd(NegLo, Hi);
3686 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, UpperBound);
3689 if (Lo == Hi) // Trivially true.
3690 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, V, V);
3692 // V < Min || V >= Hi -> V > Hi-1
3693 Hi = SubOne(cast<ConstantInt>(Hi), Context);
3694 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3695 ICmpInst::Predicate pred = (isSigned ?
3696 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3697 return new ICmpInst(*Context, pred, V, Hi);
3700 // Emit V-Lo >u Hi-1-Lo
3701 // Note that Hi has already had one subtracted from it, above.
3702 ConstantInt *NegLo = cast<ConstantInt>(Context->getConstantExprNeg(Lo));
3703 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3704 InsertNewInstBefore(Add, IB);
3705 Constant *LowerBound = Context->getConstantExprAdd(NegLo, Hi);
3706 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add, LowerBound);
3709 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3710 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3711 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3712 // not, since all 1s are not contiguous.
3713 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3714 const APInt& V = Val->getValue();
3715 uint32_t BitWidth = Val->getType()->getBitWidth();
3716 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3718 // look for the first zero bit after the run of ones
3719 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3720 // look for the first non-zero bit
3721 ME = V.getActiveBits();
3725 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3726 /// where isSub determines whether the operator is a sub. If we can fold one of
3727 /// the following xforms:
3729 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3730 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3731 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3733 /// return (A +/- B).
3735 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3736 ConstantInt *Mask, bool isSub,
3738 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3739 if (!LHSI || LHSI->getNumOperands() != 2 ||
3740 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3742 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3744 switch (LHSI->getOpcode()) {
3746 case Instruction::And:
3747 if (Context->getConstantExprAnd(N, Mask) == Mask) {
3748 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3749 if ((Mask->getValue().countLeadingZeros() +
3750 Mask->getValue().countPopulation()) ==
3751 Mask->getValue().getBitWidth())
3754 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3755 // part, we don't need any explicit masks to take them out of A. If that
3756 // is all N is, ignore it.
3757 uint32_t MB = 0, ME = 0;
3758 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3759 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3760 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3761 if (MaskedValueIsZero(RHS, Mask))
3766 case Instruction::Or:
3767 case Instruction::Xor:
3768 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3769 if ((Mask->getValue().countLeadingZeros() +
3770 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3771 && Context->getConstantExprAnd(N, Mask)->isNullValue())
3778 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3780 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3781 return InsertNewInstBefore(New, I);
3784 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3785 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3786 ICmpInst *LHS, ICmpInst *RHS) {
3788 ConstantInt *LHSCst, *RHSCst;
3789 ICmpInst::Predicate LHSCC, RHSCC;
3791 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3792 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3793 m_ConstantInt(LHSCst)), *Context) ||
3794 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3795 m_ConstantInt(RHSCst)), *Context))
3798 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3799 // where C is a power of 2
3800 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3801 LHSCst->getValue().isPowerOf2()) {
3802 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3803 InsertNewInstBefore(NewOr, I);
3804 return new ICmpInst(*Context, LHSCC, NewOr, LHSCst);
3807 // From here on, we only handle:
3808 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3809 if (Val != Val2) return 0;
3811 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3812 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3813 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3814 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3815 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3818 // We can't fold (ugt x, C) & (sgt x, C2).
3819 if (!PredicatesFoldable(LHSCC, RHSCC))
3822 // Ensure that the larger constant is on the RHS.
3824 if (ICmpInst::isSignedPredicate(LHSCC) ||
3825 (ICmpInst::isEquality(LHSCC) &&
3826 ICmpInst::isSignedPredicate(RHSCC)))
3827 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3829 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3832 std::swap(LHS, RHS);
3833 std::swap(LHSCst, RHSCst);
3834 std::swap(LHSCC, RHSCC);
3837 // At this point, we know we have have two icmp instructions
3838 // comparing a value against two constants and and'ing the result
3839 // together. Because of the above check, we know that we only have
3840 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3841 // (from the FoldICmpLogical check above), that the two constants
3842 // are not equal and that the larger constant is on the RHS
3843 assert(LHSCst != RHSCst && "Compares not folded above?");
3846 default: llvm_unreachable("Unknown integer condition code!");
3847 case ICmpInst::ICMP_EQ:
3849 default: llvm_unreachable("Unknown integer condition code!");
3850 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3851 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3852 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3853 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3854 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3855 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3856 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3857 return ReplaceInstUsesWith(I, LHS);
3859 case ICmpInst::ICMP_NE:
3861 default: llvm_unreachable("Unknown integer condition code!");
3862 case ICmpInst::ICMP_ULT:
3863 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X u< 14) -> X < 13
3864 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Val, LHSCst);
3865 break; // (X != 13 & X u< 15) -> no change
3866 case ICmpInst::ICMP_SLT:
3867 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X s< 14) -> X < 13
3868 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Val, LHSCst);
3869 break; // (X != 13 & X s< 15) -> no change
3870 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3871 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3872 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3873 return ReplaceInstUsesWith(I, RHS);
3874 case ICmpInst::ICMP_NE:
3875 if (LHSCst == SubOne(RHSCst, Context)){// (X != 13 & X != 14) -> X-13 >u 1
3876 Constant *AddCST = Context->getConstantExprNeg(LHSCst);
3877 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3878 Val->getName()+".off");
3879 InsertNewInstBefore(Add, I);
3880 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add,
3881 Context->getConstantInt(Add->getType(), 1));
3883 break; // (X != 13 & X != 15) -> no change
3886 case ICmpInst::ICMP_ULT:
3888 default: llvm_unreachable("Unknown integer condition code!");
3889 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3890 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3891 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3892 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3894 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3895 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3896 return ReplaceInstUsesWith(I, LHS);
3897 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3901 case ICmpInst::ICMP_SLT:
3903 default: llvm_unreachable("Unknown integer condition code!");
3904 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3905 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3906 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3907 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3909 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3910 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3911 return ReplaceInstUsesWith(I, LHS);
3912 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3916 case ICmpInst::ICMP_UGT:
3918 default: llvm_unreachable("Unknown integer condition code!");
3919 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3920 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3921 return ReplaceInstUsesWith(I, RHS);
3922 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3924 case ICmpInst::ICMP_NE:
3925 if (RHSCst == AddOne(LHSCst, Context)) // (X u> 13 & X != 14) -> X u> 14
3926 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3927 break; // (X u> 13 & X != 15) -> no change
3928 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3929 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3930 RHSCst, false, true, I);
3931 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3935 case ICmpInst::ICMP_SGT:
3937 default: llvm_unreachable("Unknown integer condition code!");
3938 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3939 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3940 return ReplaceInstUsesWith(I, RHS);
3941 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3943 case ICmpInst::ICMP_NE:
3944 if (RHSCst == AddOne(LHSCst, Context)) // (X s> 13 & X != 14) -> X s> 14
3945 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3946 break; // (X s> 13 & X != 15) -> no change
3947 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3948 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3949 RHSCst, true, true, I);
3950 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3960 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3961 bool Changed = SimplifyCommutative(I);
3962 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3964 if (isa<UndefValue>(Op1)) // X & undef -> 0
3965 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3969 return ReplaceInstUsesWith(I, Op1);
3971 // See if we can simplify any instructions used by the instruction whose sole
3972 // purpose is to compute bits we don't care about.
3973 if (SimplifyDemandedInstructionBits(I))
3975 if (isa<VectorType>(I.getType())) {
3976 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3977 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3978 return ReplaceInstUsesWith(I, I.getOperand(0));
3979 } else if (isa<ConstantAggregateZero>(Op1)) {
3980 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3984 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3985 const APInt& AndRHSMask = AndRHS->getValue();
3986 APInt NotAndRHS(~AndRHSMask);
3988 // Optimize a variety of ((val OP C1) & C2) combinations...
3989 if (isa<BinaryOperator>(Op0)) {
3990 Instruction *Op0I = cast<Instruction>(Op0);
3991 Value *Op0LHS = Op0I->getOperand(0);
3992 Value *Op0RHS = Op0I->getOperand(1);
3993 switch (Op0I->getOpcode()) {
3994 case Instruction::Xor:
3995 case Instruction::Or:
3996 // If the mask is only needed on one incoming arm, push it up.
3997 if (Op0I->hasOneUse()) {
3998 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3999 // Not masking anything out for the LHS, move to RHS.
4000 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4001 Op0RHS->getName()+".masked");
4002 InsertNewInstBefore(NewRHS, I);
4003 return BinaryOperator::Create(
4004 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4006 if (!isa<Constant>(Op0RHS) &&
4007 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4008 // Not masking anything out for the RHS, move to LHS.
4009 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4010 Op0LHS->getName()+".masked");
4011 InsertNewInstBefore(NewLHS, I);
4012 return BinaryOperator::Create(
4013 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4018 case Instruction::Add:
4019 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4020 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4021 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4022 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4023 return BinaryOperator::CreateAnd(V, AndRHS);
4024 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4025 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4028 case Instruction::Sub:
4029 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4030 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4031 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4032 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4033 return BinaryOperator::CreateAnd(V, AndRHS);
4035 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4036 // has 1's for all bits that the subtraction with A might affect.
4037 if (Op0I->hasOneUse()) {
4038 uint32_t BitWidth = AndRHSMask.getBitWidth();
4039 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4040 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4042 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4043 if (!(A && A->isZero()) && // avoid infinite recursion.
4044 MaskedValueIsZero(Op0LHS, Mask)) {
4045 Instruction *NewNeg = BinaryOperator::CreateNeg(*Context, Op0RHS);
4046 InsertNewInstBefore(NewNeg, I);
4047 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4052 case Instruction::Shl:
4053 case Instruction::LShr:
4054 // (1 << x) & 1 --> zext(x == 0)
4055 // (1 >> x) & 1 --> zext(x == 0)
4056 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4057 Instruction *NewICmp = new ICmpInst(*Context, ICmpInst::ICMP_EQ,
4058 Op0RHS, Context->getNullValue(I.getType()));
4059 InsertNewInstBefore(NewICmp, I);
4060 return new ZExtInst(NewICmp, I.getType());
4065 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4066 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4068 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4069 // If this is an integer truncation or change from signed-to-unsigned, and
4070 // if the source is an and/or with immediate, transform it. This
4071 // frequently occurs for bitfield accesses.
4072 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4073 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4074 CastOp->getNumOperands() == 2)
4075 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4076 if (CastOp->getOpcode() == Instruction::And) {
4077 // Change: and (cast (and X, C1) to T), C2
4078 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4079 // This will fold the two constants together, which may allow
4080 // other simplifications.
4081 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4082 CastOp->getOperand(0), I.getType(),
4083 CastOp->getName()+".shrunk");
4084 NewCast = InsertNewInstBefore(NewCast, I);
4085 // trunc_or_bitcast(C1)&C2
4087 Context->getConstantExprTruncOrBitCast(AndCI,I.getType());
4088 C3 = Context->getConstantExprAnd(C3, AndRHS);
4089 return BinaryOperator::CreateAnd(NewCast, C3);
4090 } else if (CastOp->getOpcode() == Instruction::Or) {
4091 // Change: and (cast (or X, C1) to T), C2
4092 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4094 Context->getConstantExprTruncOrBitCast(AndCI,I.getType());
4095 if (Context->getConstantExprAnd(C3, AndRHS) == AndRHS)
4097 return ReplaceInstUsesWith(I, AndRHS);
4103 // Try to fold constant and into select arguments.
4104 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4105 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4107 if (isa<PHINode>(Op0))
4108 if (Instruction *NV = FoldOpIntoPhi(I))
4112 Value *Op0NotVal = dyn_castNotVal(Op0, Context);
4113 Value *Op1NotVal = dyn_castNotVal(Op1, Context);
4115 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4116 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
4118 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4119 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4120 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4121 I.getName()+".demorgan");
4122 InsertNewInstBefore(Or, I);
4123 return BinaryOperator::CreateNot(*Context, Or);
4127 Value *A = 0, *B = 0, *C = 0, *D = 0;
4128 if (match(Op0, m_Or(m_Value(A), m_Value(B)), *Context)) {
4129 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4130 return ReplaceInstUsesWith(I, Op1);
4132 // (A|B) & ~(A&B) -> A^B
4133 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4134 if ((A == C && B == D) || (A == D && B == C))
4135 return BinaryOperator::CreateXor(A, B);
4139 if (match(Op1, m_Or(m_Value(A), m_Value(B)), *Context)) {
4140 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4141 return ReplaceInstUsesWith(I, Op0);
4143 // ~(A&B) & (A|B) -> A^B
4144 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4145 if ((A == C && B == D) || (A == D && B == C))
4146 return BinaryOperator::CreateXor(A, B);
4150 if (Op0->hasOneUse() &&
4151 match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4152 if (A == Op1) { // (A^B)&A -> A&(A^B)
4153 I.swapOperands(); // Simplify below
4154 std::swap(Op0, Op1);
4155 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4156 cast<BinaryOperator>(Op0)->swapOperands();
4157 I.swapOperands(); // Simplify below
4158 std::swap(Op0, Op1);
4162 if (Op1->hasOneUse() &&
4163 match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4164 if (B == Op0) { // B&(A^B) -> B&(B^A)
4165 cast<BinaryOperator>(Op1)->swapOperands();
4168 if (A == Op0) { // A&(A^B) -> A & ~B
4169 Instruction *NotB = BinaryOperator::CreateNot(*Context, B, "tmp");
4170 InsertNewInstBefore(NotB, I);
4171 return BinaryOperator::CreateAnd(A, NotB);
4175 // (A&((~A)|B)) -> A&B
4176 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A)), *Context) ||
4177 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1))), *Context))
4178 return BinaryOperator::CreateAnd(A, Op1);
4179 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A)), *Context) ||
4180 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0))), *Context))
4181 return BinaryOperator::CreateAnd(A, Op0);
4184 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4185 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4186 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4189 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4190 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4194 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4195 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4196 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4197 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4198 const Type *SrcTy = Op0C->getOperand(0)->getType();
4199 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4200 // Only do this if the casts both really cause code to be generated.
4201 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4203 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4205 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4206 Op1C->getOperand(0),
4208 InsertNewInstBefore(NewOp, I);
4209 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4213 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4214 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4215 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4216 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4217 SI0->getOperand(1) == SI1->getOperand(1) &&
4218 (SI0->hasOneUse() || SI1->hasOneUse())) {
4219 Instruction *NewOp =
4220 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4222 SI0->getName()), I);
4223 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4224 SI1->getOperand(1));
4228 // If and'ing two fcmp, try combine them into one.
4229 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4230 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4231 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4232 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4233 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4234 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4235 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4236 // If either of the constants are nans, then the whole thing returns
4238 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4239 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4240 return new FCmpInst(*Context, FCmpInst::FCMP_ORD,
4241 LHS->getOperand(0), RHS->getOperand(0));
4244 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4245 FCmpInst::Predicate Op0CC, Op1CC;
4246 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS),
4247 m_Value(Op0RHS)), *Context) &&
4248 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS),
4249 m_Value(Op1RHS)), *Context)) {
4250 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4251 // Swap RHS operands to match LHS.
4252 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4253 std::swap(Op1LHS, Op1RHS);
4255 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4256 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4258 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4260 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4261 Op1CC == FCmpInst::FCMP_FALSE)
4262 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4263 else if (Op0CC == FCmpInst::FCMP_TRUE)
4264 return ReplaceInstUsesWith(I, Op1);
4265 else if (Op1CC == FCmpInst::FCMP_TRUE)
4266 return ReplaceInstUsesWith(I, Op0);
4269 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4270 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4272 std::swap(Op0, Op1);
4273 std::swap(Op0Pred, Op1Pred);
4274 std::swap(Op0Ordered, Op1Ordered);
4277 // uno && ueq -> uno && (uno || eq) -> ueq
4278 // ord && olt -> ord && (ord && lt) -> olt
4279 if (Op0Ordered == Op1Ordered)
4280 return ReplaceInstUsesWith(I, Op1);
4281 // uno && oeq -> uno && (ord && eq) -> false
4282 // uno && ord -> false
4284 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4285 // ord && ueq -> ord && (uno || eq) -> oeq
4286 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4287 Op0LHS, Op0RHS, Context));
4295 return Changed ? &I : 0;
4298 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4299 /// capable of providing pieces of a bswap. The subexpression provides pieces
4300 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4301 /// the expression came from the corresponding "byte swapped" byte in some other
4302 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4303 /// we know that the expression deposits the low byte of %X into the high byte
4304 /// of the bswap result and that all other bytes are zero. This expression is
4305 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4308 /// This function returns true if the match was unsuccessful and false if so.
4309 /// On entry to the function the "OverallLeftShift" is a signed integer value
4310 /// indicating the number of bytes that the subexpression is later shifted. For
4311 /// example, if the expression is later right shifted by 16 bits, the
4312 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4313 /// byte of ByteValues is actually being set.
4315 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4316 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4317 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4318 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4319 /// always in the local (OverallLeftShift) coordinate space.
4321 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4322 SmallVector<Value*, 8> &ByteValues) {
4323 if (Instruction *I = dyn_cast<Instruction>(V)) {
4324 // If this is an or instruction, it may be an inner node of the bswap.
4325 if (I->getOpcode() == Instruction::Or) {
4326 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4328 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4332 // If this is a logical shift by a constant multiple of 8, recurse with
4333 // OverallLeftShift and ByteMask adjusted.
4334 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4336 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4337 // Ensure the shift amount is defined and of a byte value.
4338 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4341 unsigned ByteShift = ShAmt >> 3;
4342 if (I->getOpcode() == Instruction::Shl) {
4343 // X << 2 -> collect(X, +2)
4344 OverallLeftShift += ByteShift;
4345 ByteMask >>= ByteShift;
4347 // X >>u 2 -> collect(X, -2)
4348 OverallLeftShift -= ByteShift;
4349 ByteMask <<= ByteShift;
4350 ByteMask &= (~0U >> (32-ByteValues.size()));
4353 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4354 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4356 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4360 // If this is a logical 'and' with a mask that clears bytes, clear the
4361 // corresponding bytes in ByteMask.
4362 if (I->getOpcode() == Instruction::And &&
4363 isa<ConstantInt>(I->getOperand(1))) {
4364 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4365 unsigned NumBytes = ByteValues.size();
4366 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4367 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4369 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4370 // If this byte is masked out by a later operation, we don't care what
4372 if ((ByteMask & (1 << i)) == 0)
4375 // If the AndMask is all zeros for this byte, clear the bit.
4376 APInt MaskB = AndMask & Byte;
4378 ByteMask &= ~(1U << i);
4382 // If the AndMask is not all ones for this byte, it's not a bytezap.
4386 // Otherwise, this byte is kept.
4389 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4394 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4395 // the input value to the bswap. Some observations: 1) if more than one byte
4396 // is demanded from this input, then it could not be successfully assembled
4397 // into a byteswap. At least one of the two bytes would not be aligned with
4398 // their ultimate destination.
4399 if (!isPowerOf2_32(ByteMask)) return true;
4400 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4402 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4403 // is demanded, it needs to go into byte 0 of the result. This means that the
4404 // byte needs to be shifted until it lands in the right byte bucket. The
4405 // shift amount depends on the position: if the byte is coming from the high
4406 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4407 // low part, it must be shifted left.
4408 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4409 if (InputByteNo < ByteValues.size()/2) {
4410 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4413 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4417 // If the destination byte value is already defined, the values are or'd
4418 // together, which isn't a bswap (unless it's an or of the same bits).
4419 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4421 ByteValues[DestByteNo] = V;
4425 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4426 /// If so, insert the new bswap intrinsic and return it.
4427 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4428 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4429 if (!ITy || ITy->getBitWidth() % 16 ||
4430 // ByteMask only allows up to 32-byte values.
4431 ITy->getBitWidth() > 32*8)
4432 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4434 /// ByteValues - For each byte of the result, we keep track of which value
4435 /// defines each byte.
4436 SmallVector<Value*, 8> ByteValues;
4437 ByteValues.resize(ITy->getBitWidth()/8);
4439 // Try to find all the pieces corresponding to the bswap.
4440 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4441 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4444 // Check to see if all of the bytes come from the same value.
4445 Value *V = ByteValues[0];
4446 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4448 // Check to make sure that all of the bytes come from the same value.
4449 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4450 if (ByteValues[i] != V)
4452 const Type *Tys[] = { ITy };
4453 Module *M = I.getParent()->getParent()->getParent();
4454 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4455 return CallInst::Create(F, V);
4458 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4459 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4460 /// we can simplify this expression to "cond ? C : D or B".
4461 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4463 LLVMContext *Context) {
4464 // If A is not a select of -1/0, this cannot match.
4466 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond)), *Context))
4469 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4470 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4471 return SelectInst::Create(Cond, C, B);
4472 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4473 return SelectInst::Create(Cond, C, B);
4474 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4475 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4476 return SelectInst::Create(Cond, C, D);
4477 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4478 return SelectInst::Create(Cond, C, D);
4482 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4483 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4484 ICmpInst *LHS, ICmpInst *RHS) {
4486 ConstantInt *LHSCst, *RHSCst;
4487 ICmpInst::Predicate LHSCC, RHSCC;
4489 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4490 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4491 m_ConstantInt(LHSCst)), *Context) ||
4492 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4493 m_ConstantInt(RHSCst)), *Context))
4496 // From here on, we only handle:
4497 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4498 if (Val != Val2) return 0;
4500 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4501 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4502 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4503 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4504 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4507 // We can't fold (ugt x, C) | (sgt x, C2).
4508 if (!PredicatesFoldable(LHSCC, RHSCC))
4511 // Ensure that the larger constant is on the RHS.
4513 if (ICmpInst::isSignedPredicate(LHSCC) ||
4514 (ICmpInst::isEquality(LHSCC) &&
4515 ICmpInst::isSignedPredicate(RHSCC)))
4516 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4518 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4521 std::swap(LHS, RHS);
4522 std::swap(LHSCst, RHSCst);
4523 std::swap(LHSCC, RHSCC);
4526 // At this point, we know we have have two icmp instructions
4527 // comparing a value against two constants and or'ing the result
4528 // together. Because of the above check, we know that we only have
4529 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4530 // FoldICmpLogical check above), that the two constants are not
4532 assert(LHSCst != RHSCst && "Compares not folded above?");
4535 default: llvm_unreachable("Unknown integer condition code!");
4536 case ICmpInst::ICMP_EQ:
4538 default: llvm_unreachable("Unknown integer condition code!");
4539 case ICmpInst::ICMP_EQ:
4540 if (LHSCst == SubOne(RHSCst, Context)) {
4541 // (X == 13 | X == 14) -> X-13 <u 2
4542 Constant *AddCST = Context->getConstantExprNeg(LHSCst);
4543 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4544 Val->getName()+".off");
4545 InsertNewInstBefore(Add, I);
4546 AddCST = Context->getConstantExprSub(AddOne(RHSCst, Context), LHSCst);
4547 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, AddCST);
4549 break; // (X == 13 | X == 15) -> no change
4550 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4551 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4553 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4554 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4555 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4556 return ReplaceInstUsesWith(I, RHS);
4559 case ICmpInst::ICMP_NE:
4561 default: llvm_unreachable("Unknown integer condition code!");
4562 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4563 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4564 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4565 return ReplaceInstUsesWith(I, LHS);
4566 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4567 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4568 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4569 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4572 case ICmpInst::ICMP_ULT:
4574 default: llvm_unreachable("Unknown integer condition code!");
4575 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4577 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4578 // If RHSCst is [us]MAXINT, it is always false. Not handling
4579 // this can cause overflow.
4580 if (RHSCst->isMaxValue(false))
4581 return ReplaceInstUsesWith(I, LHS);
4582 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4584 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4586 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4587 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4588 return ReplaceInstUsesWith(I, RHS);
4589 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4593 case ICmpInst::ICMP_SLT:
4595 default: llvm_unreachable("Unknown integer condition code!");
4596 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4598 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4599 // If RHSCst is [us]MAXINT, it is always false. Not handling
4600 // this can cause overflow.
4601 if (RHSCst->isMaxValue(true))
4602 return ReplaceInstUsesWith(I, LHS);
4603 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4605 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4607 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4608 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4609 return ReplaceInstUsesWith(I, RHS);
4610 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4614 case ICmpInst::ICMP_UGT:
4616 default: llvm_unreachable("Unknown integer condition code!");
4617 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4618 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4619 return ReplaceInstUsesWith(I, LHS);
4620 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4622 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4623 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4624 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4625 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4629 case ICmpInst::ICMP_SGT:
4631 default: llvm_unreachable("Unknown integer condition code!");
4632 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4633 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4634 return ReplaceInstUsesWith(I, LHS);
4635 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4637 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4638 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4639 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4640 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4648 /// FoldOrWithConstants - This helper function folds:
4650 /// ((A | B) & C1) | (B & C2)
4656 /// when the XOR of the two constants is "all ones" (-1).
4657 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4658 Value *A, Value *B, Value *C) {
4659 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4663 ConstantInt *CI2 = 0;
4664 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)), *Context)) return 0;
4666 APInt Xor = CI1->getValue() ^ CI2->getValue();
4667 if (!Xor.isAllOnesValue()) return 0;
4669 if (V1 == A || V1 == B) {
4670 Instruction *NewOp =
4671 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4672 return BinaryOperator::CreateOr(NewOp, V1);
4678 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4679 bool Changed = SimplifyCommutative(I);
4680 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4682 if (isa<UndefValue>(Op1)) // X | undef -> -1
4683 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4687 return ReplaceInstUsesWith(I, Op0);
4689 // See if we can simplify any instructions used by the instruction whose sole
4690 // purpose is to compute bits we don't care about.
4691 if (SimplifyDemandedInstructionBits(I))
4693 if (isa<VectorType>(I.getType())) {
4694 if (isa<ConstantAggregateZero>(Op1)) {
4695 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4696 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4697 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4698 return ReplaceInstUsesWith(I, I.getOperand(1));
4703 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4704 ConstantInt *C1 = 0; Value *X = 0;
4705 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4706 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1)), *Context) &&
4708 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4709 InsertNewInstBefore(Or, I);
4711 return BinaryOperator::CreateAnd(Or,
4712 Context->getConstantInt(RHS->getValue() | C1->getValue()));
4715 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4716 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1)), *Context) &&
4718 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4719 InsertNewInstBefore(Or, I);
4721 return BinaryOperator::CreateXor(Or,
4722 Context->getConstantInt(C1->getValue() & ~RHS->getValue()));
4725 // Try to fold constant and into select arguments.
4726 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4727 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4729 if (isa<PHINode>(Op0))
4730 if (Instruction *NV = FoldOpIntoPhi(I))
4734 Value *A = 0, *B = 0;
4735 ConstantInt *C1 = 0, *C2 = 0;
4737 if (match(Op0, m_And(m_Value(A), m_Value(B)), *Context))
4738 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4739 return ReplaceInstUsesWith(I, Op1);
4740 if (match(Op1, m_And(m_Value(A), m_Value(B)), *Context))
4741 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4742 return ReplaceInstUsesWith(I, Op0);
4744 // (A | B) | C and A | (B | C) -> bswap if possible.
4745 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4746 if (match(Op0, m_Or(m_Value(), m_Value()), *Context) ||
4747 match(Op1, m_Or(m_Value(), m_Value()), *Context) ||
4748 (match(Op0, m_Shift(m_Value(), m_Value()), *Context) &&
4749 match(Op1, m_Shift(m_Value(), m_Value()), *Context))) {
4750 if (Instruction *BSwap = MatchBSwap(I))
4754 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4755 if (Op0->hasOneUse() &&
4756 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4757 MaskedValueIsZero(Op1, C1->getValue())) {
4758 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4759 InsertNewInstBefore(NOr, I);
4761 return BinaryOperator::CreateXor(NOr, C1);
4764 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4765 if (Op1->hasOneUse() &&
4766 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4767 MaskedValueIsZero(Op0, C1->getValue())) {
4768 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4769 InsertNewInstBefore(NOr, I);
4771 return BinaryOperator::CreateXor(NOr, C1);
4775 Value *C = 0, *D = 0;
4776 if (match(Op0, m_And(m_Value(A), m_Value(C)), *Context) &&
4777 match(Op1, m_And(m_Value(B), m_Value(D)), *Context)) {
4778 Value *V1 = 0, *V2 = 0, *V3 = 0;
4779 C1 = dyn_cast<ConstantInt>(C);
4780 C2 = dyn_cast<ConstantInt>(D);
4781 if (C1 && C2) { // (A & C1)|(B & C2)
4782 // If we have: ((V + N) & C1) | (V & C2)
4783 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4784 // replace with V+N.
4785 if (C1->getValue() == ~C2->getValue()) {
4786 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4787 match(A, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4788 // Add commutes, try both ways.
4789 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4790 return ReplaceInstUsesWith(I, A);
4791 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4792 return ReplaceInstUsesWith(I, A);
4794 // Or commutes, try both ways.
4795 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4796 match(B, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4797 // Add commutes, try both ways.
4798 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4799 return ReplaceInstUsesWith(I, B);
4800 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4801 return ReplaceInstUsesWith(I, B);
4804 V1 = 0; V2 = 0; V3 = 0;
4807 // Check to see if we have any common things being and'ed. If so, find the
4808 // terms for V1 & (V2|V3).
4809 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4810 if (A == B) // (A & C)|(A & D) == A & (C|D)
4811 V1 = A, V2 = C, V3 = D;
4812 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4813 V1 = A, V2 = B, V3 = C;
4814 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4815 V1 = C, V2 = A, V3 = D;
4816 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4817 V1 = C, V2 = A, V3 = B;
4821 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4822 return BinaryOperator::CreateAnd(V1, Or);
4826 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4827 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4829 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4831 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4833 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4836 // ((A&~B)|(~A&B)) -> A^B
4837 if ((match(C, m_Not(m_Specific(D)), *Context) &&
4838 match(B, m_Not(m_Specific(A)), *Context)))
4839 return BinaryOperator::CreateXor(A, D);
4840 // ((~B&A)|(~A&B)) -> A^B
4841 if ((match(A, m_Not(m_Specific(D)), *Context) &&
4842 match(B, m_Not(m_Specific(C)), *Context)))
4843 return BinaryOperator::CreateXor(C, D);
4844 // ((A&~B)|(B&~A)) -> A^B
4845 if ((match(C, m_Not(m_Specific(B)), *Context) &&
4846 match(D, m_Not(m_Specific(A)), *Context)))
4847 return BinaryOperator::CreateXor(A, B);
4848 // ((~B&A)|(B&~A)) -> A^B
4849 if ((match(A, m_Not(m_Specific(B)), *Context) &&
4850 match(D, m_Not(m_Specific(C)), *Context)))
4851 return BinaryOperator::CreateXor(C, B);
4854 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4855 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4856 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4857 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4858 SI0->getOperand(1) == SI1->getOperand(1) &&
4859 (SI0->hasOneUse() || SI1->hasOneUse())) {
4860 Instruction *NewOp =
4861 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4863 SI0->getName()), I);
4864 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4865 SI1->getOperand(1));
4869 // ((A|B)&1)|(B&-2) -> (A&1) | B
4870 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4871 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4872 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4873 if (Ret) return Ret;
4875 // (B&-2)|((A|B)&1) -> (A&1) | B
4876 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4877 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4878 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4879 if (Ret) return Ret;
4882 if (match(Op0, m_Not(m_Value(A)), *Context)) { // ~A | Op1
4883 if (A == Op1) // ~A | A == -1
4884 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4888 // Note, A is still live here!
4889 if (match(Op1, m_Not(m_Value(B)), *Context)) { // Op0 | ~B
4891 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4893 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4894 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4895 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4896 I.getName()+".demorgan"), I);
4897 return BinaryOperator::CreateNot(*Context, And);
4901 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4902 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4903 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4906 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4907 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4911 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4912 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4913 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4914 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4915 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4916 !isa<ICmpInst>(Op1C->getOperand(0))) {
4917 const Type *SrcTy = Op0C->getOperand(0)->getType();
4918 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4919 // Only do this if the casts both really cause code to be
4921 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4923 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4925 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4926 Op1C->getOperand(0),
4928 InsertNewInstBefore(NewOp, I);
4929 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4936 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4937 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4938 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4939 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4940 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4941 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4942 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4943 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4944 // If either of the constants are nans, then the whole thing returns
4946 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4947 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4949 // Otherwise, no need to compare the two constants, compare the
4951 return new FCmpInst(*Context, FCmpInst::FCMP_UNO,
4952 LHS->getOperand(0), RHS->getOperand(0));
4955 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4956 FCmpInst::Predicate Op0CC, Op1CC;
4957 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS),
4958 m_Value(Op0RHS)), *Context) &&
4959 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS),
4960 m_Value(Op1RHS)), *Context)) {
4961 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4962 // Swap RHS operands to match LHS.
4963 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4964 std::swap(Op1LHS, Op1RHS);
4966 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4967 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4969 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4971 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4972 Op1CC == FCmpInst::FCMP_TRUE)
4973 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4974 else if (Op0CC == FCmpInst::FCMP_FALSE)
4975 return ReplaceInstUsesWith(I, Op1);
4976 else if (Op1CC == FCmpInst::FCMP_FALSE)
4977 return ReplaceInstUsesWith(I, Op0);
4980 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4981 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4982 if (Op0Ordered == Op1Ordered) {
4983 // If both are ordered or unordered, return a new fcmp with
4984 // or'ed predicates.
4985 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4986 Op0LHS, Op0RHS, Context);
4987 if (Instruction *I = dyn_cast<Instruction>(RV))
4989 // Otherwise, it's a constant boolean value...
4990 return ReplaceInstUsesWith(I, RV);
4998 return Changed ? &I : 0;
5003 // XorSelf - Implements: X ^ X --> 0
5006 XorSelf(Value *rhs) : RHS(rhs) {}
5007 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5008 Instruction *apply(BinaryOperator &Xor) const {
5015 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5016 bool Changed = SimplifyCommutative(I);
5017 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5019 if (isa<UndefValue>(Op1)) {
5020 if (isa<UndefValue>(Op0))
5021 // Handle undef ^ undef -> 0 special case. This is a common
5023 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
5024 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5027 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5028 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1), Context)) {
5029 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5030 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
5033 // See if we can simplify any instructions used by the instruction whose sole
5034 // purpose is to compute bits we don't care about.
5035 if (SimplifyDemandedInstructionBits(I))
5037 if (isa<VectorType>(I.getType()))
5038 if (isa<ConstantAggregateZero>(Op1))
5039 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5041 // Is this a ~ operation?
5042 if (Value *NotOp = dyn_castNotVal(&I, Context)) {
5043 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5044 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5045 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5046 if (Op0I->getOpcode() == Instruction::And ||
5047 Op0I->getOpcode() == Instruction::Or) {
5048 if (dyn_castNotVal(Op0I->getOperand(1), Context)) Op0I->swapOperands();
5049 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0), Context)) {
5051 BinaryOperator::CreateNot(*Context, Op0I->getOperand(1),
5052 Op0I->getOperand(1)->getName()+".not");
5053 InsertNewInstBefore(NotY, I);
5054 if (Op0I->getOpcode() == Instruction::And)
5055 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5057 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5064 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5065 if (RHS == Context->getConstantIntTrue() && Op0->hasOneUse()) {
5066 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5067 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5068 return new ICmpInst(*Context, ICI->getInversePredicate(),
5069 ICI->getOperand(0), ICI->getOperand(1));
5071 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5072 return new FCmpInst(*Context, FCI->getInversePredicate(),
5073 FCI->getOperand(0), FCI->getOperand(1));
5076 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5077 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5078 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5079 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5080 Instruction::CastOps Opcode = Op0C->getOpcode();
5081 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
5082 if (RHS == Context->getConstantExprCast(Opcode,
5083 Context->getConstantIntTrue(),
5084 Op0C->getDestTy())) {
5085 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
5087 CI->getOpcode(), CI->getInversePredicate(),
5088 CI->getOperand(0), CI->getOperand(1)), I);
5089 NewCI->takeName(CI);
5090 return CastInst::Create(Opcode, NewCI, Op0C->getType());
5097 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5098 // ~(c-X) == X-c-1 == X+(-c-1)
5099 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5100 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5101 Constant *NegOp0I0C = Context->getConstantExprNeg(Op0I0C);
5102 Constant *ConstantRHS = Context->getConstantExprSub(NegOp0I0C,
5103 Context->getConstantInt(I.getType(), 1));
5104 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5107 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5108 if (Op0I->getOpcode() == Instruction::Add) {
5109 // ~(X-c) --> (-c-1)-X
5110 if (RHS->isAllOnesValue()) {
5111 Constant *NegOp0CI = Context->getConstantExprNeg(Op0CI);
5112 return BinaryOperator::CreateSub(
5113 Context->getConstantExprSub(NegOp0CI,
5114 Context->getConstantInt(I.getType(), 1)),
5115 Op0I->getOperand(0));
5116 } else if (RHS->getValue().isSignBit()) {
5117 // (X + C) ^ signbit -> (X + C + signbit)
5119 Context->getConstantInt(RHS->getValue() + Op0CI->getValue());
5120 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5123 } else if (Op0I->getOpcode() == Instruction::Or) {
5124 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5125 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5126 Constant *NewRHS = Context->getConstantExprOr(Op0CI, RHS);
5127 // Anything in both C1 and C2 is known to be zero, remove it from
5129 Constant *CommonBits = Context->getConstantExprAnd(Op0CI, RHS);
5130 NewRHS = Context->getConstantExprAnd(NewRHS,
5131 Context->getConstantExprNot(CommonBits));
5132 AddToWorkList(Op0I);
5133 I.setOperand(0, Op0I->getOperand(0));
5134 I.setOperand(1, NewRHS);
5141 // Try to fold constant and into select arguments.
5142 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5143 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5145 if (isa<PHINode>(Op0))
5146 if (Instruction *NV = FoldOpIntoPhi(I))
5150 if (Value *X = dyn_castNotVal(Op0, Context)) // ~A ^ A == -1
5152 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
5154 if (Value *X = dyn_castNotVal(Op1, Context)) // A ^ ~A == -1
5156 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
5159 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5162 if (match(Op1I, m_Or(m_Value(A), m_Value(B)), *Context)) {
5163 if (A == Op0) { // B^(B|A) == (A|B)^B
5164 Op1I->swapOperands();
5166 std::swap(Op0, Op1);
5167 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5168 I.swapOperands(); // Simplified below.
5169 std::swap(Op0, Op1);
5171 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)), *Context)) {
5172 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5173 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)), *Context)) {
5174 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5175 } else if (match(Op1I, m_And(m_Value(A), m_Value(B)), *Context) &&
5177 if (A == Op0) { // A^(A&B) -> A^(B&A)
5178 Op1I->swapOperands();
5181 if (B == Op0) { // A^(B&A) -> (B&A)^A
5182 I.swapOperands(); // Simplified below.
5183 std::swap(Op0, Op1);
5188 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5191 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5192 Op0I->hasOneUse()) {
5193 if (A == Op1) // (B|A)^B == (A|B)^B
5195 if (B == Op1) { // (A|B)^B == A & ~B
5197 InsertNewInstBefore(BinaryOperator::CreateNot(*Context,
5199 return BinaryOperator::CreateAnd(A, NotB);
5201 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)), *Context)) {
5202 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5203 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)), *Context)) {
5204 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5205 } else if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5207 if (A == Op1) // (A&B)^A -> (B&A)^A
5209 if (B == Op1 && // (B&A)^A == ~B & A
5210 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5212 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, A, "tmp"), I);
5213 return BinaryOperator::CreateAnd(N, Op1);
5218 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5219 if (Op0I && Op1I && Op0I->isShift() &&
5220 Op0I->getOpcode() == Op1I->getOpcode() &&
5221 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5222 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5223 Instruction *NewOp =
5224 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5225 Op1I->getOperand(0),
5226 Op0I->getName()), I);
5227 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5228 Op1I->getOperand(1));
5232 Value *A, *B, *C, *D;
5233 // (A & B)^(A | B) -> A ^ B
5234 if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5235 match(Op1I, m_Or(m_Value(C), m_Value(D)), *Context)) {
5236 if ((A == C && B == D) || (A == D && B == C))
5237 return BinaryOperator::CreateXor(A, B);
5239 // (A | B)^(A & B) -> A ^ B
5240 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5241 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5242 if ((A == C && B == D) || (A == D && B == C))
5243 return BinaryOperator::CreateXor(A, B);
5247 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5248 match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5249 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5250 // (X & Y)^(X & Y) -> (Y^Z) & X
5251 Value *X = 0, *Y = 0, *Z = 0;
5253 X = A, Y = B, Z = D;
5255 X = A, Y = B, Z = C;
5257 X = B, Y = A, Z = D;
5259 X = B, Y = A, Z = C;
5262 Instruction *NewOp =
5263 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5264 return BinaryOperator::CreateAnd(NewOp, X);
5269 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5270 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5271 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
5274 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5275 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5276 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5277 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5278 const Type *SrcTy = Op0C->getOperand(0)->getType();
5279 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5280 // Only do this if the casts both really cause code to be generated.
5281 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5283 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5285 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5286 Op1C->getOperand(0),
5288 InsertNewInstBefore(NewOp, I);
5289 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5294 return Changed ? &I : 0;
5297 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5298 LLVMContext *Context) {
5299 return cast<ConstantInt>(Context->getConstantExprExtractElement(V, Idx));
5302 static bool HasAddOverflow(ConstantInt *Result,
5303 ConstantInt *In1, ConstantInt *In2,
5306 if (In2->getValue().isNegative())
5307 return Result->getValue().sgt(In1->getValue());
5309 return Result->getValue().slt(In1->getValue());
5311 return Result->getValue().ult(In1->getValue());
5314 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5315 /// overflowed for this type.
5316 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5317 Constant *In2, LLVMContext *Context,
5318 bool IsSigned = false) {
5319 Result = Context->getConstantExprAdd(In1, In2);
5321 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5322 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5323 Constant *Idx = Context->getConstantInt(Type::Int32Ty, i);
5324 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5325 ExtractElement(In1, Idx, Context),
5326 ExtractElement(In2, Idx, Context),
5333 return HasAddOverflow(cast<ConstantInt>(Result),
5334 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5338 static bool HasSubOverflow(ConstantInt *Result,
5339 ConstantInt *In1, ConstantInt *In2,
5342 if (In2->getValue().isNegative())
5343 return Result->getValue().slt(In1->getValue());
5345 return Result->getValue().sgt(In1->getValue());
5347 return Result->getValue().ugt(In1->getValue());
5350 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5351 /// overflowed for this type.
5352 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5353 Constant *In2, LLVMContext *Context,
5354 bool IsSigned = false) {
5355 Result = Context->getConstantExprSub(In1, In2);
5357 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5358 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5359 Constant *Idx = Context->getConstantInt(Type::Int32Ty, i);
5360 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5361 ExtractElement(In1, Idx, Context),
5362 ExtractElement(In2, Idx, Context),
5369 return HasSubOverflow(cast<ConstantInt>(Result),
5370 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5374 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5375 /// code necessary to compute the offset from the base pointer (without adding
5376 /// in the base pointer). Return the result as a signed integer of intptr size.
5377 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5378 TargetData &TD = IC.getTargetData();
5379 gep_type_iterator GTI = gep_type_begin(GEP);
5380 const Type *IntPtrTy = TD.getIntPtrType();
5381 LLVMContext *Context = IC.getContext();
5382 Value *Result = Context->getNullValue(IntPtrTy);
5384 // Build a mask for high order bits.
5385 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5386 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5388 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5391 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5392 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5393 if (OpC->isZero()) continue;
5395 // Handle a struct index, which adds its field offset to the pointer.
5396 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5397 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5399 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5401 Context->getConstantInt(RC->getValue() + APInt(IntPtrWidth, Size));
5403 Result = IC.InsertNewInstBefore(
5404 BinaryOperator::CreateAdd(Result,
5405 Context->getConstantInt(IntPtrTy, Size),
5406 GEP->getName()+".offs"), I);
5410 Constant *Scale = Context->getConstantInt(IntPtrTy, Size);
5412 Context->getConstantExprIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5413 Scale = Context->getConstantExprMul(OC, Scale);
5414 if (Constant *RC = dyn_cast<Constant>(Result))
5415 Result = Context->getConstantExprAdd(RC, Scale);
5417 // Emit an add instruction.
5418 Result = IC.InsertNewInstBefore(
5419 BinaryOperator::CreateAdd(Result, Scale,
5420 GEP->getName()+".offs"), I);
5424 // Convert to correct type.
5425 if (Op->getType() != IntPtrTy) {
5426 if (Constant *OpC = dyn_cast<Constant>(Op))
5427 Op = Context->getConstantExprIntegerCast(OpC, IntPtrTy, true);
5429 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5431 Op->getName()+".c"), I);
5434 Constant *Scale = Context->getConstantInt(IntPtrTy, Size);
5435 if (Constant *OpC = dyn_cast<Constant>(Op))
5436 Op = Context->getConstantExprMul(OpC, Scale);
5437 else // We'll let instcombine(mul) convert this to a shl if possible.
5438 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5439 GEP->getName()+".idx"), I);
5442 // Emit an add instruction.
5443 if (isa<Constant>(Op) && isa<Constant>(Result))
5444 Result = Context->getConstantExprAdd(cast<Constant>(Op),
5445 cast<Constant>(Result));
5447 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5448 GEP->getName()+".offs"), I);
5454 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5455 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5456 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5457 /// be complex, and scales are involved. The above expression would also be
5458 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5459 /// This later form is less amenable to optimization though, and we are allowed
5460 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5462 /// If we can't emit an optimized form for this expression, this returns null.
5464 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5466 TargetData &TD = IC.getTargetData();
5467 gep_type_iterator GTI = gep_type_begin(GEP);
5469 // Check to see if this gep only has a single variable index. If so, and if
5470 // any constant indices are a multiple of its scale, then we can compute this
5471 // in terms of the scale of the variable index. For example, if the GEP
5472 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5473 // because the expression will cross zero at the same point.
5474 unsigned i, e = GEP->getNumOperands();
5476 for (i = 1; i != e; ++i, ++GTI) {
5477 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5478 // Compute the aggregate offset of constant indices.
5479 if (CI->isZero()) continue;
5481 // Handle a struct index, which adds its field offset to the pointer.
5482 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5483 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5485 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5486 Offset += Size*CI->getSExtValue();
5489 // Found our variable index.
5494 // If there are no variable indices, we must have a constant offset, just
5495 // evaluate it the general way.
5496 if (i == e) return 0;
5498 Value *VariableIdx = GEP->getOperand(i);
5499 // Determine the scale factor of the variable element. For example, this is
5500 // 4 if the variable index is into an array of i32.
5501 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5503 // Verify that there are no other variable indices. If so, emit the hard way.
5504 for (++i, ++GTI; i != e; ++i, ++GTI) {
5505 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5508 // Compute the aggregate offset of constant indices.
5509 if (CI->isZero()) continue;
5511 // Handle a struct index, which adds its field offset to the pointer.
5512 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5513 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5515 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5516 Offset += Size*CI->getSExtValue();
5520 // Okay, we know we have a single variable index, which must be a
5521 // pointer/array/vector index. If there is no offset, life is simple, return
5523 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5525 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5526 // we don't need to bother extending: the extension won't affect where the
5527 // computation crosses zero.
5528 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5529 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5530 VariableIdx->getNameStart(), &I);
5534 // Otherwise, there is an index. The computation we will do will be modulo
5535 // the pointer size, so get it.
5536 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5538 Offset &= PtrSizeMask;
5539 VariableScale &= PtrSizeMask;
5541 // To do this transformation, any constant index must be a multiple of the
5542 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5543 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5544 // multiple of the variable scale.
5545 int64_t NewOffs = Offset / (int64_t)VariableScale;
5546 if (Offset != NewOffs*(int64_t)VariableScale)
5549 // Okay, we can do this evaluation. Start by converting the index to intptr.
5550 const Type *IntPtrTy = TD.getIntPtrType();
5551 if (VariableIdx->getType() != IntPtrTy)
5552 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5554 VariableIdx->getNameStart(), &I);
5555 Constant *OffsetVal = IC.getContext()->getConstantInt(IntPtrTy, NewOffs);
5556 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5560 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5561 /// else. At this point we know that the GEP is on the LHS of the comparison.
5562 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5563 ICmpInst::Predicate Cond,
5565 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5567 // Look through bitcasts.
5568 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5569 RHS = BCI->getOperand(0);
5571 Value *PtrBase = GEPLHS->getOperand(0);
5572 if (PtrBase == RHS) {
5573 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5574 // This transformation (ignoring the base and scales) is valid because we
5575 // know pointers can't overflow. See if we can output an optimized form.
5576 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5578 // If not, synthesize the offset the hard way.
5580 Offset = EmitGEPOffset(GEPLHS, I, *this);
5581 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), Offset,
5582 Context->getNullValue(Offset->getType()));
5583 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5584 // If the base pointers are different, but the indices are the same, just
5585 // compare the base pointer.
5586 if (PtrBase != GEPRHS->getOperand(0)) {
5587 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5588 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5589 GEPRHS->getOperand(0)->getType();
5591 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5592 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5593 IndicesTheSame = false;
5597 // If all indices are the same, just compare the base pointers.
5599 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond),
5600 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5602 // Otherwise, the base pointers are different and the indices are
5603 // different, bail out.
5607 // If one of the GEPs has all zero indices, recurse.
5608 bool AllZeros = true;
5609 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5610 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5611 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5616 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5617 ICmpInst::getSwappedPredicate(Cond), I);
5619 // If the other GEP has all zero indices, recurse.
5621 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5622 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5623 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5628 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5630 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5631 // If the GEPs only differ by one index, compare it.
5632 unsigned NumDifferences = 0; // Keep track of # differences.
5633 unsigned DiffOperand = 0; // The operand that differs.
5634 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5635 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5636 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5637 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5638 // Irreconcilable differences.
5642 if (NumDifferences++) break;
5647 if (NumDifferences == 0) // SAME GEP?
5648 return ReplaceInstUsesWith(I, // No comparison is needed here.
5649 Context->getConstantInt(Type::Int1Ty,
5650 ICmpInst::isTrueWhenEqual(Cond)));
5652 else if (NumDifferences == 1) {
5653 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5654 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5655 // Make sure we do a signed comparison here.
5656 return new ICmpInst(*Context,
5657 ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5661 // Only lower this if the icmp is the only user of the GEP or if we expect
5662 // the result to fold to a constant!
5663 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5664 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5665 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5666 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5667 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5668 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), L, R);
5674 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5676 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5679 if (!isa<ConstantFP>(RHSC)) return 0;
5680 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5682 // Get the width of the mantissa. We don't want to hack on conversions that
5683 // might lose information from the integer, e.g. "i64 -> float"
5684 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5685 if (MantissaWidth == -1) return 0; // Unknown.
5687 // Check to see that the input is converted from an integer type that is small
5688 // enough that preserves all bits. TODO: check here for "known" sign bits.
5689 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5690 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5692 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5693 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5697 // If the conversion would lose info, don't hack on this.
5698 if ((int)InputSize > MantissaWidth)
5701 // Otherwise, we can potentially simplify the comparison. We know that it
5702 // will always come through as an integer value and we know the constant is
5703 // not a NAN (it would have been previously simplified).
5704 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5706 ICmpInst::Predicate Pred;
5707 switch (I.getPredicate()) {
5708 default: llvm_unreachable("Unexpected predicate!");
5709 case FCmpInst::FCMP_UEQ:
5710 case FCmpInst::FCMP_OEQ:
5711 Pred = ICmpInst::ICMP_EQ;
5713 case FCmpInst::FCMP_UGT:
5714 case FCmpInst::FCMP_OGT:
5715 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5717 case FCmpInst::FCMP_UGE:
5718 case FCmpInst::FCMP_OGE:
5719 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5721 case FCmpInst::FCMP_ULT:
5722 case FCmpInst::FCMP_OLT:
5723 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5725 case FCmpInst::FCMP_ULE:
5726 case FCmpInst::FCMP_OLE:
5727 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5729 case FCmpInst::FCMP_UNE:
5730 case FCmpInst::FCMP_ONE:
5731 Pred = ICmpInst::ICMP_NE;
5733 case FCmpInst::FCMP_ORD:
5734 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5735 case FCmpInst::FCMP_UNO:
5736 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5739 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5741 // Now we know that the APFloat is a normal number, zero or inf.
5743 // See if the FP constant is too large for the integer. For example,
5744 // comparing an i8 to 300.0.
5745 unsigned IntWidth = IntTy->getScalarSizeInBits();
5748 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5749 // and large values.
5750 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5751 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5752 APFloat::rmNearestTiesToEven);
5753 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5754 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5755 Pred == ICmpInst::ICMP_SLE)
5756 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5757 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5760 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5761 // +INF and large values.
5762 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5763 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5764 APFloat::rmNearestTiesToEven);
5765 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5766 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5767 Pred == ICmpInst::ICMP_ULE)
5768 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5769 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5774 // See if the RHS value is < SignedMin.
5775 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5776 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5777 APFloat::rmNearestTiesToEven);
5778 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5779 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5780 Pred == ICmpInst::ICMP_SGE)
5781 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5782 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5786 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5787 // [0, UMAX], but it may still be fractional. See if it is fractional by
5788 // casting the FP value to the integer value and back, checking for equality.
5789 // Don't do this for zero, because -0.0 is not fractional.
5790 Constant *RHSInt = LHSUnsigned
5791 ? Context->getConstantExprFPToUI(RHSC, IntTy)
5792 : Context->getConstantExprFPToSI(RHSC, IntTy);
5793 if (!RHS.isZero()) {
5794 bool Equal = LHSUnsigned
5795 ? Context->getConstantExprUIToFP(RHSInt, RHSC->getType()) == RHSC
5796 : Context->getConstantExprSIToFP(RHSInt, RHSC->getType()) == RHSC;
5798 // If we had a comparison against a fractional value, we have to adjust
5799 // the compare predicate and sometimes the value. RHSC is rounded towards
5800 // zero at this point.
5802 default: llvm_unreachable("Unexpected integer comparison!");
5803 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5804 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5805 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5806 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5807 case ICmpInst::ICMP_ULE:
5808 // (float)int <= 4.4 --> int <= 4
5809 // (float)int <= -4.4 --> false
5810 if (RHS.isNegative())
5811 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5813 case ICmpInst::ICMP_SLE:
5814 // (float)int <= 4.4 --> int <= 4
5815 // (float)int <= -4.4 --> int < -4
5816 if (RHS.isNegative())
5817 Pred = ICmpInst::ICMP_SLT;
5819 case ICmpInst::ICMP_ULT:
5820 // (float)int < -4.4 --> false
5821 // (float)int < 4.4 --> int <= 4
5822 if (RHS.isNegative())
5823 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5824 Pred = ICmpInst::ICMP_ULE;
5826 case ICmpInst::ICMP_SLT:
5827 // (float)int < -4.4 --> int < -4
5828 // (float)int < 4.4 --> int <= 4
5829 if (!RHS.isNegative())
5830 Pred = ICmpInst::ICMP_SLE;
5832 case ICmpInst::ICMP_UGT:
5833 // (float)int > 4.4 --> int > 4
5834 // (float)int > -4.4 --> true
5835 if (RHS.isNegative())
5836 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5838 case ICmpInst::ICMP_SGT:
5839 // (float)int > 4.4 --> int > 4
5840 // (float)int > -4.4 --> int >= -4
5841 if (RHS.isNegative())
5842 Pred = ICmpInst::ICMP_SGE;
5844 case ICmpInst::ICMP_UGE:
5845 // (float)int >= -4.4 --> true
5846 // (float)int >= 4.4 --> int > 4
5847 if (!RHS.isNegative())
5848 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5849 Pred = ICmpInst::ICMP_UGT;
5851 case ICmpInst::ICMP_SGE:
5852 // (float)int >= -4.4 --> int >= -4
5853 // (float)int >= 4.4 --> int > 4
5854 if (!RHS.isNegative())
5855 Pred = ICmpInst::ICMP_SGT;
5861 // Lower this FP comparison into an appropriate integer version of the
5863 return new ICmpInst(*Context, Pred, LHSI->getOperand(0), RHSInt);
5866 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5867 bool Changed = SimplifyCompare(I);
5868 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5870 // Fold trivial predicates.
5871 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5872 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5873 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5874 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5876 // Simplify 'fcmp pred X, X'
5878 switch (I.getPredicate()) {
5879 default: llvm_unreachable("Unknown predicate!");
5880 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5881 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5882 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5883 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5884 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5885 case FCmpInst::FCMP_OLT: // True if ordered and less than
5886 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5887 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5889 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5890 case FCmpInst::FCMP_ULT: // True if unordered or less than
5891 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5892 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5893 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5894 I.setPredicate(FCmpInst::FCMP_UNO);
5895 I.setOperand(1, Context->getNullValue(Op0->getType()));
5898 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5899 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5900 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5901 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5902 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5903 I.setPredicate(FCmpInst::FCMP_ORD);
5904 I.setOperand(1, Context->getNullValue(Op0->getType()));
5909 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5910 return ReplaceInstUsesWith(I, Context->getUndef(Type::Int1Ty));
5912 // Handle fcmp with constant RHS
5913 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5914 // If the constant is a nan, see if we can fold the comparison based on it.
5915 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5916 if (CFP->getValueAPF().isNaN()) {
5917 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5918 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5919 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5920 "Comparison must be either ordered or unordered!");
5921 // True if unordered.
5922 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5926 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5927 switch (LHSI->getOpcode()) {
5928 case Instruction::PHI:
5929 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5930 // block. If in the same block, we're encouraging jump threading. If
5931 // not, we are just pessimizing the code by making an i1 phi.
5932 if (LHSI->getParent() == I.getParent())
5933 if (Instruction *NV = FoldOpIntoPhi(I))
5936 case Instruction::SIToFP:
5937 case Instruction::UIToFP:
5938 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5941 case Instruction::Select:
5942 // If either operand of the select is a constant, we can fold the
5943 // comparison into the select arms, which will cause one to be
5944 // constant folded and the select turned into a bitwise or.
5945 Value *Op1 = 0, *Op2 = 0;
5946 if (LHSI->hasOneUse()) {
5947 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5948 // Fold the known value into the constant operand.
5949 Op1 = Context->getConstantExprCompare(I.getPredicate(), C, RHSC);
5950 // Insert a new FCmp of the other select operand.
5951 Op2 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5952 LHSI->getOperand(2), RHSC,
5954 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5955 // Fold the known value into the constant operand.
5956 Op2 = Context->getConstantExprCompare(I.getPredicate(), C, RHSC);
5957 // Insert a new FCmp of the other select operand.
5958 Op1 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5959 LHSI->getOperand(1), RHSC,
5965 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5970 return Changed ? &I : 0;
5973 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5974 bool Changed = SimplifyCompare(I);
5975 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5976 const Type *Ty = Op0->getType();
5980 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
5981 I.isTrueWhenEqual()));
5983 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5984 return ReplaceInstUsesWith(I, Context->getUndef(Type::Int1Ty));
5986 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5987 // addresses never equal each other! We already know that Op0 != Op1.
5988 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5989 isa<ConstantPointerNull>(Op0)) &&
5990 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5991 isa<ConstantPointerNull>(Op1)))
5992 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
5993 !I.isTrueWhenEqual()));
5995 // icmp's with boolean values can always be turned into bitwise operations
5996 if (Ty == Type::Int1Ty) {
5997 switch (I.getPredicate()) {
5998 default: llvm_unreachable("Invalid icmp instruction!");
5999 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
6000 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
6001 InsertNewInstBefore(Xor, I);
6002 return BinaryOperator::CreateNot(*Context, Xor);
6004 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6005 return BinaryOperator::CreateXor(Op0, Op1);
6007 case ICmpInst::ICMP_UGT:
6008 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6010 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6011 Instruction *Not = BinaryOperator::CreateNot(*Context,
6012 Op0, I.getName()+"tmp");
6013 InsertNewInstBefore(Not, I);
6014 return BinaryOperator::CreateAnd(Not, Op1);
6016 case ICmpInst::ICMP_SGT:
6017 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6019 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6020 Instruction *Not = BinaryOperator::CreateNot(*Context,
6021 Op1, I.getName()+"tmp");
6022 InsertNewInstBefore(Not, I);
6023 return BinaryOperator::CreateAnd(Not, Op0);
6025 case ICmpInst::ICMP_UGE:
6026 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6028 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6029 Instruction *Not = BinaryOperator::CreateNot(*Context,
6030 Op0, I.getName()+"tmp");
6031 InsertNewInstBefore(Not, I);
6032 return BinaryOperator::CreateOr(Not, Op1);
6034 case ICmpInst::ICMP_SGE:
6035 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6037 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6038 Instruction *Not = BinaryOperator::CreateNot(*Context,
6039 Op1, I.getName()+"tmp");
6040 InsertNewInstBefore(Not, I);
6041 return BinaryOperator::CreateOr(Not, Op0);
6046 unsigned BitWidth = 0;
6048 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6049 else if (Ty->isIntOrIntVector())
6050 BitWidth = Ty->getScalarSizeInBits();
6052 bool isSignBit = false;
6054 // See if we are doing a comparison with a constant.
6055 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6056 Value *A = 0, *B = 0;
6058 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6059 if (I.isEquality() && CI->isNullValue() &&
6060 match(Op0, m_Sub(m_Value(A), m_Value(B)), *Context)) {
6061 // (icmp cond A B) if cond is equality
6062 return new ICmpInst(*Context, I.getPredicate(), A, B);
6065 // If we have an icmp le or icmp ge instruction, turn it into the
6066 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6067 // them being folded in the code below.
6068 switch (I.getPredicate()) {
6070 case ICmpInst::ICMP_ULE:
6071 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6072 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6073 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Op0,
6074 AddOne(CI, Context));
6075 case ICmpInst::ICMP_SLE:
6076 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6077 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6078 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6079 AddOne(CI, Context));
6080 case ICmpInst::ICMP_UGE:
6081 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6082 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6083 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Op0,
6084 SubOne(CI, Context));
6085 case ICmpInst::ICMP_SGE:
6086 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6087 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6088 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6089 SubOne(CI, Context));
6092 // If this comparison is a normal comparison, it demands all
6093 // bits, if it is a sign bit comparison, it only demands the sign bit.
6095 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6098 // See if we can fold the comparison based on range information we can get
6099 // by checking whether bits are known to be zero or one in the input.
6100 if (BitWidth != 0) {
6101 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6102 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6104 if (SimplifyDemandedBits(I.getOperandUse(0),
6105 isSignBit ? APInt::getSignBit(BitWidth)
6106 : APInt::getAllOnesValue(BitWidth),
6107 Op0KnownZero, Op0KnownOne, 0))
6109 if (SimplifyDemandedBits(I.getOperandUse(1),
6110 APInt::getAllOnesValue(BitWidth),
6111 Op1KnownZero, Op1KnownOne, 0))
6114 // Given the known and unknown bits, compute a range that the LHS could be
6115 // in. Compute the Min, Max and RHS values based on the known bits. For the
6116 // EQ and NE we use unsigned values.
6117 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6118 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6119 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6120 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6122 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6125 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6127 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6131 // If Min and Max are known to be the same, then SimplifyDemandedBits
6132 // figured out that the LHS is a constant. Just constant fold this now so
6133 // that code below can assume that Min != Max.
6134 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6135 return new ICmpInst(*Context, I.getPredicate(),
6136 Context->getConstantInt(Op0Min), Op1);
6137 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6138 return new ICmpInst(*Context, I.getPredicate(), Op0,
6139 Context->getConstantInt(Op1Min));
6141 // Based on the range information we know about the LHS, see if we can
6142 // simplify this comparison. For example, (x&4) < 8 is always true.
6143 switch (I.getPredicate()) {
6144 default: llvm_unreachable("Unknown icmp opcode!");
6145 case ICmpInst::ICMP_EQ:
6146 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6147 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6149 case ICmpInst::ICMP_NE:
6150 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6151 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6153 case ICmpInst::ICMP_ULT:
6154 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6155 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6156 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6157 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6158 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6159 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6160 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6161 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6162 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6163 SubOne(CI, Context));
6165 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6166 if (CI->isMinValue(true))
6167 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6168 Context->getAllOnesValue(Op0->getType()));
6171 case ICmpInst::ICMP_UGT:
6172 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6173 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6174 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6175 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6177 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6178 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6179 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6180 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6181 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6182 AddOne(CI, Context));
6184 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6185 if (CI->isMaxValue(true))
6186 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6187 Context->getNullValue(Op0->getType()));
6190 case ICmpInst::ICMP_SLT:
6191 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6192 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6193 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6194 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6195 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6196 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6197 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6198 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6199 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6200 SubOne(CI, Context));
6203 case ICmpInst::ICMP_SGT:
6204 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6205 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6206 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6207 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6209 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6210 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6211 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6212 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6213 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6214 AddOne(CI, Context));
6217 case ICmpInst::ICMP_SGE:
6218 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6219 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6220 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6221 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6222 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6224 case ICmpInst::ICMP_SLE:
6225 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6226 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6227 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6228 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6229 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6231 case ICmpInst::ICMP_UGE:
6232 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6233 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6234 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6235 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6236 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6238 case ICmpInst::ICMP_ULE:
6239 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6240 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6241 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6242 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6243 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6247 // Turn a signed comparison into an unsigned one if both operands
6248 // are known to have the same sign.
6249 if (I.isSignedPredicate() &&
6250 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6251 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6252 return new ICmpInst(*Context, I.getUnsignedPredicate(), Op0, Op1);
6255 // Test if the ICmpInst instruction is used exclusively by a select as
6256 // part of a minimum or maximum operation. If so, refrain from doing
6257 // any other folding. This helps out other analyses which understand
6258 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6259 // and CodeGen. And in this case, at least one of the comparison
6260 // operands has at least one user besides the compare (the select),
6261 // which would often largely negate the benefit of folding anyway.
6263 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6264 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6265 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6268 // See if we are doing a comparison between a constant and an instruction that
6269 // can be folded into the comparison.
6270 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6271 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6272 // instruction, see if that instruction also has constants so that the
6273 // instruction can be folded into the icmp
6274 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6275 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6279 // Handle icmp with constant (but not simple integer constant) RHS
6280 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6281 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6282 switch (LHSI->getOpcode()) {
6283 case Instruction::GetElementPtr:
6284 if (RHSC->isNullValue()) {
6285 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6286 bool isAllZeros = true;
6287 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6288 if (!isa<Constant>(LHSI->getOperand(i)) ||
6289 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6294 return new ICmpInst(*Context, I.getPredicate(), LHSI->getOperand(0),
6295 Context->getNullValue(LHSI->getOperand(0)->getType()));
6299 case Instruction::PHI:
6300 // Only fold icmp into the PHI if the phi and fcmp are in the same
6301 // block. If in the same block, we're encouraging jump threading. If
6302 // not, we are just pessimizing the code by making an i1 phi.
6303 if (LHSI->getParent() == I.getParent())
6304 if (Instruction *NV = FoldOpIntoPhi(I))
6307 case Instruction::Select: {
6308 // If either operand of the select is a constant, we can fold the
6309 // comparison into the select arms, which will cause one to be
6310 // constant folded and the select turned into a bitwise or.
6311 Value *Op1 = 0, *Op2 = 0;
6312 if (LHSI->hasOneUse()) {
6313 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6314 // Fold the known value into the constant operand.
6315 Op1 = Context->getConstantExprICmp(I.getPredicate(), C, RHSC);
6316 // Insert a new ICmp of the other select operand.
6317 Op2 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6318 LHSI->getOperand(2), RHSC,
6320 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6321 // Fold the known value into the constant operand.
6322 Op2 = Context->getConstantExprICmp(I.getPredicate(), C, RHSC);
6323 // Insert a new ICmp of the other select operand.
6324 Op1 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6325 LHSI->getOperand(1), RHSC,
6331 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6334 case Instruction::Malloc:
6335 // If we have (malloc != null), and if the malloc has a single use, we
6336 // can assume it is successful and remove the malloc.
6337 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6338 AddToWorkList(LHSI);
6339 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
6340 !I.isTrueWhenEqual()));
6346 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6347 if (User *GEP = dyn_castGetElementPtr(Op0))
6348 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6350 if (User *GEP = dyn_castGetElementPtr(Op1))
6351 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6352 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6355 // Test to see if the operands of the icmp are casted versions of other
6356 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6358 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6359 if (isa<PointerType>(Op0->getType()) &&
6360 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6361 // We keep moving the cast from the left operand over to the right
6362 // operand, where it can often be eliminated completely.
6363 Op0 = CI->getOperand(0);
6365 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6366 // so eliminate it as well.
6367 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6368 Op1 = CI2->getOperand(0);
6370 // If Op1 is a constant, we can fold the cast into the constant.
6371 if (Op0->getType() != Op1->getType()) {
6372 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6373 Op1 = Context->getConstantExprBitCast(Op1C, Op0->getType());
6375 // Otherwise, cast the RHS right before the icmp
6376 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6379 return new ICmpInst(*Context, I.getPredicate(), Op0, Op1);
6383 if (isa<CastInst>(Op0)) {
6384 // Handle the special case of: icmp (cast bool to X), <cst>
6385 // This comes up when you have code like
6388 // For generality, we handle any zero-extension of any operand comparison
6389 // with a constant or another cast from the same type.
6390 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6391 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6395 // See if it's the same type of instruction on the left and right.
6396 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6397 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6398 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6399 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6400 switch (Op0I->getOpcode()) {
6402 case Instruction::Add:
6403 case Instruction::Sub:
6404 case Instruction::Xor:
6405 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6406 return new ICmpInst(*Context, I.getPredicate(), Op0I->getOperand(0),
6407 Op1I->getOperand(0));
6408 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6409 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6410 if (CI->getValue().isSignBit()) {
6411 ICmpInst::Predicate Pred = I.isSignedPredicate()
6412 ? I.getUnsignedPredicate()
6413 : I.getSignedPredicate();
6414 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6415 Op1I->getOperand(0));
6418 if (CI->getValue().isMaxSignedValue()) {
6419 ICmpInst::Predicate Pred = I.isSignedPredicate()
6420 ? I.getUnsignedPredicate()
6421 : I.getSignedPredicate();
6422 Pred = I.getSwappedPredicate(Pred);
6423 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6424 Op1I->getOperand(0));
6428 case Instruction::Mul:
6429 if (!I.isEquality())
6432 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6433 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6434 // Mask = -1 >> count-trailing-zeros(Cst).
6435 if (!CI->isZero() && !CI->isOne()) {
6436 const APInt &AP = CI->getValue();
6437 ConstantInt *Mask = Context->getConstantInt(
6438 APInt::getLowBitsSet(AP.getBitWidth(),
6440 AP.countTrailingZeros()));
6441 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6443 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6445 InsertNewInstBefore(And1, I);
6446 InsertNewInstBefore(And2, I);
6447 return new ICmpInst(*Context, I.getPredicate(), And1, And2);
6456 // ~x < ~y --> y < x
6458 if (match(Op0, m_Not(m_Value(A)), *Context) &&
6459 match(Op1, m_Not(m_Value(B)), *Context))
6460 return new ICmpInst(*Context, I.getPredicate(), B, A);
6463 if (I.isEquality()) {
6464 Value *A, *B, *C, *D;
6466 // -x == -y --> x == y
6467 if (match(Op0, m_Neg(m_Value(A)), *Context) &&
6468 match(Op1, m_Neg(m_Value(B)), *Context))
6469 return new ICmpInst(*Context, I.getPredicate(), A, B);
6471 if (match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
6472 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6473 Value *OtherVal = A == Op1 ? B : A;
6474 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6475 Context->getNullValue(A->getType()));
6478 if (match(Op1, m_Xor(m_Value(C), m_Value(D)), *Context)) {
6479 // A^c1 == C^c2 --> A == C^(c1^c2)
6480 ConstantInt *C1, *C2;
6481 if (match(B, m_ConstantInt(C1), *Context) &&
6482 match(D, m_ConstantInt(C2), *Context) && Op1->hasOneUse()) {
6484 Context->getConstantInt(C1->getValue() ^ C2->getValue());
6485 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6486 return new ICmpInst(*Context, I.getPredicate(), A,
6487 InsertNewInstBefore(Xor, I));
6490 // A^B == A^D -> B == D
6491 if (A == C) return new ICmpInst(*Context, I.getPredicate(), B, D);
6492 if (A == D) return new ICmpInst(*Context, I.getPredicate(), B, C);
6493 if (B == C) return new ICmpInst(*Context, I.getPredicate(), A, D);
6494 if (B == D) return new ICmpInst(*Context, I.getPredicate(), A, C);
6498 if (match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context) &&
6499 (A == Op0 || B == Op0)) {
6500 // A == (A^B) -> B == 0
6501 Value *OtherVal = A == Op0 ? B : A;
6502 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6503 Context->getNullValue(A->getType()));
6506 // (A-B) == A -> B == 0
6507 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B)), *Context))
6508 return new ICmpInst(*Context, I.getPredicate(), B,
6509 Context->getNullValue(B->getType()));
6511 // A == (A-B) -> B == 0
6512 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B)), *Context))
6513 return new ICmpInst(*Context, I.getPredicate(), B,
6514 Context->getNullValue(B->getType()));
6516 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6517 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6518 match(Op0, m_And(m_Value(A), m_Value(B)), *Context) &&
6519 match(Op1, m_And(m_Value(C), m_Value(D)), *Context)) {
6520 Value *X = 0, *Y = 0, *Z = 0;
6523 X = B; Y = D; Z = A;
6524 } else if (A == D) {
6525 X = B; Y = C; Z = A;
6526 } else if (B == C) {
6527 X = A; Y = D; Z = B;
6528 } else if (B == D) {
6529 X = A; Y = C; Z = B;
6532 if (X) { // Build (X^Y) & Z
6533 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6534 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6535 I.setOperand(0, Op1);
6536 I.setOperand(1, Context->getNullValue(Op1->getType()));
6541 return Changed ? &I : 0;
6545 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6546 /// and CmpRHS are both known to be integer constants.
6547 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6548 ConstantInt *DivRHS) {
6549 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6550 const APInt &CmpRHSV = CmpRHS->getValue();
6552 // FIXME: If the operand types don't match the type of the divide
6553 // then don't attempt this transform. The code below doesn't have the
6554 // logic to deal with a signed divide and an unsigned compare (and
6555 // vice versa). This is because (x /s C1) <s C2 produces different
6556 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6557 // (x /u C1) <u C2. Simply casting the operands and result won't
6558 // work. :( The if statement below tests that condition and bails
6560 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6561 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6563 if (DivRHS->isZero())
6564 return 0; // The ProdOV computation fails on divide by zero.
6565 if (DivIsSigned && DivRHS->isAllOnesValue())
6566 return 0; // The overflow computation also screws up here
6567 if (DivRHS->isOne())
6568 return 0; // Not worth bothering, and eliminates some funny cases
6571 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6572 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6573 // C2 (CI). By solving for X we can turn this into a range check
6574 // instead of computing a divide.
6575 Constant *Prod = Context->getConstantExprMul(CmpRHS, DivRHS);
6577 // Determine if the product overflows by seeing if the product is
6578 // not equal to the divide. Make sure we do the same kind of divide
6579 // as in the LHS instruction that we're folding.
6580 bool ProdOV = (DivIsSigned ? Context->getConstantExprSDiv(Prod, DivRHS) :
6581 Context->getConstantExprUDiv(Prod, DivRHS)) != CmpRHS;
6583 // Get the ICmp opcode
6584 ICmpInst::Predicate Pred = ICI.getPredicate();
6586 // Figure out the interval that is being checked. For example, a comparison
6587 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6588 // Compute this interval based on the constants involved and the signedness of
6589 // the compare/divide. This computes a half-open interval, keeping track of
6590 // whether either value in the interval overflows. After analysis each
6591 // overflow variable is set to 0 if it's corresponding bound variable is valid
6592 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6593 int LoOverflow = 0, HiOverflow = 0;
6594 Constant *LoBound = 0, *HiBound = 0;
6596 if (!DivIsSigned) { // udiv
6597 // e.g. X/5 op 3 --> [15, 20)
6599 HiOverflow = LoOverflow = ProdOV;
6601 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6602 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6603 if (CmpRHSV == 0) { // (X / pos) op 0
6604 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6605 LoBound = cast<ConstantInt>(Context->getConstantExprNeg(SubOne(DivRHS,
6608 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6609 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6610 HiOverflow = LoOverflow = ProdOV;
6612 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6613 } else { // (X / pos) op neg
6614 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6615 HiBound = AddOne(Prod, Context);
6616 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6618 ConstantInt* DivNeg =
6619 cast<ConstantInt>(Context->getConstantExprNeg(DivRHS));
6620 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6624 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6625 if (CmpRHSV == 0) { // (X / neg) op 0
6626 // e.g. X/-5 op 0 --> [-4, 5)
6627 LoBound = AddOne(DivRHS, Context);
6628 HiBound = cast<ConstantInt>(Context->getConstantExprNeg(DivRHS));
6629 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6630 HiOverflow = 1; // [INTMIN+1, overflow)
6631 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6633 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6634 // e.g. X/-5 op 3 --> [-19, -14)
6635 HiBound = AddOne(Prod, Context);
6636 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6638 LoOverflow = AddWithOverflow(LoBound, HiBound,
6639 DivRHS, Context, true) ? -1 : 0;
6640 } else { // (X / neg) op neg
6641 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6642 LoOverflow = HiOverflow = ProdOV;
6644 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6647 // Dividing by a negative swaps the condition. LT <-> GT
6648 Pred = ICmpInst::getSwappedPredicate(Pred);
6651 Value *X = DivI->getOperand(0);
6653 default: llvm_unreachable("Unhandled icmp opcode!");
6654 case ICmpInst::ICMP_EQ:
6655 if (LoOverflow && HiOverflow)
6656 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6657 else if (HiOverflow)
6658 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6659 ICmpInst::ICMP_UGE, X, LoBound);
6660 else if (LoOverflow)
6661 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6662 ICmpInst::ICMP_ULT, X, HiBound);
6664 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6665 case ICmpInst::ICMP_NE:
6666 if (LoOverflow && HiOverflow)
6667 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6668 else if (HiOverflow)
6669 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6670 ICmpInst::ICMP_ULT, X, LoBound);
6671 else if (LoOverflow)
6672 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6673 ICmpInst::ICMP_UGE, X, HiBound);
6675 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6676 case ICmpInst::ICMP_ULT:
6677 case ICmpInst::ICMP_SLT:
6678 if (LoOverflow == +1) // Low bound is greater than input range.
6679 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6680 if (LoOverflow == -1) // Low bound is less than input range.
6681 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6682 return new ICmpInst(*Context, Pred, X, LoBound);
6683 case ICmpInst::ICMP_UGT:
6684 case ICmpInst::ICMP_SGT:
6685 if (HiOverflow == +1) // High bound greater than input range.
6686 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6687 else if (HiOverflow == -1) // High bound less than input range.
6688 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6689 if (Pred == ICmpInst::ICMP_UGT)
6690 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, X, HiBound);
6692 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, X, HiBound);
6697 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6699 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6702 const APInt &RHSV = RHS->getValue();
6704 switch (LHSI->getOpcode()) {
6705 case Instruction::Trunc:
6706 if (ICI.isEquality() && LHSI->hasOneUse()) {
6707 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6708 // of the high bits truncated out of x are known.
6709 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6710 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6711 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6712 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6713 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6715 // If all the high bits are known, we can do this xform.
6716 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6717 // Pull in the high bits from known-ones set.
6718 APInt NewRHS(RHS->getValue());
6719 NewRHS.zext(SrcBits);
6721 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6722 Context->getConstantInt(NewRHS));
6727 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6728 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6729 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6731 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6732 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6733 Value *CompareVal = LHSI->getOperand(0);
6735 // If the sign bit of the XorCST is not set, there is no change to
6736 // the operation, just stop using the Xor.
6737 if (!XorCST->getValue().isNegative()) {
6738 ICI.setOperand(0, CompareVal);
6739 AddToWorkList(LHSI);
6743 // Was the old condition true if the operand is positive?
6744 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6746 // If so, the new one isn't.
6747 isTrueIfPositive ^= true;
6749 if (isTrueIfPositive)
6750 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, CompareVal,
6751 SubOne(RHS, Context));
6753 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, CompareVal,
6754 AddOne(RHS, Context));
6757 if (LHSI->hasOneUse()) {
6758 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6759 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6760 const APInt &SignBit = XorCST->getValue();
6761 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6762 ? ICI.getUnsignedPredicate()
6763 : ICI.getSignedPredicate();
6764 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6765 Context->getConstantInt(RHSV ^ SignBit));
6768 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6769 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6770 const APInt &NotSignBit = XorCST->getValue();
6771 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6772 ? ICI.getUnsignedPredicate()
6773 : ICI.getSignedPredicate();
6774 Pred = ICI.getSwappedPredicate(Pred);
6775 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6776 Context->getConstantInt(RHSV ^ NotSignBit));
6781 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6782 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6783 LHSI->getOperand(0)->hasOneUse()) {
6784 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6786 // If the LHS is an AND of a truncating cast, we can widen the
6787 // and/compare to be the input width without changing the value
6788 // produced, eliminating a cast.
6789 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6790 // We can do this transformation if either the AND constant does not
6791 // have its sign bit set or if it is an equality comparison.
6792 // Extending a relational comparison when we're checking the sign
6793 // bit would not work.
6794 if (Cast->hasOneUse() &&
6795 (ICI.isEquality() ||
6796 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6798 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6799 APInt NewCST = AndCST->getValue();
6800 NewCST.zext(BitWidth);
6802 NewCI.zext(BitWidth);
6803 Instruction *NewAnd =
6804 BinaryOperator::CreateAnd(Cast->getOperand(0),
6805 Context->getConstantInt(NewCST),LHSI->getName());
6806 InsertNewInstBefore(NewAnd, ICI);
6807 return new ICmpInst(*Context, ICI.getPredicate(), NewAnd,
6808 Context->getConstantInt(NewCI));
6812 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6813 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6814 // happens a LOT in code produced by the C front-end, for bitfield
6816 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6817 if (Shift && !Shift->isShift())
6821 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6822 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6823 const Type *AndTy = AndCST->getType(); // Type of the and.
6825 // We can fold this as long as we can't shift unknown bits
6826 // into the mask. This can only happen with signed shift
6827 // rights, as they sign-extend.
6829 bool CanFold = Shift->isLogicalShift();
6831 // To test for the bad case of the signed shr, see if any
6832 // of the bits shifted in could be tested after the mask.
6833 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6834 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6836 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6837 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6838 AndCST->getValue()) == 0)
6844 if (Shift->getOpcode() == Instruction::Shl)
6845 NewCst = Context->getConstantExprLShr(RHS, ShAmt);
6847 NewCst = Context->getConstantExprShl(RHS, ShAmt);
6849 // Check to see if we are shifting out any of the bits being
6851 if (Context->getConstantExpr(Shift->getOpcode(),
6852 NewCst, ShAmt) != RHS) {
6853 // If we shifted bits out, the fold is not going to work out.
6854 // As a special case, check to see if this means that the
6855 // result is always true or false now.
6856 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6857 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6858 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6859 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6861 ICI.setOperand(1, NewCst);
6862 Constant *NewAndCST;
6863 if (Shift->getOpcode() == Instruction::Shl)
6864 NewAndCST = Context->getConstantExprLShr(AndCST, ShAmt);
6866 NewAndCST = Context->getConstantExprShl(AndCST, ShAmt);
6867 LHSI->setOperand(1, NewAndCST);
6868 LHSI->setOperand(0, Shift->getOperand(0));
6869 AddToWorkList(Shift); // Shift is dead.
6870 AddUsesToWorkList(ICI);
6876 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6877 // preferable because it allows the C<<Y expression to be hoisted out
6878 // of a loop if Y is invariant and X is not.
6879 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6880 ICI.isEquality() && !Shift->isArithmeticShift() &&
6881 !isa<Constant>(Shift->getOperand(0))) {
6884 if (Shift->getOpcode() == Instruction::LShr) {
6885 NS = BinaryOperator::CreateShl(AndCST,
6886 Shift->getOperand(1), "tmp");
6888 // Insert a logical shift.
6889 NS = BinaryOperator::CreateLShr(AndCST,
6890 Shift->getOperand(1), "tmp");
6892 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6894 // Compute X & (C << Y).
6895 Instruction *NewAnd =
6896 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6897 InsertNewInstBefore(NewAnd, ICI);
6899 ICI.setOperand(0, NewAnd);
6905 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6906 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6909 uint32_t TypeBits = RHSV.getBitWidth();
6911 // Check that the shift amount is in range. If not, don't perform
6912 // undefined shifts. When the shift is visited it will be
6914 if (ShAmt->uge(TypeBits))
6917 if (ICI.isEquality()) {
6918 // If we are comparing against bits always shifted out, the
6919 // comparison cannot succeed.
6921 Context->getConstantExprShl(Context->getConstantExprLShr(RHS, ShAmt),
6923 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6924 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6925 Constant *Cst = Context->getConstantInt(Type::Int1Ty, IsICMP_NE);
6926 return ReplaceInstUsesWith(ICI, Cst);
6929 if (LHSI->hasOneUse()) {
6930 // Otherwise strength reduce the shift into an and.
6931 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6933 Context->getConstantInt(APInt::getLowBitsSet(TypeBits,
6934 TypeBits-ShAmtVal));
6937 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6938 Mask, LHSI->getName()+".mask");
6939 Value *And = InsertNewInstBefore(AndI, ICI);
6940 return new ICmpInst(*Context, ICI.getPredicate(), And,
6941 Context->getConstantInt(RHSV.lshr(ShAmtVal)));
6945 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6946 bool TrueIfSigned = false;
6947 if (LHSI->hasOneUse() &&
6948 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6949 // (X << 31) <s 0 --> (X&1) != 0
6950 Constant *Mask = Context->getConstantInt(APInt(TypeBits, 1) <<
6951 (TypeBits-ShAmt->getZExtValue()-1));
6953 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6954 Mask, LHSI->getName()+".mask");
6955 Value *And = InsertNewInstBefore(AndI, ICI);
6957 return new ICmpInst(*Context,
6958 TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6959 And, Context->getNullValue(And->getType()));
6964 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6965 case Instruction::AShr: {
6966 // Only handle equality comparisons of shift-by-constant.
6967 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6968 if (!ShAmt || !ICI.isEquality()) break;
6970 // Check that the shift amount is in range. If not, don't perform
6971 // undefined shifts. When the shift is visited it will be
6973 uint32_t TypeBits = RHSV.getBitWidth();
6974 if (ShAmt->uge(TypeBits))
6977 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6979 // If we are comparing against bits always shifted out, the
6980 // comparison cannot succeed.
6981 APInt Comp = RHSV << ShAmtVal;
6982 if (LHSI->getOpcode() == Instruction::LShr)
6983 Comp = Comp.lshr(ShAmtVal);
6985 Comp = Comp.ashr(ShAmtVal);
6987 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6988 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6989 Constant *Cst = Context->getConstantInt(Type::Int1Ty, IsICMP_NE);
6990 return ReplaceInstUsesWith(ICI, Cst);
6993 // Otherwise, check to see if the bits shifted out are known to be zero.
6994 // If so, we can compare against the unshifted value:
6995 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6996 if (LHSI->hasOneUse() &&
6997 MaskedValueIsZero(LHSI->getOperand(0),
6998 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6999 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
7000 Context->getConstantExprShl(RHS, ShAmt));
7003 if (LHSI->hasOneUse()) {
7004 // Otherwise strength reduce the shift into an and.
7005 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7006 Constant *Mask = Context->getConstantInt(Val);
7009 BinaryOperator::CreateAnd(LHSI->getOperand(0),
7010 Mask, LHSI->getName()+".mask");
7011 Value *And = InsertNewInstBefore(AndI, ICI);
7012 return new ICmpInst(*Context, ICI.getPredicate(), And,
7013 Context->getConstantExprShl(RHS, ShAmt));
7018 case Instruction::SDiv:
7019 case Instruction::UDiv:
7020 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7021 // Fold this div into the comparison, producing a range check.
7022 // Determine, based on the divide type, what the range is being
7023 // checked. If there is an overflow on the low or high side, remember
7024 // it, otherwise compute the range [low, hi) bounding the new value.
7025 // See: InsertRangeTest above for the kinds of replacements possible.
7026 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7027 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7032 case Instruction::Add:
7033 // Fold: icmp pred (add, X, C1), C2
7035 if (!ICI.isEquality()) {
7036 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7038 const APInt &LHSV = LHSC->getValue();
7040 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7043 if (ICI.isSignedPredicate()) {
7044 if (CR.getLower().isSignBit()) {
7045 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7046 Context->getConstantInt(CR.getUpper()));
7047 } else if (CR.getUpper().isSignBit()) {
7048 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7049 Context->getConstantInt(CR.getLower()));
7052 if (CR.getLower().isMinValue()) {
7053 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7054 Context->getConstantInt(CR.getUpper()));
7055 } else if (CR.getUpper().isMinValue()) {
7056 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7057 Context->getConstantInt(CR.getLower()));
7064 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7065 if (ICI.isEquality()) {
7066 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7068 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7069 // the second operand is a constant, simplify a bit.
7070 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7071 switch (BO->getOpcode()) {
7072 case Instruction::SRem:
7073 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7074 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7075 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7076 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7077 Instruction *NewRem =
7078 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
7080 InsertNewInstBefore(NewRem, ICI);
7081 return new ICmpInst(*Context, ICI.getPredicate(), NewRem,
7082 Context->getNullValue(BO->getType()));
7086 case Instruction::Add:
7087 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7088 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7089 if (BO->hasOneUse())
7090 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7091 Context->getConstantExprSub(RHS, BOp1C));
7092 } else if (RHSV == 0) {
7093 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7094 // efficiently invertible, or if the add has just this one use.
7095 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7097 if (Value *NegVal = dyn_castNegVal(BOp1, Context))
7098 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, NegVal);
7099 else if (Value *NegVal = dyn_castNegVal(BOp0, Context))
7100 return new ICmpInst(*Context, ICI.getPredicate(), NegVal, BOp1);
7101 else if (BO->hasOneUse()) {
7102 Instruction *Neg = BinaryOperator::CreateNeg(*Context, BOp1);
7103 InsertNewInstBefore(Neg, ICI);
7105 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, Neg);
7109 case Instruction::Xor:
7110 // For the xor case, we can xor two constants together, eliminating
7111 // the explicit xor.
7112 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7113 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7114 Context->getConstantExprXor(RHS, BOC));
7117 case Instruction::Sub:
7118 // Replace (([sub|xor] A, B) != 0) with (A != B)
7120 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7124 case Instruction::Or:
7125 // If bits are being or'd in that are not present in the constant we
7126 // are comparing against, then the comparison could never succeed!
7127 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7128 Constant *NotCI = Context->getConstantExprNot(RHS);
7129 if (!Context->getConstantExprAnd(BOC, NotCI)->isNullValue())
7130 return ReplaceInstUsesWith(ICI,
7131 Context->getConstantInt(Type::Int1Ty,
7136 case Instruction::And:
7137 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7138 // If bits are being compared against that are and'd out, then the
7139 // comparison can never succeed!
7140 if ((RHSV & ~BOC->getValue()) != 0)
7141 return ReplaceInstUsesWith(ICI,
7142 Context->getConstantInt(Type::Int1Ty,
7145 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7146 if (RHS == BOC && RHSV.isPowerOf2())
7147 return new ICmpInst(*Context, isICMP_NE ? ICmpInst::ICMP_EQ :
7148 ICmpInst::ICMP_NE, LHSI,
7149 Context->getNullValue(RHS->getType()));
7151 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7152 if (BOC->getValue().isSignBit()) {
7153 Value *X = BO->getOperand(0);
7154 Constant *Zero = Context->getNullValue(X->getType());
7155 ICmpInst::Predicate pred = isICMP_NE ?
7156 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7157 return new ICmpInst(*Context, pred, X, Zero);
7160 // ((X & ~7) == 0) --> X < 8
7161 if (RHSV == 0 && isHighOnes(BOC)) {
7162 Value *X = BO->getOperand(0);
7163 Constant *NegX = Context->getConstantExprNeg(BOC);
7164 ICmpInst::Predicate pred = isICMP_NE ?
7165 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7166 return new ICmpInst(*Context, pred, X, NegX);
7171 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7172 // Handle icmp {eq|ne} <intrinsic>, intcst.
7173 if (II->getIntrinsicID() == Intrinsic::bswap) {
7175 ICI.setOperand(0, II->getOperand(1));
7176 ICI.setOperand(1, Context->getConstantInt(RHSV.byteSwap()));
7184 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7185 /// We only handle extending casts so far.
7187 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7188 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7189 Value *LHSCIOp = LHSCI->getOperand(0);
7190 const Type *SrcTy = LHSCIOp->getType();
7191 const Type *DestTy = LHSCI->getType();
7194 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7195 // integer type is the same size as the pointer type.
7196 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
7197 getTargetData().getPointerSizeInBits() ==
7198 cast<IntegerType>(DestTy)->getBitWidth()) {
7200 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7201 RHSOp = Context->getConstantExprIntToPtr(RHSC, SrcTy);
7202 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7203 RHSOp = RHSC->getOperand(0);
7204 // If the pointer types don't match, insert a bitcast.
7205 if (LHSCIOp->getType() != RHSOp->getType())
7206 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
7210 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSOp);
7213 // The code below only handles extension cast instructions, so far.
7215 if (LHSCI->getOpcode() != Instruction::ZExt &&
7216 LHSCI->getOpcode() != Instruction::SExt)
7219 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7220 bool isSignedCmp = ICI.isSignedPredicate();
7222 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7223 // Not an extension from the same type?
7224 RHSCIOp = CI->getOperand(0);
7225 if (RHSCIOp->getType() != LHSCIOp->getType())
7228 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7229 // and the other is a zext), then we can't handle this.
7230 if (CI->getOpcode() != LHSCI->getOpcode())
7233 // Deal with equality cases early.
7234 if (ICI.isEquality())
7235 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7237 // A signed comparison of sign extended values simplifies into a
7238 // signed comparison.
7239 if (isSignedCmp && isSignedExt)
7240 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7242 // The other three cases all fold into an unsigned comparison.
7243 return new ICmpInst(*Context, ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7246 // If we aren't dealing with a constant on the RHS, exit early
7247 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7251 // Compute the constant that would happen if we truncated to SrcTy then
7252 // reextended to DestTy.
7253 Constant *Res1 = Context->getConstantExprTrunc(CI, SrcTy);
7254 Constant *Res2 = Context->getConstantExprCast(LHSCI->getOpcode(),
7257 // If the re-extended constant didn't change...
7259 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7260 // For example, we might have:
7261 // %A = sext i16 %X to i32
7262 // %B = icmp ugt i32 %A, 1330
7263 // It is incorrect to transform this into
7264 // %B = icmp ugt i16 %X, 1330
7265 // because %A may have negative value.
7267 // However, we allow this when the compare is EQ/NE, because they are
7269 if (isSignedExt == isSignedCmp || ICI.isEquality())
7270 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, Res1);
7274 // The re-extended constant changed so the constant cannot be represented
7275 // in the shorter type. Consequently, we cannot emit a simple comparison.
7277 // First, handle some easy cases. We know the result cannot be equal at this
7278 // point so handle the ICI.isEquality() cases
7279 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7280 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
7281 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7282 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
7284 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7285 // should have been folded away previously and not enter in here.
7288 // We're performing a signed comparison.
7289 if (cast<ConstantInt>(CI)->getValue().isNegative())
7290 Result = Context->getConstantIntFalse(); // X < (small) --> false
7292 Result = Context->getConstantIntTrue(); // X < (large) --> true
7294 // We're performing an unsigned comparison.
7296 // We're performing an unsigned comp with a sign extended value.
7297 // This is true if the input is >= 0. [aka >s -1]
7298 Constant *NegOne = Context->getAllOnesValue(SrcTy);
7299 Result = InsertNewInstBefore(new ICmpInst(*Context, ICmpInst::ICMP_SGT,
7300 LHSCIOp, NegOne, ICI.getName()), ICI);
7302 // Unsigned extend & unsigned compare -> always true.
7303 Result = Context->getConstantIntTrue();
7307 // Finally, return the value computed.
7308 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7309 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7310 return ReplaceInstUsesWith(ICI, Result);
7312 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7313 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7314 "ICmp should be folded!");
7315 if (Constant *CI = dyn_cast<Constant>(Result))
7316 return ReplaceInstUsesWith(ICI, Context->getConstantExprNot(CI));
7317 return BinaryOperator::CreateNot(*Context, Result);
7320 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7321 return commonShiftTransforms(I);
7324 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7325 return commonShiftTransforms(I);
7328 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7329 if (Instruction *R = commonShiftTransforms(I))
7332 Value *Op0 = I.getOperand(0);
7334 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7335 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7336 if (CSI->isAllOnesValue())
7337 return ReplaceInstUsesWith(I, CSI);
7339 // See if we can turn a signed shr into an unsigned shr.
7340 if (MaskedValueIsZero(Op0,
7341 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7342 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7344 // Arithmetic shifting an all-sign-bit value is a no-op.
7345 unsigned NumSignBits = ComputeNumSignBits(Op0);
7346 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7347 return ReplaceInstUsesWith(I, Op0);
7352 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7353 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7354 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7356 // shl X, 0 == X and shr X, 0 == X
7357 // shl 0, X == 0 and shr 0, X == 0
7358 if (Op1 == Context->getNullValue(Op1->getType()) ||
7359 Op0 == Context->getNullValue(Op0->getType()))
7360 return ReplaceInstUsesWith(I, Op0);
7362 if (isa<UndefValue>(Op0)) {
7363 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7364 return ReplaceInstUsesWith(I, Op0);
7365 else // undef << X -> 0, undef >>u X -> 0
7366 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7368 if (isa<UndefValue>(Op1)) {
7369 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7370 return ReplaceInstUsesWith(I, Op0);
7371 else // X << undef, X >>u undef -> 0
7372 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7375 // See if we can fold away this shift.
7376 if (SimplifyDemandedInstructionBits(I))
7379 // Try to fold constant and into select arguments.
7380 if (isa<Constant>(Op0))
7381 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7382 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7385 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7386 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7391 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7392 BinaryOperator &I) {
7393 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7395 // See if we can simplify any instructions used by the instruction whose sole
7396 // purpose is to compute bits we don't care about.
7397 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7399 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7402 if (Op1->uge(TypeBits)) {
7403 if (I.getOpcode() != Instruction::AShr)
7404 return ReplaceInstUsesWith(I, Context->getNullValue(Op0->getType()));
7406 I.setOperand(1, Context->getConstantInt(I.getType(), TypeBits-1));
7411 // ((X*C1) << C2) == (X * (C1 << C2))
7412 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7413 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7414 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7415 return BinaryOperator::CreateMul(BO->getOperand(0),
7416 Context->getConstantExprShl(BOOp, Op1));
7418 // Try to fold constant and into select arguments.
7419 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7420 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7422 if (isa<PHINode>(Op0))
7423 if (Instruction *NV = FoldOpIntoPhi(I))
7426 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7427 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7428 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7429 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7430 // place. Don't try to do this transformation in this case. Also, we
7431 // require that the input operand is a shift-by-constant so that we have
7432 // confidence that the shifts will get folded together. We could do this
7433 // xform in more cases, but it is unlikely to be profitable.
7434 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7435 isa<ConstantInt>(TrOp->getOperand(1))) {
7436 // Okay, we'll do this xform. Make the shift of shift.
7437 Constant *ShAmt = Context->getConstantExprZExt(Op1, TrOp->getType());
7438 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7440 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7442 // For logical shifts, the truncation has the effect of making the high
7443 // part of the register be zeros. Emulate this by inserting an AND to
7444 // clear the top bits as needed. This 'and' will usually be zapped by
7445 // other xforms later if dead.
7446 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7447 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7448 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7450 // The mask we constructed says what the trunc would do if occurring
7451 // between the shifts. We want to know the effect *after* the second
7452 // shift. We know that it is a logical shift by a constant, so adjust the
7453 // mask as appropriate.
7454 if (I.getOpcode() == Instruction::Shl)
7455 MaskV <<= Op1->getZExtValue();
7457 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7458 MaskV = MaskV.lshr(Op1->getZExtValue());
7462 BinaryOperator::CreateAnd(NSh, Context->getConstantInt(MaskV),
7464 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7466 // Return the value truncated to the interesting size.
7467 return new TruncInst(And, I.getType());
7471 if (Op0->hasOneUse()) {
7472 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7473 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7476 switch (Op0BO->getOpcode()) {
7478 case Instruction::Add:
7479 case Instruction::And:
7480 case Instruction::Or:
7481 case Instruction::Xor: {
7482 // These operators commute.
7483 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7484 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7485 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7486 m_Specific(Op1)), *Context)){
7487 Instruction *YS = BinaryOperator::CreateShl(
7488 Op0BO->getOperand(0), Op1,
7490 InsertNewInstBefore(YS, I); // (Y << C)
7492 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7493 Op0BO->getOperand(1)->getName());
7494 InsertNewInstBefore(X, I); // (X + (Y << C))
7495 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7496 return BinaryOperator::CreateAnd(X, Context->getConstantInt(
7497 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7500 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7501 Value *Op0BOOp1 = Op0BO->getOperand(1);
7502 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7504 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7505 m_ConstantInt(CC)), *Context) &&
7506 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7507 Instruction *YS = BinaryOperator::CreateShl(
7508 Op0BO->getOperand(0), Op1,
7510 InsertNewInstBefore(YS, I); // (Y << C)
7512 BinaryOperator::CreateAnd(V1,
7513 Context->getConstantExprShl(CC, Op1),
7514 V1->getName()+".mask");
7515 InsertNewInstBefore(XM, I); // X & (CC << C)
7517 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7522 case Instruction::Sub: {
7523 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7524 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7525 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7526 m_Specific(Op1)), *Context)){
7527 Instruction *YS = BinaryOperator::CreateShl(
7528 Op0BO->getOperand(1), Op1,
7530 InsertNewInstBefore(YS, I); // (Y << C)
7532 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7533 Op0BO->getOperand(0)->getName());
7534 InsertNewInstBefore(X, I); // (X + (Y << C))
7535 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7536 return BinaryOperator::CreateAnd(X, Context->getConstantInt(
7537 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7540 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7541 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7542 match(Op0BO->getOperand(0),
7543 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7544 m_ConstantInt(CC)), *Context) && V2 == Op1 &&
7545 cast<BinaryOperator>(Op0BO->getOperand(0))
7546 ->getOperand(0)->hasOneUse()) {
7547 Instruction *YS = BinaryOperator::CreateShl(
7548 Op0BO->getOperand(1), Op1,
7550 InsertNewInstBefore(YS, I); // (Y << C)
7552 BinaryOperator::CreateAnd(V1,
7553 Context->getConstantExprShl(CC, Op1),
7554 V1->getName()+".mask");
7555 InsertNewInstBefore(XM, I); // X & (CC << C)
7557 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7565 // If the operand is an bitwise operator with a constant RHS, and the
7566 // shift is the only use, we can pull it out of the shift.
7567 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7568 bool isValid = true; // Valid only for And, Or, Xor
7569 bool highBitSet = false; // Transform if high bit of constant set?
7571 switch (Op0BO->getOpcode()) {
7572 default: isValid = false; break; // Do not perform transform!
7573 case Instruction::Add:
7574 isValid = isLeftShift;
7576 case Instruction::Or:
7577 case Instruction::Xor:
7580 case Instruction::And:
7585 // If this is a signed shift right, and the high bit is modified
7586 // by the logical operation, do not perform the transformation.
7587 // The highBitSet boolean indicates the value of the high bit of
7588 // the constant which would cause it to be modified for this
7591 if (isValid && I.getOpcode() == Instruction::AShr)
7592 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7595 Constant *NewRHS = Context->getConstantExpr(I.getOpcode(), Op0C, Op1);
7597 Instruction *NewShift =
7598 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7599 InsertNewInstBefore(NewShift, I);
7600 NewShift->takeName(Op0BO);
7602 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7609 // Find out if this is a shift of a shift by a constant.
7610 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7611 if (ShiftOp && !ShiftOp->isShift())
7614 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7615 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7616 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7617 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7618 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7619 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7620 Value *X = ShiftOp->getOperand(0);
7622 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7624 const IntegerType *Ty = cast<IntegerType>(I.getType());
7626 // Check for (X << c1) << c2 and (X >> c1) >> c2
7627 if (I.getOpcode() == ShiftOp->getOpcode()) {
7628 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7630 if (AmtSum >= TypeBits) {
7631 if (I.getOpcode() != Instruction::AShr)
7632 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7633 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7636 return BinaryOperator::Create(I.getOpcode(), X,
7637 Context->getConstantInt(Ty, AmtSum));
7638 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7639 I.getOpcode() == Instruction::AShr) {
7640 if (AmtSum >= TypeBits)
7641 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7643 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7644 return BinaryOperator::CreateLShr(X, Context->getConstantInt(Ty, AmtSum));
7645 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7646 I.getOpcode() == Instruction::LShr) {
7647 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7648 if (AmtSum >= TypeBits)
7649 AmtSum = TypeBits-1;
7651 Instruction *Shift =
7652 BinaryOperator::CreateAShr(X, Context->getConstantInt(Ty, AmtSum));
7653 InsertNewInstBefore(Shift, I);
7655 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7656 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7659 // Okay, if we get here, one shift must be left, and the other shift must be
7660 // right. See if the amounts are equal.
7661 if (ShiftAmt1 == ShiftAmt2) {
7662 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7663 if (I.getOpcode() == Instruction::Shl) {
7664 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7665 return BinaryOperator::CreateAnd(X, Context->getConstantInt(Mask));
7667 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7668 if (I.getOpcode() == Instruction::LShr) {
7669 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7670 return BinaryOperator::CreateAnd(X, Context->getConstantInt(Mask));
7672 // We can simplify ((X << C) >>s C) into a trunc + sext.
7673 // NOTE: we could do this for any C, but that would make 'unusual' integer
7674 // types. For now, just stick to ones well-supported by the code
7676 const Type *SExtType = 0;
7677 switch (Ty->getBitWidth() - ShiftAmt1) {
7684 SExtType = Context->getIntegerType(Ty->getBitWidth() - ShiftAmt1);
7689 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7690 InsertNewInstBefore(NewTrunc, I);
7691 return new SExtInst(NewTrunc, Ty);
7693 // Otherwise, we can't handle it yet.
7694 } else if (ShiftAmt1 < ShiftAmt2) {
7695 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7697 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7698 if (I.getOpcode() == Instruction::Shl) {
7699 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7700 ShiftOp->getOpcode() == Instruction::AShr);
7701 Instruction *Shift =
7702 BinaryOperator::CreateShl(X, Context->getConstantInt(Ty, ShiftDiff));
7703 InsertNewInstBefore(Shift, I);
7705 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7706 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7709 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7710 if (I.getOpcode() == Instruction::LShr) {
7711 assert(ShiftOp->getOpcode() == Instruction::Shl);
7712 Instruction *Shift =
7713 BinaryOperator::CreateLShr(X, Context->getConstantInt(Ty, ShiftDiff));
7714 InsertNewInstBefore(Shift, I);
7716 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7717 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7720 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7722 assert(ShiftAmt2 < ShiftAmt1);
7723 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7725 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7726 if (I.getOpcode() == Instruction::Shl) {
7727 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7728 ShiftOp->getOpcode() == Instruction::AShr);
7729 Instruction *Shift =
7730 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7731 Context->getConstantInt(Ty, ShiftDiff));
7732 InsertNewInstBefore(Shift, I);
7734 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7735 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7738 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7739 if (I.getOpcode() == Instruction::LShr) {
7740 assert(ShiftOp->getOpcode() == Instruction::Shl);
7741 Instruction *Shift =
7742 BinaryOperator::CreateShl(X, Context->getConstantInt(Ty, ShiftDiff));
7743 InsertNewInstBefore(Shift, I);
7745 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7746 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7749 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7756 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7757 /// expression. If so, decompose it, returning some value X, such that Val is
7760 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7761 int &Offset, LLVMContext *Context) {
7762 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7763 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7764 Offset = CI->getZExtValue();
7766 return Context->getConstantInt(Type::Int32Ty, 0);
7767 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7768 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7769 if (I->getOpcode() == Instruction::Shl) {
7770 // This is a value scaled by '1 << the shift amt'.
7771 Scale = 1U << RHS->getZExtValue();
7773 return I->getOperand(0);
7774 } else if (I->getOpcode() == Instruction::Mul) {
7775 // This value is scaled by 'RHS'.
7776 Scale = RHS->getZExtValue();
7778 return I->getOperand(0);
7779 } else if (I->getOpcode() == Instruction::Add) {
7780 // We have X+C. Check to see if we really have (X*C2)+C1,
7781 // where C1 is divisible by C2.
7784 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7786 Offset += RHS->getZExtValue();
7793 // Otherwise, we can't look past this.
7800 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7801 /// try to eliminate the cast by moving the type information into the alloc.
7802 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7803 AllocationInst &AI) {
7804 const PointerType *PTy = cast<PointerType>(CI.getType());
7806 // Remove any uses of AI that are dead.
7807 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7809 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7810 Instruction *User = cast<Instruction>(*UI++);
7811 if (isInstructionTriviallyDead(User)) {
7812 while (UI != E && *UI == User)
7813 ++UI; // If this instruction uses AI more than once, don't break UI.
7816 DOUT << "IC: DCE: " << *User;
7817 EraseInstFromFunction(*User);
7821 // Get the type really allocated and the type casted to.
7822 const Type *AllocElTy = AI.getAllocatedType();
7823 const Type *CastElTy = PTy->getElementType();
7824 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7826 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7827 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7828 if (CastElTyAlign < AllocElTyAlign) return 0;
7830 // If the allocation has multiple uses, only promote it if we are strictly
7831 // increasing the alignment of the resultant allocation. If we keep it the
7832 // same, we open the door to infinite loops of various kinds. (A reference
7833 // from a dbg.declare doesn't count as a use for this purpose.)
7834 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7835 CastElTyAlign == AllocElTyAlign) return 0;
7837 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7838 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7839 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7841 // See if we can satisfy the modulus by pulling a scale out of the array
7843 unsigned ArraySizeScale;
7845 Value *NumElements = // See if the array size is a decomposable linear expr.
7846 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7847 ArrayOffset, Context);
7849 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7851 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7852 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7854 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7859 // If the allocation size is constant, form a constant mul expression
7860 Amt = Context->getConstantInt(Type::Int32Ty, Scale);
7861 if (isa<ConstantInt>(NumElements))
7862 Amt = Context->getConstantExprMul(cast<ConstantInt>(NumElements),
7863 cast<ConstantInt>(Amt));
7864 // otherwise multiply the amount and the number of elements
7866 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7867 Amt = InsertNewInstBefore(Tmp, AI);
7871 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7872 Value *Off = Context->getConstantInt(Type::Int32Ty, Offset, true);
7873 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7874 Amt = InsertNewInstBefore(Tmp, AI);
7877 AllocationInst *New;
7878 if (isa<MallocInst>(AI))
7879 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7881 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7882 InsertNewInstBefore(New, AI);
7885 // If the allocation has one real use plus a dbg.declare, just remove the
7887 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7888 EraseInstFromFunction(*DI);
7890 // If the allocation has multiple real uses, insert a cast and change all
7891 // things that used it to use the new cast. This will also hack on CI, but it
7893 else if (!AI.hasOneUse()) {
7894 AddUsesToWorkList(AI);
7895 // New is the allocation instruction, pointer typed. AI is the original
7896 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7897 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7898 InsertNewInstBefore(NewCast, AI);
7899 AI.replaceAllUsesWith(NewCast);
7901 return ReplaceInstUsesWith(CI, New);
7904 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7905 /// and return it as type Ty without inserting any new casts and without
7906 /// changing the computed value. This is used by code that tries to decide
7907 /// whether promoting or shrinking integer operations to wider or smaller types
7908 /// will allow us to eliminate a truncate or extend.
7910 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7911 /// extension operation if Ty is larger.
7913 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7914 /// should return true if trunc(V) can be computed by computing V in the smaller
7915 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7916 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7917 /// efficiently truncated.
7919 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7920 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7921 /// the final result.
7922 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7924 int &NumCastsRemoved){
7925 // We can always evaluate constants in another type.
7926 if (isa<Constant>(V))
7929 Instruction *I = dyn_cast<Instruction>(V);
7930 if (!I) return false;
7932 const Type *OrigTy = V->getType();
7934 // If this is an extension or truncate, we can often eliminate it.
7935 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7936 // If this is a cast from the destination type, we can trivially eliminate
7937 // it, and this will remove a cast overall.
7938 if (I->getOperand(0)->getType() == Ty) {
7939 // If the first operand is itself a cast, and is eliminable, do not count
7940 // this as an eliminable cast. We would prefer to eliminate those two
7942 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7948 // We can't extend or shrink something that has multiple uses: doing so would
7949 // require duplicating the instruction in general, which isn't profitable.
7950 if (!I->hasOneUse()) return false;
7952 unsigned Opc = I->getOpcode();
7954 case Instruction::Add:
7955 case Instruction::Sub:
7956 case Instruction::Mul:
7957 case Instruction::And:
7958 case Instruction::Or:
7959 case Instruction::Xor:
7960 // These operators can all arbitrarily be extended or truncated.
7961 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7963 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7966 case Instruction::UDiv:
7967 case Instruction::URem: {
7968 // UDiv and URem can be truncated if all the truncated bits are zero.
7969 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7970 uint32_t BitWidth = Ty->getScalarSizeInBits();
7971 if (BitWidth < OrigBitWidth) {
7972 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7973 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7974 MaskedValueIsZero(I->getOperand(1), Mask)) {
7975 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7977 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7983 case Instruction::Shl:
7984 // If we are truncating the result of this SHL, and if it's a shift of a
7985 // constant amount, we can always perform a SHL in a smaller type.
7986 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7987 uint32_t BitWidth = Ty->getScalarSizeInBits();
7988 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7989 CI->getLimitedValue(BitWidth) < BitWidth)
7990 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7994 case Instruction::LShr:
7995 // If this is a truncate of a logical shr, we can truncate it to a smaller
7996 // lshr iff we know that the bits we would otherwise be shifting in are
7998 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7999 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
8000 uint32_t BitWidth = Ty->getScalarSizeInBits();
8001 if (BitWidth < OrigBitWidth &&
8002 MaskedValueIsZero(I->getOperand(0),
8003 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
8004 CI->getLimitedValue(BitWidth) < BitWidth) {
8005 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8010 case Instruction::ZExt:
8011 case Instruction::SExt:
8012 case Instruction::Trunc:
8013 // If this is the same kind of case as our original (e.g. zext+zext), we
8014 // can safely replace it. Note that replacing it does not reduce the number
8015 // of casts in the input.
8019 // sext (zext ty1), ty2 -> zext ty2
8020 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8023 case Instruction::Select: {
8024 SelectInst *SI = cast<SelectInst>(I);
8025 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8027 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8030 case Instruction::PHI: {
8031 // We can change a phi if we can change all operands.
8032 PHINode *PN = cast<PHINode>(I);
8033 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8034 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8040 // TODO: Can handle more cases here.
8047 /// EvaluateInDifferentType - Given an expression that
8048 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8049 /// evaluate the expression.
8050 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8052 if (Constant *C = dyn_cast<Constant>(V))
8053 return Context->getConstantExprIntegerCast(C, Ty,
8054 isSigned /*Sext or ZExt*/);
8056 // Otherwise, it must be an instruction.
8057 Instruction *I = cast<Instruction>(V);
8058 Instruction *Res = 0;
8059 unsigned Opc = I->getOpcode();
8061 case Instruction::Add:
8062 case Instruction::Sub:
8063 case Instruction::Mul:
8064 case Instruction::And:
8065 case Instruction::Or:
8066 case Instruction::Xor:
8067 case Instruction::AShr:
8068 case Instruction::LShr:
8069 case Instruction::Shl:
8070 case Instruction::UDiv:
8071 case Instruction::URem: {
8072 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8073 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8074 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8077 case Instruction::Trunc:
8078 case Instruction::ZExt:
8079 case Instruction::SExt:
8080 // If the source type of the cast is the type we're trying for then we can
8081 // just return the source. There's no need to insert it because it is not
8083 if (I->getOperand(0)->getType() == Ty)
8084 return I->getOperand(0);
8086 // Otherwise, must be the same type of cast, so just reinsert a new one.
8087 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8090 case Instruction::Select: {
8091 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8092 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8093 Res = SelectInst::Create(I->getOperand(0), True, False);
8096 case Instruction::PHI: {
8097 PHINode *OPN = cast<PHINode>(I);
8098 PHINode *NPN = PHINode::Create(Ty);
8099 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8100 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8101 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8107 // TODO: Can handle more cases here.
8108 llvm_unreachable("Unreachable!");
8113 return InsertNewInstBefore(Res, *I);
8116 /// @brief Implement the transforms common to all CastInst visitors.
8117 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8118 Value *Src = CI.getOperand(0);
8120 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8121 // eliminate it now.
8122 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8123 if (Instruction::CastOps opc =
8124 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8125 // The first cast (CSrc) is eliminable so we need to fix up or replace
8126 // the second cast (CI). CSrc will then have a good chance of being dead.
8127 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8131 // If we are casting a select then fold the cast into the select
8132 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8133 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8136 // If we are casting a PHI then fold the cast into the PHI
8137 if (isa<PHINode>(Src))
8138 if (Instruction *NV = FoldOpIntoPhi(CI))
8144 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8145 /// or not there is a sequence of GEP indices into the type that will land us at
8146 /// the specified offset. If so, fill them into NewIndices and return the
8147 /// resultant element type, otherwise return null.
8148 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8149 SmallVectorImpl<Value*> &NewIndices,
8150 const TargetData *TD,
8151 LLVMContext *Context) {
8152 if (!Ty->isSized()) return 0;
8154 // Start with the index over the outer type. Note that the type size
8155 // might be zero (even if the offset isn't zero) if the indexed type
8156 // is something like [0 x {int, int}]
8157 const Type *IntPtrTy = TD->getIntPtrType();
8158 int64_t FirstIdx = 0;
8159 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8160 FirstIdx = Offset/TySize;
8161 Offset -= FirstIdx*TySize;
8163 // Handle hosts where % returns negative instead of values [0..TySize).
8167 assert(Offset >= 0);
8169 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8172 NewIndices.push_back(Context->getConstantInt(IntPtrTy, FirstIdx));
8174 // Index into the types. If we fail, set OrigBase to null.
8176 // Indexing into tail padding between struct/array elements.
8177 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8180 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8181 const StructLayout *SL = TD->getStructLayout(STy);
8182 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8183 "Offset must stay within the indexed type");
8185 unsigned Elt = SL->getElementContainingOffset(Offset);
8186 NewIndices.push_back(Context->getConstantInt(Type::Int32Ty, Elt));
8188 Offset -= SL->getElementOffset(Elt);
8189 Ty = STy->getElementType(Elt);
8190 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8191 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8192 assert(EltSize && "Cannot index into a zero-sized array");
8193 NewIndices.push_back(Context->getConstantInt(IntPtrTy,Offset/EltSize));
8195 Ty = AT->getElementType();
8197 // Otherwise, we can't index into the middle of this atomic type, bail.
8205 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8206 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8207 Value *Src = CI.getOperand(0);
8209 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8210 // If casting the result of a getelementptr instruction with no offset, turn
8211 // this into a cast of the original pointer!
8212 if (GEP->hasAllZeroIndices()) {
8213 // Changing the cast operand is usually not a good idea but it is safe
8214 // here because the pointer operand is being replaced with another
8215 // pointer operand so the opcode doesn't need to change.
8217 CI.setOperand(0, GEP->getOperand(0));
8221 // If the GEP has a single use, and the base pointer is a bitcast, and the
8222 // GEP computes a constant offset, see if we can convert these three
8223 // instructions into fewer. This typically happens with unions and other
8224 // non-type-safe code.
8225 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8226 if (GEP->hasAllConstantIndices()) {
8227 // We are guaranteed to get a constant from EmitGEPOffset.
8228 ConstantInt *OffsetV =
8229 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8230 int64_t Offset = OffsetV->getSExtValue();
8232 // Get the base pointer input of the bitcast, and the type it points to.
8233 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8234 const Type *GEPIdxTy =
8235 cast<PointerType>(OrigBase->getType())->getElementType();
8236 SmallVector<Value*, 8> NewIndices;
8237 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8238 // If we were able to index down into an element, create the GEP
8239 // and bitcast the result. This eliminates one bitcast, potentially
8241 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
8243 NewIndices.end(), "");
8244 InsertNewInstBefore(NGEP, CI);
8245 NGEP->takeName(GEP);
8247 if (isa<BitCastInst>(CI))
8248 return new BitCastInst(NGEP, CI.getType());
8249 assert(isa<PtrToIntInst>(CI));
8250 return new PtrToIntInst(NGEP, CI.getType());
8256 return commonCastTransforms(CI);
8259 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8260 /// type like i42. We don't want to introduce operations on random non-legal
8261 /// integer types where they don't already exist in the code. In the future,
8262 /// we should consider making this based off target-data, so that 32-bit targets
8263 /// won't get i64 operations etc.
8264 static bool isSafeIntegerType(const Type *Ty) {
8265 switch (Ty->getPrimitiveSizeInBits()) {
8276 /// commonIntCastTransforms - This function implements the common transforms
8277 /// for trunc, zext, and sext.
8278 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8279 if (Instruction *Result = commonCastTransforms(CI))
8282 Value *Src = CI.getOperand(0);
8283 const Type *SrcTy = Src->getType();
8284 const Type *DestTy = CI.getType();
8285 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8286 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8288 // See if we can simplify any instructions used by the LHS whose sole
8289 // purpose is to compute bits we don't care about.
8290 if (SimplifyDemandedInstructionBits(CI))
8293 // If the source isn't an instruction or has more than one use then we
8294 // can't do anything more.
8295 Instruction *SrcI = dyn_cast<Instruction>(Src);
8296 if (!SrcI || !Src->hasOneUse())
8299 // Attempt to propagate the cast into the instruction for int->int casts.
8300 int NumCastsRemoved = 0;
8301 // Only do this if the dest type is a simple type, don't convert the
8302 // expression tree to something weird like i93 unless the source is also
8304 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8305 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8306 CanEvaluateInDifferentType(SrcI, DestTy,
8307 CI.getOpcode(), NumCastsRemoved)) {
8308 // If this cast is a truncate, evaluting in a different type always
8309 // eliminates the cast, so it is always a win. If this is a zero-extension,
8310 // we need to do an AND to maintain the clear top-part of the computation,
8311 // so we require that the input have eliminated at least one cast. If this
8312 // is a sign extension, we insert two new casts (to do the extension) so we
8313 // require that two casts have been eliminated.
8314 bool DoXForm = false;
8315 bool JustReplace = false;
8316 switch (CI.getOpcode()) {
8318 // All the others use floating point so we shouldn't actually
8319 // get here because of the check above.
8320 llvm_unreachable("Unknown cast type");
8321 case Instruction::Trunc:
8324 case Instruction::ZExt: {
8325 DoXForm = NumCastsRemoved >= 1;
8326 if (!DoXForm && 0) {
8327 // If it's unnecessary to issue an AND to clear the high bits, it's
8328 // always profitable to do this xform.
8329 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8330 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8331 if (MaskedValueIsZero(TryRes, Mask))
8332 return ReplaceInstUsesWith(CI, TryRes);
8334 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8335 if (TryI->use_empty())
8336 EraseInstFromFunction(*TryI);
8340 case Instruction::SExt: {
8341 DoXForm = NumCastsRemoved >= 2;
8342 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8343 // If we do not have to emit the truncate + sext pair, then it's always
8344 // profitable to do this xform.
8346 // It's not safe to eliminate the trunc + sext pair if one of the
8347 // eliminated cast is a truncate. e.g.
8348 // t2 = trunc i32 t1 to i16
8349 // t3 = sext i16 t2 to i32
8352 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8353 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8354 if (NumSignBits > (DestBitSize - SrcBitSize))
8355 return ReplaceInstUsesWith(CI, TryRes);
8357 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8358 if (TryI->use_empty())
8359 EraseInstFromFunction(*TryI);
8366 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
8368 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8369 CI.getOpcode() == Instruction::SExt);
8371 // Just replace this cast with the result.
8372 return ReplaceInstUsesWith(CI, Res);
8374 assert(Res->getType() == DestTy);
8375 switch (CI.getOpcode()) {
8376 default: llvm_unreachable("Unknown cast type!");
8377 case Instruction::Trunc:
8378 // Just replace this cast with the result.
8379 return ReplaceInstUsesWith(CI, Res);
8380 case Instruction::ZExt: {
8381 assert(SrcBitSize < DestBitSize && "Not a zext?");
8383 // If the high bits are already zero, just replace this cast with the
8385 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8386 if (MaskedValueIsZero(Res, Mask))
8387 return ReplaceInstUsesWith(CI, Res);
8389 // We need to emit an AND to clear the high bits.
8390 Constant *C = Context->getConstantInt(APInt::getLowBitsSet(DestBitSize,
8392 return BinaryOperator::CreateAnd(Res, C);
8394 case Instruction::SExt: {
8395 // If the high bits are already filled with sign bit, just replace this
8396 // cast with the result.
8397 unsigned NumSignBits = ComputeNumSignBits(Res);
8398 if (NumSignBits > (DestBitSize - SrcBitSize))
8399 return ReplaceInstUsesWith(CI, Res);
8401 // We need to emit a cast to truncate, then a cast to sext.
8402 return CastInst::Create(Instruction::SExt,
8403 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8410 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8411 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8413 switch (SrcI->getOpcode()) {
8414 case Instruction::Add:
8415 case Instruction::Mul:
8416 case Instruction::And:
8417 case Instruction::Or:
8418 case Instruction::Xor:
8419 // If we are discarding information, rewrite.
8420 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8421 // Don't insert two casts unless at least one can be eliminated.
8422 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8423 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8424 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8425 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8426 return BinaryOperator::Create(
8427 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8431 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8432 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8433 SrcI->getOpcode() == Instruction::Xor &&
8434 Op1 == Context->getConstantIntTrue() &&
8435 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8436 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8437 return BinaryOperator::CreateXor(New,
8438 Context->getConstantInt(CI.getType(), 1));
8442 case Instruction::Shl: {
8443 // Canonicalize trunc inside shl, if we can.
8444 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8445 if (CI && DestBitSize < SrcBitSize &&
8446 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8447 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8448 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8449 return BinaryOperator::CreateShl(Op0c, Op1c);
8457 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8458 if (Instruction *Result = commonIntCastTransforms(CI))
8461 Value *Src = CI.getOperand(0);
8462 const Type *Ty = CI.getType();
8463 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8464 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8466 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8467 if (DestBitWidth == 1) {
8468 Constant *One = Context->getConstantInt(Src->getType(), 1);
8469 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8470 Value *Zero = Context->getNullValue(Src->getType());
8471 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Src, Zero);
8474 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8475 ConstantInt *ShAmtV = 0;
8477 if (Src->hasOneUse() &&
8478 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)), *Context)) {
8479 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8481 // Get a mask for the bits shifting in.
8482 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8483 if (MaskedValueIsZero(ShiftOp, Mask)) {
8484 if (ShAmt >= DestBitWidth) // All zeros.
8485 return ReplaceInstUsesWith(CI, Context->getNullValue(Ty));
8487 // Okay, we can shrink this. Truncate the input, then return a new
8489 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8490 Value *V2 = Context->getConstantExprTrunc(ShAmtV, Ty);
8491 return BinaryOperator::CreateLShr(V1, V2);
8498 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8499 /// in order to eliminate the icmp.
8500 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8502 // If we are just checking for a icmp eq of a single bit and zext'ing it
8503 // to an integer, then shift the bit to the appropriate place and then
8504 // cast to integer to avoid the comparison.
8505 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8506 const APInt &Op1CV = Op1C->getValue();
8508 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8509 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8510 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8511 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8512 if (!DoXform) return ICI;
8514 Value *In = ICI->getOperand(0);
8515 Value *Sh = Context->getConstantInt(In->getType(),
8516 In->getType()->getScalarSizeInBits()-1);
8517 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8518 In->getName()+".lobit"),
8520 if (In->getType() != CI.getType())
8521 In = CastInst::CreateIntegerCast(In, CI.getType(),
8522 false/*ZExt*/, "tmp", &CI);
8524 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8525 Constant *One = Context->getConstantInt(In->getType(), 1);
8526 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8527 In->getName()+".not"),
8531 return ReplaceInstUsesWith(CI, In);
8536 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8537 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8538 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8539 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8540 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8541 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8542 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8543 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8544 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8545 // This only works for EQ and NE
8546 ICI->isEquality()) {
8547 // If Op1C some other power of two, convert:
8548 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8549 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8550 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8551 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8553 APInt KnownZeroMask(~KnownZero);
8554 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8555 if (!DoXform) return ICI;
8557 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8558 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8559 // (X&4) == 2 --> false
8560 // (X&4) != 2 --> true
8561 Constant *Res = Context->getConstantInt(Type::Int1Ty, isNE);
8562 Res = Context->getConstantExprZExt(Res, CI.getType());
8563 return ReplaceInstUsesWith(CI, Res);
8566 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8567 Value *In = ICI->getOperand(0);
8569 // Perform a logical shr by shiftamt.
8570 // Insert the shift to put the result in the low bit.
8571 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8572 Context->getConstantInt(In->getType(), ShiftAmt),
8573 In->getName()+".lobit"), CI);
8576 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8577 Constant *One = Context->getConstantInt(In->getType(), 1);
8578 In = BinaryOperator::CreateXor(In, One, "tmp");
8579 InsertNewInstBefore(cast<Instruction>(In), CI);
8582 if (CI.getType() == In->getType())
8583 return ReplaceInstUsesWith(CI, In);
8585 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8593 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8594 // If one of the common conversion will work ..
8595 if (Instruction *Result = commonIntCastTransforms(CI))
8598 Value *Src = CI.getOperand(0);
8600 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8601 // types and if the sizes are just right we can convert this into a logical
8602 // 'and' which will be much cheaper than the pair of casts.
8603 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8604 // Get the sizes of the types involved. We know that the intermediate type
8605 // will be smaller than A or C, but don't know the relation between A and C.
8606 Value *A = CSrc->getOperand(0);
8607 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8608 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8609 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8610 // If we're actually extending zero bits, then if
8611 // SrcSize < DstSize: zext(a & mask)
8612 // SrcSize == DstSize: a & mask
8613 // SrcSize > DstSize: trunc(a) & mask
8614 if (SrcSize < DstSize) {
8615 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8616 Constant *AndConst = Context->getConstantInt(A->getType(), AndValue);
8618 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8619 InsertNewInstBefore(And, CI);
8620 return new ZExtInst(And, CI.getType());
8621 } else if (SrcSize == DstSize) {
8622 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8623 return BinaryOperator::CreateAnd(A, Context->getConstantInt(A->getType(),
8625 } else if (SrcSize > DstSize) {
8626 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8627 InsertNewInstBefore(Trunc, CI);
8628 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8629 return BinaryOperator::CreateAnd(Trunc,
8630 Context->getConstantInt(Trunc->getType(),
8635 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8636 return transformZExtICmp(ICI, CI);
8638 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8639 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8640 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8641 // of the (zext icmp) will be transformed.
8642 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8643 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8644 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8645 (transformZExtICmp(LHS, CI, false) ||
8646 transformZExtICmp(RHS, CI, false))) {
8647 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8648 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8649 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8653 // zext(trunc(t) & C) -> (t & zext(C)).
8654 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8655 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8656 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8657 Value *TI0 = TI->getOperand(0);
8658 if (TI0->getType() == CI.getType())
8660 BinaryOperator::CreateAnd(TI0,
8661 Context->getConstantExprZExt(C, CI.getType()));
8664 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8665 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8666 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8667 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8668 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8669 And->getOperand(1) == C)
8670 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8671 Value *TI0 = TI->getOperand(0);
8672 if (TI0->getType() == CI.getType()) {
8673 Constant *ZC = Context->getConstantExprZExt(C, CI.getType());
8674 Instruction *NewAnd = BinaryOperator::CreateAnd(TI0, ZC, "tmp");
8675 InsertNewInstBefore(NewAnd, *And);
8676 return BinaryOperator::CreateXor(NewAnd, ZC);
8683 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8684 if (Instruction *I = commonIntCastTransforms(CI))
8687 Value *Src = CI.getOperand(0);
8689 // Canonicalize sign-extend from i1 to a select.
8690 if (Src->getType() == Type::Int1Ty)
8691 return SelectInst::Create(Src,
8692 Context->getAllOnesValue(CI.getType()),
8693 Context->getNullValue(CI.getType()));
8695 // See if the value being truncated is already sign extended. If so, just
8696 // eliminate the trunc/sext pair.
8697 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8698 Value *Op = cast<User>(Src)->getOperand(0);
8699 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8700 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8701 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8702 unsigned NumSignBits = ComputeNumSignBits(Op);
8704 if (OpBits == DestBits) {
8705 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8706 // bits, it is already ready.
8707 if (NumSignBits > DestBits-MidBits)
8708 return ReplaceInstUsesWith(CI, Op);
8709 } else if (OpBits < DestBits) {
8710 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8711 // bits, just sext from i32.
8712 if (NumSignBits > OpBits-MidBits)
8713 return new SExtInst(Op, CI.getType(), "tmp");
8715 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8716 // bits, just truncate to i32.
8717 if (NumSignBits > OpBits-MidBits)
8718 return new TruncInst(Op, CI.getType(), "tmp");
8722 // If the input is a shl/ashr pair of a same constant, then this is a sign
8723 // extension from a smaller value. If we could trust arbitrary bitwidth
8724 // integers, we could turn this into a truncate to the smaller bit and then
8725 // use a sext for the whole extension. Since we don't, look deeper and check
8726 // for a truncate. If the source and dest are the same type, eliminate the
8727 // trunc and extend and just do shifts. For example, turn:
8728 // %a = trunc i32 %i to i8
8729 // %b = shl i8 %a, 6
8730 // %c = ashr i8 %b, 6
8731 // %d = sext i8 %c to i32
8733 // %a = shl i32 %i, 30
8734 // %d = ashr i32 %a, 30
8736 ConstantInt *BA = 0, *CA = 0;
8737 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8738 m_ConstantInt(CA)), *Context) &&
8739 BA == CA && isa<TruncInst>(A)) {
8740 Value *I = cast<TruncInst>(A)->getOperand(0);
8741 if (I->getType() == CI.getType()) {
8742 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8743 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8744 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8745 Constant *ShAmtV = Context->getConstantInt(CI.getType(), ShAmt);
8746 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8748 return BinaryOperator::CreateAShr(I, ShAmtV);
8755 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8756 /// in the specified FP type without changing its value.
8757 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8758 LLVMContext *Context) {
8760 APFloat F = CFP->getValueAPF();
8761 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8763 return Context->getConstantFP(F);
8767 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8768 /// through it until we get the source value.
8769 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8770 if (Instruction *I = dyn_cast<Instruction>(V))
8771 if (I->getOpcode() == Instruction::FPExt)
8772 return LookThroughFPExtensions(I->getOperand(0), Context);
8774 // If this value is a constant, return the constant in the smallest FP type
8775 // that can accurately represent it. This allows us to turn
8776 // (float)((double)X+2.0) into x+2.0f.
8777 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8778 if (CFP->getType() == Type::PPC_FP128Ty)
8779 return V; // No constant folding of this.
8780 // See if the value can be truncated to float and then reextended.
8781 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8783 if (CFP->getType() == Type::DoubleTy)
8784 return V; // Won't shrink.
8785 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8787 // Don't try to shrink to various long double types.
8793 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8794 if (Instruction *I = commonCastTransforms(CI))
8797 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8798 // smaller than the destination type, we can eliminate the truncate by doing
8799 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8800 // many builtins (sqrt, etc).
8801 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8802 if (OpI && OpI->hasOneUse()) {
8803 switch (OpI->getOpcode()) {
8805 case Instruction::FAdd:
8806 case Instruction::FSub:
8807 case Instruction::FMul:
8808 case Instruction::FDiv:
8809 case Instruction::FRem:
8810 const Type *SrcTy = OpI->getType();
8811 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8812 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8813 if (LHSTrunc->getType() != SrcTy &&
8814 RHSTrunc->getType() != SrcTy) {
8815 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8816 // If the source types were both smaller than the destination type of
8817 // the cast, do this xform.
8818 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8819 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8820 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8822 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8824 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8833 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8834 return commonCastTransforms(CI);
8837 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8838 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8840 return commonCastTransforms(FI);
8842 // fptoui(uitofp(X)) --> X
8843 // fptoui(sitofp(X)) --> X
8844 // This is safe if the intermediate type has enough bits in its mantissa to
8845 // accurately represent all values of X. For example, do not do this with
8846 // i64->float->i64. This is also safe for sitofp case, because any negative
8847 // 'X' value would cause an undefined result for the fptoui.
8848 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8849 OpI->getOperand(0)->getType() == FI.getType() &&
8850 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8851 OpI->getType()->getFPMantissaWidth())
8852 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8854 return commonCastTransforms(FI);
8857 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8858 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8860 return commonCastTransforms(FI);
8862 // fptosi(sitofp(X)) --> X
8863 // fptosi(uitofp(X)) --> X
8864 // This is safe if the intermediate type has enough bits in its mantissa to
8865 // accurately represent all values of X. For example, do not do this with
8866 // i64->float->i64. This is also safe for sitofp case, because any negative
8867 // 'X' value would cause an undefined result for the fptoui.
8868 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8869 OpI->getOperand(0)->getType() == FI.getType() &&
8870 (int)FI.getType()->getScalarSizeInBits() <=
8871 OpI->getType()->getFPMantissaWidth())
8872 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8874 return commonCastTransforms(FI);
8877 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8878 return commonCastTransforms(CI);
8881 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8882 return commonCastTransforms(CI);
8885 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8886 // If the destination integer type is smaller than the intptr_t type for
8887 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8888 // trunc to be exposed to other transforms. Don't do this for extending
8889 // ptrtoint's, because we don't know if the target sign or zero extends its
8891 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8892 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8893 TD->getIntPtrType(),
8895 return new TruncInst(P, CI.getType());
8898 return commonPointerCastTransforms(CI);
8901 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8902 // If the source integer type is larger than the intptr_t type for
8903 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8904 // allows the trunc to be exposed to other transforms. Don't do this for
8905 // extending inttoptr's, because we don't know if the target sign or zero
8906 // extends to pointers.
8907 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
8908 TD->getPointerSizeInBits()) {
8909 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8910 TD->getIntPtrType(),
8912 return new IntToPtrInst(P, CI.getType());
8915 if (Instruction *I = commonCastTransforms(CI))
8918 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8919 if (!DestPointee->isSized()) return 0;
8921 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8924 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8925 m_ConstantInt(Cst)), *Context)) {
8926 // If the source and destination operands have the same type, see if this
8927 // is a single-index GEP.
8928 if (X->getType() == CI.getType()) {
8929 // Get the size of the pointee type.
8930 uint64_t Size = TD->getTypeAllocSize(DestPointee);
8932 // Convert the constant to intptr type.
8933 APInt Offset = Cst->getValue();
8934 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8936 // If Offset is evenly divisible by Size, we can do this xform.
8937 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8938 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8939 GetElementPtrInst *GEP =
8940 GetElementPtrInst::Create(X, Context->getConstantInt(Offset));
8941 // A gep synthesized from inttoptr+add+ptrtoint must be assumed to
8942 // potentially overflow, in the absense of further analysis.
8943 cast<GEPOperator>(GEP)->setHasNoPointerOverflow(false);
8947 // TODO: Could handle other cases, e.g. where add is indexing into field of
8949 } else if (CI.getOperand(0)->hasOneUse() &&
8950 match(CI.getOperand(0), m_Add(m_Value(X),
8951 m_ConstantInt(Cst)), *Context)) {
8952 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8953 // "inttoptr+GEP" instead of "add+intptr".
8955 // Get the size of the pointee type.
8956 uint64_t Size = TD->getTypeAllocSize(DestPointee);
8958 // Convert the constant to intptr type.
8959 APInt Offset = Cst->getValue();
8960 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8962 // If Offset is evenly divisible by Size, we can do this xform.
8963 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8964 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8966 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8968 GetElementPtrInst *GEP =
8969 GetElementPtrInst::Create(P, Context->getConstantInt(Offset), "tmp");
8970 // A gep synthesized from inttoptr+add+ptrtoint must be assumed to
8971 // potentially overflow, in the absense of further analysis.
8972 cast<GEPOperator>(GEP)->setHasNoPointerOverflow(false);
8979 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8980 // If the operands are integer typed then apply the integer transforms,
8981 // otherwise just apply the common ones.
8982 Value *Src = CI.getOperand(0);
8983 const Type *SrcTy = Src->getType();
8984 const Type *DestTy = CI.getType();
8986 if (isa<PointerType>(SrcTy)) {
8987 if (Instruction *I = commonPointerCastTransforms(CI))
8990 if (Instruction *Result = commonCastTransforms(CI))
8995 // Get rid of casts from one type to the same type. These are useless and can
8996 // be replaced by the operand.
8997 if (DestTy == Src->getType())
8998 return ReplaceInstUsesWith(CI, Src);
9000 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
9001 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
9002 const Type *DstElTy = DstPTy->getElementType();
9003 const Type *SrcElTy = SrcPTy->getElementType();
9005 // If the address spaces don't match, don't eliminate the bitcast, which is
9006 // required for changing types.
9007 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
9010 // If we are casting a malloc or alloca to a pointer to a type of the same
9011 // size, rewrite the allocation instruction to allocate the "right" type.
9012 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
9013 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
9016 // If the source and destination are pointers, and this cast is equivalent
9017 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
9018 // This can enhance SROA and other transforms that want type-safe pointers.
9019 Constant *ZeroUInt = Context->getNullValue(Type::Int32Ty);
9020 unsigned NumZeros = 0;
9021 while (SrcElTy != DstElTy &&
9022 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
9023 SrcElTy->getNumContainedTypes() /* not "{}" */) {
9024 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
9028 // If we found a path from the src to dest, create the getelementptr now.
9029 if (SrcElTy == DstElTy) {
9030 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
9031 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
9032 ((Instruction*) NULL));
9036 if (const VectorType *DestVTy = dyn_cast<VectorType>(DestTy)) {
9037 if (DestVTy->getNumElements() == 1) {
9038 if (!isa<VectorType>(SrcTy)) {
9039 Value *Elem = InsertCastBefore(Instruction::BitCast, Src,
9040 DestVTy->getElementType(), CI);
9041 return InsertElementInst::Create(Context->getUndef(DestTy), Elem,
9042 Context->getNullValue(Type::Int32Ty));
9044 // FIXME: Canonicalize bitcast(insertelement) -> insertelement(bitcast)
9048 if (const VectorType *SrcVTy = dyn_cast<VectorType>(SrcTy)) {
9049 if (SrcVTy->getNumElements() == 1) {
9050 if (!isa<VectorType>(DestTy)) {
9052 new ExtractElementInst(Src, Context->getNullValue(Type::Int32Ty));
9053 InsertNewInstBefore(Elem, CI);
9054 return CastInst::Create(Instruction::BitCast, Elem, DestTy);
9059 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9060 if (SVI->hasOneUse()) {
9061 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9062 // a bitconvert to a vector with the same # elts.
9063 if (isa<VectorType>(DestTy) &&
9064 cast<VectorType>(DestTy)->getNumElements() ==
9065 SVI->getType()->getNumElements() &&
9066 SVI->getType()->getNumElements() ==
9067 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9069 // If either of the operands is a cast from CI.getType(), then
9070 // evaluating the shuffle in the casted destination's type will allow
9071 // us to eliminate at least one cast.
9072 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9073 Tmp->getOperand(0)->getType() == DestTy) ||
9074 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9075 Tmp->getOperand(0)->getType() == DestTy)) {
9076 Value *LHS = InsertCastBefore(Instruction::BitCast,
9077 SVI->getOperand(0), DestTy, CI);
9078 Value *RHS = InsertCastBefore(Instruction::BitCast,
9079 SVI->getOperand(1), DestTy, CI);
9080 // Return a new shuffle vector. Use the same element ID's, as we
9081 // know the vector types match #elts.
9082 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9090 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9092 /// %D = select %cond, %C, %A
9094 /// %C = select %cond, %B, 0
9097 /// Assuming that the specified instruction is an operand to the select, return
9098 /// a bitmask indicating which operands of this instruction are foldable if they
9099 /// equal the other incoming value of the select.
9101 static unsigned GetSelectFoldableOperands(Instruction *I) {
9102 switch (I->getOpcode()) {
9103 case Instruction::Add:
9104 case Instruction::Mul:
9105 case Instruction::And:
9106 case Instruction::Or:
9107 case Instruction::Xor:
9108 return 3; // Can fold through either operand.
9109 case Instruction::Sub: // Can only fold on the amount subtracted.
9110 case Instruction::Shl: // Can only fold on the shift amount.
9111 case Instruction::LShr:
9112 case Instruction::AShr:
9115 return 0; // Cannot fold
9119 /// GetSelectFoldableConstant - For the same transformation as the previous
9120 /// function, return the identity constant that goes into the select.
9121 static Constant *GetSelectFoldableConstant(Instruction *I,
9122 LLVMContext *Context) {
9123 switch (I->getOpcode()) {
9124 default: llvm_unreachable("This cannot happen!");
9125 case Instruction::Add:
9126 case Instruction::Sub:
9127 case Instruction::Or:
9128 case Instruction::Xor:
9129 case Instruction::Shl:
9130 case Instruction::LShr:
9131 case Instruction::AShr:
9132 return Context->getNullValue(I->getType());
9133 case Instruction::And:
9134 return Context->getAllOnesValue(I->getType());
9135 case Instruction::Mul:
9136 return Context->getConstantInt(I->getType(), 1);
9140 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9141 /// have the same opcode and only one use each. Try to simplify this.
9142 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9144 if (TI->getNumOperands() == 1) {
9145 // If this is a non-volatile load or a cast from the same type,
9148 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9151 return 0; // unknown unary op.
9154 // Fold this by inserting a select from the input values.
9155 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9156 FI->getOperand(0), SI.getName()+".v");
9157 InsertNewInstBefore(NewSI, SI);
9158 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9162 // Only handle binary operators here.
9163 if (!isa<BinaryOperator>(TI))
9166 // Figure out if the operations have any operands in common.
9167 Value *MatchOp, *OtherOpT, *OtherOpF;
9169 if (TI->getOperand(0) == FI->getOperand(0)) {
9170 MatchOp = TI->getOperand(0);
9171 OtherOpT = TI->getOperand(1);
9172 OtherOpF = FI->getOperand(1);
9173 MatchIsOpZero = true;
9174 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9175 MatchOp = TI->getOperand(1);
9176 OtherOpT = TI->getOperand(0);
9177 OtherOpF = FI->getOperand(0);
9178 MatchIsOpZero = false;
9179 } else if (!TI->isCommutative()) {
9181 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9182 MatchOp = TI->getOperand(0);
9183 OtherOpT = TI->getOperand(1);
9184 OtherOpF = FI->getOperand(0);
9185 MatchIsOpZero = true;
9186 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9187 MatchOp = TI->getOperand(1);
9188 OtherOpT = TI->getOperand(0);
9189 OtherOpF = FI->getOperand(1);
9190 MatchIsOpZero = true;
9195 // If we reach here, they do have operations in common.
9196 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9197 OtherOpF, SI.getName()+".v");
9198 InsertNewInstBefore(NewSI, SI);
9200 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9202 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9204 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9206 llvm_unreachable("Shouldn't get here");
9210 static bool isSelect01(Constant *C1, Constant *C2) {
9211 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9214 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9217 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9220 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9221 /// facilitate further optimization.
9222 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9224 // See the comment above GetSelectFoldableOperands for a description of the
9225 // transformation we are doing here.
9226 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9227 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9228 !isa<Constant>(FalseVal)) {
9229 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9230 unsigned OpToFold = 0;
9231 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9233 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9238 Constant *C = GetSelectFoldableConstant(TVI, Context);
9239 Value *OOp = TVI->getOperand(2-OpToFold);
9240 // Avoid creating select between 2 constants unless it's selecting
9242 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9243 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9244 InsertNewInstBefore(NewSel, SI);
9245 NewSel->takeName(TVI);
9246 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9247 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9248 llvm_unreachable("Unknown instruction!!");
9255 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9256 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9257 !isa<Constant>(TrueVal)) {
9258 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9259 unsigned OpToFold = 0;
9260 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9262 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9267 Constant *C = GetSelectFoldableConstant(FVI, Context);
9268 Value *OOp = FVI->getOperand(2-OpToFold);
9269 // Avoid creating select between 2 constants unless it's selecting
9271 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9272 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9273 InsertNewInstBefore(NewSel, SI);
9274 NewSel->takeName(FVI);
9275 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9276 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9277 llvm_unreachable("Unknown instruction!!");
9287 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9288 /// ICmpInst as its first operand.
9290 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9292 bool Changed = false;
9293 ICmpInst::Predicate Pred = ICI->getPredicate();
9294 Value *CmpLHS = ICI->getOperand(0);
9295 Value *CmpRHS = ICI->getOperand(1);
9296 Value *TrueVal = SI.getTrueValue();
9297 Value *FalseVal = SI.getFalseValue();
9299 // Check cases where the comparison is with a constant that
9300 // can be adjusted to fit the min/max idiom. We may edit ICI in
9301 // place here, so make sure the select is the only user.
9302 if (ICI->hasOneUse())
9303 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9306 case ICmpInst::ICMP_ULT:
9307 case ICmpInst::ICMP_SLT: {
9308 // X < MIN ? T : F --> F
9309 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9310 return ReplaceInstUsesWith(SI, FalseVal);
9311 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9312 Constant *AdjustedRHS = SubOne(CI, Context);
9313 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9314 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9315 Pred = ICmpInst::getSwappedPredicate(Pred);
9316 CmpRHS = AdjustedRHS;
9317 std::swap(FalseVal, TrueVal);
9318 ICI->setPredicate(Pred);
9319 ICI->setOperand(1, CmpRHS);
9320 SI.setOperand(1, TrueVal);
9321 SI.setOperand(2, FalseVal);
9326 case ICmpInst::ICMP_UGT:
9327 case ICmpInst::ICMP_SGT: {
9328 // X > MAX ? T : F --> F
9329 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9330 return ReplaceInstUsesWith(SI, FalseVal);
9331 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9332 Constant *AdjustedRHS = AddOne(CI, Context);
9333 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9334 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9335 Pred = ICmpInst::getSwappedPredicate(Pred);
9336 CmpRHS = AdjustedRHS;
9337 std::swap(FalseVal, TrueVal);
9338 ICI->setPredicate(Pred);
9339 ICI->setOperand(1, CmpRHS);
9340 SI.setOperand(1, TrueVal);
9341 SI.setOperand(2, FalseVal);
9348 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9349 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9350 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9351 if (match(TrueVal, m_ConstantInt<-1>(), *Context) &&
9352 match(FalseVal, m_ConstantInt<0>(), *Context))
9353 Pred = ICI->getPredicate();
9354 else if (match(TrueVal, m_ConstantInt<0>(), *Context) &&
9355 match(FalseVal, m_ConstantInt<-1>(), *Context))
9356 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9358 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9359 // If we are just checking for a icmp eq of a single bit and zext'ing it
9360 // to an integer, then shift the bit to the appropriate place and then
9361 // cast to integer to avoid the comparison.
9362 const APInt &Op1CV = CI->getValue();
9364 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9365 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9366 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9367 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9368 Value *In = ICI->getOperand(0);
9369 Value *Sh = Context->getConstantInt(In->getType(),
9370 In->getType()->getScalarSizeInBits()-1);
9371 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9372 In->getName()+".lobit"),
9374 if (In->getType() != SI.getType())
9375 In = CastInst::CreateIntegerCast(In, SI.getType(),
9376 true/*SExt*/, "tmp", ICI);
9378 if (Pred == ICmpInst::ICMP_SGT)
9379 In = InsertNewInstBefore(BinaryOperator::CreateNot(*Context, In,
9380 In->getName()+".not"), *ICI);
9382 return ReplaceInstUsesWith(SI, In);
9387 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9388 // Transform (X == Y) ? X : Y -> Y
9389 if (Pred == ICmpInst::ICMP_EQ)
9390 return ReplaceInstUsesWith(SI, FalseVal);
9391 // Transform (X != Y) ? X : Y -> X
9392 if (Pred == ICmpInst::ICMP_NE)
9393 return ReplaceInstUsesWith(SI, TrueVal);
9394 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9396 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9397 // Transform (X == Y) ? Y : X -> X
9398 if (Pred == ICmpInst::ICMP_EQ)
9399 return ReplaceInstUsesWith(SI, FalseVal);
9400 // Transform (X != Y) ? Y : X -> Y
9401 if (Pred == ICmpInst::ICMP_NE)
9402 return ReplaceInstUsesWith(SI, TrueVal);
9403 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9406 /// NOTE: if we wanted to, this is where to detect integer ABS
9408 return Changed ? &SI : 0;
9411 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9412 Value *CondVal = SI.getCondition();
9413 Value *TrueVal = SI.getTrueValue();
9414 Value *FalseVal = SI.getFalseValue();
9416 // select true, X, Y -> X
9417 // select false, X, Y -> Y
9418 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9419 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9421 // select C, X, X -> X
9422 if (TrueVal == FalseVal)
9423 return ReplaceInstUsesWith(SI, TrueVal);
9425 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9426 return ReplaceInstUsesWith(SI, FalseVal);
9427 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9428 return ReplaceInstUsesWith(SI, TrueVal);
9429 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9430 if (isa<Constant>(TrueVal))
9431 return ReplaceInstUsesWith(SI, TrueVal);
9433 return ReplaceInstUsesWith(SI, FalseVal);
9436 if (SI.getType() == Type::Int1Ty) {
9437 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9438 if (C->getZExtValue()) {
9439 // Change: A = select B, true, C --> A = or B, C
9440 return BinaryOperator::CreateOr(CondVal, FalseVal);
9442 // Change: A = select B, false, C --> A = and !B, C
9444 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9445 "not."+CondVal->getName()), SI);
9446 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9448 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9449 if (C->getZExtValue() == false) {
9450 // Change: A = select B, C, false --> A = and B, C
9451 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9453 // Change: A = select B, C, true --> A = or !B, C
9455 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9456 "not."+CondVal->getName()), SI);
9457 return BinaryOperator::CreateOr(NotCond, TrueVal);
9461 // select a, b, a -> a&b
9462 // select a, a, b -> a|b
9463 if (CondVal == TrueVal)
9464 return BinaryOperator::CreateOr(CondVal, FalseVal);
9465 else if (CondVal == FalseVal)
9466 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9469 // Selecting between two integer constants?
9470 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9471 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9472 // select C, 1, 0 -> zext C to int
9473 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9474 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9475 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9476 // select C, 0, 1 -> zext !C to int
9478 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9479 "not."+CondVal->getName()), SI);
9480 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9483 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9484 // If one of the constants is zero (we know they can't both be) and we
9485 // have an icmp instruction with zero, and we have an 'and' with the
9486 // non-constant value, eliminate this whole mess. This corresponds to
9487 // cases like this: ((X & 27) ? 27 : 0)
9488 if (TrueValC->isZero() || FalseValC->isZero())
9489 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9490 cast<Constant>(IC->getOperand(1))->isNullValue())
9491 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9492 if (ICA->getOpcode() == Instruction::And &&
9493 isa<ConstantInt>(ICA->getOperand(1)) &&
9494 (ICA->getOperand(1) == TrueValC ||
9495 ICA->getOperand(1) == FalseValC) &&
9496 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9497 // Okay, now we know that everything is set up, we just don't
9498 // know whether we have a icmp_ne or icmp_eq and whether the
9499 // true or false val is the zero.
9500 bool ShouldNotVal = !TrueValC->isZero();
9501 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9504 V = InsertNewInstBefore(BinaryOperator::Create(
9505 Instruction::Xor, V, ICA->getOperand(1)), SI);
9506 return ReplaceInstUsesWith(SI, V);
9511 // See if we are selecting two values based on a comparison of the two values.
9512 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9513 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9514 // Transform (X == Y) ? X : Y -> Y
9515 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9516 // This is not safe in general for floating point:
9517 // consider X== -0, Y== +0.
9518 // It becomes safe if either operand is a nonzero constant.
9519 ConstantFP *CFPt, *CFPf;
9520 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9521 !CFPt->getValueAPF().isZero()) ||
9522 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9523 !CFPf->getValueAPF().isZero()))
9524 return ReplaceInstUsesWith(SI, FalseVal);
9526 // Transform (X != Y) ? X : Y -> X
9527 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9528 return ReplaceInstUsesWith(SI, TrueVal);
9529 // NOTE: if we wanted to, this is where to detect MIN/MAX
9531 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9532 // Transform (X == Y) ? Y : X -> X
9533 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9534 // This is not safe in general for floating point:
9535 // consider X== -0, Y== +0.
9536 // It becomes safe if either operand is a nonzero constant.
9537 ConstantFP *CFPt, *CFPf;
9538 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9539 !CFPt->getValueAPF().isZero()) ||
9540 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9541 !CFPf->getValueAPF().isZero()))
9542 return ReplaceInstUsesWith(SI, FalseVal);
9544 // Transform (X != Y) ? Y : X -> Y
9545 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9546 return ReplaceInstUsesWith(SI, TrueVal);
9547 // NOTE: if we wanted to, this is where to detect MIN/MAX
9549 // NOTE: if we wanted to, this is where to detect ABS
9552 // See if we are selecting two values based on a comparison of the two values.
9553 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9554 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9557 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9558 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9559 if (TI->hasOneUse() && FI->hasOneUse()) {
9560 Instruction *AddOp = 0, *SubOp = 0;
9562 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9563 if (TI->getOpcode() == FI->getOpcode())
9564 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9567 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9568 // even legal for FP.
9569 if ((TI->getOpcode() == Instruction::Sub &&
9570 FI->getOpcode() == Instruction::Add) ||
9571 (TI->getOpcode() == Instruction::FSub &&
9572 FI->getOpcode() == Instruction::FAdd)) {
9573 AddOp = FI; SubOp = TI;
9574 } else if ((FI->getOpcode() == Instruction::Sub &&
9575 TI->getOpcode() == Instruction::Add) ||
9576 (FI->getOpcode() == Instruction::FSub &&
9577 TI->getOpcode() == Instruction::FAdd)) {
9578 AddOp = TI; SubOp = FI;
9582 Value *OtherAddOp = 0;
9583 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9584 OtherAddOp = AddOp->getOperand(1);
9585 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9586 OtherAddOp = AddOp->getOperand(0);
9590 // So at this point we know we have (Y -> OtherAddOp):
9591 // select C, (add X, Y), (sub X, Z)
9592 Value *NegVal; // Compute -Z
9593 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9594 NegVal = Context->getConstantExprNeg(C);
9596 NegVal = InsertNewInstBefore(
9597 BinaryOperator::CreateNeg(*Context, SubOp->getOperand(1),
9601 Value *NewTrueOp = OtherAddOp;
9602 Value *NewFalseOp = NegVal;
9604 std::swap(NewTrueOp, NewFalseOp);
9605 Instruction *NewSel =
9606 SelectInst::Create(CondVal, NewTrueOp,
9607 NewFalseOp, SI.getName() + ".p");
9609 NewSel = InsertNewInstBefore(NewSel, SI);
9610 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9615 // See if we can fold the select into one of our operands.
9616 if (SI.getType()->isInteger()) {
9617 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9622 if (BinaryOperator::isNot(CondVal)) {
9623 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9624 SI.setOperand(1, FalseVal);
9625 SI.setOperand(2, TrueVal);
9632 /// EnforceKnownAlignment - If the specified pointer points to an object that
9633 /// we control, modify the object's alignment to PrefAlign. This isn't
9634 /// often possible though. If alignment is important, a more reliable approach
9635 /// is to simply align all global variables and allocation instructions to
9636 /// their preferred alignment from the beginning.
9638 static unsigned EnforceKnownAlignment(Value *V,
9639 unsigned Align, unsigned PrefAlign) {
9641 User *U = dyn_cast<User>(V);
9642 if (!U) return Align;
9644 switch (Operator::getOpcode(U)) {
9646 case Instruction::BitCast:
9647 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9648 case Instruction::GetElementPtr: {
9649 // If all indexes are zero, it is just the alignment of the base pointer.
9650 bool AllZeroOperands = true;
9651 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9652 if (!isa<Constant>(*i) ||
9653 !cast<Constant>(*i)->isNullValue()) {
9654 AllZeroOperands = false;
9658 if (AllZeroOperands) {
9659 // Treat this like a bitcast.
9660 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9666 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9667 // If there is a large requested alignment and we can, bump up the alignment
9669 if (!GV->isDeclaration()) {
9670 if (GV->getAlignment() >= PrefAlign)
9671 Align = GV->getAlignment();
9673 GV->setAlignment(PrefAlign);
9677 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9678 // If there is a requested alignment and if this is an alloca, round up. We
9679 // don't do this for malloc, because some systems can't respect the request.
9680 if (isa<AllocaInst>(AI)) {
9681 if (AI->getAlignment() >= PrefAlign)
9682 Align = AI->getAlignment();
9684 AI->setAlignment(PrefAlign);
9693 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9694 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9695 /// and it is more than the alignment of the ultimate object, see if we can
9696 /// increase the alignment of the ultimate object, making this check succeed.
9697 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9698 unsigned PrefAlign) {
9699 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9700 sizeof(PrefAlign) * CHAR_BIT;
9701 APInt Mask = APInt::getAllOnesValue(BitWidth);
9702 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9703 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9704 unsigned TrailZ = KnownZero.countTrailingOnes();
9705 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9707 if (PrefAlign > Align)
9708 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9710 // We don't need to make any adjustment.
9714 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9715 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9716 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9717 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9718 unsigned CopyAlign = MI->getAlignment();
9720 if (CopyAlign < MinAlign) {
9721 MI->setAlignment(Context->getConstantInt(MI->getAlignmentType(),
9726 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9728 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9729 if (MemOpLength == 0) return 0;
9731 // Source and destination pointer types are always "i8*" for intrinsic. See
9732 // if the size is something we can handle with a single primitive load/store.
9733 // A single load+store correctly handles overlapping memory in the memmove
9735 unsigned Size = MemOpLength->getZExtValue();
9736 if (Size == 0) return MI; // Delete this mem transfer.
9738 if (Size > 8 || (Size&(Size-1)))
9739 return 0; // If not 1/2/4/8 bytes, exit.
9741 // Use an integer load+store unless we can find something better.
9743 Context->getPointerTypeUnqual(Context->getIntegerType(Size<<3));
9745 // Memcpy forces the use of i8* for the source and destination. That means
9746 // that if you're using memcpy to move one double around, you'll get a cast
9747 // from double* to i8*. We'd much rather use a double load+store rather than
9748 // an i64 load+store, here because this improves the odds that the source or
9749 // dest address will be promotable. See if we can find a better type than the
9750 // integer datatype.
9751 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9752 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9753 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9754 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9755 // down through these levels if so.
9756 while (!SrcETy->isSingleValueType()) {
9757 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9758 if (STy->getNumElements() == 1)
9759 SrcETy = STy->getElementType(0);
9762 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9763 if (ATy->getNumElements() == 1)
9764 SrcETy = ATy->getElementType();
9771 if (SrcETy->isSingleValueType())
9772 NewPtrTy = Context->getPointerTypeUnqual(SrcETy);
9777 // If the memcpy/memmove provides better alignment info than we can
9779 SrcAlign = std::max(SrcAlign, CopyAlign);
9780 DstAlign = std::max(DstAlign, CopyAlign);
9782 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9783 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9784 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9785 InsertNewInstBefore(L, *MI);
9786 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9788 // Set the size of the copy to 0, it will be deleted on the next iteration.
9789 MI->setOperand(3, Context->getNullValue(MemOpLength->getType()));
9793 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9794 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9795 if (MI->getAlignment() < Alignment) {
9796 MI->setAlignment(Context->getConstantInt(MI->getAlignmentType(),
9801 // Extract the length and alignment and fill if they are constant.
9802 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9803 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9804 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9806 uint64_t Len = LenC->getZExtValue();
9807 Alignment = MI->getAlignment();
9809 // If the length is zero, this is a no-op
9810 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9812 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9813 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9814 const Type *ITy = Context->getIntegerType(Len*8); // n=1 -> i8.
9816 Value *Dest = MI->getDest();
9817 Dest = InsertBitCastBefore(Dest, Context->getPointerTypeUnqual(ITy), *MI);
9819 // Alignment 0 is identity for alignment 1 for memset, but not store.
9820 if (Alignment == 0) Alignment = 1;
9822 // Extract the fill value and store.
9823 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9824 InsertNewInstBefore(new StoreInst(Context->getConstantInt(ITy, Fill),
9825 Dest, false, Alignment), *MI);
9827 // Set the size of the copy to 0, it will be deleted on the next iteration.
9828 MI->setLength(Context->getNullValue(LenC->getType()));
9836 /// visitCallInst - CallInst simplification. This mostly only handles folding
9837 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9838 /// the heavy lifting.
9840 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9841 // If the caller function is nounwind, mark the call as nounwind, even if the
9843 if (CI.getParent()->getParent()->doesNotThrow() &&
9844 !CI.doesNotThrow()) {
9845 CI.setDoesNotThrow();
9851 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9852 if (!II) return visitCallSite(&CI);
9854 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9856 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9857 bool Changed = false;
9859 // memmove/cpy/set of zero bytes is a noop.
9860 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9861 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9863 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9864 if (CI->getZExtValue() == 1) {
9865 // Replace the instruction with just byte operations. We would
9866 // transform other cases to loads/stores, but we don't know if
9867 // alignment is sufficient.
9871 // If we have a memmove and the source operation is a constant global,
9872 // then the source and dest pointers can't alias, so we can change this
9873 // into a call to memcpy.
9874 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9875 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9876 if (GVSrc->isConstant()) {
9877 Module *M = CI.getParent()->getParent()->getParent();
9878 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9880 Tys[0] = CI.getOperand(3)->getType();
9882 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9886 // memmove(x,x,size) -> noop.
9887 if (MMI->getSource() == MMI->getDest())
9888 return EraseInstFromFunction(CI);
9891 // If we can determine a pointer alignment that is bigger than currently
9892 // set, update the alignment.
9893 if (isa<MemTransferInst>(MI)) {
9894 if (Instruction *I = SimplifyMemTransfer(MI))
9896 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9897 if (Instruction *I = SimplifyMemSet(MSI))
9901 if (Changed) return II;
9904 switch (II->getIntrinsicID()) {
9906 case Intrinsic::bswap:
9907 // bswap(bswap(x)) -> x
9908 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9909 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9910 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9912 case Intrinsic::ppc_altivec_lvx:
9913 case Intrinsic::ppc_altivec_lvxl:
9914 case Intrinsic::x86_sse_loadu_ps:
9915 case Intrinsic::x86_sse2_loadu_pd:
9916 case Intrinsic::x86_sse2_loadu_dq:
9917 // Turn PPC lvx -> load if the pointer is known aligned.
9918 // Turn X86 loadups -> load if the pointer is known aligned.
9919 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9920 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9921 Context->getPointerTypeUnqual(II->getType()),
9923 return new LoadInst(Ptr);
9926 case Intrinsic::ppc_altivec_stvx:
9927 case Intrinsic::ppc_altivec_stvxl:
9928 // Turn stvx -> store if the pointer is known aligned.
9929 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9930 const Type *OpPtrTy =
9931 Context->getPointerTypeUnqual(II->getOperand(1)->getType());
9932 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9933 return new StoreInst(II->getOperand(1), Ptr);
9936 case Intrinsic::x86_sse_storeu_ps:
9937 case Intrinsic::x86_sse2_storeu_pd:
9938 case Intrinsic::x86_sse2_storeu_dq:
9939 // Turn X86 storeu -> store if the pointer is known aligned.
9940 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9941 const Type *OpPtrTy =
9942 Context->getPointerTypeUnqual(II->getOperand(2)->getType());
9943 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9944 return new StoreInst(II->getOperand(2), Ptr);
9948 case Intrinsic::x86_sse_cvttss2si: {
9949 // These intrinsics only demands the 0th element of its input vector. If
9950 // we can simplify the input based on that, do so now.
9952 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9953 APInt DemandedElts(VWidth, 1);
9954 APInt UndefElts(VWidth, 0);
9955 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9957 II->setOperand(1, V);
9963 case Intrinsic::ppc_altivec_vperm:
9964 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9965 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9966 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9968 // Check that all of the elements are integer constants or undefs.
9969 bool AllEltsOk = true;
9970 for (unsigned i = 0; i != 16; ++i) {
9971 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9972 !isa<UndefValue>(Mask->getOperand(i))) {
9979 // Cast the input vectors to byte vectors.
9980 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9981 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9982 Value *Result = Context->getUndef(Op0->getType());
9984 // Only extract each element once.
9985 Value *ExtractedElts[32];
9986 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9988 for (unsigned i = 0; i != 16; ++i) {
9989 if (isa<UndefValue>(Mask->getOperand(i)))
9991 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9992 Idx &= 31; // Match the hardware behavior.
9994 if (ExtractedElts[Idx] == 0) {
9996 new ExtractElementInst(Idx < 16 ? Op0 : Op1,
9997 Context->getConstantInt(Type::Int32Ty, Idx&15, false), "tmp");
9998 InsertNewInstBefore(Elt, CI);
9999 ExtractedElts[Idx] = Elt;
10002 // Insert this value into the result vector.
10003 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
10004 Context->getConstantInt(Type::Int32Ty, i, false),
10006 InsertNewInstBefore(cast<Instruction>(Result), CI);
10008 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
10013 case Intrinsic::stackrestore: {
10014 // If the save is right next to the restore, remove the restore. This can
10015 // happen when variable allocas are DCE'd.
10016 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
10017 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
10018 BasicBlock::iterator BI = SS;
10020 return EraseInstFromFunction(CI);
10024 // Scan down this block to see if there is another stack restore in the
10025 // same block without an intervening call/alloca.
10026 BasicBlock::iterator BI = II;
10027 TerminatorInst *TI = II->getParent()->getTerminator();
10028 bool CannotRemove = false;
10029 for (++BI; &*BI != TI; ++BI) {
10030 if (isa<AllocaInst>(BI)) {
10031 CannotRemove = true;
10034 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
10035 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
10036 // If there is a stackrestore below this one, remove this one.
10037 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10038 return EraseInstFromFunction(CI);
10039 // Otherwise, ignore the intrinsic.
10041 // If we found a non-intrinsic call, we can't remove the stack
10043 CannotRemove = true;
10049 // If the stack restore is in a return/unwind block and if there are no
10050 // allocas or calls between the restore and the return, nuke the restore.
10051 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10052 return EraseInstFromFunction(CI);
10057 return visitCallSite(II);
10060 // InvokeInst simplification
10062 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10063 return visitCallSite(&II);
10066 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10067 /// passed through the varargs area, we can eliminate the use of the cast.
10068 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10069 const CastInst * const CI,
10070 const TargetData * const TD,
10072 if (!CI->isLosslessCast())
10075 // The size of ByVal arguments is derived from the type, so we
10076 // can't change to a type with a different size. If the size were
10077 // passed explicitly we could avoid this check.
10078 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10081 const Type* SrcTy =
10082 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10083 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10084 if (!SrcTy->isSized() || !DstTy->isSized())
10086 if (TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10091 // visitCallSite - Improvements for call and invoke instructions.
10093 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10094 bool Changed = false;
10096 // If the callee is a constexpr cast of a function, attempt to move the cast
10097 // to the arguments of the call/invoke.
10098 if (transformConstExprCastCall(CS)) return 0;
10100 Value *Callee = CS.getCalledValue();
10102 if (Function *CalleeF = dyn_cast<Function>(Callee))
10103 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10104 Instruction *OldCall = CS.getInstruction();
10105 // If the call and callee calling conventions don't match, this call must
10106 // be unreachable, as the call is undefined.
10107 new StoreInst(Context->getConstantIntTrue(),
10108 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)),
10110 if (!OldCall->use_empty())
10111 OldCall->replaceAllUsesWith(Context->getUndef(OldCall->getType()));
10112 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10113 return EraseInstFromFunction(*OldCall);
10117 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10118 // This instruction is not reachable, just remove it. We insert a store to
10119 // undef so that we know that this code is not reachable, despite the fact
10120 // that we can't modify the CFG here.
10121 new StoreInst(Context->getConstantIntTrue(),
10122 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)),
10123 CS.getInstruction());
10125 if (!CS.getInstruction()->use_empty())
10126 CS.getInstruction()->
10127 replaceAllUsesWith(Context->getUndef(CS.getInstruction()->getType()));
10129 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10130 // Don't break the CFG, insert a dummy cond branch.
10131 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10132 Context->getConstantIntTrue(), II);
10134 return EraseInstFromFunction(*CS.getInstruction());
10137 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10138 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10139 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10140 return transformCallThroughTrampoline(CS);
10142 const PointerType *PTy = cast<PointerType>(Callee->getType());
10143 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10144 if (FTy->isVarArg()) {
10145 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10146 // See if we can optimize any arguments passed through the varargs area of
10148 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10149 E = CS.arg_end(); I != E; ++I, ++ix) {
10150 CastInst *CI = dyn_cast<CastInst>(*I);
10151 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10152 *I = CI->getOperand(0);
10158 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10159 // Inline asm calls cannot throw - mark them 'nounwind'.
10160 CS.setDoesNotThrow();
10164 return Changed ? CS.getInstruction() : 0;
10167 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10168 // attempt to move the cast to the arguments of the call/invoke.
10170 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10171 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10172 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10173 if (CE->getOpcode() != Instruction::BitCast ||
10174 !isa<Function>(CE->getOperand(0)))
10176 Function *Callee = cast<Function>(CE->getOperand(0));
10177 Instruction *Caller = CS.getInstruction();
10178 const AttrListPtr &CallerPAL = CS.getAttributes();
10180 // Okay, this is a cast from a function to a different type. Unless doing so
10181 // would cause a type conversion of one of our arguments, change this call to
10182 // be a direct call with arguments casted to the appropriate types.
10184 const FunctionType *FT = Callee->getFunctionType();
10185 const Type *OldRetTy = Caller->getType();
10186 const Type *NewRetTy = FT->getReturnType();
10188 if (isa<StructType>(NewRetTy))
10189 return false; // TODO: Handle multiple return values.
10191 // Check to see if we are changing the return type...
10192 if (OldRetTy != NewRetTy) {
10193 if (Callee->isDeclaration() &&
10194 // Conversion is ok if changing from one pointer type to another or from
10195 // a pointer to an integer of the same size.
10196 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
10197 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
10198 return false; // Cannot transform this return value.
10200 if (!Caller->use_empty() &&
10201 // void -> non-void is handled specially
10202 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
10203 return false; // Cannot transform this return value.
10205 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10206 Attributes RAttrs = CallerPAL.getRetAttributes();
10207 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10208 return false; // Attribute not compatible with transformed value.
10211 // If the callsite is an invoke instruction, and the return value is used by
10212 // a PHI node in a successor, we cannot change the return type of the call
10213 // because there is no place to put the cast instruction (without breaking
10214 // the critical edge). Bail out in this case.
10215 if (!Caller->use_empty())
10216 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10217 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10219 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10220 if (PN->getParent() == II->getNormalDest() ||
10221 PN->getParent() == II->getUnwindDest())
10225 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10226 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10228 CallSite::arg_iterator AI = CS.arg_begin();
10229 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10230 const Type *ParamTy = FT->getParamType(i);
10231 const Type *ActTy = (*AI)->getType();
10233 if (!CastInst::isCastable(ActTy, ParamTy))
10234 return false; // Cannot transform this parameter value.
10236 if (CallerPAL.getParamAttributes(i + 1)
10237 & Attribute::typeIncompatible(ParamTy))
10238 return false; // Attribute not compatible with transformed value.
10240 // Converting from one pointer type to another or between a pointer and an
10241 // integer of the same size is safe even if we do not have a body.
10242 bool isConvertible = ActTy == ParamTy ||
10243 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
10244 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
10245 if (Callee->isDeclaration() && !isConvertible) return false;
10248 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10249 Callee->isDeclaration())
10250 return false; // Do not delete arguments unless we have a function body.
10252 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10253 !CallerPAL.isEmpty())
10254 // In this case we have more arguments than the new function type, but we
10255 // won't be dropping them. Check that these extra arguments have attributes
10256 // that are compatible with being a vararg call argument.
10257 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10258 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10260 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10261 if (PAttrs & Attribute::VarArgsIncompatible)
10265 // Okay, we decided that this is a safe thing to do: go ahead and start
10266 // inserting cast instructions as necessary...
10267 std::vector<Value*> Args;
10268 Args.reserve(NumActualArgs);
10269 SmallVector<AttributeWithIndex, 8> attrVec;
10270 attrVec.reserve(NumCommonArgs);
10272 // Get any return attributes.
10273 Attributes RAttrs = CallerPAL.getRetAttributes();
10275 // If the return value is not being used, the type may not be compatible
10276 // with the existing attributes. Wipe out any problematic attributes.
10277 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10279 // Add the new return attributes.
10281 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10283 AI = CS.arg_begin();
10284 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10285 const Type *ParamTy = FT->getParamType(i);
10286 if ((*AI)->getType() == ParamTy) {
10287 Args.push_back(*AI);
10289 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10290 false, ParamTy, false);
10291 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
10292 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
10295 // Add any parameter attributes.
10296 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10297 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10300 // If the function takes more arguments than the call was taking, add them
10302 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10303 Args.push_back(Context->getNullValue(FT->getParamType(i)));
10305 // If we are removing arguments to the function, emit an obnoxious warning...
10306 if (FT->getNumParams() < NumActualArgs) {
10307 if (!FT->isVarArg()) {
10308 cerr << "WARNING: While resolving call to function '"
10309 << Callee->getName() << "' arguments were dropped!\n";
10311 // Add all of the arguments in their promoted form to the arg list...
10312 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10313 const Type *PTy = getPromotedType((*AI)->getType());
10314 if (PTy != (*AI)->getType()) {
10315 // Must promote to pass through va_arg area!
10316 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
10318 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
10319 InsertNewInstBefore(Cast, *Caller);
10320 Args.push_back(Cast);
10322 Args.push_back(*AI);
10325 // Add any parameter attributes.
10326 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10327 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10332 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10333 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10335 if (NewRetTy == Type::VoidTy)
10336 Caller->setName(""); // Void type should not have a name.
10338 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
10341 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10342 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10343 Args.begin(), Args.end(),
10344 Caller->getName(), Caller);
10345 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10346 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10348 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10349 Caller->getName(), Caller);
10350 CallInst *CI = cast<CallInst>(Caller);
10351 if (CI->isTailCall())
10352 cast<CallInst>(NC)->setTailCall();
10353 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10354 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10357 // Insert a cast of the return type as necessary.
10359 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10360 if (NV->getType() != Type::VoidTy) {
10361 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10363 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10365 // If this is an invoke instruction, we should insert it after the first
10366 // non-phi, instruction in the normal successor block.
10367 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10368 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10369 InsertNewInstBefore(NC, *I);
10371 // Otherwise, it's a call, just insert cast right after the call instr
10372 InsertNewInstBefore(NC, *Caller);
10374 AddUsersToWorkList(*Caller);
10376 NV = Context->getUndef(Caller->getType());
10380 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10381 Caller->replaceAllUsesWith(NV);
10382 Caller->eraseFromParent();
10383 RemoveFromWorkList(Caller);
10387 // transformCallThroughTrampoline - Turn a call to a function created by the
10388 // init_trampoline intrinsic into a direct call to the underlying function.
10390 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10391 Value *Callee = CS.getCalledValue();
10392 const PointerType *PTy = cast<PointerType>(Callee->getType());
10393 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10394 const AttrListPtr &Attrs = CS.getAttributes();
10396 // If the call already has the 'nest' attribute somewhere then give up -
10397 // otherwise 'nest' would occur twice after splicing in the chain.
10398 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10401 IntrinsicInst *Tramp =
10402 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10404 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10405 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10406 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10408 const AttrListPtr &NestAttrs = NestF->getAttributes();
10409 if (!NestAttrs.isEmpty()) {
10410 unsigned NestIdx = 1;
10411 const Type *NestTy = 0;
10412 Attributes NestAttr = Attribute::None;
10414 // Look for a parameter marked with the 'nest' attribute.
10415 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10416 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10417 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10418 // Record the parameter type and any other attributes.
10420 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10425 Instruction *Caller = CS.getInstruction();
10426 std::vector<Value*> NewArgs;
10427 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10429 SmallVector<AttributeWithIndex, 8> NewAttrs;
10430 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10432 // Insert the nest argument into the call argument list, which may
10433 // mean appending it. Likewise for attributes.
10435 // Add any result attributes.
10436 if (Attributes Attr = Attrs.getRetAttributes())
10437 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10441 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10443 if (Idx == NestIdx) {
10444 // Add the chain argument and attributes.
10445 Value *NestVal = Tramp->getOperand(3);
10446 if (NestVal->getType() != NestTy)
10447 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10448 NewArgs.push_back(NestVal);
10449 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10455 // Add the original argument and attributes.
10456 NewArgs.push_back(*I);
10457 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10459 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10465 // Add any function attributes.
10466 if (Attributes Attr = Attrs.getFnAttributes())
10467 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10469 // The trampoline may have been bitcast to a bogus type (FTy).
10470 // Handle this by synthesizing a new function type, equal to FTy
10471 // with the chain parameter inserted.
10473 std::vector<const Type*> NewTypes;
10474 NewTypes.reserve(FTy->getNumParams()+1);
10476 // Insert the chain's type into the list of parameter types, which may
10477 // mean appending it.
10480 FunctionType::param_iterator I = FTy->param_begin(),
10481 E = FTy->param_end();
10484 if (Idx == NestIdx)
10485 // Add the chain's type.
10486 NewTypes.push_back(NestTy);
10491 // Add the original type.
10492 NewTypes.push_back(*I);
10498 // Replace the trampoline call with a direct call. Let the generic
10499 // code sort out any function type mismatches.
10500 FunctionType *NewFTy =
10501 Context->getFunctionType(FTy->getReturnType(), NewTypes,
10503 Constant *NewCallee =
10504 NestF->getType() == Context->getPointerTypeUnqual(NewFTy) ?
10505 NestF : Context->getConstantExprBitCast(NestF,
10506 Context->getPointerTypeUnqual(NewFTy));
10507 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
10509 Instruction *NewCaller;
10510 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10511 NewCaller = InvokeInst::Create(NewCallee,
10512 II->getNormalDest(), II->getUnwindDest(),
10513 NewArgs.begin(), NewArgs.end(),
10514 Caller->getName(), Caller);
10515 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10516 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10518 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10519 Caller->getName(), Caller);
10520 if (cast<CallInst>(Caller)->isTailCall())
10521 cast<CallInst>(NewCaller)->setTailCall();
10522 cast<CallInst>(NewCaller)->
10523 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10524 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10526 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10527 Caller->replaceAllUsesWith(NewCaller);
10528 Caller->eraseFromParent();
10529 RemoveFromWorkList(Caller);
10534 // Replace the trampoline call with a direct call. Since there is no 'nest'
10535 // parameter, there is no need to adjust the argument list. Let the generic
10536 // code sort out any function type mismatches.
10537 Constant *NewCallee =
10538 NestF->getType() == PTy ? NestF :
10539 Context->getConstantExprBitCast(NestF, PTy);
10540 CS.setCalledFunction(NewCallee);
10541 return CS.getInstruction();
10544 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10545 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10546 /// and a single binop.
10547 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10548 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10549 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10550 unsigned Opc = FirstInst->getOpcode();
10551 Value *LHSVal = FirstInst->getOperand(0);
10552 Value *RHSVal = FirstInst->getOperand(1);
10554 const Type *LHSType = LHSVal->getType();
10555 const Type *RHSType = RHSVal->getType();
10557 // Scan to see if all operands are the same opcode, all have one use, and all
10558 // kill their operands (i.e. the operands have one use).
10559 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10560 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10561 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10562 // Verify type of the LHS matches so we don't fold cmp's of different
10563 // types or GEP's with different index types.
10564 I->getOperand(0)->getType() != LHSType ||
10565 I->getOperand(1)->getType() != RHSType)
10568 // If they are CmpInst instructions, check their predicates
10569 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10570 if (cast<CmpInst>(I)->getPredicate() !=
10571 cast<CmpInst>(FirstInst)->getPredicate())
10574 // Keep track of which operand needs a phi node.
10575 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10576 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10579 // Otherwise, this is safe to transform!
10581 Value *InLHS = FirstInst->getOperand(0);
10582 Value *InRHS = FirstInst->getOperand(1);
10583 PHINode *NewLHS = 0, *NewRHS = 0;
10585 NewLHS = PHINode::Create(LHSType,
10586 FirstInst->getOperand(0)->getName() + ".pn");
10587 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10588 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10589 InsertNewInstBefore(NewLHS, PN);
10594 NewRHS = PHINode::Create(RHSType,
10595 FirstInst->getOperand(1)->getName() + ".pn");
10596 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10597 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10598 InsertNewInstBefore(NewRHS, PN);
10602 // Add all operands to the new PHIs.
10603 if (NewLHS || NewRHS) {
10604 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10605 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10607 Value *NewInLHS = InInst->getOperand(0);
10608 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10611 Value *NewInRHS = InInst->getOperand(1);
10612 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10617 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10618 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10619 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10620 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10624 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10625 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10627 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10628 FirstInst->op_end());
10629 // This is true if all GEP bases are allocas and if all indices into them are
10631 bool AllBasePointersAreAllocas = true;
10633 // Scan to see if all operands are the same opcode, all have one use, and all
10634 // kill their operands (i.e. the operands have one use).
10635 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10636 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10637 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10638 GEP->getNumOperands() != FirstInst->getNumOperands())
10641 // Keep track of whether or not all GEPs are of alloca pointers.
10642 if (AllBasePointersAreAllocas &&
10643 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10644 !GEP->hasAllConstantIndices()))
10645 AllBasePointersAreAllocas = false;
10647 // Compare the operand lists.
10648 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10649 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10652 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10653 // if one of the PHIs has a constant for the index. The index may be
10654 // substantially cheaper to compute for the constants, so making it a
10655 // variable index could pessimize the path. This also handles the case
10656 // for struct indices, which must always be constant.
10657 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10658 isa<ConstantInt>(GEP->getOperand(op)))
10661 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10663 FixedOperands[op] = 0; // Needs a PHI.
10667 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10668 // bother doing this transformation. At best, this will just save a bit of
10669 // offset calculation, but all the predecessors will have to materialize the
10670 // stack address into a register anyway. We'd actually rather *clone* the
10671 // load up into the predecessors so that we have a load of a gep of an alloca,
10672 // which can usually all be folded into the load.
10673 if (AllBasePointersAreAllocas)
10676 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10677 // that is variable.
10678 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10680 bool HasAnyPHIs = false;
10681 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10682 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10683 Value *FirstOp = FirstInst->getOperand(i);
10684 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10685 FirstOp->getName()+".pn");
10686 InsertNewInstBefore(NewPN, PN);
10688 NewPN->reserveOperandSpace(e);
10689 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10690 OperandPhis[i] = NewPN;
10691 FixedOperands[i] = NewPN;
10696 // Add all operands to the new PHIs.
10698 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10699 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10700 BasicBlock *InBB = PN.getIncomingBlock(i);
10702 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10703 if (PHINode *OpPhi = OperandPhis[op])
10704 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10708 Value *Base = FixedOperands[0];
10709 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10710 FixedOperands.end());
10714 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10715 /// sink the load out of the block that defines it. This means that it must be
10716 /// obvious the value of the load is not changed from the point of the load to
10717 /// the end of the block it is in.
10719 /// Finally, it is safe, but not profitable, to sink a load targetting a
10720 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10722 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10723 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10725 for (++BBI; BBI != E; ++BBI)
10726 if (BBI->mayWriteToMemory())
10729 // Check for non-address taken alloca. If not address-taken already, it isn't
10730 // profitable to do this xform.
10731 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10732 bool isAddressTaken = false;
10733 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10735 if (isa<LoadInst>(UI)) continue;
10736 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10737 // If storing TO the alloca, then the address isn't taken.
10738 if (SI->getOperand(1) == AI) continue;
10740 isAddressTaken = true;
10744 if (!isAddressTaken && AI->isStaticAlloca())
10748 // If this load is a load from a GEP with a constant offset from an alloca,
10749 // then we don't want to sink it. In its present form, it will be
10750 // load [constant stack offset]. Sinking it will cause us to have to
10751 // materialize the stack addresses in each predecessor in a register only to
10752 // do a shared load from register in the successor.
10753 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10754 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10755 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10762 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10763 // operator and they all are only used by the PHI, PHI together their
10764 // inputs, and do the operation once, to the result of the PHI.
10765 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10766 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10768 // Scan the instruction, looking for input operations that can be folded away.
10769 // If all input operands to the phi are the same instruction (e.g. a cast from
10770 // the same type or "+42") we can pull the operation through the PHI, reducing
10771 // code size and simplifying code.
10772 Constant *ConstantOp = 0;
10773 const Type *CastSrcTy = 0;
10774 bool isVolatile = false;
10775 if (isa<CastInst>(FirstInst)) {
10776 CastSrcTy = FirstInst->getOperand(0)->getType();
10777 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10778 // Can fold binop, compare or shift here if the RHS is a constant,
10779 // otherwise call FoldPHIArgBinOpIntoPHI.
10780 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10781 if (ConstantOp == 0)
10782 return FoldPHIArgBinOpIntoPHI(PN);
10783 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10784 isVolatile = LI->isVolatile();
10785 // We can't sink the load if the loaded value could be modified between the
10786 // load and the PHI.
10787 if (LI->getParent() != PN.getIncomingBlock(0) ||
10788 !isSafeAndProfitableToSinkLoad(LI))
10791 // If the PHI is of volatile loads and the load block has multiple
10792 // successors, sinking it would remove a load of the volatile value from
10793 // the path through the other successor.
10795 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10798 } else if (isa<GetElementPtrInst>(FirstInst)) {
10799 return FoldPHIArgGEPIntoPHI(PN);
10801 return 0; // Cannot fold this operation.
10804 // Check to see if all arguments are the same operation.
10805 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10806 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10807 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10808 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10811 if (I->getOperand(0)->getType() != CastSrcTy)
10812 return 0; // Cast operation must match.
10813 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10814 // We can't sink the load if the loaded value could be modified between
10815 // the load and the PHI.
10816 if (LI->isVolatile() != isVolatile ||
10817 LI->getParent() != PN.getIncomingBlock(i) ||
10818 !isSafeAndProfitableToSinkLoad(LI))
10821 // If the PHI is of volatile loads and the load block has multiple
10822 // successors, sinking it would remove a load of the volatile value from
10823 // the path through the other successor.
10825 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10828 } else if (I->getOperand(1) != ConstantOp) {
10833 // Okay, they are all the same operation. Create a new PHI node of the
10834 // correct type, and PHI together all of the LHS's of the instructions.
10835 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10836 PN.getName()+".in");
10837 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10839 Value *InVal = FirstInst->getOperand(0);
10840 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10842 // Add all operands to the new PHI.
10843 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10844 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10845 if (NewInVal != InVal)
10847 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10852 // The new PHI unions all of the same values together. This is really
10853 // common, so we handle it intelligently here for compile-time speed.
10857 InsertNewInstBefore(NewPN, PN);
10861 // Insert and return the new operation.
10862 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10863 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10864 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10865 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10866 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10867 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10868 PhiVal, ConstantOp);
10869 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10871 // If this was a volatile load that we are merging, make sure to loop through
10872 // and mark all the input loads as non-volatile. If we don't do this, we will
10873 // insert a new volatile load and the old ones will not be deletable.
10875 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10876 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10878 return new LoadInst(PhiVal, "", isVolatile);
10881 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10883 static bool DeadPHICycle(PHINode *PN,
10884 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10885 if (PN->use_empty()) return true;
10886 if (!PN->hasOneUse()) return false;
10888 // Remember this node, and if we find the cycle, return.
10889 if (!PotentiallyDeadPHIs.insert(PN))
10892 // Don't scan crazily complex things.
10893 if (PotentiallyDeadPHIs.size() == 16)
10896 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10897 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10902 /// PHIsEqualValue - Return true if this phi node is always equal to
10903 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10904 /// z = some value; x = phi (y, z); y = phi (x, z)
10905 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10906 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10907 // See if we already saw this PHI node.
10908 if (!ValueEqualPHIs.insert(PN))
10911 // Don't scan crazily complex things.
10912 if (ValueEqualPHIs.size() == 16)
10915 // Scan the operands to see if they are either phi nodes or are equal to
10917 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10918 Value *Op = PN->getIncomingValue(i);
10919 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10920 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10922 } else if (Op != NonPhiInVal)
10930 // PHINode simplification
10932 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10933 // If LCSSA is around, don't mess with Phi nodes
10934 if (MustPreserveLCSSA) return 0;
10936 if (Value *V = PN.hasConstantValue())
10937 return ReplaceInstUsesWith(PN, V);
10939 // If all PHI operands are the same operation, pull them through the PHI,
10940 // reducing code size.
10941 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10942 isa<Instruction>(PN.getIncomingValue(1)) &&
10943 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10944 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10945 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10946 // than themselves more than once.
10947 PN.getIncomingValue(0)->hasOneUse())
10948 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10951 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10952 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10953 // PHI)... break the cycle.
10954 if (PN.hasOneUse()) {
10955 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10956 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10957 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10958 PotentiallyDeadPHIs.insert(&PN);
10959 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10960 return ReplaceInstUsesWith(PN, Context->getUndef(PN.getType()));
10963 // If this phi has a single use, and if that use just computes a value for
10964 // the next iteration of a loop, delete the phi. This occurs with unused
10965 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10966 // common case here is good because the only other things that catch this
10967 // are induction variable analysis (sometimes) and ADCE, which is only run
10969 if (PHIUser->hasOneUse() &&
10970 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10971 PHIUser->use_back() == &PN) {
10972 return ReplaceInstUsesWith(PN, Context->getUndef(PN.getType()));
10976 // We sometimes end up with phi cycles that non-obviously end up being the
10977 // same value, for example:
10978 // z = some value; x = phi (y, z); y = phi (x, z)
10979 // where the phi nodes don't necessarily need to be in the same block. Do a
10980 // quick check to see if the PHI node only contains a single non-phi value, if
10981 // so, scan to see if the phi cycle is actually equal to that value.
10983 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10984 // Scan for the first non-phi operand.
10985 while (InValNo != NumOperandVals &&
10986 isa<PHINode>(PN.getIncomingValue(InValNo)))
10989 if (InValNo != NumOperandVals) {
10990 Value *NonPhiInVal = PN.getOperand(InValNo);
10992 // Scan the rest of the operands to see if there are any conflicts, if so
10993 // there is no need to recursively scan other phis.
10994 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10995 Value *OpVal = PN.getIncomingValue(InValNo);
10996 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
11000 // If we scanned over all operands, then we have one unique value plus
11001 // phi values. Scan PHI nodes to see if they all merge in each other or
11003 if (InValNo == NumOperandVals) {
11004 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
11005 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
11006 return ReplaceInstUsesWith(PN, NonPhiInVal);
11013 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
11014 Instruction *InsertPoint,
11015 InstCombiner *IC) {
11016 unsigned PtrSize = DTy->getScalarSizeInBits();
11017 unsigned VTySize = V->getType()->getScalarSizeInBits();
11018 // We must cast correctly to the pointer type. Ensure that we
11019 // sign extend the integer value if it is smaller as this is
11020 // used for address computation.
11021 Instruction::CastOps opcode =
11022 (VTySize < PtrSize ? Instruction::SExt :
11023 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
11024 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
11028 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
11029 Value *PtrOp = GEP.getOperand(0);
11030 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
11031 // If so, eliminate the noop.
11032 if (GEP.getNumOperands() == 1)
11033 return ReplaceInstUsesWith(GEP, PtrOp);
11035 if (isa<UndefValue>(GEP.getOperand(0)))
11036 return ReplaceInstUsesWith(GEP, Context->getUndef(GEP.getType()));
11038 bool HasZeroPointerIndex = false;
11039 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
11040 HasZeroPointerIndex = C->isNullValue();
11042 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11043 return ReplaceInstUsesWith(GEP, PtrOp);
11045 // Eliminate unneeded casts for indices.
11046 bool MadeChange = false;
11048 gep_type_iterator GTI = gep_type_begin(GEP);
11049 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
11050 i != e; ++i, ++GTI) {
11051 if (isa<SequentialType>(*GTI)) {
11052 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
11053 if (CI->getOpcode() == Instruction::ZExt ||
11054 CI->getOpcode() == Instruction::SExt) {
11055 const Type *SrcTy = CI->getOperand(0)->getType();
11056 // We can eliminate a cast from i32 to i64 iff the target
11057 // is a 32-bit pointer target.
11058 if (SrcTy->getScalarSizeInBits() >= TD->getPointerSizeInBits()) {
11060 *i = CI->getOperand(0);
11064 // If we are using a wider index than needed for this platform, shrink it
11065 // to what we need. If narrower, sign-extend it to what we need.
11066 // If the incoming value needs a cast instruction,
11067 // insert it. This explicit cast can make subsequent optimizations more
11070 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
11071 if (Constant *C = dyn_cast<Constant>(Op)) {
11072 *i = Context->getConstantExprTrunc(C, TD->getIntPtrType());
11075 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
11080 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
11081 if (Constant *C = dyn_cast<Constant>(Op)) {
11082 *i = Context->getConstantExprSExt(C, TD->getIntPtrType());
11085 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
11093 if (MadeChange) return &GEP;
11095 // Combine Indices - If the source pointer to this getelementptr instruction
11096 // is a getelementptr instruction, combine the indices of the two
11097 // getelementptr instructions into a single instruction.
11099 SmallVector<Value*, 8> SrcGEPOperands;
11100 if (User *Src = dyn_castGetElementPtr(PtrOp))
11101 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
11103 if (!SrcGEPOperands.empty()) {
11104 // Note that if our source is a gep chain itself that we wait for that
11105 // chain to be resolved before we perform this transformation. This
11106 // avoids us creating a TON of code in some cases.
11108 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
11109 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
11110 return 0; // Wait until our source is folded to completion.
11112 SmallVector<Value*, 8> Indices;
11114 // Find out whether the last index in the source GEP is a sequential idx.
11115 bool EndsWithSequential = false;
11116 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
11117 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
11118 EndsWithSequential = !isa<StructType>(*I);
11120 // Can we combine the two pointer arithmetics offsets?
11121 if (EndsWithSequential) {
11122 // Replace: gep (gep %P, long B), long A, ...
11123 // With: T = long A+B; gep %P, T, ...
11125 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
11126 if (SO1 == Context->getNullValue(SO1->getType())) {
11128 } else if (GO1 == Context->getNullValue(GO1->getType())) {
11131 // If they aren't the same type, convert both to an integer of the
11132 // target's pointer size.
11133 if (SO1->getType() != GO1->getType()) {
11134 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
11136 Context->getConstantExprIntegerCast(SO1C, GO1->getType(), true);
11137 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
11139 Context->getConstantExprIntegerCast(GO1C, SO1->getType(), true);
11141 unsigned PS = TD->getPointerSizeInBits();
11142 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
11143 // Convert GO1 to SO1's type.
11144 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
11146 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
11147 // Convert SO1 to GO1's type.
11148 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
11150 const Type *PT = TD->getIntPtrType();
11151 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
11152 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
11156 if (isa<Constant>(SO1) && isa<Constant>(GO1))
11157 Sum = Context->getConstantExprAdd(cast<Constant>(SO1),
11158 cast<Constant>(GO1));
11160 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11161 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
11165 // Recycle the GEP we already have if possible.
11166 if (SrcGEPOperands.size() == 2) {
11167 GEP.setOperand(0, SrcGEPOperands[0]);
11168 GEP.setOperand(1, Sum);
11171 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11172 SrcGEPOperands.end()-1);
11173 Indices.push_back(Sum);
11174 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
11176 } else if (isa<Constant>(*GEP.idx_begin()) &&
11177 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11178 SrcGEPOperands.size() != 1) {
11179 // Otherwise we can do the fold if the first index of the GEP is a zero
11180 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11181 SrcGEPOperands.end());
11182 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
11185 if (!Indices.empty())
11186 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
11187 Indices.end(), GEP.getName());
11189 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
11190 // GEP of global variable. If all of the indices for this GEP are
11191 // constants, we can promote this to a constexpr instead of an instruction.
11193 // Scan for nonconstants...
11194 SmallVector<Constant*, 8> Indices;
11195 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
11196 for (; I != E && isa<Constant>(*I); ++I)
11197 Indices.push_back(cast<Constant>(*I));
11199 if (I == E) { // If they are all constants...
11200 Constant *CE = Context->getConstantExprGetElementPtr(GV,
11201 &Indices[0],Indices.size());
11203 // Replace all uses of the GEP with the new constexpr...
11204 return ReplaceInstUsesWith(GEP, CE);
11206 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
11207 if (!isa<PointerType>(X->getType())) {
11208 // Not interesting. Source pointer must be a cast from pointer.
11209 } else if (HasZeroPointerIndex) {
11210 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11211 // into : GEP [10 x i8]* X, i32 0, ...
11213 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11214 // into : GEP i8* X, ...
11216 // This occurs when the program declares an array extern like "int X[];"
11217 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11218 const PointerType *XTy = cast<PointerType>(X->getType());
11219 if (const ArrayType *CATy =
11220 dyn_cast<ArrayType>(CPTy->getElementType())) {
11221 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11222 if (CATy->getElementType() == XTy->getElementType()) {
11223 // -> GEP i8* X, ...
11224 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11225 return GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11227 } else if (const ArrayType *XATy =
11228 dyn_cast<ArrayType>(XTy->getElementType())) {
11229 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11230 if (CATy->getElementType() == XATy->getElementType()) {
11231 // -> GEP [10 x i8]* X, i32 0, ...
11232 // At this point, we know that the cast source type is a pointer
11233 // to an array of the same type as the destination pointer
11234 // array. Because the array type is never stepped over (there
11235 // is a leading zero) we can fold the cast into this GEP.
11236 GEP.setOperand(0, X);
11241 } else if (GEP.getNumOperands() == 2) {
11242 // Transform things like:
11243 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11244 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11245 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11246 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11247 if (isa<ArrayType>(SrcElTy) &&
11248 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11249 TD->getTypeAllocSize(ResElTy)) {
11251 Idx[0] = Context->getNullValue(Type::Int32Ty);
11252 Idx[1] = GEP.getOperand(1);
11253 Value *V = InsertNewInstBefore(
11254 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
11255 // V and GEP are both pointer types --> BitCast
11256 return new BitCastInst(V, GEP.getType());
11259 // Transform things like:
11260 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11261 // (where tmp = 8*tmp2) into:
11262 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11264 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
11265 uint64_t ArrayEltSize =
11266 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11268 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11269 // allow either a mul, shift, or constant here.
11271 ConstantInt *Scale = 0;
11272 if (ArrayEltSize == 1) {
11273 NewIdx = GEP.getOperand(1);
11275 Context->getConstantInt(cast<IntegerType>(NewIdx->getType()), 1);
11276 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11277 NewIdx = Context->getConstantInt(CI->getType(), 1);
11279 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11280 if (Inst->getOpcode() == Instruction::Shl &&
11281 isa<ConstantInt>(Inst->getOperand(1))) {
11282 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11283 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11284 Scale = Context->getConstantInt(cast<IntegerType>(Inst->getType()),
11286 NewIdx = Inst->getOperand(0);
11287 } else if (Inst->getOpcode() == Instruction::Mul &&
11288 isa<ConstantInt>(Inst->getOperand(1))) {
11289 Scale = cast<ConstantInt>(Inst->getOperand(1));
11290 NewIdx = Inst->getOperand(0);
11294 // If the index will be to exactly the right offset with the scale taken
11295 // out, perform the transformation. Note, we don't know whether Scale is
11296 // signed or not. We'll use unsigned version of division/modulo
11297 // operation after making sure Scale doesn't have the sign bit set.
11298 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11299 Scale->getZExtValue() % ArrayEltSize == 0) {
11300 Scale = Context->getConstantInt(Scale->getType(),
11301 Scale->getZExtValue() / ArrayEltSize);
11302 if (Scale->getZExtValue() != 1) {
11304 Context->getConstantExprIntegerCast(Scale, NewIdx->getType(),
11306 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
11307 NewIdx = InsertNewInstBefore(Sc, GEP);
11310 // Insert the new GEP instruction.
11312 Idx[0] = Context->getNullValue(Type::Int32Ty);
11314 Instruction *NewGEP =
11315 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11316 NewGEP = InsertNewInstBefore(NewGEP, GEP);
11317 // The NewGEP must be pointer typed, so must the old one -> BitCast
11318 return new BitCastInst(NewGEP, GEP.getType());
11324 /// See if we can simplify:
11325 /// X = bitcast A to B*
11326 /// Y = gep X, <...constant indices...>
11327 /// into a gep of the original struct. This is important for SROA and alias
11328 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11329 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11330 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11331 // Determine how much the GEP moves the pointer. We are guaranteed to get
11332 // a constant back from EmitGEPOffset.
11333 ConstantInt *OffsetV =
11334 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11335 int64_t Offset = OffsetV->getSExtValue();
11337 // If this GEP instruction doesn't move the pointer, just replace the GEP
11338 // with a bitcast of the real input to the dest type.
11340 // If the bitcast is of an allocation, and the allocation will be
11341 // converted to match the type of the cast, don't touch this.
11342 if (isa<AllocationInst>(BCI->getOperand(0))) {
11343 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11344 if (Instruction *I = visitBitCast(*BCI)) {
11347 BCI->getParent()->getInstList().insert(BCI, I);
11348 ReplaceInstUsesWith(*BCI, I);
11353 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11356 // Otherwise, if the offset is non-zero, we need to find out if there is a
11357 // field at Offset in 'A's type. If so, we can pull the cast through the
11359 SmallVector<Value*, 8> NewIndices;
11361 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11362 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11363 Instruction *NGEP =
11364 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11366 if (NGEP->getType() == GEP.getType()) return NGEP;
11367 InsertNewInstBefore(NGEP, GEP);
11368 NGEP->takeName(&GEP);
11369 return new BitCastInst(NGEP, GEP.getType());
11377 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11378 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11379 if (AI.isArrayAllocation()) { // Check C != 1
11380 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11381 const Type *NewTy =
11382 Context->getArrayType(AI.getAllocatedType(), C->getZExtValue());
11383 AllocationInst *New = 0;
11385 // Create and insert the replacement instruction...
11386 if (isa<MallocInst>(AI))
11387 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11389 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11390 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11393 InsertNewInstBefore(New, AI);
11395 // Scan to the end of the allocation instructions, to skip over a block of
11396 // allocas if possible...also skip interleaved debug info
11398 BasicBlock::iterator It = New;
11399 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11401 // Now that I is pointing to the first non-allocation-inst in the block,
11402 // insert our getelementptr instruction...
11404 Value *NullIdx = Context->getNullValue(Type::Int32Ty);
11408 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11409 New->getName()+".sub", It);
11411 // Now make everything use the getelementptr instead of the original
11413 return ReplaceInstUsesWith(AI, V);
11414 } else if (isa<UndefValue>(AI.getArraySize())) {
11415 return ReplaceInstUsesWith(AI, Context->getNullValue(AI.getType()));
11419 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11420 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11421 // Note that we only do this for alloca's, because malloc should allocate
11422 // and return a unique pointer, even for a zero byte allocation.
11423 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11424 return ReplaceInstUsesWith(AI, Context->getNullValue(AI.getType()));
11426 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11427 if (AI.getAlignment() == 0)
11428 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11434 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11435 Value *Op = FI.getOperand(0);
11437 // free undef -> unreachable.
11438 if (isa<UndefValue>(Op)) {
11439 // Insert a new store to null because we cannot modify the CFG here.
11440 new StoreInst(Context->getConstantIntTrue(),
11441 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)), &FI);
11442 return EraseInstFromFunction(FI);
11445 // If we have 'free null' delete the instruction. This can happen in stl code
11446 // when lots of inlining happens.
11447 if (isa<ConstantPointerNull>(Op))
11448 return EraseInstFromFunction(FI);
11450 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11451 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11452 FI.setOperand(0, CI->getOperand(0));
11456 // Change free (gep X, 0,0,0,0) into free(X)
11457 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11458 if (GEPI->hasAllZeroIndices()) {
11459 AddToWorkList(GEPI);
11460 FI.setOperand(0, GEPI->getOperand(0));
11465 // Change free(malloc) into nothing, if the malloc has a single use.
11466 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11467 if (MI->hasOneUse()) {
11468 EraseInstFromFunction(FI);
11469 return EraseInstFromFunction(*MI);
11476 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11477 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11478 const TargetData *TD) {
11479 User *CI = cast<User>(LI.getOperand(0));
11480 Value *CastOp = CI->getOperand(0);
11481 LLVMContext *Context = IC.getContext();
11484 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11485 // Instead of loading constant c string, use corresponding integer value
11486 // directly if string length is small enough.
11488 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11489 unsigned len = Str.length();
11490 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11491 unsigned numBits = Ty->getPrimitiveSizeInBits();
11492 // Replace LI with immediate integer store.
11493 if ((numBits >> 3) == len + 1) {
11494 APInt StrVal(numBits, 0);
11495 APInt SingleChar(numBits, 0);
11496 if (TD->isLittleEndian()) {
11497 for (signed i = len-1; i >= 0; i--) {
11498 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11499 StrVal = (StrVal << 8) | SingleChar;
11502 for (unsigned i = 0; i < len; i++) {
11503 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11504 StrVal = (StrVal << 8) | SingleChar;
11506 // Append NULL at the end.
11508 StrVal = (StrVal << 8) | SingleChar;
11510 Value *NL = Context->getConstantInt(StrVal);
11511 return IC.ReplaceInstUsesWith(LI, NL);
11517 const PointerType *DestTy = cast<PointerType>(CI->getType());
11518 const Type *DestPTy = DestTy->getElementType();
11519 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11521 // If the address spaces don't match, don't eliminate the cast.
11522 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11525 const Type *SrcPTy = SrcTy->getElementType();
11527 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11528 isa<VectorType>(DestPTy)) {
11529 // If the source is an array, the code below will not succeed. Check to
11530 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11532 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11533 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11534 if (ASrcTy->getNumElements() != 0) {
11536 Idxs[0] = Idxs[1] = Context->getNullValue(Type::Int32Ty);
11537 CastOp = Context->getConstantExprGetElementPtr(CSrc, Idxs, 2);
11538 SrcTy = cast<PointerType>(CastOp->getType());
11539 SrcPTy = SrcTy->getElementType();
11542 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11543 isa<VectorType>(SrcPTy)) &&
11544 // Do not allow turning this into a load of an integer, which is then
11545 // casted to a pointer, this pessimizes pointer analysis a lot.
11546 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11547 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11548 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11550 // Okay, we are casting from one integer or pointer type to another of
11551 // the same size. Instead of casting the pointer before the load, cast
11552 // the result of the loaded value.
11553 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11555 LI.isVolatile()),LI);
11556 // Now cast the result of the load.
11557 return new BitCastInst(NewLoad, LI.getType());
11564 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11565 Value *Op = LI.getOperand(0);
11567 // Attempt to improve the alignment.
11568 unsigned KnownAlign =
11569 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11571 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11572 LI.getAlignment()))
11573 LI.setAlignment(KnownAlign);
11575 // load (cast X) --> cast (load X) iff safe
11576 if (isa<CastInst>(Op))
11577 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11580 // None of the following transforms are legal for volatile loads.
11581 if (LI.isVolatile()) return 0;
11583 // Do really simple store-to-load forwarding and load CSE, to catch cases
11584 // where there are several consequtive memory accesses to the same location,
11585 // separated by a few arithmetic operations.
11586 BasicBlock::iterator BBI = &LI;
11587 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11588 return ReplaceInstUsesWith(LI, AvailableVal);
11590 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11591 const Value *GEPI0 = GEPI->getOperand(0);
11592 // TODO: Consider a target hook for valid address spaces for this xform.
11593 if (isa<ConstantPointerNull>(GEPI0) &&
11594 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11595 // Insert a new store to null instruction before the load to indicate
11596 // that this code is not reachable. We do this instead of inserting
11597 // an unreachable instruction directly because we cannot modify the
11599 new StoreInst(Context->getUndef(LI.getType()),
11600 Context->getNullValue(Op->getType()), &LI);
11601 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11605 if (Constant *C = dyn_cast<Constant>(Op)) {
11606 // load null/undef -> undef
11607 // TODO: Consider a target hook for valid address spaces for this xform.
11608 if (isa<UndefValue>(C) || (C->isNullValue() &&
11609 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11610 // Insert a new store to null instruction before the load to indicate that
11611 // this code is not reachable. We do this instead of inserting an
11612 // unreachable instruction directly because we cannot modify the CFG.
11613 new StoreInst(Context->getUndef(LI.getType()),
11614 Context->getNullValue(Op->getType()), &LI);
11615 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11618 // Instcombine load (constant global) into the value loaded.
11619 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11620 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11621 return ReplaceInstUsesWith(LI, GV->getInitializer());
11623 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11624 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11625 if (CE->getOpcode() == Instruction::GetElementPtr) {
11626 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11627 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11629 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11631 return ReplaceInstUsesWith(LI, V);
11632 if (CE->getOperand(0)->isNullValue()) {
11633 // Insert a new store to null instruction before the load to indicate
11634 // that this code is not reachable. We do this instead of inserting
11635 // an unreachable instruction directly because we cannot modify the
11637 new StoreInst(Context->getUndef(LI.getType()),
11638 Context->getNullValue(Op->getType()), &LI);
11639 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11642 } else if (CE->isCast()) {
11643 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11649 // If this load comes from anywhere in a constant global, and if the global
11650 // is all undef or zero, we know what it loads.
11651 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11652 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11653 if (GV->getInitializer()->isNullValue())
11654 return ReplaceInstUsesWith(LI, Context->getNullValue(LI.getType()));
11655 else if (isa<UndefValue>(GV->getInitializer()))
11656 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11660 if (Op->hasOneUse()) {
11661 // Change select and PHI nodes to select values instead of addresses: this
11662 // helps alias analysis out a lot, allows many others simplifications, and
11663 // exposes redundancy in the code.
11665 // Note that we cannot do the transformation unless we know that the
11666 // introduced loads cannot trap! Something like this is valid as long as
11667 // the condition is always false: load (select bool %C, int* null, int* %G),
11668 // but it would not be valid if we transformed it to load from null
11669 // unconditionally.
11671 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11672 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11673 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11674 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11675 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11676 SI->getOperand(1)->getName()+".val"), LI);
11677 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11678 SI->getOperand(2)->getName()+".val"), LI);
11679 return SelectInst::Create(SI->getCondition(), V1, V2);
11682 // load (select (cond, null, P)) -> load P
11683 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11684 if (C->isNullValue()) {
11685 LI.setOperand(0, SI->getOperand(2));
11689 // load (select (cond, P, null)) -> load P
11690 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11691 if (C->isNullValue()) {
11692 LI.setOperand(0, SI->getOperand(1));
11700 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11701 /// when possible. This makes it generally easy to do alias analysis and/or
11702 /// SROA/mem2reg of the memory object.
11703 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11704 User *CI = cast<User>(SI.getOperand(1));
11705 Value *CastOp = CI->getOperand(0);
11706 LLVMContext *Context = IC.getContext();
11708 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11709 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11710 if (SrcTy == 0) return 0;
11712 const Type *SrcPTy = SrcTy->getElementType();
11714 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11717 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11718 /// to its first element. This allows us to handle things like:
11719 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11720 /// on 32-bit hosts.
11721 SmallVector<Value*, 4> NewGEPIndices;
11723 // If the source is an array, the code below will not succeed. Check to
11724 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11726 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11727 // Index through pointer.
11728 Constant *Zero = Context->getNullValue(Type::Int32Ty);
11729 NewGEPIndices.push_back(Zero);
11732 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11733 if (!STy->getNumElements()) /* Struct can be empty {} */
11735 NewGEPIndices.push_back(Zero);
11736 SrcPTy = STy->getElementType(0);
11737 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11738 NewGEPIndices.push_back(Zero);
11739 SrcPTy = ATy->getElementType();
11745 SrcTy = Context->getPointerType(SrcPTy, SrcTy->getAddressSpace());
11748 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11751 // If the pointers point into different address spaces or if they point to
11752 // values with different sizes, we can't do the transformation.
11753 if (SrcTy->getAddressSpace() !=
11754 cast<PointerType>(CI->getType())->getAddressSpace() ||
11755 IC.getTargetData().getTypeSizeInBits(SrcPTy) !=
11756 IC.getTargetData().getTypeSizeInBits(DestPTy))
11759 // Okay, we are casting from one integer or pointer type to another of
11760 // the same size. Instead of casting the pointer before
11761 // the store, cast the value to be stored.
11763 Value *SIOp0 = SI.getOperand(0);
11764 Instruction::CastOps opcode = Instruction::BitCast;
11765 const Type* CastSrcTy = SIOp0->getType();
11766 const Type* CastDstTy = SrcPTy;
11767 if (isa<PointerType>(CastDstTy)) {
11768 if (CastSrcTy->isInteger())
11769 opcode = Instruction::IntToPtr;
11770 } else if (isa<IntegerType>(CastDstTy)) {
11771 if (isa<PointerType>(SIOp0->getType()))
11772 opcode = Instruction::PtrToInt;
11775 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11776 // emit a GEP to index into its first field.
11777 if (!NewGEPIndices.empty()) {
11778 if (Constant *C = dyn_cast<Constant>(CastOp))
11779 CastOp = Context->getConstantExprGetElementPtr(C, &NewGEPIndices[0],
11780 NewGEPIndices.size());
11782 CastOp = IC.InsertNewInstBefore(
11783 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11784 NewGEPIndices.end()), SI);
11787 if (Constant *C = dyn_cast<Constant>(SIOp0))
11788 NewCast = Context->getConstantExprCast(opcode, C, CastDstTy);
11790 NewCast = IC.InsertNewInstBefore(
11791 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11793 return new StoreInst(NewCast, CastOp);
11796 /// equivalentAddressValues - Test if A and B will obviously have the same
11797 /// value. This includes recognizing that %t0 and %t1 will have the same
11798 /// value in code like this:
11799 /// %t0 = getelementptr \@a, 0, 3
11800 /// store i32 0, i32* %t0
11801 /// %t1 = getelementptr \@a, 0, 3
11802 /// %t2 = load i32* %t1
11804 static bool equivalentAddressValues(Value *A, Value *B) {
11805 // Test if the values are trivially equivalent.
11806 if (A == B) return true;
11808 // Test if the values come form identical arithmetic instructions.
11809 if (isa<BinaryOperator>(A) ||
11810 isa<CastInst>(A) ||
11812 isa<GetElementPtrInst>(A))
11813 if (Instruction *BI = dyn_cast<Instruction>(B))
11814 if (cast<Instruction>(A)->isIdenticalTo(BI))
11817 // Otherwise they may not be equivalent.
11821 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11822 // return the llvm.dbg.declare.
11823 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11824 if (!V->hasNUses(2))
11826 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11828 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11830 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11831 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11838 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11839 Value *Val = SI.getOperand(0);
11840 Value *Ptr = SI.getOperand(1);
11842 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11843 EraseInstFromFunction(SI);
11848 // If the RHS is an alloca with a single use, zapify the store, making the
11850 // If the RHS is an alloca with a two uses, the other one being a
11851 // llvm.dbg.declare, zapify the store and the declare, making the
11852 // alloca dead. We must do this to prevent declare's from affecting
11854 if (!SI.isVolatile()) {
11855 if (Ptr->hasOneUse()) {
11856 if (isa<AllocaInst>(Ptr)) {
11857 EraseInstFromFunction(SI);
11861 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11862 if (isa<AllocaInst>(GEP->getOperand(0))) {
11863 if (GEP->getOperand(0)->hasOneUse()) {
11864 EraseInstFromFunction(SI);
11868 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11869 EraseInstFromFunction(*DI);
11870 EraseInstFromFunction(SI);
11877 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11878 EraseInstFromFunction(*DI);
11879 EraseInstFromFunction(SI);
11885 // Attempt to improve the alignment.
11886 unsigned KnownAlign =
11887 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11889 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11890 SI.getAlignment()))
11891 SI.setAlignment(KnownAlign);
11893 // Do really simple DSE, to catch cases where there are several consecutive
11894 // stores to the same location, separated by a few arithmetic operations. This
11895 // situation often occurs with bitfield accesses.
11896 BasicBlock::iterator BBI = &SI;
11897 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11900 // Don't count debug info directives, lest they affect codegen,
11901 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11902 // It is necessary for correctness to skip those that feed into a
11903 // llvm.dbg.declare, as these are not present when debugging is off.
11904 if (isa<DbgInfoIntrinsic>(BBI) ||
11905 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11910 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11911 // Prev store isn't volatile, and stores to the same location?
11912 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11913 SI.getOperand(1))) {
11916 EraseInstFromFunction(*PrevSI);
11922 // If this is a load, we have to stop. However, if the loaded value is from
11923 // the pointer we're loading and is producing the pointer we're storing,
11924 // then *this* store is dead (X = load P; store X -> P).
11925 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11926 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11927 !SI.isVolatile()) {
11928 EraseInstFromFunction(SI);
11932 // Otherwise, this is a load from some other location. Stores before it
11933 // may not be dead.
11937 // Don't skip over loads or things that can modify memory.
11938 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11943 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11945 // store X, null -> turns into 'unreachable' in SimplifyCFG
11946 if (isa<ConstantPointerNull>(Ptr) &&
11947 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11948 if (!isa<UndefValue>(Val)) {
11949 SI.setOperand(0, Context->getUndef(Val->getType()));
11950 if (Instruction *U = dyn_cast<Instruction>(Val))
11951 AddToWorkList(U); // Dropped a use.
11954 return 0; // Do not modify these!
11957 // store undef, Ptr -> noop
11958 if (isa<UndefValue>(Val)) {
11959 EraseInstFromFunction(SI);
11964 // If the pointer destination is a cast, see if we can fold the cast into the
11966 if (isa<CastInst>(Ptr))
11967 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11969 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11971 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11975 // If this store is the last instruction in the basic block (possibly
11976 // excepting debug info instructions and the pointer bitcasts that feed
11977 // into them), and if the block ends with an unconditional branch, try
11978 // to move it to the successor block.
11982 } while (isa<DbgInfoIntrinsic>(BBI) ||
11983 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11984 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11985 if (BI->isUnconditional())
11986 if (SimplifyStoreAtEndOfBlock(SI))
11987 return 0; // xform done!
11992 /// SimplifyStoreAtEndOfBlock - Turn things like:
11993 /// if () { *P = v1; } else { *P = v2 }
11994 /// into a phi node with a store in the successor.
11996 /// Simplify things like:
11997 /// *P = v1; if () { *P = v2; }
11998 /// into a phi node with a store in the successor.
12000 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
12001 BasicBlock *StoreBB = SI.getParent();
12003 // Check to see if the successor block has exactly two incoming edges. If
12004 // so, see if the other predecessor contains a store to the same location.
12005 // if so, insert a PHI node (if needed) and move the stores down.
12006 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
12008 // Determine whether Dest has exactly two predecessors and, if so, compute
12009 // the other predecessor.
12010 pred_iterator PI = pred_begin(DestBB);
12011 BasicBlock *OtherBB = 0;
12012 if (*PI != StoreBB)
12015 if (PI == pred_end(DestBB))
12018 if (*PI != StoreBB) {
12023 if (++PI != pred_end(DestBB))
12026 // Bail out if all the relevant blocks aren't distinct (this can happen,
12027 // for example, if SI is in an infinite loop)
12028 if (StoreBB == DestBB || OtherBB == DestBB)
12031 // Verify that the other block ends in a branch and is not otherwise empty.
12032 BasicBlock::iterator BBI = OtherBB->getTerminator();
12033 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
12034 if (!OtherBr || BBI == OtherBB->begin())
12037 // If the other block ends in an unconditional branch, check for the 'if then
12038 // else' case. there is an instruction before the branch.
12039 StoreInst *OtherStore = 0;
12040 if (OtherBr->isUnconditional()) {
12042 // Skip over debugging info.
12043 while (isa<DbgInfoIntrinsic>(BBI) ||
12044 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12045 if (BBI==OtherBB->begin())
12049 // If this isn't a store, or isn't a store to the same location, bail out.
12050 OtherStore = dyn_cast<StoreInst>(BBI);
12051 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
12054 // Otherwise, the other block ended with a conditional branch. If one of the
12055 // destinations is StoreBB, then we have the if/then case.
12056 if (OtherBr->getSuccessor(0) != StoreBB &&
12057 OtherBr->getSuccessor(1) != StoreBB)
12060 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12061 // if/then triangle. See if there is a store to the same ptr as SI that
12062 // lives in OtherBB.
12064 // Check to see if we find the matching store.
12065 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12066 if (OtherStore->getOperand(1) != SI.getOperand(1))
12070 // If we find something that may be using or overwriting the stored
12071 // value, or if we run out of instructions, we can't do the xform.
12072 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12073 BBI == OtherBB->begin())
12077 // In order to eliminate the store in OtherBr, we have to
12078 // make sure nothing reads or overwrites the stored value in
12080 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12081 // FIXME: This should really be AA driven.
12082 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12087 // Insert a PHI node now if we need it.
12088 Value *MergedVal = OtherStore->getOperand(0);
12089 if (MergedVal != SI.getOperand(0)) {
12090 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12091 PN->reserveOperandSpace(2);
12092 PN->addIncoming(SI.getOperand(0), SI.getParent());
12093 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12094 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12097 // Advance to a place where it is safe to insert the new store and
12099 BBI = DestBB->getFirstNonPHI();
12100 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12101 OtherStore->isVolatile()), *BBI);
12103 // Nuke the old stores.
12104 EraseInstFromFunction(SI);
12105 EraseInstFromFunction(*OtherStore);
12111 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12112 // Change br (not X), label True, label False to: br X, label False, True
12114 BasicBlock *TrueDest;
12115 BasicBlock *FalseDest;
12116 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest), *Context) &&
12117 !isa<Constant>(X)) {
12118 // Swap Destinations and condition...
12119 BI.setCondition(X);
12120 BI.setSuccessor(0, FalseDest);
12121 BI.setSuccessor(1, TrueDest);
12125 // Cannonicalize fcmp_one -> fcmp_oeq
12126 FCmpInst::Predicate FPred; Value *Y;
12127 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12128 TrueDest, FalseDest), *Context))
12129 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12130 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
12131 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
12132 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
12133 Instruction *NewSCC = new FCmpInst(I, NewPred, X, Y, "");
12134 NewSCC->takeName(I);
12135 // Swap Destinations and condition...
12136 BI.setCondition(NewSCC);
12137 BI.setSuccessor(0, FalseDest);
12138 BI.setSuccessor(1, TrueDest);
12139 RemoveFromWorkList(I);
12140 I->eraseFromParent();
12141 AddToWorkList(NewSCC);
12145 // Cannonicalize icmp_ne -> icmp_eq
12146 ICmpInst::Predicate IPred;
12147 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12148 TrueDest, FalseDest), *Context))
12149 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12150 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12151 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
12152 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
12153 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
12154 Instruction *NewSCC = new ICmpInst(I, NewPred, X, Y, "");
12155 NewSCC->takeName(I);
12156 // Swap Destinations and condition...
12157 BI.setCondition(NewSCC);
12158 BI.setSuccessor(0, FalseDest);
12159 BI.setSuccessor(1, TrueDest);
12160 RemoveFromWorkList(I);
12161 I->eraseFromParent();;
12162 AddToWorkList(NewSCC);
12169 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12170 Value *Cond = SI.getCondition();
12171 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12172 if (I->getOpcode() == Instruction::Add)
12173 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12174 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12175 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12177 Context->getConstantExprSub(cast<Constant>(SI.getOperand(i)),
12179 SI.setOperand(0, I->getOperand(0));
12187 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12188 Value *Agg = EV.getAggregateOperand();
12190 if (!EV.hasIndices())
12191 return ReplaceInstUsesWith(EV, Agg);
12193 if (Constant *C = dyn_cast<Constant>(Agg)) {
12194 if (isa<UndefValue>(C))
12195 return ReplaceInstUsesWith(EV, Context->getUndef(EV.getType()));
12197 if (isa<ConstantAggregateZero>(C))
12198 return ReplaceInstUsesWith(EV, Context->getNullValue(EV.getType()));
12200 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12201 // Extract the element indexed by the first index out of the constant
12202 Value *V = C->getOperand(*EV.idx_begin());
12203 if (EV.getNumIndices() > 1)
12204 // Extract the remaining indices out of the constant indexed by the
12206 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12208 return ReplaceInstUsesWith(EV, V);
12210 return 0; // Can't handle other constants
12212 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12213 // We're extracting from an insertvalue instruction, compare the indices
12214 const unsigned *exti, *exte, *insi, *inse;
12215 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12216 exte = EV.idx_end(), inse = IV->idx_end();
12217 exti != exte && insi != inse;
12219 if (*insi != *exti)
12220 // The insert and extract both reference distinctly different elements.
12221 // This means the extract is not influenced by the insert, and we can
12222 // replace the aggregate operand of the extract with the aggregate
12223 // operand of the insert. i.e., replace
12224 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12225 // %E = extractvalue { i32, { i32 } } %I, 0
12227 // %E = extractvalue { i32, { i32 } } %A, 0
12228 return ExtractValueInst::Create(IV->getAggregateOperand(),
12229 EV.idx_begin(), EV.idx_end());
12231 if (exti == exte && insi == inse)
12232 // Both iterators are at the end: Index lists are identical. Replace
12233 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12234 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12236 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12237 if (exti == exte) {
12238 // The extract list is a prefix of the insert list. i.e. replace
12239 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12240 // %E = extractvalue { i32, { i32 } } %I, 1
12242 // %X = extractvalue { i32, { i32 } } %A, 1
12243 // %E = insertvalue { i32 } %X, i32 42, 0
12244 // by switching the order of the insert and extract (though the
12245 // insertvalue should be left in, since it may have other uses).
12246 Value *NewEV = InsertNewInstBefore(
12247 ExtractValueInst::Create(IV->getAggregateOperand(),
12248 EV.idx_begin(), EV.idx_end()),
12250 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12254 // The insert list is a prefix of the extract list
12255 // We can simply remove the common indices from the extract and make it
12256 // operate on the inserted value instead of the insertvalue result.
12258 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12259 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12261 // %E extractvalue { i32 } { i32 42 }, 0
12262 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12265 // Can't simplify extracts from other values. Note that nested extracts are
12266 // already simplified implicitely by the above (extract ( extract (insert) )
12267 // will be translated into extract ( insert ( extract ) ) first and then just
12268 // the value inserted, if appropriate).
12272 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12273 /// is to leave as a vector operation.
12274 static bool CheapToScalarize(Value *V, bool isConstant) {
12275 if (isa<ConstantAggregateZero>(V))
12277 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12278 if (isConstant) return true;
12279 // If all elts are the same, we can extract.
12280 Constant *Op0 = C->getOperand(0);
12281 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12282 if (C->getOperand(i) != Op0)
12286 Instruction *I = dyn_cast<Instruction>(V);
12287 if (!I) return false;
12289 // Insert element gets simplified to the inserted element or is deleted if
12290 // this is constant idx extract element and its a constant idx insertelt.
12291 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12292 isa<ConstantInt>(I->getOperand(2)))
12294 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12296 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12297 if (BO->hasOneUse() &&
12298 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12299 CheapToScalarize(BO->getOperand(1), isConstant)))
12301 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12302 if (CI->hasOneUse() &&
12303 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12304 CheapToScalarize(CI->getOperand(1), isConstant)))
12310 /// Read and decode a shufflevector mask.
12312 /// It turns undef elements into values that are larger than the number of
12313 /// elements in the input.
12314 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12315 unsigned NElts = SVI->getType()->getNumElements();
12316 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12317 return std::vector<unsigned>(NElts, 0);
12318 if (isa<UndefValue>(SVI->getOperand(2)))
12319 return std::vector<unsigned>(NElts, 2*NElts);
12321 std::vector<unsigned> Result;
12322 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12323 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12324 if (isa<UndefValue>(*i))
12325 Result.push_back(NElts*2); // undef -> 8
12327 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12331 /// FindScalarElement - Given a vector and an element number, see if the scalar
12332 /// value is already around as a register, for example if it were inserted then
12333 /// extracted from the vector.
12334 static Value *FindScalarElement(Value *V, unsigned EltNo,
12335 LLVMContext *Context) {
12336 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12337 const VectorType *PTy = cast<VectorType>(V->getType());
12338 unsigned Width = PTy->getNumElements();
12339 if (EltNo >= Width) // Out of range access.
12340 return Context->getUndef(PTy->getElementType());
12342 if (isa<UndefValue>(V))
12343 return Context->getUndef(PTy->getElementType());
12344 else if (isa<ConstantAggregateZero>(V))
12345 return Context->getNullValue(PTy->getElementType());
12346 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12347 return CP->getOperand(EltNo);
12348 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12349 // If this is an insert to a variable element, we don't know what it is.
12350 if (!isa<ConstantInt>(III->getOperand(2)))
12352 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12354 // If this is an insert to the element we are looking for, return the
12356 if (EltNo == IIElt)
12357 return III->getOperand(1);
12359 // Otherwise, the insertelement doesn't modify the value, recurse on its
12361 return FindScalarElement(III->getOperand(0), EltNo, Context);
12362 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12363 unsigned LHSWidth =
12364 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12365 unsigned InEl = getShuffleMask(SVI)[EltNo];
12366 if (InEl < LHSWidth)
12367 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12368 else if (InEl < LHSWidth*2)
12369 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12371 return Context->getUndef(PTy->getElementType());
12374 // Otherwise, we don't know.
12378 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12379 // If vector val is undef, replace extract with scalar undef.
12380 if (isa<UndefValue>(EI.getOperand(0)))
12381 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12383 // If vector val is constant 0, replace extract with scalar 0.
12384 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12385 return ReplaceInstUsesWith(EI, Context->getNullValue(EI.getType()));
12387 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12388 // If vector val is constant with all elements the same, replace EI with
12389 // that element. When the elements are not identical, we cannot replace yet
12390 // (we do that below, but only when the index is constant).
12391 Constant *op0 = C->getOperand(0);
12392 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12393 if (C->getOperand(i) != op0) {
12398 return ReplaceInstUsesWith(EI, op0);
12401 // If extracting a specified index from the vector, see if we can recursively
12402 // find a previously computed scalar that was inserted into the vector.
12403 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12404 unsigned IndexVal = IdxC->getZExtValue();
12405 unsigned VectorWidth =
12406 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12408 // If this is extracting an invalid index, turn this into undef, to avoid
12409 // crashing the code below.
12410 if (IndexVal >= VectorWidth)
12411 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12413 // This instruction only demands the single element from the input vector.
12414 // If the input vector has a single use, simplify it based on this use
12416 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12417 APInt UndefElts(VectorWidth, 0);
12418 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12419 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12420 DemandedMask, UndefElts)) {
12421 EI.setOperand(0, V);
12426 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12427 return ReplaceInstUsesWith(EI, Elt);
12429 // If the this extractelement is directly using a bitcast from a vector of
12430 // the same number of elements, see if we can find the source element from
12431 // it. In this case, we will end up needing to bitcast the scalars.
12432 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12433 if (const VectorType *VT =
12434 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12435 if (VT->getNumElements() == VectorWidth)
12436 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12437 IndexVal, Context))
12438 return new BitCastInst(Elt, EI.getType());
12442 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12443 if (I->hasOneUse()) {
12444 // Push extractelement into predecessor operation if legal and
12445 // profitable to do so
12446 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12447 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12448 if (CheapToScalarize(BO, isConstantElt)) {
12449 ExtractElementInst *newEI0 =
12450 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
12451 EI.getName()+".lhs");
12452 ExtractElementInst *newEI1 =
12453 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
12454 EI.getName()+".rhs");
12455 InsertNewInstBefore(newEI0, EI);
12456 InsertNewInstBefore(newEI1, EI);
12457 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12459 } else if (isa<LoadInst>(I)) {
12461 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12462 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12463 Context->getPointerType(EI.getType(), AS),EI);
12464 GetElementPtrInst *GEP =
12465 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12466 InsertNewInstBefore(GEP, EI);
12467 return new LoadInst(GEP);
12470 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12471 // Extracting the inserted element?
12472 if (IE->getOperand(2) == EI.getOperand(1))
12473 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12474 // If the inserted and extracted elements are constants, they must not
12475 // be the same value, extract from the pre-inserted value instead.
12476 if (isa<Constant>(IE->getOperand(2)) &&
12477 isa<Constant>(EI.getOperand(1))) {
12478 AddUsesToWorkList(EI);
12479 EI.setOperand(0, IE->getOperand(0));
12482 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12483 // If this is extracting an element from a shufflevector, figure out where
12484 // it came from and extract from the appropriate input element instead.
12485 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12486 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12488 unsigned LHSWidth =
12489 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12491 if (SrcIdx < LHSWidth)
12492 Src = SVI->getOperand(0);
12493 else if (SrcIdx < LHSWidth*2) {
12494 SrcIdx -= LHSWidth;
12495 Src = SVI->getOperand(1);
12497 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12499 return new ExtractElementInst(Src,
12500 Context->getConstantInt(Type::Int32Ty, SrcIdx, false));
12503 // FIXME: Canonicalize extractelement(bitcast) -> bitcast(extractelement)
12508 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12509 /// elements from either LHS or RHS, return the shuffle mask and true.
12510 /// Otherwise, return false.
12511 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12512 std::vector<Constant*> &Mask,
12513 LLVMContext *Context) {
12514 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12515 "Invalid CollectSingleShuffleElements");
12516 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12518 if (isa<UndefValue>(V)) {
12519 Mask.assign(NumElts, Context->getUndef(Type::Int32Ty));
12521 } else if (V == LHS) {
12522 for (unsigned i = 0; i != NumElts; ++i)
12523 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i));
12525 } else if (V == RHS) {
12526 for (unsigned i = 0; i != NumElts; ++i)
12527 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i+NumElts));
12529 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12530 // If this is an insert of an extract from some other vector, include it.
12531 Value *VecOp = IEI->getOperand(0);
12532 Value *ScalarOp = IEI->getOperand(1);
12533 Value *IdxOp = IEI->getOperand(2);
12535 if (!isa<ConstantInt>(IdxOp))
12537 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12539 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12540 // Okay, we can handle this if the vector we are insertinting into is
12541 // transitively ok.
12542 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12543 // If so, update the mask to reflect the inserted undef.
12544 Mask[InsertedIdx] = Context->getUndef(Type::Int32Ty);
12547 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12548 if (isa<ConstantInt>(EI->getOperand(1)) &&
12549 EI->getOperand(0)->getType() == V->getType()) {
12550 unsigned ExtractedIdx =
12551 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12553 // This must be extracting from either LHS or RHS.
12554 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12555 // Okay, we can handle this if the vector we are insertinting into is
12556 // transitively ok.
12557 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12558 // If so, update the mask to reflect the inserted value.
12559 if (EI->getOperand(0) == LHS) {
12560 Mask[InsertedIdx % NumElts] =
12561 Context->getConstantInt(Type::Int32Ty, ExtractedIdx);
12563 assert(EI->getOperand(0) == RHS);
12564 Mask[InsertedIdx % NumElts] =
12565 Context->getConstantInt(Type::Int32Ty, ExtractedIdx+NumElts);
12574 // TODO: Handle shufflevector here!
12579 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12580 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12581 /// that computes V and the LHS value of the shuffle.
12582 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12583 Value *&RHS, LLVMContext *Context) {
12584 assert(isa<VectorType>(V->getType()) &&
12585 (RHS == 0 || V->getType() == RHS->getType()) &&
12586 "Invalid shuffle!");
12587 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12589 if (isa<UndefValue>(V)) {
12590 Mask.assign(NumElts, Context->getUndef(Type::Int32Ty));
12592 } else if (isa<ConstantAggregateZero>(V)) {
12593 Mask.assign(NumElts, Context->getConstantInt(Type::Int32Ty, 0));
12595 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12596 // If this is an insert of an extract from some other vector, include it.
12597 Value *VecOp = IEI->getOperand(0);
12598 Value *ScalarOp = IEI->getOperand(1);
12599 Value *IdxOp = IEI->getOperand(2);
12601 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12602 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12603 EI->getOperand(0)->getType() == V->getType()) {
12604 unsigned ExtractedIdx =
12605 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12606 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12608 // Either the extracted from or inserted into vector must be RHSVec,
12609 // otherwise we'd end up with a shuffle of three inputs.
12610 if (EI->getOperand(0) == RHS || RHS == 0) {
12611 RHS = EI->getOperand(0);
12612 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12613 Mask[InsertedIdx % NumElts] =
12614 Context->getConstantInt(Type::Int32Ty, NumElts+ExtractedIdx);
12618 if (VecOp == RHS) {
12619 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12621 // Everything but the extracted element is replaced with the RHS.
12622 for (unsigned i = 0; i != NumElts; ++i) {
12623 if (i != InsertedIdx)
12624 Mask[i] = Context->getConstantInt(Type::Int32Ty, NumElts+i);
12629 // If this insertelement is a chain that comes from exactly these two
12630 // vectors, return the vector and the effective shuffle.
12631 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12633 return EI->getOperand(0);
12638 // TODO: Handle shufflevector here!
12640 // Otherwise, can't do anything fancy. Return an identity vector.
12641 for (unsigned i = 0; i != NumElts; ++i)
12642 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i));
12646 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12647 Value *VecOp = IE.getOperand(0);
12648 Value *ScalarOp = IE.getOperand(1);
12649 Value *IdxOp = IE.getOperand(2);
12651 // Inserting an undef or into an undefined place, remove this.
12652 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12653 ReplaceInstUsesWith(IE, VecOp);
12655 // If the inserted element was extracted from some other vector, and if the
12656 // indexes are constant, try to turn this into a shufflevector operation.
12657 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12658 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12659 EI->getOperand(0)->getType() == IE.getType()) {
12660 unsigned NumVectorElts = IE.getType()->getNumElements();
12661 unsigned ExtractedIdx =
12662 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12663 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12665 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12666 return ReplaceInstUsesWith(IE, VecOp);
12668 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12669 return ReplaceInstUsesWith(IE, Context->getUndef(IE.getType()));
12671 // If we are extracting a value from a vector, then inserting it right
12672 // back into the same place, just use the input vector.
12673 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12674 return ReplaceInstUsesWith(IE, VecOp);
12676 // We could theoretically do this for ANY input. However, doing so could
12677 // turn chains of insertelement instructions into a chain of shufflevector
12678 // instructions, and right now we do not merge shufflevectors. As such,
12679 // only do this in a situation where it is clear that there is benefit.
12680 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12681 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12682 // the values of VecOp, except then one read from EIOp0.
12683 // Build a new shuffle mask.
12684 std::vector<Constant*> Mask;
12685 if (isa<UndefValue>(VecOp))
12686 Mask.assign(NumVectorElts, Context->getUndef(Type::Int32Ty));
12688 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12689 Mask.assign(NumVectorElts, Context->getConstantInt(Type::Int32Ty,
12692 Mask[InsertedIdx] =
12693 Context->getConstantInt(Type::Int32Ty, ExtractedIdx);
12694 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12695 Context->getConstantVector(Mask));
12698 // If this insertelement isn't used by some other insertelement, turn it
12699 // (and any insertelements it points to), into one big shuffle.
12700 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12701 std::vector<Constant*> Mask;
12703 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12704 if (RHS == 0) RHS = Context->getUndef(LHS->getType());
12705 // We now have a shuffle of LHS, RHS, Mask.
12706 return new ShuffleVectorInst(LHS, RHS,
12707 Context->getConstantVector(Mask));
12712 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12713 APInt UndefElts(VWidth, 0);
12714 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12715 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12722 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12723 Value *LHS = SVI.getOperand(0);
12724 Value *RHS = SVI.getOperand(1);
12725 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12727 bool MadeChange = false;
12729 // Undefined shuffle mask -> undefined value.
12730 if (isa<UndefValue>(SVI.getOperand(2)))
12731 return ReplaceInstUsesWith(SVI, Context->getUndef(SVI.getType()));
12733 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12735 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12738 APInt UndefElts(VWidth, 0);
12739 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12740 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12741 LHS = SVI.getOperand(0);
12742 RHS = SVI.getOperand(1);
12746 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12747 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12748 if (LHS == RHS || isa<UndefValue>(LHS)) {
12749 if (isa<UndefValue>(LHS) && LHS == RHS) {
12750 // shuffle(undef,undef,mask) -> undef.
12751 return ReplaceInstUsesWith(SVI, LHS);
12754 // Remap any references to RHS to use LHS.
12755 std::vector<Constant*> Elts;
12756 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12757 if (Mask[i] >= 2*e)
12758 Elts.push_back(Context->getUndef(Type::Int32Ty));
12760 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12761 (Mask[i] < e && isa<UndefValue>(LHS))) {
12762 Mask[i] = 2*e; // Turn into undef.
12763 Elts.push_back(Context->getUndef(Type::Int32Ty));
12765 Mask[i] = Mask[i] % e; // Force to LHS.
12766 Elts.push_back(Context->getConstantInt(Type::Int32Ty, Mask[i]));
12770 SVI.setOperand(0, SVI.getOperand(1));
12771 SVI.setOperand(1, Context->getUndef(RHS->getType()));
12772 SVI.setOperand(2, Context->getConstantVector(Elts));
12773 LHS = SVI.getOperand(0);
12774 RHS = SVI.getOperand(1);
12778 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12779 bool isLHSID = true, isRHSID = true;
12781 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12782 if (Mask[i] >= e*2) continue; // Ignore undef values.
12783 // Is this an identity shuffle of the LHS value?
12784 isLHSID &= (Mask[i] == i);
12786 // Is this an identity shuffle of the RHS value?
12787 isRHSID &= (Mask[i]-e == i);
12790 // Eliminate identity shuffles.
12791 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12792 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12794 // If the LHS is a shufflevector itself, see if we can combine it with this
12795 // one without producing an unusual shuffle. Here we are really conservative:
12796 // we are absolutely afraid of producing a shuffle mask not in the input
12797 // program, because the code gen may not be smart enough to turn a merged
12798 // shuffle into two specific shuffles: it may produce worse code. As such,
12799 // we only merge two shuffles if the result is one of the two input shuffle
12800 // masks. In this case, merging the shuffles just removes one instruction,
12801 // which we know is safe. This is good for things like turning:
12802 // (splat(splat)) -> splat.
12803 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12804 if (isa<UndefValue>(RHS)) {
12805 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12807 std::vector<unsigned> NewMask;
12808 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12809 if (Mask[i] >= 2*e)
12810 NewMask.push_back(2*e);
12812 NewMask.push_back(LHSMask[Mask[i]]);
12814 // If the result mask is equal to the src shuffle or this shuffle mask, do
12815 // the replacement.
12816 if (NewMask == LHSMask || NewMask == Mask) {
12817 unsigned LHSInNElts =
12818 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12819 std::vector<Constant*> Elts;
12820 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12821 if (NewMask[i] >= LHSInNElts*2) {
12822 Elts.push_back(Context->getUndef(Type::Int32Ty));
12824 Elts.push_back(Context->getConstantInt(Type::Int32Ty, NewMask[i]));
12827 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12828 LHSSVI->getOperand(1),
12829 Context->getConstantVector(Elts));
12834 return MadeChange ? &SVI : 0;
12840 /// TryToSinkInstruction - Try to move the specified instruction from its
12841 /// current block into the beginning of DestBlock, which can only happen if it's
12842 /// safe to move the instruction past all of the instructions between it and the
12843 /// end of its block.
12844 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12845 assert(I->hasOneUse() && "Invariants didn't hold!");
12847 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12848 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12851 // Do not sink alloca instructions out of the entry block.
12852 if (isa<AllocaInst>(I) && I->getParent() ==
12853 &DestBlock->getParent()->getEntryBlock())
12856 // We can only sink load instructions if there is nothing between the load and
12857 // the end of block that could change the value.
12858 if (I->mayReadFromMemory()) {
12859 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12861 if (Scan->mayWriteToMemory())
12865 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12867 CopyPrecedingStopPoint(I, InsertPos);
12868 I->moveBefore(InsertPos);
12874 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12875 /// all reachable code to the worklist.
12877 /// This has a couple of tricks to make the code faster and more powerful. In
12878 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12879 /// them to the worklist (this significantly speeds up instcombine on code where
12880 /// many instructions are dead or constant). Additionally, if we find a branch
12881 /// whose condition is a known constant, we only visit the reachable successors.
12883 static void AddReachableCodeToWorklist(BasicBlock *BB,
12884 SmallPtrSet<BasicBlock*, 64> &Visited,
12886 const TargetData *TD) {
12887 SmallVector<BasicBlock*, 256> Worklist;
12888 Worklist.push_back(BB);
12890 while (!Worklist.empty()) {
12891 BB = Worklist.back();
12892 Worklist.pop_back();
12894 // We have now visited this block! If we've already been here, ignore it.
12895 if (!Visited.insert(BB)) continue;
12897 DbgInfoIntrinsic *DBI_Prev = NULL;
12898 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12899 Instruction *Inst = BBI++;
12901 // DCE instruction if trivially dead.
12902 if (isInstructionTriviallyDead(Inst)) {
12904 DOUT << "IC: DCE: " << *Inst;
12905 Inst->eraseFromParent();
12909 // ConstantProp instruction if trivially constant.
12910 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12911 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12912 Inst->replaceAllUsesWith(C);
12914 Inst->eraseFromParent();
12918 // If there are two consecutive llvm.dbg.stoppoint calls then
12919 // it is likely that the optimizer deleted code in between these
12921 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12924 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12925 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12926 IC.RemoveFromWorkList(DBI_Prev);
12927 DBI_Prev->eraseFromParent();
12929 DBI_Prev = DBI_Next;
12934 IC.AddToWorkList(Inst);
12937 // Recursively visit successors. If this is a branch or switch on a
12938 // constant, only visit the reachable successor.
12939 TerminatorInst *TI = BB->getTerminator();
12940 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12941 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12942 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12943 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12944 Worklist.push_back(ReachableBB);
12947 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12948 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12949 // See if this is an explicit destination.
12950 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12951 if (SI->getCaseValue(i) == Cond) {
12952 BasicBlock *ReachableBB = SI->getSuccessor(i);
12953 Worklist.push_back(ReachableBB);
12957 // Otherwise it is the default destination.
12958 Worklist.push_back(SI->getSuccessor(0));
12963 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12964 Worklist.push_back(TI->getSuccessor(i));
12968 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12969 bool Changed = false;
12970 TD = &getAnalysis<TargetData>();
12972 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12973 << F.getNameStr() << "\n");
12976 // Do a depth-first traversal of the function, populate the worklist with
12977 // the reachable instructions. Ignore blocks that are not reachable. Keep
12978 // track of which blocks we visit.
12979 SmallPtrSet<BasicBlock*, 64> Visited;
12980 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12982 // Do a quick scan over the function. If we find any blocks that are
12983 // unreachable, remove any instructions inside of them. This prevents
12984 // the instcombine code from having to deal with some bad special cases.
12985 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12986 if (!Visited.count(BB)) {
12987 Instruction *Term = BB->getTerminator();
12988 while (Term != BB->begin()) { // Remove instrs bottom-up
12989 BasicBlock::iterator I = Term; --I;
12991 DOUT << "IC: DCE: " << *I;
12992 // A debug intrinsic shouldn't force another iteration if we weren't
12993 // going to do one without it.
12994 if (!isa<DbgInfoIntrinsic>(I)) {
12998 if (!I->use_empty())
12999 I->replaceAllUsesWith(Context->getUndef(I->getType()));
13000 I->eraseFromParent();
13005 while (!Worklist.empty()) {
13006 Instruction *I = RemoveOneFromWorkList();
13007 if (I == 0) continue; // skip null values.
13009 // Check to see if we can DCE the instruction.
13010 if (isInstructionTriviallyDead(I)) {
13011 // Add operands to the worklist.
13012 if (I->getNumOperands() < 4)
13013 AddUsesToWorkList(*I);
13016 DOUT << "IC: DCE: " << *I;
13018 I->eraseFromParent();
13019 RemoveFromWorkList(I);
13024 // Instruction isn't dead, see if we can constant propagate it.
13025 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
13026 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
13028 // Add operands to the worklist.
13029 AddUsesToWorkList(*I);
13030 ReplaceInstUsesWith(*I, C);
13033 I->eraseFromParent();
13034 RemoveFromWorkList(I);
13040 // See if we can constant fold its operands.
13041 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
13042 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
13043 if (Constant *NewC = ConstantFoldConstantExpression(CE,
13044 F.getContext(), TD))
13051 // See if we can trivially sink this instruction to a successor basic block.
13052 if (I->hasOneUse()) {
13053 BasicBlock *BB = I->getParent();
13054 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
13055 if (UserParent != BB) {
13056 bool UserIsSuccessor = false;
13057 // See if the user is one of our successors.
13058 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13059 if (*SI == UserParent) {
13060 UserIsSuccessor = true;
13064 // If the user is one of our immediate successors, and if that successor
13065 // only has us as a predecessors (we'd have to split the critical edge
13066 // otherwise), we can keep going.
13067 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
13068 next(pred_begin(UserParent)) == pred_end(UserParent))
13069 // Okay, the CFG is simple enough, try to sink this instruction.
13070 Changed |= TryToSinkInstruction(I, UserParent);
13074 // Now that we have an instruction, try combining it to simplify it...
13078 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
13079 if (Instruction *Result = visit(*I)) {
13081 // Should we replace the old instruction with a new one?
13083 DOUT << "IC: Old = " << *I
13084 << " New = " << *Result;
13086 // Everything uses the new instruction now.
13087 I->replaceAllUsesWith(Result);
13089 // Push the new instruction and any users onto the worklist.
13090 AddToWorkList(Result);
13091 AddUsersToWorkList(*Result);
13093 // Move the name to the new instruction first.
13094 Result->takeName(I);
13096 // Insert the new instruction into the basic block...
13097 BasicBlock *InstParent = I->getParent();
13098 BasicBlock::iterator InsertPos = I;
13100 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13101 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13104 InstParent->getInstList().insert(InsertPos, Result);
13106 // Make sure that we reprocess all operands now that we reduced their
13108 AddUsesToWorkList(*I);
13110 // Instructions can end up on the worklist more than once. Make sure
13111 // we do not process an instruction that has been deleted.
13112 RemoveFromWorkList(I);
13114 // Erase the old instruction.
13115 InstParent->getInstList().erase(I);
13118 DOUT << "IC: Mod = " << OrigI
13119 << " New = " << *I;
13122 // If the instruction was modified, it's possible that it is now dead.
13123 // if so, remove it.
13124 if (isInstructionTriviallyDead(I)) {
13125 // Make sure we process all operands now that we are reducing their
13127 AddUsesToWorkList(*I);
13129 // Instructions may end up in the worklist more than once. Erase all
13130 // occurrences of this instruction.
13131 RemoveFromWorkList(I);
13132 I->eraseFromParent();
13135 AddUsersToWorkList(*I);
13142 assert(WorklistMap.empty() && "Worklist empty, but map not?");
13144 // Do an explicit clear, this shrinks the map if needed.
13145 WorklistMap.clear();
13150 bool InstCombiner::runOnFunction(Function &F) {
13151 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13153 bool EverMadeChange = false;
13155 // Iterate while there is work to do.
13156 unsigned Iteration = 0;
13157 while (DoOneIteration(F, Iteration++))
13158 EverMadeChange = true;
13159 return EverMadeChange;
13162 FunctionPass *llvm::createInstructionCombiningPass() {
13163 return new InstCombiner();