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 // TODO: If Op1 is undef and Op0 is finite, return zero.
2656 if (!I.getType()->isFPOrFPVector() &&
2657 isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2658 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2660 // Simplify mul instructions with a constant RHS...
2661 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2662 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2664 // ((X << C1)*C2) == (X * (C2 << C1))
2665 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2666 if (SI->getOpcode() == Instruction::Shl)
2667 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2668 return BinaryOperator::CreateMul(SI->getOperand(0),
2669 Context->getConstantExprShl(CI, ShOp));
2672 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2673 if (CI->equalsInt(1)) // X * 1 == X
2674 return ReplaceInstUsesWith(I, Op0);
2675 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2676 return BinaryOperator::CreateNeg(*Context, Op0, I.getName());
2678 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2679 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2680 return BinaryOperator::CreateShl(Op0,
2681 Context->getConstantInt(Op0->getType(), Val.logBase2()));
2683 } else if (isa<VectorType>(Op1->getType())) {
2684 if (Op1->isNullValue())
2685 return ReplaceInstUsesWith(I, Op1);
2687 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2688 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2689 return BinaryOperator::CreateNeg(*Context, Op0, I.getName());
2691 // As above, vector X*splat(1.0) -> X in all defined cases.
2692 if (Constant *Splat = Op1V->getSplatValue()) {
2693 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2694 if (CI->equalsInt(1))
2695 return ReplaceInstUsesWith(I, Op0);
2700 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2701 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2702 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2703 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2704 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2706 InsertNewInstBefore(Add, I);
2707 Value *C1C2 = Context->getConstantExprMul(Op1,
2708 cast<Constant>(Op0I->getOperand(1)));
2709 return BinaryOperator::CreateAdd(Add, C1C2);
2713 // Try to fold constant mul into select arguments.
2714 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2715 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2718 if (isa<PHINode>(Op0))
2719 if (Instruction *NV = FoldOpIntoPhi(I))
2723 if (Value *Op0v = dyn_castNegVal(Op0, Context)) // -X * -Y = X*Y
2724 if (Value *Op1v = dyn_castNegVal(I.getOperand(1), Context))
2725 return BinaryOperator::CreateMul(Op0v, Op1v);
2727 // (X / Y) * Y = X - (X % Y)
2728 // (X / Y) * -Y = (X % Y) - X
2730 Value *Op1 = I.getOperand(1);
2731 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2733 (BO->getOpcode() != Instruction::UDiv &&
2734 BO->getOpcode() != Instruction::SDiv)) {
2736 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2738 Value *Neg = dyn_castNegVal(Op1, Context);
2739 if (BO && BO->hasOneUse() &&
2740 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2741 (BO->getOpcode() == Instruction::UDiv ||
2742 BO->getOpcode() == Instruction::SDiv)) {
2743 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2746 if (BO->getOpcode() == Instruction::UDiv)
2747 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2749 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2751 InsertNewInstBefore(Rem, I);
2755 return BinaryOperator::CreateSub(Op0BO, Rem);
2757 return BinaryOperator::CreateSub(Rem, Op0BO);
2761 if (I.getType() == Type::Int1Ty)
2762 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2764 // If one of the operands of the multiply is a cast from a boolean value, then
2765 // we know the bool is either zero or one, so this is a 'masking' multiply.
2766 // See if we can simplify things based on how the boolean was originally
2768 CastInst *BoolCast = 0;
2769 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2770 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2773 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2774 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2777 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2778 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2779 const Type *SCOpTy = SCIOp0->getType();
2782 // If the icmp is true iff the sign bit of X is set, then convert this
2783 // multiply into a shift/and combination.
2784 if (isa<ConstantInt>(SCIOp1) &&
2785 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2787 // Shift the X value right to turn it into "all signbits".
2788 Constant *Amt = Context->getConstantInt(SCIOp0->getType(),
2789 SCOpTy->getPrimitiveSizeInBits()-1);
2791 InsertNewInstBefore(
2792 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2793 BoolCast->getOperand(0)->getName()+
2796 // If the multiply type is not the same as the source type, sign extend
2797 // or truncate to the multiply type.
2798 if (I.getType() != V->getType()) {
2799 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2800 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2801 Instruction::CastOps opcode =
2802 (SrcBits == DstBits ? Instruction::BitCast :
2803 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2804 V = InsertCastBefore(opcode, V, I.getType(), I);
2807 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2808 return BinaryOperator::CreateAnd(V, OtherOp);
2813 return Changed ? &I : 0;
2816 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2817 bool Changed = SimplifyCommutative(I);
2818 Value *Op0 = I.getOperand(0);
2820 // Simplify mul instructions with a constant RHS...
2821 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2822 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2823 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2824 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2825 if (Op1F->isExactlyValue(1.0))
2826 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2827 } else if (isa<VectorType>(Op1->getType())) {
2828 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2829 // As above, vector X*splat(1.0) -> X in all defined cases.
2830 if (Constant *Splat = Op1V->getSplatValue()) {
2831 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2832 if (F->isExactlyValue(1.0))
2833 return ReplaceInstUsesWith(I, Op0);
2838 // Try to fold constant mul into select arguments.
2839 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2840 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2843 if (isa<PHINode>(Op0))
2844 if (Instruction *NV = FoldOpIntoPhi(I))
2848 if (Value *Op0v = dyn_castFNegVal(Op0, Context)) // -X * -Y = X*Y
2849 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1), Context))
2850 return BinaryOperator::CreateFMul(Op0v, Op1v);
2852 return Changed ? &I : 0;
2855 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2857 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2858 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2860 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2861 int NonNullOperand = -1;
2862 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2863 if (ST->isNullValue())
2865 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2866 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2867 if (ST->isNullValue())
2870 if (NonNullOperand == -1)
2873 Value *SelectCond = SI->getOperand(0);
2875 // Change the div/rem to use 'Y' instead of the select.
2876 I.setOperand(1, SI->getOperand(NonNullOperand));
2878 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2879 // problem. However, the select, or the condition of the select may have
2880 // multiple uses. Based on our knowledge that the operand must be non-zero,
2881 // propagate the known value for the select into other uses of it, and
2882 // propagate a known value of the condition into its other users.
2884 // If the select and condition only have a single use, don't bother with this,
2886 if (SI->use_empty() && SelectCond->hasOneUse())
2889 // Scan the current block backward, looking for other uses of SI.
2890 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2892 while (BBI != BBFront) {
2894 // If we found a call to a function, we can't assume it will return, so
2895 // information from below it cannot be propagated above it.
2896 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2899 // Replace uses of the select or its condition with the known values.
2900 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2903 *I = SI->getOperand(NonNullOperand);
2905 } else if (*I == SelectCond) {
2906 *I = NonNullOperand == 1 ? Context->getConstantIntTrue() :
2907 Context->getConstantIntFalse();
2912 // If we past the instruction, quit looking for it.
2915 if (&*BBI == SelectCond)
2918 // If we ran out of things to eliminate, break out of the loop.
2919 if (SelectCond == 0 && SI == 0)
2927 /// This function implements the transforms on div instructions that work
2928 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2929 /// used by the visitors to those instructions.
2930 /// @brief Transforms common to all three div instructions
2931 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2932 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2934 // undef / X -> 0 for integer.
2935 // undef / X -> undef for FP (the undef could be a snan).
2936 if (isa<UndefValue>(Op0)) {
2937 if (Op0->getType()->isFPOrFPVector())
2938 return ReplaceInstUsesWith(I, Op0);
2939 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2942 // X / undef -> undef
2943 if (isa<UndefValue>(Op1))
2944 return ReplaceInstUsesWith(I, Op1);
2949 /// This function implements the transforms common to both integer division
2950 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2951 /// division instructions.
2952 /// @brief Common integer divide transforms
2953 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2954 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2956 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2958 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2959 Constant *CI = Context->getConstantInt(Ty->getElementType(), 1);
2960 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2961 return ReplaceInstUsesWith(I, Context->getConstantVector(Elts));
2964 Constant *CI = Context->getConstantInt(I.getType(), 1);
2965 return ReplaceInstUsesWith(I, CI);
2968 if (Instruction *Common = commonDivTransforms(I))
2971 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2972 // This does not apply for fdiv.
2973 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2976 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2978 if (RHS->equalsInt(1))
2979 return ReplaceInstUsesWith(I, Op0);
2981 // (X / C1) / C2 -> X / (C1*C2)
2982 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2983 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2984 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2985 if (MultiplyOverflows(RHS, LHSRHS,
2986 I.getOpcode()==Instruction::SDiv, Context))
2987 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2989 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2990 Context->getConstantExprMul(RHS, LHSRHS));
2993 if (!RHS->isZero()) { // avoid X udiv 0
2994 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2995 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2997 if (isa<PHINode>(Op0))
2998 if (Instruction *NV = FoldOpIntoPhi(I))
3003 // 0 / X == 0, we don't need to preserve faults!
3004 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3005 if (LHS->equalsInt(0))
3006 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3008 // It can't be division by zero, hence it must be division by one.
3009 if (I.getType() == Type::Int1Ty)
3010 return ReplaceInstUsesWith(I, Op0);
3012 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3013 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3016 return ReplaceInstUsesWith(I, Op0);
3022 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3023 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3025 // Handle the integer div common cases
3026 if (Instruction *Common = commonIDivTransforms(I))
3029 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3030 // X udiv C^2 -> X >> C
3031 // Check to see if this is an unsigned division with an exact power of 2,
3032 // if so, convert to a right shift.
3033 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3034 return BinaryOperator::CreateLShr(Op0,
3035 Context->getConstantInt(Op0->getType(), C->getValue().logBase2()));
3037 // X udiv C, where C >= signbit
3038 if (C->getValue().isNegative()) {
3039 Value *IC = InsertNewInstBefore(new ICmpInst(*Context,
3040 ICmpInst::ICMP_ULT, Op0, C),
3042 return SelectInst::Create(IC, Context->getNullValue(I.getType()),
3043 Context->getConstantInt(I.getType(), 1));
3047 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3048 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3049 if (RHSI->getOpcode() == Instruction::Shl &&
3050 isa<ConstantInt>(RHSI->getOperand(0))) {
3051 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3052 if (C1.isPowerOf2()) {
3053 Value *N = RHSI->getOperand(1);
3054 const Type *NTy = N->getType();
3055 if (uint32_t C2 = C1.logBase2()) {
3056 Constant *C2V = Context->getConstantInt(NTy, C2);
3057 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3059 return BinaryOperator::CreateLShr(Op0, N);
3064 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3065 // where C1&C2 are powers of two.
3066 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3067 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3068 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3069 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3070 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3071 // Compute the shift amounts
3072 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3073 // Construct the "on true" case of the select
3074 Constant *TC = Context->getConstantInt(Op0->getType(), TSA);
3075 Instruction *TSI = BinaryOperator::CreateLShr(
3076 Op0, TC, SI->getName()+".t");
3077 TSI = InsertNewInstBefore(TSI, I);
3079 // Construct the "on false" case of the select
3080 Constant *FC = Context->getConstantInt(Op0->getType(), FSA);
3081 Instruction *FSI = BinaryOperator::CreateLShr(
3082 Op0, FC, SI->getName()+".f");
3083 FSI = InsertNewInstBefore(FSI, I);
3085 // construct the select instruction and return it.
3086 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3092 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3093 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3095 // Handle the integer div common cases
3096 if (Instruction *Common = commonIDivTransforms(I))
3099 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3101 if (RHS->isAllOnesValue())
3102 return BinaryOperator::CreateNeg(*Context, Op0);
3105 // If the sign bits of both operands are zero (i.e. we can prove they are
3106 // unsigned inputs), turn this into a udiv.
3107 if (I.getType()->isInteger()) {
3108 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3109 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3110 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3111 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3118 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3119 return commonDivTransforms(I);
3122 /// This function implements the transforms on rem instructions that work
3123 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3124 /// is used by the visitors to those instructions.
3125 /// @brief Transforms common to all three rem instructions
3126 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3127 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3129 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3130 if (I.getType()->isFPOrFPVector())
3131 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3132 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3134 if (isa<UndefValue>(Op1))
3135 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3137 // Handle cases involving: rem X, (select Cond, Y, Z)
3138 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3144 /// This function implements the transforms common to both integer remainder
3145 /// instructions (urem and srem). It is called by the visitors to those integer
3146 /// remainder instructions.
3147 /// @brief Common integer remainder transforms
3148 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3149 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3151 if (Instruction *common = commonRemTransforms(I))
3154 // 0 % X == 0 for integer, we don't need to preserve faults!
3155 if (Constant *LHS = dyn_cast<Constant>(Op0))
3156 if (LHS->isNullValue())
3157 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3159 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3160 // X % 0 == undef, we don't need to preserve faults!
3161 if (RHS->equalsInt(0))
3162 return ReplaceInstUsesWith(I, Context->getUndef(I.getType()));
3164 if (RHS->equalsInt(1)) // X % 1 == 0
3165 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3167 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3168 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3169 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3171 } else if (isa<PHINode>(Op0I)) {
3172 if (Instruction *NV = FoldOpIntoPhi(I))
3176 // See if we can fold away this rem instruction.
3177 if (SimplifyDemandedInstructionBits(I))
3185 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3186 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3188 if (Instruction *common = commonIRemTransforms(I))
3191 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3192 // X urem C^2 -> X and C
3193 // Check to see if this is an unsigned remainder with an exact power of 2,
3194 // if so, convert to a bitwise and.
3195 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3196 if (C->getValue().isPowerOf2())
3197 return BinaryOperator::CreateAnd(Op0, SubOne(C, Context));
3200 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3201 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3202 if (RHSI->getOpcode() == Instruction::Shl &&
3203 isa<ConstantInt>(RHSI->getOperand(0))) {
3204 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3205 Constant *N1 = Context->getAllOnesValue(I.getType());
3206 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3208 return BinaryOperator::CreateAnd(Op0, Add);
3213 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3214 // where C1&C2 are powers of two.
3215 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3216 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3217 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3218 // STO == 0 and SFO == 0 handled above.
3219 if ((STO->getValue().isPowerOf2()) &&
3220 (SFO->getValue().isPowerOf2())) {
3221 Value *TrueAnd = InsertNewInstBefore(
3222 BinaryOperator::CreateAnd(Op0, SubOne(STO, Context),
3223 SI->getName()+".t"), I);
3224 Value *FalseAnd = InsertNewInstBefore(
3225 BinaryOperator::CreateAnd(Op0, SubOne(SFO, Context),
3226 SI->getName()+".f"), I);
3227 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3235 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3236 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3238 // Handle the integer rem common cases
3239 if (Instruction *common = commonIRemTransforms(I))
3242 if (Value *RHSNeg = dyn_castNegVal(Op1, Context))
3243 if (!isa<Constant>(RHSNeg) ||
3244 (isa<ConstantInt>(RHSNeg) &&
3245 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3247 AddUsesToWorkList(I);
3248 I.setOperand(1, RHSNeg);
3252 // If the sign bits of both operands are zero (i.e. we can prove they are
3253 // unsigned inputs), turn this into a urem.
3254 if (I.getType()->isInteger()) {
3255 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3256 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3257 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3258 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3262 // If it's a constant vector, flip any negative values positive.
3263 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3264 unsigned VWidth = RHSV->getNumOperands();
3266 bool hasNegative = false;
3267 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3268 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3269 if (RHS->getValue().isNegative())
3273 std::vector<Constant *> Elts(VWidth);
3274 for (unsigned i = 0; i != VWidth; ++i) {
3275 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3276 if (RHS->getValue().isNegative())
3277 Elts[i] = cast<ConstantInt>(Context->getConstantExprNeg(RHS));
3283 Constant *NewRHSV = Context->getConstantVector(Elts);
3284 if (NewRHSV != RHSV) {
3285 AddUsesToWorkList(I);
3286 I.setOperand(1, NewRHSV);
3295 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3296 return commonRemTransforms(I);
3299 // isOneBitSet - Return true if there is exactly one bit set in the specified
3301 static bool isOneBitSet(const ConstantInt *CI) {
3302 return CI->getValue().isPowerOf2();
3305 // isHighOnes - Return true if the constant is of the form 1+0+.
3306 // This is the same as lowones(~X).
3307 static bool isHighOnes(const ConstantInt *CI) {
3308 return (~CI->getValue() + 1).isPowerOf2();
3311 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3312 /// are carefully arranged to allow folding of expressions such as:
3314 /// (A < B) | (A > B) --> (A != B)
3316 /// Note that this is only valid if the first and second predicates have the
3317 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3319 /// Three bits are used to represent the condition, as follows:
3324 /// <=> Value Definition
3325 /// 000 0 Always false
3332 /// 111 7 Always true
3334 static unsigned getICmpCode(const ICmpInst *ICI) {
3335 switch (ICI->getPredicate()) {
3337 case ICmpInst::ICMP_UGT: return 1; // 001
3338 case ICmpInst::ICMP_SGT: return 1; // 001
3339 case ICmpInst::ICMP_EQ: return 2; // 010
3340 case ICmpInst::ICMP_UGE: return 3; // 011
3341 case ICmpInst::ICMP_SGE: return 3; // 011
3342 case ICmpInst::ICMP_ULT: return 4; // 100
3343 case ICmpInst::ICMP_SLT: return 4; // 100
3344 case ICmpInst::ICMP_NE: return 5; // 101
3345 case ICmpInst::ICMP_ULE: return 6; // 110
3346 case ICmpInst::ICMP_SLE: return 6; // 110
3349 llvm_unreachable("Invalid ICmp predicate!");
3354 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3355 /// predicate into a three bit mask. It also returns whether it is an ordered
3356 /// predicate by reference.
3357 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3360 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3361 case FCmpInst::FCMP_UNO: return 0; // 000
3362 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3363 case FCmpInst::FCMP_UGT: return 1; // 001
3364 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3365 case FCmpInst::FCMP_UEQ: return 2; // 010
3366 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3367 case FCmpInst::FCMP_UGE: return 3; // 011
3368 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3369 case FCmpInst::FCMP_ULT: return 4; // 100
3370 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3371 case FCmpInst::FCMP_UNE: return 5; // 101
3372 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3373 case FCmpInst::FCMP_ULE: return 6; // 110
3376 // Not expecting FCMP_FALSE and FCMP_TRUE;
3377 llvm_unreachable("Unexpected FCmp predicate!");
3382 /// getICmpValue - This is the complement of getICmpCode, which turns an
3383 /// opcode and two operands into either a constant true or false, or a brand
3384 /// new ICmp instruction. The sign is passed in to determine which kind
3385 /// of predicate to use in the new icmp instruction.
3386 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3387 LLVMContext *Context) {
3389 default: llvm_unreachable("Illegal ICmp code!");
3390 case 0: return Context->getConstantIntFalse();
3393 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, LHS, RHS);
3395 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, LHS, RHS);
3396 case 2: return new ICmpInst(*Context, ICmpInst::ICMP_EQ, LHS, RHS);
3399 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHS, RHS);
3401 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHS, RHS);
3404 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHS, RHS);
3406 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHS, RHS);
3407 case 5: return new ICmpInst(*Context, ICmpInst::ICMP_NE, LHS, RHS);
3410 return new ICmpInst(*Context, ICmpInst::ICMP_SLE, LHS, RHS);
3412 return new ICmpInst(*Context, ICmpInst::ICMP_ULE, LHS, RHS);
3413 case 7: return Context->getConstantIntTrue();
3417 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3418 /// opcode and two operands into either a FCmp instruction. isordered is passed
3419 /// in to determine which kind of predicate to use in the new fcmp instruction.
3420 static Value *getFCmpValue(bool isordered, unsigned code,
3421 Value *LHS, Value *RHS, LLVMContext *Context) {
3423 default: llvm_unreachable("Illegal FCmp code!");
3426 return new FCmpInst(*Context, FCmpInst::FCMP_ORD, LHS, RHS);
3428 return new FCmpInst(*Context, FCmpInst::FCMP_UNO, LHS, RHS);
3431 return new FCmpInst(*Context, FCmpInst::FCMP_OGT, LHS, RHS);
3433 return new FCmpInst(*Context, FCmpInst::FCMP_UGT, LHS, RHS);
3436 return new FCmpInst(*Context, FCmpInst::FCMP_OEQ, LHS, RHS);
3438 return new FCmpInst(*Context, FCmpInst::FCMP_UEQ, LHS, RHS);
3441 return new FCmpInst(*Context, FCmpInst::FCMP_OGE, LHS, RHS);
3443 return new FCmpInst(*Context, FCmpInst::FCMP_UGE, LHS, RHS);
3446 return new FCmpInst(*Context, FCmpInst::FCMP_OLT, LHS, RHS);
3448 return new FCmpInst(*Context, FCmpInst::FCMP_ULT, LHS, RHS);
3451 return new FCmpInst(*Context, FCmpInst::FCMP_ONE, LHS, RHS);
3453 return new FCmpInst(*Context, FCmpInst::FCMP_UNE, LHS, RHS);
3456 return new FCmpInst(*Context, FCmpInst::FCMP_OLE, LHS, RHS);
3458 return new FCmpInst(*Context, FCmpInst::FCMP_ULE, LHS, RHS);
3459 case 7: return Context->getConstantIntTrue();
3463 /// PredicatesFoldable - Return true if both predicates match sign or if at
3464 /// least one of them is an equality comparison (which is signless).
3465 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3466 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3467 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3468 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3472 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3473 struct FoldICmpLogical {
3476 ICmpInst::Predicate pred;
3477 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3478 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3479 pred(ICI->getPredicate()) {}
3480 bool shouldApply(Value *V) const {
3481 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3482 if (PredicatesFoldable(pred, ICI->getPredicate()))
3483 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3484 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3487 Instruction *apply(Instruction &Log) const {
3488 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3489 if (ICI->getOperand(0) != LHS) {
3490 assert(ICI->getOperand(1) == LHS);
3491 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3494 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3495 unsigned LHSCode = getICmpCode(ICI);
3496 unsigned RHSCode = getICmpCode(RHSICI);
3498 switch (Log.getOpcode()) {
3499 case Instruction::And: Code = LHSCode & RHSCode; break;
3500 case Instruction::Or: Code = LHSCode | RHSCode; break;
3501 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3502 default: llvm_unreachable("Illegal logical opcode!"); return 0;
3505 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3506 ICmpInst::isSignedPredicate(ICI->getPredicate());
3508 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3509 if (Instruction *I = dyn_cast<Instruction>(RV))
3511 // Otherwise, it's a constant boolean value...
3512 return IC.ReplaceInstUsesWith(Log, RV);
3515 } // end anonymous namespace
3517 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3518 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3519 // guaranteed to be a binary operator.
3520 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3522 ConstantInt *AndRHS,
3523 BinaryOperator &TheAnd) {
3524 Value *X = Op->getOperand(0);
3525 Constant *Together = 0;
3527 Together = Context->getConstantExprAnd(AndRHS, OpRHS);
3529 switch (Op->getOpcode()) {
3530 case Instruction::Xor:
3531 if (Op->hasOneUse()) {
3532 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3533 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3534 InsertNewInstBefore(And, TheAnd);
3536 return BinaryOperator::CreateXor(And, Together);
3539 case Instruction::Or:
3540 if (Together == AndRHS) // (X | C) & C --> C
3541 return ReplaceInstUsesWith(TheAnd, AndRHS);
3543 if (Op->hasOneUse() && Together != OpRHS) {
3544 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3545 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3546 InsertNewInstBefore(Or, TheAnd);
3548 return BinaryOperator::CreateAnd(Or, AndRHS);
3551 case Instruction::Add:
3552 if (Op->hasOneUse()) {
3553 // Adding a one to a single bit bit-field should be turned into an XOR
3554 // of the bit. First thing to check is to see if this AND is with a
3555 // single bit constant.
3556 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3558 // If there is only one bit set...
3559 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3560 // Ok, at this point, we know that we are masking the result of the
3561 // ADD down to exactly one bit. If the constant we are adding has
3562 // no bits set below this bit, then we can eliminate the ADD.
3563 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3565 // Check to see if any bits below the one bit set in AndRHSV are set.
3566 if ((AddRHS & (AndRHSV-1)) == 0) {
3567 // If not, the only thing that can effect the output of the AND is
3568 // the bit specified by AndRHSV. If that bit is set, the effect of
3569 // the XOR is to toggle the bit. If it is clear, then the ADD has
3571 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3572 TheAnd.setOperand(0, X);
3575 // Pull the XOR out of the AND.
3576 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3577 InsertNewInstBefore(NewAnd, TheAnd);
3578 NewAnd->takeName(Op);
3579 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3586 case Instruction::Shl: {
3587 // We know that the AND will not produce any of the bits shifted in, so if
3588 // the anded constant includes them, clear them now!
3590 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3591 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3592 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3593 ConstantInt *CI = Context->getConstantInt(AndRHS->getValue() & ShlMask);
3595 if (CI->getValue() == ShlMask) {
3596 // Masking out bits that the shift already masks
3597 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3598 } else if (CI != AndRHS) { // Reducing bits set in and.
3599 TheAnd.setOperand(1, CI);
3604 case Instruction::LShr:
3606 // We know that the AND will not produce any of the bits shifted in, so if
3607 // the anded constant includes them, clear them now! This only applies to
3608 // unsigned shifts, because a signed shr may bring in set bits!
3610 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3611 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3612 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3613 ConstantInt *CI = Context->getConstantInt(AndRHS->getValue() & ShrMask);
3615 if (CI->getValue() == ShrMask) {
3616 // Masking out bits that the shift already masks.
3617 return ReplaceInstUsesWith(TheAnd, Op);
3618 } else if (CI != AndRHS) {
3619 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3624 case Instruction::AShr:
3626 // See if this is shifting in some sign extension, then masking it out
3628 if (Op->hasOneUse()) {
3629 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3630 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3631 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3632 Constant *C = Context->getConstantInt(AndRHS->getValue() & ShrMask);
3633 if (C == AndRHS) { // Masking out bits shifted in.
3634 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3635 // Make the argument unsigned.
3636 Value *ShVal = Op->getOperand(0);
3637 ShVal = InsertNewInstBefore(
3638 BinaryOperator::CreateLShr(ShVal, OpRHS,
3639 Op->getName()), TheAnd);
3640 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3649 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3650 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3651 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3652 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3653 /// insert new instructions.
3654 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3655 bool isSigned, bool Inside,
3657 assert(cast<ConstantInt>(Context->getConstantExprICmp((isSigned ?
3658 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3659 "Lo is not <= Hi in range emission code!");
3662 if (Lo == Hi) // Trivially false.
3663 return new ICmpInst(*Context, ICmpInst::ICMP_NE, V, V);
3665 // V >= Min && V < Hi --> V < Hi
3666 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3667 ICmpInst::Predicate pred = (isSigned ?
3668 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3669 return new ICmpInst(*Context, pred, V, Hi);
3672 // Emit V-Lo <u Hi-Lo
3673 Constant *NegLo = Context->getConstantExprNeg(Lo);
3674 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3675 InsertNewInstBefore(Add, IB);
3676 Constant *UpperBound = Context->getConstantExprAdd(NegLo, Hi);
3677 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, UpperBound);
3680 if (Lo == Hi) // Trivially true.
3681 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, V, V);
3683 // V < Min || V >= Hi -> V > Hi-1
3684 Hi = SubOne(cast<ConstantInt>(Hi), Context);
3685 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3686 ICmpInst::Predicate pred = (isSigned ?
3687 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3688 return new ICmpInst(*Context, pred, V, Hi);
3691 // Emit V-Lo >u Hi-1-Lo
3692 // Note that Hi has already had one subtracted from it, above.
3693 ConstantInt *NegLo = cast<ConstantInt>(Context->getConstantExprNeg(Lo));
3694 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3695 InsertNewInstBefore(Add, IB);
3696 Constant *LowerBound = Context->getConstantExprAdd(NegLo, Hi);
3697 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add, LowerBound);
3700 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3701 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3702 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3703 // not, since all 1s are not contiguous.
3704 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3705 const APInt& V = Val->getValue();
3706 uint32_t BitWidth = Val->getType()->getBitWidth();
3707 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3709 // look for the first zero bit after the run of ones
3710 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3711 // look for the first non-zero bit
3712 ME = V.getActiveBits();
3716 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3717 /// where isSub determines whether the operator is a sub. If we can fold one of
3718 /// the following xforms:
3720 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3721 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3722 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3724 /// return (A +/- B).
3726 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3727 ConstantInt *Mask, bool isSub,
3729 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3730 if (!LHSI || LHSI->getNumOperands() != 2 ||
3731 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3733 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3735 switch (LHSI->getOpcode()) {
3737 case Instruction::And:
3738 if (Context->getConstantExprAnd(N, Mask) == Mask) {
3739 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3740 if ((Mask->getValue().countLeadingZeros() +
3741 Mask->getValue().countPopulation()) ==
3742 Mask->getValue().getBitWidth())
3745 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3746 // part, we don't need any explicit masks to take them out of A. If that
3747 // is all N is, ignore it.
3748 uint32_t MB = 0, ME = 0;
3749 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3750 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3751 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3752 if (MaskedValueIsZero(RHS, Mask))
3757 case Instruction::Or:
3758 case Instruction::Xor:
3759 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3760 if ((Mask->getValue().countLeadingZeros() +
3761 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3762 && Context->getConstantExprAnd(N, Mask)->isNullValue())
3769 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3771 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3772 return InsertNewInstBefore(New, I);
3775 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3776 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3777 ICmpInst *LHS, ICmpInst *RHS) {
3779 ConstantInt *LHSCst, *RHSCst;
3780 ICmpInst::Predicate LHSCC, RHSCC;
3782 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3783 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3784 m_ConstantInt(LHSCst)), *Context) ||
3785 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3786 m_ConstantInt(RHSCst)), *Context))
3789 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3790 // where C is a power of 2
3791 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3792 LHSCst->getValue().isPowerOf2()) {
3793 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3794 InsertNewInstBefore(NewOr, I);
3795 return new ICmpInst(*Context, LHSCC, NewOr, LHSCst);
3798 // From here on, we only handle:
3799 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3800 if (Val != Val2) return 0;
3802 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3803 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3804 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3805 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3806 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3809 // We can't fold (ugt x, C) & (sgt x, C2).
3810 if (!PredicatesFoldable(LHSCC, RHSCC))
3813 // Ensure that the larger constant is on the RHS.
3815 if (ICmpInst::isSignedPredicate(LHSCC) ||
3816 (ICmpInst::isEquality(LHSCC) &&
3817 ICmpInst::isSignedPredicate(RHSCC)))
3818 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3820 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3823 std::swap(LHS, RHS);
3824 std::swap(LHSCst, RHSCst);
3825 std::swap(LHSCC, RHSCC);
3828 // At this point, we know we have have two icmp instructions
3829 // comparing a value against two constants and and'ing the result
3830 // together. Because of the above check, we know that we only have
3831 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3832 // (from the FoldICmpLogical check above), that the two constants
3833 // are not equal and that the larger constant is on the RHS
3834 assert(LHSCst != RHSCst && "Compares not folded above?");
3837 default: llvm_unreachable("Unknown integer condition code!");
3838 case ICmpInst::ICMP_EQ:
3840 default: llvm_unreachable("Unknown integer condition code!");
3841 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3842 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3843 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3844 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3845 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3846 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3847 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3848 return ReplaceInstUsesWith(I, LHS);
3850 case ICmpInst::ICMP_NE:
3852 default: llvm_unreachable("Unknown integer condition code!");
3853 case ICmpInst::ICMP_ULT:
3854 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X u< 14) -> X < 13
3855 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Val, LHSCst);
3856 break; // (X != 13 & X u< 15) -> no change
3857 case ICmpInst::ICMP_SLT:
3858 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X s< 14) -> X < 13
3859 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Val, LHSCst);
3860 break; // (X != 13 & X s< 15) -> no change
3861 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3862 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3863 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3864 return ReplaceInstUsesWith(I, RHS);
3865 case ICmpInst::ICMP_NE:
3866 if (LHSCst == SubOne(RHSCst, Context)){// (X != 13 & X != 14) -> X-13 >u 1
3867 Constant *AddCST = Context->getConstantExprNeg(LHSCst);
3868 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3869 Val->getName()+".off");
3870 InsertNewInstBefore(Add, I);
3871 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add,
3872 Context->getConstantInt(Add->getType(), 1));
3874 break; // (X != 13 & X != 15) -> no change
3877 case ICmpInst::ICMP_ULT:
3879 default: llvm_unreachable("Unknown integer condition code!");
3880 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3881 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3882 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3883 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3885 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3886 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3887 return ReplaceInstUsesWith(I, LHS);
3888 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3892 case ICmpInst::ICMP_SLT:
3894 default: llvm_unreachable("Unknown integer condition code!");
3895 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3896 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3897 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3898 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3900 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3901 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3902 return ReplaceInstUsesWith(I, LHS);
3903 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3907 case ICmpInst::ICMP_UGT:
3909 default: llvm_unreachable("Unknown integer condition code!");
3910 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3911 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3912 return ReplaceInstUsesWith(I, RHS);
3913 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3915 case ICmpInst::ICMP_NE:
3916 if (RHSCst == AddOne(LHSCst, Context)) // (X u> 13 & X != 14) -> X u> 14
3917 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3918 break; // (X u> 13 & X != 15) -> no change
3919 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3920 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3921 RHSCst, false, true, I);
3922 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3926 case ICmpInst::ICMP_SGT:
3928 default: llvm_unreachable("Unknown integer condition code!");
3929 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3930 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3931 return ReplaceInstUsesWith(I, RHS);
3932 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3934 case ICmpInst::ICMP_NE:
3935 if (RHSCst == AddOne(LHSCst, Context)) // (X s> 13 & X != 14) -> X s> 14
3936 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3937 break; // (X s> 13 & X != 15) -> no change
3938 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3939 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3940 RHSCst, true, true, I);
3941 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3951 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3952 bool Changed = SimplifyCommutative(I);
3953 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3955 if (isa<UndefValue>(Op1)) // X & undef -> 0
3956 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3960 return ReplaceInstUsesWith(I, Op1);
3962 // See if we can simplify any instructions used by the instruction whose sole
3963 // purpose is to compute bits we don't care about.
3964 if (SimplifyDemandedInstructionBits(I))
3966 if (isa<VectorType>(I.getType())) {
3967 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3968 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3969 return ReplaceInstUsesWith(I, I.getOperand(0));
3970 } else if (isa<ConstantAggregateZero>(Op1)) {
3971 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3975 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3976 const APInt& AndRHSMask = AndRHS->getValue();
3977 APInt NotAndRHS(~AndRHSMask);
3979 // Optimize a variety of ((val OP C1) & C2) combinations...
3980 if (isa<BinaryOperator>(Op0)) {
3981 Instruction *Op0I = cast<Instruction>(Op0);
3982 Value *Op0LHS = Op0I->getOperand(0);
3983 Value *Op0RHS = Op0I->getOperand(1);
3984 switch (Op0I->getOpcode()) {
3985 case Instruction::Xor:
3986 case Instruction::Or:
3987 // If the mask is only needed on one incoming arm, push it up.
3988 if (Op0I->hasOneUse()) {
3989 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3990 // Not masking anything out for the LHS, move to RHS.
3991 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
3992 Op0RHS->getName()+".masked");
3993 InsertNewInstBefore(NewRHS, I);
3994 return BinaryOperator::Create(
3995 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3997 if (!isa<Constant>(Op0RHS) &&
3998 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3999 // Not masking anything out for the RHS, move to LHS.
4000 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4001 Op0LHS->getName()+".masked");
4002 InsertNewInstBefore(NewLHS, I);
4003 return BinaryOperator::Create(
4004 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4009 case Instruction::Add:
4010 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4011 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4012 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4013 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4014 return BinaryOperator::CreateAnd(V, AndRHS);
4015 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4016 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4019 case Instruction::Sub:
4020 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4021 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4022 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4023 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4024 return BinaryOperator::CreateAnd(V, AndRHS);
4026 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4027 // has 1's for all bits that the subtraction with A might affect.
4028 if (Op0I->hasOneUse()) {
4029 uint32_t BitWidth = AndRHSMask.getBitWidth();
4030 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4031 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4033 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4034 if (!(A && A->isZero()) && // avoid infinite recursion.
4035 MaskedValueIsZero(Op0LHS, Mask)) {
4036 Instruction *NewNeg = BinaryOperator::CreateNeg(*Context, Op0RHS);
4037 InsertNewInstBefore(NewNeg, I);
4038 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4043 case Instruction::Shl:
4044 case Instruction::LShr:
4045 // (1 << x) & 1 --> zext(x == 0)
4046 // (1 >> x) & 1 --> zext(x == 0)
4047 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4048 Instruction *NewICmp = new ICmpInst(*Context, ICmpInst::ICMP_EQ,
4049 Op0RHS, Context->getNullValue(I.getType()));
4050 InsertNewInstBefore(NewICmp, I);
4051 return new ZExtInst(NewICmp, I.getType());
4056 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4057 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4059 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4060 // If this is an integer truncation or change from signed-to-unsigned, and
4061 // if the source is an and/or with immediate, transform it. This
4062 // frequently occurs for bitfield accesses.
4063 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4064 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4065 CastOp->getNumOperands() == 2)
4066 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4067 if (CastOp->getOpcode() == Instruction::And) {
4068 // Change: and (cast (and X, C1) to T), C2
4069 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4070 // This will fold the two constants together, which may allow
4071 // other simplifications.
4072 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4073 CastOp->getOperand(0), I.getType(),
4074 CastOp->getName()+".shrunk");
4075 NewCast = InsertNewInstBefore(NewCast, I);
4076 // trunc_or_bitcast(C1)&C2
4078 Context->getConstantExprTruncOrBitCast(AndCI,I.getType());
4079 C3 = Context->getConstantExprAnd(C3, AndRHS);
4080 return BinaryOperator::CreateAnd(NewCast, C3);
4081 } else if (CastOp->getOpcode() == Instruction::Or) {
4082 // Change: and (cast (or X, C1) to T), C2
4083 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4085 Context->getConstantExprTruncOrBitCast(AndCI,I.getType());
4086 if (Context->getConstantExprAnd(C3, AndRHS) == AndRHS)
4088 return ReplaceInstUsesWith(I, AndRHS);
4094 // Try to fold constant and into select arguments.
4095 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4096 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4098 if (isa<PHINode>(Op0))
4099 if (Instruction *NV = FoldOpIntoPhi(I))
4103 Value *Op0NotVal = dyn_castNotVal(Op0, Context);
4104 Value *Op1NotVal = dyn_castNotVal(Op1, Context);
4106 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4107 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
4109 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4110 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4111 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4112 I.getName()+".demorgan");
4113 InsertNewInstBefore(Or, I);
4114 return BinaryOperator::CreateNot(*Context, Or);
4118 Value *A = 0, *B = 0, *C = 0, *D = 0;
4119 if (match(Op0, m_Or(m_Value(A), m_Value(B)), *Context)) {
4120 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4121 return ReplaceInstUsesWith(I, Op1);
4123 // (A|B) & ~(A&B) -> A^B
4124 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4125 if ((A == C && B == D) || (A == D && B == C))
4126 return BinaryOperator::CreateXor(A, B);
4130 if (match(Op1, m_Or(m_Value(A), m_Value(B)), *Context)) {
4131 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4132 return ReplaceInstUsesWith(I, Op0);
4134 // ~(A&B) & (A|B) -> A^B
4135 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4136 if ((A == C && B == D) || (A == D && B == C))
4137 return BinaryOperator::CreateXor(A, B);
4141 if (Op0->hasOneUse() &&
4142 match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4143 if (A == Op1) { // (A^B)&A -> A&(A^B)
4144 I.swapOperands(); // Simplify below
4145 std::swap(Op0, Op1);
4146 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4147 cast<BinaryOperator>(Op0)->swapOperands();
4148 I.swapOperands(); // Simplify below
4149 std::swap(Op0, Op1);
4153 if (Op1->hasOneUse() &&
4154 match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4155 if (B == Op0) { // B&(A^B) -> B&(B^A)
4156 cast<BinaryOperator>(Op1)->swapOperands();
4159 if (A == Op0) { // A&(A^B) -> A & ~B
4160 Instruction *NotB = BinaryOperator::CreateNot(*Context, B, "tmp");
4161 InsertNewInstBefore(NotB, I);
4162 return BinaryOperator::CreateAnd(A, NotB);
4166 // (A&((~A)|B)) -> A&B
4167 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A)), *Context) ||
4168 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1))), *Context))
4169 return BinaryOperator::CreateAnd(A, Op1);
4170 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A)), *Context) ||
4171 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0))), *Context))
4172 return BinaryOperator::CreateAnd(A, Op0);
4175 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4176 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4177 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4180 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4181 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4185 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4186 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4187 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4188 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4189 const Type *SrcTy = Op0C->getOperand(0)->getType();
4190 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4191 // Only do this if the casts both really cause code to be generated.
4192 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4194 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4196 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4197 Op1C->getOperand(0),
4199 InsertNewInstBefore(NewOp, I);
4200 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4204 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4205 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4206 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4207 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4208 SI0->getOperand(1) == SI1->getOperand(1) &&
4209 (SI0->hasOneUse() || SI1->hasOneUse())) {
4210 Instruction *NewOp =
4211 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4213 SI0->getName()), I);
4214 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4215 SI1->getOperand(1));
4219 // If and'ing two fcmp, try combine them into one.
4220 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4221 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4222 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4223 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4224 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4225 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4226 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4227 // If either of the constants are nans, then the whole thing returns
4229 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4230 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4231 return new FCmpInst(*Context, FCmpInst::FCMP_ORD,
4232 LHS->getOperand(0), RHS->getOperand(0));
4235 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4236 FCmpInst::Predicate Op0CC, Op1CC;
4237 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS),
4238 m_Value(Op0RHS)), *Context) &&
4239 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS),
4240 m_Value(Op1RHS)), *Context)) {
4241 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4242 // Swap RHS operands to match LHS.
4243 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4244 std::swap(Op1LHS, Op1RHS);
4246 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4247 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4249 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4251 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4252 Op1CC == FCmpInst::FCMP_FALSE)
4253 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4254 else if (Op0CC == FCmpInst::FCMP_TRUE)
4255 return ReplaceInstUsesWith(I, Op1);
4256 else if (Op1CC == FCmpInst::FCMP_TRUE)
4257 return ReplaceInstUsesWith(I, Op0);
4260 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4261 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4263 std::swap(Op0, Op1);
4264 std::swap(Op0Pred, Op1Pred);
4265 std::swap(Op0Ordered, Op1Ordered);
4268 // uno && ueq -> uno && (uno || eq) -> ueq
4269 // ord && olt -> ord && (ord && lt) -> olt
4270 if (Op0Ordered == Op1Ordered)
4271 return ReplaceInstUsesWith(I, Op1);
4272 // uno && oeq -> uno && (ord && eq) -> false
4273 // uno && ord -> false
4275 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4276 // ord && ueq -> ord && (uno || eq) -> oeq
4277 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4278 Op0LHS, Op0RHS, Context));
4286 return Changed ? &I : 0;
4289 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4290 /// capable of providing pieces of a bswap. The subexpression provides pieces
4291 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4292 /// the expression came from the corresponding "byte swapped" byte in some other
4293 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4294 /// we know that the expression deposits the low byte of %X into the high byte
4295 /// of the bswap result and that all other bytes are zero. This expression is
4296 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4299 /// This function returns true if the match was unsuccessful and false if so.
4300 /// On entry to the function the "OverallLeftShift" is a signed integer value
4301 /// indicating the number of bytes that the subexpression is later shifted. For
4302 /// example, if the expression is later right shifted by 16 bits, the
4303 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4304 /// byte of ByteValues is actually being set.
4306 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4307 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4308 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4309 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4310 /// always in the local (OverallLeftShift) coordinate space.
4312 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4313 SmallVector<Value*, 8> &ByteValues) {
4314 if (Instruction *I = dyn_cast<Instruction>(V)) {
4315 // If this is an or instruction, it may be an inner node of the bswap.
4316 if (I->getOpcode() == Instruction::Or) {
4317 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4319 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4323 // If this is a logical shift by a constant multiple of 8, recurse with
4324 // OverallLeftShift and ByteMask adjusted.
4325 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4327 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4328 // Ensure the shift amount is defined and of a byte value.
4329 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4332 unsigned ByteShift = ShAmt >> 3;
4333 if (I->getOpcode() == Instruction::Shl) {
4334 // X << 2 -> collect(X, +2)
4335 OverallLeftShift += ByteShift;
4336 ByteMask >>= ByteShift;
4338 // X >>u 2 -> collect(X, -2)
4339 OverallLeftShift -= ByteShift;
4340 ByteMask <<= ByteShift;
4341 ByteMask &= (~0U >> (32-ByteValues.size()));
4344 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4345 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4347 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4351 // If this is a logical 'and' with a mask that clears bytes, clear the
4352 // corresponding bytes in ByteMask.
4353 if (I->getOpcode() == Instruction::And &&
4354 isa<ConstantInt>(I->getOperand(1))) {
4355 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4356 unsigned NumBytes = ByteValues.size();
4357 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4358 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4360 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4361 // If this byte is masked out by a later operation, we don't care what
4363 if ((ByteMask & (1 << i)) == 0)
4366 // If the AndMask is all zeros for this byte, clear the bit.
4367 APInt MaskB = AndMask & Byte;
4369 ByteMask &= ~(1U << i);
4373 // If the AndMask is not all ones for this byte, it's not a bytezap.
4377 // Otherwise, this byte is kept.
4380 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4385 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4386 // the input value to the bswap. Some observations: 1) if more than one byte
4387 // is demanded from this input, then it could not be successfully assembled
4388 // into a byteswap. At least one of the two bytes would not be aligned with
4389 // their ultimate destination.
4390 if (!isPowerOf2_32(ByteMask)) return true;
4391 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4393 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4394 // is demanded, it needs to go into byte 0 of the result. This means that the
4395 // byte needs to be shifted until it lands in the right byte bucket. The
4396 // shift amount depends on the position: if the byte is coming from the high
4397 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4398 // low part, it must be shifted left.
4399 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4400 if (InputByteNo < ByteValues.size()/2) {
4401 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4404 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4408 // If the destination byte value is already defined, the values are or'd
4409 // together, which isn't a bswap (unless it's an or of the same bits).
4410 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4412 ByteValues[DestByteNo] = V;
4416 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4417 /// If so, insert the new bswap intrinsic and return it.
4418 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4419 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4420 if (!ITy || ITy->getBitWidth() % 16 ||
4421 // ByteMask only allows up to 32-byte values.
4422 ITy->getBitWidth() > 32*8)
4423 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4425 /// ByteValues - For each byte of the result, we keep track of which value
4426 /// defines each byte.
4427 SmallVector<Value*, 8> ByteValues;
4428 ByteValues.resize(ITy->getBitWidth()/8);
4430 // Try to find all the pieces corresponding to the bswap.
4431 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4432 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4435 // Check to see if all of the bytes come from the same value.
4436 Value *V = ByteValues[0];
4437 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4439 // Check to make sure that all of the bytes come from the same value.
4440 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4441 if (ByteValues[i] != V)
4443 const Type *Tys[] = { ITy };
4444 Module *M = I.getParent()->getParent()->getParent();
4445 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4446 return CallInst::Create(F, V);
4449 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4450 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4451 /// we can simplify this expression to "cond ? C : D or B".
4452 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4454 LLVMContext *Context) {
4455 // If A is not a select of -1/0, this cannot match.
4457 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond)), *Context))
4460 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4461 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4462 return SelectInst::Create(Cond, C, B);
4463 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4464 return SelectInst::Create(Cond, C, B);
4465 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4466 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4467 return SelectInst::Create(Cond, C, D);
4468 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4469 return SelectInst::Create(Cond, C, D);
4473 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4474 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4475 ICmpInst *LHS, ICmpInst *RHS) {
4477 ConstantInt *LHSCst, *RHSCst;
4478 ICmpInst::Predicate LHSCC, RHSCC;
4480 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4481 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4482 m_ConstantInt(LHSCst)), *Context) ||
4483 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4484 m_ConstantInt(RHSCst)), *Context))
4487 // From here on, we only handle:
4488 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4489 if (Val != Val2) return 0;
4491 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4492 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4493 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4494 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4495 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4498 // We can't fold (ugt x, C) | (sgt x, C2).
4499 if (!PredicatesFoldable(LHSCC, RHSCC))
4502 // Ensure that the larger constant is on the RHS.
4504 if (ICmpInst::isSignedPredicate(LHSCC) ||
4505 (ICmpInst::isEquality(LHSCC) &&
4506 ICmpInst::isSignedPredicate(RHSCC)))
4507 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4509 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4512 std::swap(LHS, RHS);
4513 std::swap(LHSCst, RHSCst);
4514 std::swap(LHSCC, RHSCC);
4517 // At this point, we know we have have two icmp instructions
4518 // comparing a value against two constants and or'ing the result
4519 // together. Because of the above check, we know that we only have
4520 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4521 // FoldICmpLogical check above), that the two constants are not
4523 assert(LHSCst != RHSCst && "Compares not folded above?");
4526 default: llvm_unreachable("Unknown integer condition code!");
4527 case ICmpInst::ICMP_EQ:
4529 default: llvm_unreachable("Unknown integer condition code!");
4530 case ICmpInst::ICMP_EQ:
4531 if (LHSCst == SubOne(RHSCst, Context)) {
4532 // (X == 13 | X == 14) -> X-13 <u 2
4533 Constant *AddCST = Context->getConstantExprNeg(LHSCst);
4534 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4535 Val->getName()+".off");
4536 InsertNewInstBefore(Add, I);
4537 AddCST = Context->getConstantExprSub(AddOne(RHSCst, Context), LHSCst);
4538 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, AddCST);
4540 break; // (X == 13 | X == 15) -> no change
4541 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4542 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4544 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4545 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4546 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4547 return ReplaceInstUsesWith(I, RHS);
4550 case ICmpInst::ICMP_NE:
4552 default: llvm_unreachable("Unknown integer condition code!");
4553 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4554 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4555 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4556 return ReplaceInstUsesWith(I, LHS);
4557 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4558 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4559 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4560 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4563 case ICmpInst::ICMP_ULT:
4565 default: llvm_unreachable("Unknown integer condition code!");
4566 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4568 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4569 // If RHSCst is [us]MAXINT, it is always false. Not handling
4570 // this can cause overflow.
4571 if (RHSCst->isMaxValue(false))
4572 return ReplaceInstUsesWith(I, LHS);
4573 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4575 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4577 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4578 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4579 return ReplaceInstUsesWith(I, RHS);
4580 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4584 case ICmpInst::ICMP_SLT:
4586 default: llvm_unreachable("Unknown integer condition code!");
4587 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4589 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4590 // If RHSCst is [us]MAXINT, it is always false. Not handling
4591 // this can cause overflow.
4592 if (RHSCst->isMaxValue(true))
4593 return ReplaceInstUsesWith(I, LHS);
4594 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4596 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4598 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4599 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4600 return ReplaceInstUsesWith(I, RHS);
4601 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4605 case ICmpInst::ICMP_UGT:
4607 default: llvm_unreachable("Unknown integer condition code!");
4608 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4609 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4610 return ReplaceInstUsesWith(I, LHS);
4611 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4613 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4614 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4615 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4616 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4620 case ICmpInst::ICMP_SGT:
4622 default: llvm_unreachable("Unknown integer condition code!");
4623 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4624 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4625 return ReplaceInstUsesWith(I, LHS);
4626 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4628 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4629 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4630 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4631 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4639 /// FoldOrWithConstants - This helper function folds:
4641 /// ((A | B) & C1) | (B & C2)
4647 /// when the XOR of the two constants is "all ones" (-1).
4648 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4649 Value *A, Value *B, Value *C) {
4650 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4654 ConstantInt *CI2 = 0;
4655 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)), *Context)) return 0;
4657 APInt Xor = CI1->getValue() ^ CI2->getValue();
4658 if (!Xor.isAllOnesValue()) return 0;
4660 if (V1 == A || V1 == B) {
4661 Instruction *NewOp =
4662 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4663 return BinaryOperator::CreateOr(NewOp, V1);
4669 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4670 bool Changed = SimplifyCommutative(I);
4671 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4673 if (isa<UndefValue>(Op1)) // X | undef -> -1
4674 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4678 return ReplaceInstUsesWith(I, Op0);
4680 // See if we can simplify any instructions used by the instruction whose sole
4681 // purpose is to compute bits we don't care about.
4682 if (SimplifyDemandedInstructionBits(I))
4684 if (isa<VectorType>(I.getType())) {
4685 if (isa<ConstantAggregateZero>(Op1)) {
4686 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4687 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4688 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4689 return ReplaceInstUsesWith(I, I.getOperand(1));
4694 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4695 ConstantInt *C1 = 0; Value *X = 0;
4696 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4697 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1)), *Context) &&
4699 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4700 InsertNewInstBefore(Or, I);
4702 return BinaryOperator::CreateAnd(Or,
4703 Context->getConstantInt(RHS->getValue() | C1->getValue()));
4706 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4707 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1)), *Context) &&
4709 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4710 InsertNewInstBefore(Or, I);
4712 return BinaryOperator::CreateXor(Or,
4713 Context->getConstantInt(C1->getValue() & ~RHS->getValue()));
4716 // Try to fold constant and into select arguments.
4717 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4718 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4720 if (isa<PHINode>(Op0))
4721 if (Instruction *NV = FoldOpIntoPhi(I))
4725 Value *A = 0, *B = 0;
4726 ConstantInt *C1 = 0, *C2 = 0;
4728 if (match(Op0, m_And(m_Value(A), m_Value(B)), *Context))
4729 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4730 return ReplaceInstUsesWith(I, Op1);
4731 if (match(Op1, m_And(m_Value(A), m_Value(B)), *Context))
4732 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4733 return ReplaceInstUsesWith(I, Op0);
4735 // (A | B) | C and A | (B | C) -> bswap if possible.
4736 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4737 if (match(Op0, m_Or(m_Value(), m_Value()), *Context) ||
4738 match(Op1, m_Or(m_Value(), m_Value()), *Context) ||
4739 (match(Op0, m_Shift(m_Value(), m_Value()), *Context) &&
4740 match(Op1, m_Shift(m_Value(), m_Value()), *Context))) {
4741 if (Instruction *BSwap = MatchBSwap(I))
4745 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4746 if (Op0->hasOneUse() &&
4747 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4748 MaskedValueIsZero(Op1, C1->getValue())) {
4749 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4750 InsertNewInstBefore(NOr, I);
4752 return BinaryOperator::CreateXor(NOr, C1);
4755 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4756 if (Op1->hasOneUse() &&
4757 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4758 MaskedValueIsZero(Op0, C1->getValue())) {
4759 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4760 InsertNewInstBefore(NOr, I);
4762 return BinaryOperator::CreateXor(NOr, C1);
4766 Value *C = 0, *D = 0;
4767 if (match(Op0, m_And(m_Value(A), m_Value(C)), *Context) &&
4768 match(Op1, m_And(m_Value(B), m_Value(D)), *Context)) {
4769 Value *V1 = 0, *V2 = 0, *V3 = 0;
4770 C1 = dyn_cast<ConstantInt>(C);
4771 C2 = dyn_cast<ConstantInt>(D);
4772 if (C1 && C2) { // (A & C1)|(B & C2)
4773 // If we have: ((V + N) & C1) | (V & C2)
4774 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4775 // replace with V+N.
4776 if (C1->getValue() == ~C2->getValue()) {
4777 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4778 match(A, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4779 // Add commutes, try both ways.
4780 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4781 return ReplaceInstUsesWith(I, A);
4782 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4783 return ReplaceInstUsesWith(I, A);
4785 // Or commutes, try both ways.
4786 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4787 match(B, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4788 // Add commutes, try both ways.
4789 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4790 return ReplaceInstUsesWith(I, B);
4791 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4792 return ReplaceInstUsesWith(I, B);
4795 V1 = 0; V2 = 0; V3 = 0;
4798 // Check to see if we have any common things being and'ed. If so, find the
4799 // terms for V1 & (V2|V3).
4800 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4801 if (A == B) // (A & C)|(A & D) == A & (C|D)
4802 V1 = A, V2 = C, V3 = D;
4803 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4804 V1 = A, V2 = B, V3 = C;
4805 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4806 V1 = C, V2 = A, V3 = D;
4807 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4808 V1 = C, V2 = A, V3 = B;
4812 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4813 return BinaryOperator::CreateAnd(V1, Or);
4817 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4818 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4820 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4822 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4824 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4827 // ((A&~B)|(~A&B)) -> A^B
4828 if ((match(C, m_Not(m_Specific(D)), *Context) &&
4829 match(B, m_Not(m_Specific(A)), *Context)))
4830 return BinaryOperator::CreateXor(A, D);
4831 // ((~B&A)|(~A&B)) -> A^B
4832 if ((match(A, m_Not(m_Specific(D)), *Context) &&
4833 match(B, m_Not(m_Specific(C)), *Context)))
4834 return BinaryOperator::CreateXor(C, D);
4835 // ((A&~B)|(B&~A)) -> A^B
4836 if ((match(C, m_Not(m_Specific(B)), *Context) &&
4837 match(D, m_Not(m_Specific(A)), *Context)))
4838 return BinaryOperator::CreateXor(A, B);
4839 // ((~B&A)|(B&~A)) -> A^B
4840 if ((match(A, m_Not(m_Specific(B)), *Context) &&
4841 match(D, m_Not(m_Specific(C)), *Context)))
4842 return BinaryOperator::CreateXor(C, B);
4845 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4846 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4847 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4848 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4849 SI0->getOperand(1) == SI1->getOperand(1) &&
4850 (SI0->hasOneUse() || SI1->hasOneUse())) {
4851 Instruction *NewOp =
4852 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4854 SI0->getName()), I);
4855 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4856 SI1->getOperand(1));
4860 // ((A|B)&1)|(B&-2) -> (A&1) | B
4861 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4862 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4863 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4864 if (Ret) return Ret;
4866 // (B&-2)|((A|B)&1) -> (A&1) | B
4867 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4868 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4869 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4870 if (Ret) return Ret;
4873 if (match(Op0, m_Not(m_Value(A)), *Context)) { // ~A | Op1
4874 if (A == Op1) // ~A | A == -1
4875 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4879 // Note, A is still live here!
4880 if (match(Op1, m_Not(m_Value(B)), *Context)) { // Op0 | ~B
4882 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4884 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4885 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4886 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4887 I.getName()+".demorgan"), I);
4888 return BinaryOperator::CreateNot(*Context, And);
4892 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4893 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4894 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4897 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4898 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4902 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4903 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4904 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4905 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4906 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4907 !isa<ICmpInst>(Op1C->getOperand(0))) {
4908 const Type *SrcTy = Op0C->getOperand(0)->getType();
4909 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4910 // Only do this if the casts both really cause code to be
4912 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4914 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4916 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4917 Op1C->getOperand(0),
4919 InsertNewInstBefore(NewOp, I);
4920 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4927 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4928 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4929 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4930 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4931 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4932 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4933 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4934 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4935 // If either of the constants are nans, then the whole thing returns
4937 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4938 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4940 // Otherwise, no need to compare the two constants, compare the
4942 return new FCmpInst(*Context, FCmpInst::FCMP_UNO,
4943 LHS->getOperand(0), RHS->getOperand(0));
4946 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4947 FCmpInst::Predicate Op0CC, Op1CC;
4948 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS),
4949 m_Value(Op0RHS)), *Context) &&
4950 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS),
4951 m_Value(Op1RHS)), *Context)) {
4952 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4953 // Swap RHS operands to match LHS.
4954 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4955 std::swap(Op1LHS, Op1RHS);
4957 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4958 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4960 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4962 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4963 Op1CC == FCmpInst::FCMP_TRUE)
4964 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4965 else if (Op0CC == FCmpInst::FCMP_FALSE)
4966 return ReplaceInstUsesWith(I, Op1);
4967 else if (Op1CC == FCmpInst::FCMP_FALSE)
4968 return ReplaceInstUsesWith(I, Op0);
4971 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4972 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4973 if (Op0Ordered == Op1Ordered) {
4974 // If both are ordered or unordered, return a new fcmp with
4975 // or'ed predicates.
4976 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4977 Op0LHS, Op0RHS, Context);
4978 if (Instruction *I = dyn_cast<Instruction>(RV))
4980 // Otherwise, it's a constant boolean value...
4981 return ReplaceInstUsesWith(I, RV);
4989 return Changed ? &I : 0;
4994 // XorSelf - Implements: X ^ X --> 0
4997 XorSelf(Value *rhs) : RHS(rhs) {}
4998 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4999 Instruction *apply(BinaryOperator &Xor) const {
5006 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5007 bool Changed = SimplifyCommutative(I);
5008 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5010 if (isa<UndefValue>(Op1)) {
5011 if (isa<UndefValue>(Op0))
5012 // Handle undef ^ undef -> 0 special case. This is a common
5014 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
5015 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5018 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5019 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1), Context)) {
5020 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5021 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
5024 // See if we can simplify any instructions used by the instruction whose sole
5025 // purpose is to compute bits we don't care about.
5026 if (SimplifyDemandedInstructionBits(I))
5028 if (isa<VectorType>(I.getType()))
5029 if (isa<ConstantAggregateZero>(Op1))
5030 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5032 // Is this a ~ operation?
5033 if (Value *NotOp = dyn_castNotVal(&I, Context)) {
5034 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5035 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5036 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5037 if (Op0I->getOpcode() == Instruction::And ||
5038 Op0I->getOpcode() == Instruction::Or) {
5039 if (dyn_castNotVal(Op0I->getOperand(1), Context)) Op0I->swapOperands();
5040 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0), Context)) {
5042 BinaryOperator::CreateNot(*Context, Op0I->getOperand(1),
5043 Op0I->getOperand(1)->getName()+".not");
5044 InsertNewInstBefore(NotY, I);
5045 if (Op0I->getOpcode() == Instruction::And)
5046 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5048 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5055 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5056 if (RHS == Context->getConstantIntTrue() && Op0->hasOneUse()) {
5057 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5058 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5059 return new ICmpInst(*Context, ICI->getInversePredicate(),
5060 ICI->getOperand(0), ICI->getOperand(1));
5062 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5063 return new FCmpInst(*Context, FCI->getInversePredicate(),
5064 FCI->getOperand(0), FCI->getOperand(1));
5067 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5068 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5069 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5070 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5071 Instruction::CastOps Opcode = Op0C->getOpcode();
5072 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
5073 if (RHS == Context->getConstantExprCast(Opcode,
5074 Context->getConstantIntTrue(),
5075 Op0C->getDestTy())) {
5076 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
5078 CI->getOpcode(), CI->getInversePredicate(),
5079 CI->getOperand(0), CI->getOperand(1)), I);
5080 NewCI->takeName(CI);
5081 return CastInst::Create(Opcode, NewCI, Op0C->getType());
5088 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5089 // ~(c-X) == X-c-1 == X+(-c-1)
5090 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5091 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5092 Constant *NegOp0I0C = Context->getConstantExprNeg(Op0I0C);
5093 Constant *ConstantRHS = Context->getConstantExprSub(NegOp0I0C,
5094 Context->getConstantInt(I.getType(), 1));
5095 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5098 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5099 if (Op0I->getOpcode() == Instruction::Add) {
5100 // ~(X-c) --> (-c-1)-X
5101 if (RHS->isAllOnesValue()) {
5102 Constant *NegOp0CI = Context->getConstantExprNeg(Op0CI);
5103 return BinaryOperator::CreateSub(
5104 Context->getConstantExprSub(NegOp0CI,
5105 Context->getConstantInt(I.getType(), 1)),
5106 Op0I->getOperand(0));
5107 } else if (RHS->getValue().isSignBit()) {
5108 // (X + C) ^ signbit -> (X + C + signbit)
5110 Context->getConstantInt(RHS->getValue() + Op0CI->getValue());
5111 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5114 } else if (Op0I->getOpcode() == Instruction::Or) {
5115 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5116 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5117 Constant *NewRHS = Context->getConstantExprOr(Op0CI, RHS);
5118 // Anything in both C1 and C2 is known to be zero, remove it from
5120 Constant *CommonBits = Context->getConstantExprAnd(Op0CI, RHS);
5121 NewRHS = Context->getConstantExprAnd(NewRHS,
5122 Context->getConstantExprNot(CommonBits));
5123 AddToWorkList(Op0I);
5124 I.setOperand(0, Op0I->getOperand(0));
5125 I.setOperand(1, NewRHS);
5132 // Try to fold constant and into select arguments.
5133 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5134 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5136 if (isa<PHINode>(Op0))
5137 if (Instruction *NV = FoldOpIntoPhi(I))
5141 if (Value *X = dyn_castNotVal(Op0, Context)) // ~A ^ A == -1
5143 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
5145 if (Value *X = dyn_castNotVal(Op1, Context)) // A ^ ~A == -1
5147 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
5150 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5153 if (match(Op1I, m_Or(m_Value(A), m_Value(B)), *Context)) {
5154 if (A == Op0) { // B^(B|A) == (A|B)^B
5155 Op1I->swapOperands();
5157 std::swap(Op0, Op1);
5158 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5159 I.swapOperands(); // Simplified below.
5160 std::swap(Op0, Op1);
5162 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)), *Context)) {
5163 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5164 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)), *Context)) {
5165 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5166 } else if (match(Op1I, m_And(m_Value(A), m_Value(B)), *Context) &&
5168 if (A == Op0) { // A^(A&B) -> A^(B&A)
5169 Op1I->swapOperands();
5172 if (B == Op0) { // A^(B&A) -> (B&A)^A
5173 I.swapOperands(); // Simplified below.
5174 std::swap(Op0, Op1);
5179 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5182 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5183 Op0I->hasOneUse()) {
5184 if (A == Op1) // (B|A)^B == (A|B)^B
5186 if (B == Op1) { // (A|B)^B == A & ~B
5188 InsertNewInstBefore(BinaryOperator::CreateNot(*Context,
5190 return BinaryOperator::CreateAnd(A, NotB);
5192 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)), *Context)) {
5193 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5194 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)), *Context)) {
5195 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5196 } else if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5198 if (A == Op1) // (A&B)^A -> (B&A)^A
5200 if (B == Op1 && // (B&A)^A == ~B & A
5201 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5203 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, A, "tmp"), I);
5204 return BinaryOperator::CreateAnd(N, Op1);
5209 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5210 if (Op0I && Op1I && Op0I->isShift() &&
5211 Op0I->getOpcode() == Op1I->getOpcode() &&
5212 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5213 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5214 Instruction *NewOp =
5215 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5216 Op1I->getOperand(0),
5217 Op0I->getName()), I);
5218 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5219 Op1I->getOperand(1));
5223 Value *A, *B, *C, *D;
5224 // (A & B)^(A | B) -> A ^ B
5225 if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5226 match(Op1I, m_Or(m_Value(C), m_Value(D)), *Context)) {
5227 if ((A == C && B == D) || (A == D && B == C))
5228 return BinaryOperator::CreateXor(A, B);
5230 // (A | B)^(A & B) -> A ^ B
5231 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5232 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5233 if ((A == C && B == D) || (A == D && B == C))
5234 return BinaryOperator::CreateXor(A, B);
5238 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5239 match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5240 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5241 // (X & Y)^(X & Y) -> (Y^Z) & X
5242 Value *X = 0, *Y = 0, *Z = 0;
5244 X = A, Y = B, Z = D;
5246 X = A, Y = B, Z = C;
5248 X = B, Y = A, Z = D;
5250 X = B, Y = A, Z = C;
5253 Instruction *NewOp =
5254 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5255 return BinaryOperator::CreateAnd(NewOp, X);
5260 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5261 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5262 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
5265 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5266 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5267 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5268 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5269 const Type *SrcTy = Op0C->getOperand(0)->getType();
5270 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5271 // Only do this if the casts both really cause code to be generated.
5272 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5274 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5276 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5277 Op1C->getOperand(0),
5279 InsertNewInstBefore(NewOp, I);
5280 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5285 return Changed ? &I : 0;
5288 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5289 LLVMContext *Context) {
5290 return cast<ConstantInt>(Context->getConstantExprExtractElement(V, Idx));
5293 static bool HasAddOverflow(ConstantInt *Result,
5294 ConstantInt *In1, ConstantInt *In2,
5297 if (In2->getValue().isNegative())
5298 return Result->getValue().sgt(In1->getValue());
5300 return Result->getValue().slt(In1->getValue());
5302 return Result->getValue().ult(In1->getValue());
5305 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5306 /// overflowed for this type.
5307 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5308 Constant *In2, LLVMContext *Context,
5309 bool IsSigned = false) {
5310 Result = Context->getConstantExprAdd(In1, In2);
5312 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5313 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5314 Constant *Idx = Context->getConstantInt(Type::Int32Ty, i);
5315 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5316 ExtractElement(In1, Idx, Context),
5317 ExtractElement(In2, Idx, Context),
5324 return HasAddOverflow(cast<ConstantInt>(Result),
5325 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5329 static bool HasSubOverflow(ConstantInt *Result,
5330 ConstantInt *In1, ConstantInt *In2,
5333 if (In2->getValue().isNegative())
5334 return Result->getValue().slt(In1->getValue());
5336 return Result->getValue().sgt(In1->getValue());
5338 return Result->getValue().ugt(In1->getValue());
5341 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5342 /// overflowed for this type.
5343 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5344 Constant *In2, LLVMContext *Context,
5345 bool IsSigned = false) {
5346 Result = Context->getConstantExprSub(In1, In2);
5348 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5349 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5350 Constant *Idx = Context->getConstantInt(Type::Int32Ty, i);
5351 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5352 ExtractElement(In1, Idx, Context),
5353 ExtractElement(In2, Idx, Context),
5360 return HasSubOverflow(cast<ConstantInt>(Result),
5361 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5365 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5366 /// code necessary to compute the offset from the base pointer (without adding
5367 /// in the base pointer). Return the result as a signed integer of intptr size.
5368 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5369 TargetData &TD = IC.getTargetData();
5370 gep_type_iterator GTI = gep_type_begin(GEP);
5371 const Type *IntPtrTy = TD.getIntPtrType();
5372 LLVMContext *Context = IC.getContext();
5373 Value *Result = Context->getNullValue(IntPtrTy);
5375 // Build a mask for high order bits.
5376 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5377 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5379 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5382 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5383 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5384 if (OpC->isZero()) continue;
5386 // Handle a struct index, which adds its field offset to the pointer.
5387 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5388 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5390 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5392 Context->getConstantInt(RC->getValue() + APInt(IntPtrWidth, Size));
5394 Result = IC.InsertNewInstBefore(
5395 BinaryOperator::CreateAdd(Result,
5396 Context->getConstantInt(IntPtrTy, Size),
5397 GEP->getName()+".offs"), I);
5401 Constant *Scale = Context->getConstantInt(IntPtrTy, Size);
5403 Context->getConstantExprIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5404 Scale = Context->getConstantExprMul(OC, Scale);
5405 if (Constant *RC = dyn_cast<Constant>(Result))
5406 Result = Context->getConstantExprAdd(RC, Scale);
5408 // Emit an add instruction.
5409 Result = IC.InsertNewInstBefore(
5410 BinaryOperator::CreateAdd(Result, Scale,
5411 GEP->getName()+".offs"), I);
5415 // Convert to correct type.
5416 if (Op->getType() != IntPtrTy) {
5417 if (Constant *OpC = dyn_cast<Constant>(Op))
5418 Op = Context->getConstantExprIntegerCast(OpC, IntPtrTy, true);
5420 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5422 Op->getName()+".c"), I);
5425 Constant *Scale = Context->getConstantInt(IntPtrTy, Size);
5426 if (Constant *OpC = dyn_cast<Constant>(Op))
5427 Op = Context->getConstantExprMul(OpC, Scale);
5428 else // We'll let instcombine(mul) convert this to a shl if possible.
5429 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5430 GEP->getName()+".idx"), I);
5433 // Emit an add instruction.
5434 if (isa<Constant>(Op) && isa<Constant>(Result))
5435 Result = Context->getConstantExprAdd(cast<Constant>(Op),
5436 cast<Constant>(Result));
5438 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5439 GEP->getName()+".offs"), I);
5445 /// EvaluateGEPOffsetExpression - Return a value that can be used to compare
5446 /// the *offset* implied by a GEP to zero. For example, if we have &A[i], we
5447 /// want to return 'i' for "icmp ne i, 0". Note that, in general, indices can
5448 /// be complex, and scales are involved. The above expression would also be
5449 /// legal to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32).
5450 /// This later form is less amenable to optimization though, and we are allowed
5451 /// to generate the first by knowing that pointer arithmetic doesn't overflow.
5453 /// If we can't emit an optimized form for this expression, this returns null.
5455 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5457 TargetData &TD = IC.getTargetData();
5458 gep_type_iterator GTI = gep_type_begin(GEP);
5460 // Check to see if this gep only has a single variable index. If so, and if
5461 // any constant indices are a multiple of its scale, then we can compute this
5462 // in terms of the scale of the variable index. For example, if the GEP
5463 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5464 // because the expression will cross zero at the same point.
5465 unsigned i, e = GEP->getNumOperands();
5467 for (i = 1; i != e; ++i, ++GTI) {
5468 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5469 // Compute the aggregate offset of constant indices.
5470 if (CI->isZero()) continue;
5472 // Handle a struct index, which adds its field offset to the pointer.
5473 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5474 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5476 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5477 Offset += Size*CI->getSExtValue();
5480 // Found our variable index.
5485 // If there are no variable indices, we must have a constant offset, just
5486 // evaluate it the general way.
5487 if (i == e) return 0;
5489 Value *VariableIdx = GEP->getOperand(i);
5490 // Determine the scale factor of the variable element. For example, this is
5491 // 4 if the variable index is into an array of i32.
5492 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5494 // Verify that there are no other variable indices. If so, emit the hard way.
5495 for (++i, ++GTI; i != e; ++i, ++GTI) {
5496 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5499 // Compute the aggregate offset of constant indices.
5500 if (CI->isZero()) continue;
5502 // Handle a struct index, which adds its field offset to the pointer.
5503 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5504 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5506 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5507 Offset += Size*CI->getSExtValue();
5511 // Okay, we know we have a single variable index, which must be a
5512 // pointer/array/vector index. If there is no offset, life is simple, return
5514 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5516 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5517 // we don't need to bother extending: the extension won't affect where the
5518 // computation crosses zero.
5519 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5520 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5521 VariableIdx->getNameStart(), &I);
5525 // Otherwise, there is an index. The computation we will do will be modulo
5526 // the pointer size, so get it.
5527 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5529 Offset &= PtrSizeMask;
5530 VariableScale &= PtrSizeMask;
5532 // To do this transformation, any constant index must be a multiple of the
5533 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5534 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5535 // multiple of the variable scale.
5536 int64_t NewOffs = Offset / (int64_t)VariableScale;
5537 if (Offset != NewOffs*(int64_t)VariableScale)
5540 // Okay, we can do this evaluation. Start by converting the index to intptr.
5541 const Type *IntPtrTy = TD.getIntPtrType();
5542 if (VariableIdx->getType() != IntPtrTy)
5543 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5545 VariableIdx->getNameStart(), &I);
5546 Constant *OffsetVal = IC.getContext()->getConstantInt(IntPtrTy, NewOffs);
5547 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5551 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5552 /// else. At this point we know that the GEP is on the LHS of the comparison.
5553 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5554 ICmpInst::Predicate Cond,
5556 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5558 // Look through bitcasts.
5559 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5560 RHS = BCI->getOperand(0);
5562 Value *PtrBase = GEPLHS->getOperand(0);
5563 if (PtrBase == RHS) {
5564 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5565 // This transformation (ignoring the base and scales) is valid because we
5566 // know pointers can't overflow. See if we can output an optimized form.
5567 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5569 // If not, synthesize the offset the hard way.
5571 Offset = EmitGEPOffset(GEPLHS, I, *this);
5572 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), Offset,
5573 Context->getNullValue(Offset->getType()));
5574 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5575 // If the base pointers are different, but the indices are the same, just
5576 // compare the base pointer.
5577 if (PtrBase != GEPRHS->getOperand(0)) {
5578 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5579 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5580 GEPRHS->getOperand(0)->getType();
5582 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5583 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5584 IndicesTheSame = false;
5588 // If all indices are the same, just compare the base pointers.
5590 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond),
5591 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5593 // Otherwise, the base pointers are different and the indices are
5594 // different, bail out.
5598 // If one of the GEPs has all zero indices, recurse.
5599 bool AllZeros = true;
5600 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5601 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5602 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5607 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5608 ICmpInst::getSwappedPredicate(Cond), I);
5610 // If the other GEP has all zero indices, recurse.
5612 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5613 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5614 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5619 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5621 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5622 // If the GEPs only differ by one index, compare it.
5623 unsigned NumDifferences = 0; // Keep track of # differences.
5624 unsigned DiffOperand = 0; // The operand that differs.
5625 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5626 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5627 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5628 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5629 // Irreconcilable differences.
5633 if (NumDifferences++) break;
5638 if (NumDifferences == 0) // SAME GEP?
5639 return ReplaceInstUsesWith(I, // No comparison is needed here.
5640 Context->getConstantInt(Type::Int1Ty,
5641 ICmpInst::isTrueWhenEqual(Cond)));
5643 else if (NumDifferences == 1) {
5644 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5645 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5646 // Make sure we do a signed comparison here.
5647 return new ICmpInst(*Context,
5648 ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5652 // Only lower this if the icmp is the only user of the GEP or if we expect
5653 // the result to fold to a constant!
5654 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5655 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5656 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5657 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5658 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5659 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), L, R);
5665 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5667 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5670 if (!isa<ConstantFP>(RHSC)) return 0;
5671 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5673 // Get the width of the mantissa. We don't want to hack on conversions that
5674 // might lose information from the integer, e.g. "i64 -> float"
5675 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5676 if (MantissaWidth == -1) return 0; // Unknown.
5678 // Check to see that the input is converted from an integer type that is small
5679 // enough that preserves all bits. TODO: check here for "known" sign bits.
5680 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5681 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5683 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5684 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5688 // If the conversion would lose info, don't hack on this.
5689 if ((int)InputSize > MantissaWidth)
5692 // Otherwise, we can potentially simplify the comparison. We know that it
5693 // will always come through as an integer value and we know the constant is
5694 // not a NAN (it would have been previously simplified).
5695 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5697 ICmpInst::Predicate Pred;
5698 switch (I.getPredicate()) {
5699 default: llvm_unreachable("Unexpected predicate!");
5700 case FCmpInst::FCMP_UEQ:
5701 case FCmpInst::FCMP_OEQ:
5702 Pred = ICmpInst::ICMP_EQ;
5704 case FCmpInst::FCMP_UGT:
5705 case FCmpInst::FCMP_OGT:
5706 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5708 case FCmpInst::FCMP_UGE:
5709 case FCmpInst::FCMP_OGE:
5710 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5712 case FCmpInst::FCMP_ULT:
5713 case FCmpInst::FCMP_OLT:
5714 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5716 case FCmpInst::FCMP_ULE:
5717 case FCmpInst::FCMP_OLE:
5718 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5720 case FCmpInst::FCMP_UNE:
5721 case FCmpInst::FCMP_ONE:
5722 Pred = ICmpInst::ICMP_NE;
5724 case FCmpInst::FCMP_ORD:
5725 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5726 case FCmpInst::FCMP_UNO:
5727 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5730 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5732 // Now we know that the APFloat is a normal number, zero or inf.
5734 // See if the FP constant is too large for the integer. For example,
5735 // comparing an i8 to 300.0.
5736 unsigned IntWidth = IntTy->getScalarSizeInBits();
5739 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5740 // and large values.
5741 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5742 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5743 APFloat::rmNearestTiesToEven);
5744 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5745 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5746 Pred == ICmpInst::ICMP_SLE)
5747 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5748 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5751 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5752 // +INF and large values.
5753 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5754 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5755 APFloat::rmNearestTiesToEven);
5756 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5757 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5758 Pred == ICmpInst::ICMP_ULE)
5759 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5760 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5765 // See if the RHS value is < SignedMin.
5766 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5767 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5768 APFloat::rmNearestTiesToEven);
5769 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5770 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5771 Pred == ICmpInst::ICMP_SGE)
5772 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5773 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5777 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5778 // [0, UMAX], but it may still be fractional. See if it is fractional by
5779 // casting the FP value to the integer value and back, checking for equality.
5780 // Don't do this for zero, because -0.0 is not fractional.
5781 Constant *RHSInt = LHSUnsigned
5782 ? Context->getConstantExprFPToUI(RHSC, IntTy)
5783 : Context->getConstantExprFPToSI(RHSC, IntTy);
5784 if (!RHS.isZero()) {
5785 bool Equal = LHSUnsigned
5786 ? Context->getConstantExprUIToFP(RHSInt, RHSC->getType()) == RHSC
5787 : Context->getConstantExprSIToFP(RHSInt, RHSC->getType()) == RHSC;
5789 // If we had a comparison against a fractional value, we have to adjust
5790 // the compare predicate and sometimes the value. RHSC is rounded towards
5791 // zero at this point.
5793 default: llvm_unreachable("Unexpected integer comparison!");
5794 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5795 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5796 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5797 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5798 case ICmpInst::ICMP_ULE:
5799 // (float)int <= 4.4 --> int <= 4
5800 // (float)int <= -4.4 --> false
5801 if (RHS.isNegative())
5802 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5804 case ICmpInst::ICMP_SLE:
5805 // (float)int <= 4.4 --> int <= 4
5806 // (float)int <= -4.4 --> int < -4
5807 if (RHS.isNegative())
5808 Pred = ICmpInst::ICMP_SLT;
5810 case ICmpInst::ICMP_ULT:
5811 // (float)int < -4.4 --> false
5812 // (float)int < 4.4 --> int <= 4
5813 if (RHS.isNegative())
5814 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5815 Pred = ICmpInst::ICMP_ULE;
5817 case ICmpInst::ICMP_SLT:
5818 // (float)int < -4.4 --> int < -4
5819 // (float)int < 4.4 --> int <= 4
5820 if (!RHS.isNegative())
5821 Pred = ICmpInst::ICMP_SLE;
5823 case ICmpInst::ICMP_UGT:
5824 // (float)int > 4.4 --> int > 4
5825 // (float)int > -4.4 --> true
5826 if (RHS.isNegative())
5827 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5829 case ICmpInst::ICMP_SGT:
5830 // (float)int > 4.4 --> int > 4
5831 // (float)int > -4.4 --> int >= -4
5832 if (RHS.isNegative())
5833 Pred = ICmpInst::ICMP_SGE;
5835 case ICmpInst::ICMP_UGE:
5836 // (float)int >= -4.4 --> true
5837 // (float)int >= 4.4 --> int > 4
5838 if (!RHS.isNegative())
5839 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5840 Pred = ICmpInst::ICMP_UGT;
5842 case ICmpInst::ICMP_SGE:
5843 // (float)int >= -4.4 --> int >= -4
5844 // (float)int >= 4.4 --> int > 4
5845 if (!RHS.isNegative())
5846 Pred = ICmpInst::ICMP_SGT;
5852 // Lower this FP comparison into an appropriate integer version of the
5854 return new ICmpInst(*Context, Pred, LHSI->getOperand(0), RHSInt);
5857 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5858 bool Changed = SimplifyCompare(I);
5859 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5861 // Fold trivial predicates.
5862 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5863 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5864 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5865 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5867 // Simplify 'fcmp pred X, X'
5869 switch (I.getPredicate()) {
5870 default: llvm_unreachable("Unknown predicate!");
5871 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5872 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5873 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5874 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5875 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5876 case FCmpInst::FCMP_OLT: // True if ordered and less than
5877 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5878 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5880 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5881 case FCmpInst::FCMP_ULT: // True if unordered or less than
5882 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5883 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5884 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5885 I.setPredicate(FCmpInst::FCMP_UNO);
5886 I.setOperand(1, Context->getNullValue(Op0->getType()));
5889 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5890 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5891 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5892 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5893 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5894 I.setPredicate(FCmpInst::FCMP_ORD);
5895 I.setOperand(1, Context->getNullValue(Op0->getType()));
5900 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5901 return ReplaceInstUsesWith(I, Context->getUndef(Type::Int1Ty));
5903 // Handle fcmp with constant RHS
5904 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5905 // If the constant is a nan, see if we can fold the comparison based on it.
5906 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5907 if (CFP->getValueAPF().isNaN()) {
5908 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5909 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5910 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5911 "Comparison must be either ordered or unordered!");
5912 // True if unordered.
5913 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5917 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5918 switch (LHSI->getOpcode()) {
5919 case Instruction::PHI:
5920 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5921 // block. If in the same block, we're encouraging jump threading. If
5922 // not, we are just pessimizing the code by making an i1 phi.
5923 if (LHSI->getParent() == I.getParent())
5924 if (Instruction *NV = FoldOpIntoPhi(I))
5927 case Instruction::SIToFP:
5928 case Instruction::UIToFP:
5929 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5932 case Instruction::Select:
5933 // If either operand of the select is a constant, we can fold the
5934 // comparison into the select arms, which will cause one to be
5935 // constant folded and the select turned into a bitwise or.
5936 Value *Op1 = 0, *Op2 = 0;
5937 if (LHSI->hasOneUse()) {
5938 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5939 // Fold the known value into the constant operand.
5940 Op1 = Context->getConstantExprCompare(I.getPredicate(), C, RHSC);
5941 // Insert a new FCmp of the other select operand.
5942 Op2 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5943 LHSI->getOperand(2), RHSC,
5945 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5946 // Fold the known value into the constant operand.
5947 Op2 = Context->getConstantExprCompare(I.getPredicate(), C, RHSC);
5948 // Insert a new FCmp of the other select operand.
5949 Op1 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5950 LHSI->getOperand(1), RHSC,
5956 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5961 return Changed ? &I : 0;
5964 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5965 bool Changed = SimplifyCompare(I);
5966 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5967 const Type *Ty = Op0->getType();
5971 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
5972 I.isTrueWhenEqual()));
5974 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5975 return ReplaceInstUsesWith(I, Context->getUndef(Type::Int1Ty));
5977 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5978 // addresses never equal each other! We already know that Op0 != Op1.
5979 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5980 isa<ConstantPointerNull>(Op0)) &&
5981 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5982 isa<ConstantPointerNull>(Op1)))
5983 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
5984 !I.isTrueWhenEqual()));
5986 // icmp's with boolean values can always be turned into bitwise operations
5987 if (Ty == Type::Int1Ty) {
5988 switch (I.getPredicate()) {
5989 default: llvm_unreachable("Invalid icmp instruction!");
5990 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5991 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
5992 InsertNewInstBefore(Xor, I);
5993 return BinaryOperator::CreateNot(*Context, Xor);
5995 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
5996 return BinaryOperator::CreateXor(Op0, Op1);
5998 case ICmpInst::ICMP_UGT:
5999 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6001 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6002 Instruction *Not = BinaryOperator::CreateNot(*Context,
6003 Op0, I.getName()+"tmp");
6004 InsertNewInstBefore(Not, I);
6005 return BinaryOperator::CreateAnd(Not, Op1);
6007 case ICmpInst::ICMP_SGT:
6008 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6010 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6011 Instruction *Not = BinaryOperator::CreateNot(*Context,
6012 Op1, I.getName()+"tmp");
6013 InsertNewInstBefore(Not, I);
6014 return BinaryOperator::CreateAnd(Not, Op0);
6016 case ICmpInst::ICMP_UGE:
6017 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6019 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6020 Instruction *Not = BinaryOperator::CreateNot(*Context,
6021 Op0, I.getName()+"tmp");
6022 InsertNewInstBefore(Not, I);
6023 return BinaryOperator::CreateOr(Not, Op1);
6025 case ICmpInst::ICMP_SGE:
6026 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6028 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6029 Instruction *Not = BinaryOperator::CreateNot(*Context,
6030 Op1, I.getName()+"tmp");
6031 InsertNewInstBefore(Not, I);
6032 return BinaryOperator::CreateOr(Not, Op0);
6037 unsigned BitWidth = 0;
6039 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6040 else if (Ty->isIntOrIntVector())
6041 BitWidth = Ty->getScalarSizeInBits();
6043 bool isSignBit = false;
6045 // See if we are doing a comparison with a constant.
6046 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6047 Value *A = 0, *B = 0;
6049 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6050 if (I.isEquality() && CI->isNullValue() &&
6051 match(Op0, m_Sub(m_Value(A), m_Value(B)), *Context)) {
6052 // (icmp cond A B) if cond is equality
6053 return new ICmpInst(*Context, I.getPredicate(), A, B);
6056 // If we have an icmp le or icmp ge instruction, turn it into the
6057 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6058 // them being folded in the code below.
6059 switch (I.getPredicate()) {
6061 case ICmpInst::ICMP_ULE:
6062 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6063 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6064 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Op0,
6065 AddOne(CI, Context));
6066 case ICmpInst::ICMP_SLE:
6067 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6068 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6069 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6070 AddOne(CI, Context));
6071 case ICmpInst::ICMP_UGE:
6072 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6073 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6074 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Op0,
6075 SubOne(CI, Context));
6076 case ICmpInst::ICMP_SGE:
6077 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6078 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6079 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6080 SubOne(CI, Context));
6083 // If this comparison is a normal comparison, it demands all
6084 // bits, if it is a sign bit comparison, it only demands the sign bit.
6086 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6089 // See if we can fold the comparison based on range information we can get
6090 // by checking whether bits are known to be zero or one in the input.
6091 if (BitWidth != 0) {
6092 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6093 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6095 if (SimplifyDemandedBits(I.getOperandUse(0),
6096 isSignBit ? APInt::getSignBit(BitWidth)
6097 : APInt::getAllOnesValue(BitWidth),
6098 Op0KnownZero, Op0KnownOne, 0))
6100 if (SimplifyDemandedBits(I.getOperandUse(1),
6101 APInt::getAllOnesValue(BitWidth),
6102 Op1KnownZero, Op1KnownOne, 0))
6105 // Given the known and unknown bits, compute a range that the LHS could be
6106 // in. Compute the Min, Max and RHS values based on the known bits. For the
6107 // EQ and NE we use unsigned values.
6108 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6109 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6110 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6111 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6113 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6116 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6118 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6122 // If Min and Max are known to be the same, then SimplifyDemandedBits
6123 // figured out that the LHS is a constant. Just constant fold this now so
6124 // that code below can assume that Min != Max.
6125 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6126 return new ICmpInst(*Context, I.getPredicate(),
6127 Context->getConstantInt(Op0Min), Op1);
6128 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6129 return new ICmpInst(*Context, I.getPredicate(), Op0,
6130 Context->getConstantInt(Op1Min));
6132 // Based on the range information we know about the LHS, see if we can
6133 // simplify this comparison. For example, (x&4) < 8 is always true.
6134 switch (I.getPredicate()) {
6135 default: llvm_unreachable("Unknown icmp opcode!");
6136 case ICmpInst::ICMP_EQ:
6137 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6138 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6140 case ICmpInst::ICMP_NE:
6141 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6142 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6144 case ICmpInst::ICMP_ULT:
6145 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6146 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6147 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6148 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6149 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6150 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6151 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6152 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6153 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6154 SubOne(CI, Context));
6156 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6157 if (CI->isMinValue(true))
6158 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6159 Context->getAllOnesValue(Op0->getType()));
6162 case ICmpInst::ICMP_UGT:
6163 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6164 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6165 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6166 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6168 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6169 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6170 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6171 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6172 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6173 AddOne(CI, Context));
6175 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6176 if (CI->isMaxValue(true))
6177 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6178 Context->getNullValue(Op0->getType()));
6181 case ICmpInst::ICMP_SLT:
6182 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6183 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6184 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6185 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6186 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6187 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6188 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6189 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6190 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6191 SubOne(CI, Context));
6194 case ICmpInst::ICMP_SGT:
6195 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6196 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6197 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6198 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6200 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6201 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6202 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6203 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6204 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6205 AddOne(CI, Context));
6208 case ICmpInst::ICMP_SGE:
6209 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6210 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6211 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6212 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6213 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6215 case ICmpInst::ICMP_SLE:
6216 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6217 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6218 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6219 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6220 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6222 case ICmpInst::ICMP_UGE:
6223 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6224 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6225 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6226 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6227 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6229 case ICmpInst::ICMP_ULE:
6230 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6231 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6232 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6233 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6234 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6238 // Turn a signed comparison into an unsigned one if both operands
6239 // are known to have the same sign.
6240 if (I.isSignedPredicate() &&
6241 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6242 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6243 return new ICmpInst(*Context, I.getUnsignedPredicate(), Op0, Op1);
6246 // Test if the ICmpInst instruction is used exclusively by a select as
6247 // part of a minimum or maximum operation. If so, refrain from doing
6248 // any other folding. This helps out other analyses which understand
6249 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6250 // and CodeGen. And in this case, at least one of the comparison
6251 // operands has at least one user besides the compare (the select),
6252 // which would often largely negate the benefit of folding anyway.
6254 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6255 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6256 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6259 // See if we are doing a comparison between a constant and an instruction that
6260 // can be folded into the comparison.
6261 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6262 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6263 // instruction, see if that instruction also has constants so that the
6264 // instruction can be folded into the icmp
6265 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6266 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6270 // Handle icmp with constant (but not simple integer constant) RHS
6271 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6272 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6273 switch (LHSI->getOpcode()) {
6274 case Instruction::GetElementPtr:
6275 if (RHSC->isNullValue()) {
6276 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6277 bool isAllZeros = true;
6278 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6279 if (!isa<Constant>(LHSI->getOperand(i)) ||
6280 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6285 return new ICmpInst(*Context, I.getPredicate(), LHSI->getOperand(0),
6286 Context->getNullValue(LHSI->getOperand(0)->getType()));
6290 case Instruction::PHI:
6291 // Only fold icmp into the PHI if the phi and fcmp are in the same
6292 // block. If in the same block, we're encouraging jump threading. If
6293 // not, we are just pessimizing the code by making an i1 phi.
6294 if (LHSI->getParent() == I.getParent())
6295 if (Instruction *NV = FoldOpIntoPhi(I))
6298 case Instruction::Select: {
6299 // If either operand of the select is a constant, we can fold the
6300 // comparison into the select arms, which will cause one to be
6301 // constant folded and the select turned into a bitwise or.
6302 Value *Op1 = 0, *Op2 = 0;
6303 if (LHSI->hasOneUse()) {
6304 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6305 // Fold the known value into the constant operand.
6306 Op1 = Context->getConstantExprICmp(I.getPredicate(), C, RHSC);
6307 // Insert a new ICmp of the other select operand.
6308 Op2 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6309 LHSI->getOperand(2), RHSC,
6311 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6312 // Fold the known value into the constant operand.
6313 Op2 = Context->getConstantExprICmp(I.getPredicate(), C, RHSC);
6314 // Insert a new ICmp of the other select operand.
6315 Op1 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6316 LHSI->getOperand(1), RHSC,
6322 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6325 case Instruction::Malloc:
6326 // If we have (malloc != null), and if the malloc has a single use, we
6327 // can assume it is successful and remove the malloc.
6328 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6329 AddToWorkList(LHSI);
6330 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
6331 !I.isTrueWhenEqual()));
6337 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6338 if (User *GEP = dyn_castGetElementPtr(Op0))
6339 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6341 if (User *GEP = dyn_castGetElementPtr(Op1))
6342 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6343 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6346 // Test to see if the operands of the icmp are casted versions of other
6347 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6349 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6350 if (isa<PointerType>(Op0->getType()) &&
6351 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6352 // We keep moving the cast from the left operand over to the right
6353 // operand, where it can often be eliminated completely.
6354 Op0 = CI->getOperand(0);
6356 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6357 // so eliminate it as well.
6358 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6359 Op1 = CI2->getOperand(0);
6361 // If Op1 is a constant, we can fold the cast into the constant.
6362 if (Op0->getType() != Op1->getType()) {
6363 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6364 Op1 = Context->getConstantExprBitCast(Op1C, Op0->getType());
6366 // Otherwise, cast the RHS right before the icmp
6367 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6370 return new ICmpInst(*Context, I.getPredicate(), Op0, Op1);
6374 if (isa<CastInst>(Op0)) {
6375 // Handle the special case of: icmp (cast bool to X), <cst>
6376 // This comes up when you have code like
6379 // For generality, we handle any zero-extension of any operand comparison
6380 // with a constant or another cast from the same type.
6381 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6382 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6386 // See if it's the same type of instruction on the left and right.
6387 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6388 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6389 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6390 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6391 switch (Op0I->getOpcode()) {
6393 case Instruction::Add:
6394 case Instruction::Sub:
6395 case Instruction::Xor:
6396 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6397 return new ICmpInst(*Context, I.getPredicate(), Op0I->getOperand(0),
6398 Op1I->getOperand(0));
6399 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6400 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6401 if (CI->getValue().isSignBit()) {
6402 ICmpInst::Predicate Pred = I.isSignedPredicate()
6403 ? I.getUnsignedPredicate()
6404 : I.getSignedPredicate();
6405 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6406 Op1I->getOperand(0));
6409 if (CI->getValue().isMaxSignedValue()) {
6410 ICmpInst::Predicate Pred = I.isSignedPredicate()
6411 ? I.getUnsignedPredicate()
6412 : I.getSignedPredicate();
6413 Pred = I.getSwappedPredicate(Pred);
6414 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6415 Op1I->getOperand(0));
6419 case Instruction::Mul:
6420 if (!I.isEquality())
6423 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6424 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6425 // Mask = -1 >> count-trailing-zeros(Cst).
6426 if (!CI->isZero() && !CI->isOne()) {
6427 const APInt &AP = CI->getValue();
6428 ConstantInt *Mask = Context->getConstantInt(
6429 APInt::getLowBitsSet(AP.getBitWidth(),
6431 AP.countTrailingZeros()));
6432 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6434 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6436 InsertNewInstBefore(And1, I);
6437 InsertNewInstBefore(And2, I);
6438 return new ICmpInst(*Context, I.getPredicate(), And1, And2);
6447 // ~x < ~y --> y < x
6449 if (match(Op0, m_Not(m_Value(A)), *Context) &&
6450 match(Op1, m_Not(m_Value(B)), *Context))
6451 return new ICmpInst(*Context, I.getPredicate(), B, A);
6454 if (I.isEquality()) {
6455 Value *A, *B, *C, *D;
6457 // -x == -y --> x == y
6458 if (match(Op0, m_Neg(m_Value(A)), *Context) &&
6459 match(Op1, m_Neg(m_Value(B)), *Context))
6460 return new ICmpInst(*Context, I.getPredicate(), A, B);
6462 if (match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
6463 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6464 Value *OtherVal = A == Op1 ? B : A;
6465 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6466 Context->getNullValue(A->getType()));
6469 if (match(Op1, m_Xor(m_Value(C), m_Value(D)), *Context)) {
6470 // A^c1 == C^c2 --> A == C^(c1^c2)
6471 ConstantInt *C1, *C2;
6472 if (match(B, m_ConstantInt(C1), *Context) &&
6473 match(D, m_ConstantInt(C2), *Context) && Op1->hasOneUse()) {
6475 Context->getConstantInt(C1->getValue() ^ C2->getValue());
6476 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6477 return new ICmpInst(*Context, I.getPredicate(), A,
6478 InsertNewInstBefore(Xor, I));
6481 // A^B == A^D -> B == D
6482 if (A == C) return new ICmpInst(*Context, I.getPredicate(), B, D);
6483 if (A == D) return new ICmpInst(*Context, I.getPredicate(), B, C);
6484 if (B == C) return new ICmpInst(*Context, I.getPredicate(), A, D);
6485 if (B == D) return new ICmpInst(*Context, I.getPredicate(), A, C);
6489 if (match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context) &&
6490 (A == Op0 || B == Op0)) {
6491 // A == (A^B) -> B == 0
6492 Value *OtherVal = A == Op0 ? B : A;
6493 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6494 Context->getNullValue(A->getType()));
6497 // (A-B) == A -> B == 0
6498 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B)), *Context))
6499 return new ICmpInst(*Context, I.getPredicate(), B,
6500 Context->getNullValue(B->getType()));
6502 // A == (A-B) -> B == 0
6503 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B)), *Context))
6504 return new ICmpInst(*Context, I.getPredicate(), B,
6505 Context->getNullValue(B->getType()));
6507 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6508 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6509 match(Op0, m_And(m_Value(A), m_Value(B)), *Context) &&
6510 match(Op1, m_And(m_Value(C), m_Value(D)), *Context)) {
6511 Value *X = 0, *Y = 0, *Z = 0;
6514 X = B; Y = D; Z = A;
6515 } else if (A == D) {
6516 X = B; Y = C; Z = A;
6517 } else if (B == C) {
6518 X = A; Y = D; Z = B;
6519 } else if (B == D) {
6520 X = A; Y = C; Z = B;
6523 if (X) { // Build (X^Y) & Z
6524 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6525 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6526 I.setOperand(0, Op1);
6527 I.setOperand(1, Context->getNullValue(Op1->getType()));
6532 return Changed ? &I : 0;
6536 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6537 /// and CmpRHS are both known to be integer constants.
6538 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6539 ConstantInt *DivRHS) {
6540 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6541 const APInt &CmpRHSV = CmpRHS->getValue();
6543 // FIXME: If the operand types don't match the type of the divide
6544 // then don't attempt this transform. The code below doesn't have the
6545 // logic to deal with a signed divide and an unsigned compare (and
6546 // vice versa). This is because (x /s C1) <s C2 produces different
6547 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6548 // (x /u C1) <u C2. Simply casting the operands and result won't
6549 // work. :( The if statement below tests that condition and bails
6551 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6552 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6554 if (DivRHS->isZero())
6555 return 0; // The ProdOV computation fails on divide by zero.
6556 if (DivIsSigned && DivRHS->isAllOnesValue())
6557 return 0; // The overflow computation also screws up here
6558 if (DivRHS->isOne())
6559 return 0; // Not worth bothering, and eliminates some funny cases
6562 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6563 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6564 // C2 (CI). By solving for X we can turn this into a range check
6565 // instead of computing a divide.
6566 Constant *Prod = Context->getConstantExprMul(CmpRHS, DivRHS);
6568 // Determine if the product overflows by seeing if the product is
6569 // not equal to the divide. Make sure we do the same kind of divide
6570 // as in the LHS instruction that we're folding.
6571 bool ProdOV = (DivIsSigned ? Context->getConstantExprSDiv(Prod, DivRHS) :
6572 Context->getConstantExprUDiv(Prod, DivRHS)) != CmpRHS;
6574 // Get the ICmp opcode
6575 ICmpInst::Predicate Pred = ICI.getPredicate();
6577 // Figure out the interval that is being checked. For example, a comparison
6578 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6579 // Compute this interval based on the constants involved and the signedness of
6580 // the compare/divide. This computes a half-open interval, keeping track of
6581 // whether either value in the interval overflows. After analysis each
6582 // overflow variable is set to 0 if it's corresponding bound variable is valid
6583 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6584 int LoOverflow = 0, HiOverflow = 0;
6585 Constant *LoBound = 0, *HiBound = 0;
6587 if (!DivIsSigned) { // udiv
6588 // e.g. X/5 op 3 --> [15, 20)
6590 HiOverflow = LoOverflow = ProdOV;
6592 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6593 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6594 if (CmpRHSV == 0) { // (X / pos) op 0
6595 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6596 LoBound = cast<ConstantInt>(Context->getConstantExprNeg(SubOne(DivRHS,
6599 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6600 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6601 HiOverflow = LoOverflow = ProdOV;
6603 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6604 } else { // (X / pos) op neg
6605 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6606 HiBound = AddOne(Prod, Context);
6607 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6609 ConstantInt* DivNeg =
6610 cast<ConstantInt>(Context->getConstantExprNeg(DivRHS));
6611 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6615 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6616 if (CmpRHSV == 0) { // (X / neg) op 0
6617 // e.g. X/-5 op 0 --> [-4, 5)
6618 LoBound = AddOne(DivRHS, Context);
6619 HiBound = cast<ConstantInt>(Context->getConstantExprNeg(DivRHS));
6620 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6621 HiOverflow = 1; // [INTMIN+1, overflow)
6622 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6624 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6625 // e.g. X/-5 op 3 --> [-19, -14)
6626 HiBound = AddOne(Prod, Context);
6627 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6629 LoOverflow = AddWithOverflow(LoBound, HiBound,
6630 DivRHS, Context, true) ? -1 : 0;
6631 } else { // (X / neg) op neg
6632 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6633 LoOverflow = HiOverflow = ProdOV;
6635 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6638 // Dividing by a negative swaps the condition. LT <-> GT
6639 Pred = ICmpInst::getSwappedPredicate(Pred);
6642 Value *X = DivI->getOperand(0);
6644 default: llvm_unreachable("Unhandled icmp opcode!");
6645 case ICmpInst::ICMP_EQ:
6646 if (LoOverflow && HiOverflow)
6647 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6648 else if (HiOverflow)
6649 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6650 ICmpInst::ICMP_UGE, X, LoBound);
6651 else if (LoOverflow)
6652 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6653 ICmpInst::ICMP_ULT, X, HiBound);
6655 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6656 case ICmpInst::ICMP_NE:
6657 if (LoOverflow && HiOverflow)
6658 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6659 else if (HiOverflow)
6660 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6661 ICmpInst::ICMP_ULT, X, LoBound);
6662 else if (LoOverflow)
6663 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6664 ICmpInst::ICMP_UGE, X, HiBound);
6666 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6667 case ICmpInst::ICMP_ULT:
6668 case ICmpInst::ICMP_SLT:
6669 if (LoOverflow == +1) // Low bound is greater than input range.
6670 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6671 if (LoOverflow == -1) // Low bound is less than input range.
6672 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6673 return new ICmpInst(*Context, Pred, X, LoBound);
6674 case ICmpInst::ICMP_UGT:
6675 case ICmpInst::ICMP_SGT:
6676 if (HiOverflow == +1) // High bound greater than input range.
6677 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6678 else if (HiOverflow == -1) // High bound less than input range.
6679 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6680 if (Pred == ICmpInst::ICMP_UGT)
6681 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, X, HiBound);
6683 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, X, HiBound);
6688 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6690 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6693 const APInt &RHSV = RHS->getValue();
6695 switch (LHSI->getOpcode()) {
6696 case Instruction::Trunc:
6697 if (ICI.isEquality() && LHSI->hasOneUse()) {
6698 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6699 // of the high bits truncated out of x are known.
6700 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6701 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6702 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6703 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6704 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6706 // If all the high bits are known, we can do this xform.
6707 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6708 // Pull in the high bits from known-ones set.
6709 APInt NewRHS(RHS->getValue());
6710 NewRHS.zext(SrcBits);
6712 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6713 Context->getConstantInt(NewRHS));
6718 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6719 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6720 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6722 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6723 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6724 Value *CompareVal = LHSI->getOperand(0);
6726 // If the sign bit of the XorCST is not set, there is no change to
6727 // the operation, just stop using the Xor.
6728 if (!XorCST->getValue().isNegative()) {
6729 ICI.setOperand(0, CompareVal);
6730 AddToWorkList(LHSI);
6734 // Was the old condition true if the operand is positive?
6735 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6737 // If so, the new one isn't.
6738 isTrueIfPositive ^= true;
6740 if (isTrueIfPositive)
6741 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, CompareVal,
6742 SubOne(RHS, Context));
6744 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, CompareVal,
6745 AddOne(RHS, Context));
6748 if (LHSI->hasOneUse()) {
6749 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6750 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6751 const APInt &SignBit = XorCST->getValue();
6752 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6753 ? ICI.getUnsignedPredicate()
6754 : ICI.getSignedPredicate();
6755 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6756 Context->getConstantInt(RHSV ^ SignBit));
6759 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6760 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6761 const APInt &NotSignBit = XorCST->getValue();
6762 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6763 ? ICI.getUnsignedPredicate()
6764 : ICI.getSignedPredicate();
6765 Pred = ICI.getSwappedPredicate(Pred);
6766 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6767 Context->getConstantInt(RHSV ^ NotSignBit));
6772 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6773 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6774 LHSI->getOperand(0)->hasOneUse()) {
6775 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6777 // If the LHS is an AND of a truncating cast, we can widen the
6778 // and/compare to be the input width without changing the value
6779 // produced, eliminating a cast.
6780 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6781 // We can do this transformation if either the AND constant does not
6782 // have its sign bit set or if it is an equality comparison.
6783 // Extending a relational comparison when we're checking the sign
6784 // bit would not work.
6785 if (Cast->hasOneUse() &&
6786 (ICI.isEquality() ||
6787 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6789 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6790 APInt NewCST = AndCST->getValue();
6791 NewCST.zext(BitWidth);
6793 NewCI.zext(BitWidth);
6794 Instruction *NewAnd =
6795 BinaryOperator::CreateAnd(Cast->getOperand(0),
6796 Context->getConstantInt(NewCST),LHSI->getName());
6797 InsertNewInstBefore(NewAnd, ICI);
6798 return new ICmpInst(*Context, ICI.getPredicate(), NewAnd,
6799 Context->getConstantInt(NewCI));
6803 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6804 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6805 // happens a LOT in code produced by the C front-end, for bitfield
6807 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6808 if (Shift && !Shift->isShift())
6812 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6813 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6814 const Type *AndTy = AndCST->getType(); // Type of the and.
6816 // We can fold this as long as we can't shift unknown bits
6817 // into the mask. This can only happen with signed shift
6818 // rights, as they sign-extend.
6820 bool CanFold = Shift->isLogicalShift();
6822 // To test for the bad case of the signed shr, see if any
6823 // of the bits shifted in could be tested after the mask.
6824 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6825 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6827 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6828 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6829 AndCST->getValue()) == 0)
6835 if (Shift->getOpcode() == Instruction::Shl)
6836 NewCst = Context->getConstantExprLShr(RHS, ShAmt);
6838 NewCst = Context->getConstantExprShl(RHS, ShAmt);
6840 // Check to see if we are shifting out any of the bits being
6842 if (Context->getConstantExpr(Shift->getOpcode(),
6843 NewCst, ShAmt) != RHS) {
6844 // If we shifted bits out, the fold is not going to work out.
6845 // As a special case, check to see if this means that the
6846 // result is always true or false now.
6847 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6848 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6849 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6850 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6852 ICI.setOperand(1, NewCst);
6853 Constant *NewAndCST;
6854 if (Shift->getOpcode() == Instruction::Shl)
6855 NewAndCST = Context->getConstantExprLShr(AndCST, ShAmt);
6857 NewAndCST = Context->getConstantExprShl(AndCST, ShAmt);
6858 LHSI->setOperand(1, NewAndCST);
6859 LHSI->setOperand(0, Shift->getOperand(0));
6860 AddToWorkList(Shift); // Shift is dead.
6861 AddUsesToWorkList(ICI);
6867 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6868 // preferable because it allows the C<<Y expression to be hoisted out
6869 // of a loop if Y is invariant and X is not.
6870 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6871 ICI.isEquality() && !Shift->isArithmeticShift() &&
6872 !isa<Constant>(Shift->getOperand(0))) {
6875 if (Shift->getOpcode() == Instruction::LShr) {
6876 NS = BinaryOperator::CreateShl(AndCST,
6877 Shift->getOperand(1), "tmp");
6879 // Insert a logical shift.
6880 NS = BinaryOperator::CreateLShr(AndCST,
6881 Shift->getOperand(1), "tmp");
6883 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6885 // Compute X & (C << Y).
6886 Instruction *NewAnd =
6887 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6888 InsertNewInstBefore(NewAnd, ICI);
6890 ICI.setOperand(0, NewAnd);
6896 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6897 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6900 uint32_t TypeBits = RHSV.getBitWidth();
6902 // Check that the shift amount is in range. If not, don't perform
6903 // undefined shifts. When the shift is visited it will be
6905 if (ShAmt->uge(TypeBits))
6908 if (ICI.isEquality()) {
6909 // If we are comparing against bits always shifted out, the
6910 // comparison cannot succeed.
6912 Context->getConstantExprShl(Context->getConstantExprLShr(RHS, ShAmt),
6914 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6915 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6916 Constant *Cst = Context->getConstantInt(Type::Int1Ty, IsICMP_NE);
6917 return ReplaceInstUsesWith(ICI, Cst);
6920 if (LHSI->hasOneUse()) {
6921 // Otherwise strength reduce the shift into an and.
6922 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6924 Context->getConstantInt(APInt::getLowBitsSet(TypeBits,
6925 TypeBits-ShAmtVal));
6928 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6929 Mask, LHSI->getName()+".mask");
6930 Value *And = InsertNewInstBefore(AndI, ICI);
6931 return new ICmpInst(*Context, ICI.getPredicate(), And,
6932 Context->getConstantInt(RHSV.lshr(ShAmtVal)));
6936 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6937 bool TrueIfSigned = false;
6938 if (LHSI->hasOneUse() &&
6939 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6940 // (X << 31) <s 0 --> (X&1) != 0
6941 Constant *Mask = Context->getConstantInt(APInt(TypeBits, 1) <<
6942 (TypeBits-ShAmt->getZExtValue()-1));
6944 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6945 Mask, LHSI->getName()+".mask");
6946 Value *And = InsertNewInstBefore(AndI, ICI);
6948 return new ICmpInst(*Context,
6949 TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6950 And, Context->getNullValue(And->getType()));
6955 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6956 case Instruction::AShr: {
6957 // Only handle equality comparisons of shift-by-constant.
6958 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6959 if (!ShAmt || !ICI.isEquality()) break;
6961 // Check that the shift amount is in range. If not, don't perform
6962 // undefined shifts. When the shift is visited it will be
6964 uint32_t TypeBits = RHSV.getBitWidth();
6965 if (ShAmt->uge(TypeBits))
6968 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6970 // If we are comparing against bits always shifted out, the
6971 // comparison cannot succeed.
6972 APInt Comp = RHSV << ShAmtVal;
6973 if (LHSI->getOpcode() == Instruction::LShr)
6974 Comp = Comp.lshr(ShAmtVal);
6976 Comp = Comp.ashr(ShAmtVal);
6978 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6979 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6980 Constant *Cst = Context->getConstantInt(Type::Int1Ty, IsICMP_NE);
6981 return ReplaceInstUsesWith(ICI, Cst);
6984 // Otherwise, check to see if the bits shifted out are known to be zero.
6985 // If so, we can compare against the unshifted value:
6986 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6987 if (LHSI->hasOneUse() &&
6988 MaskedValueIsZero(LHSI->getOperand(0),
6989 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6990 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6991 Context->getConstantExprShl(RHS, ShAmt));
6994 if (LHSI->hasOneUse()) {
6995 // Otherwise strength reduce the shift into an and.
6996 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6997 Constant *Mask = Context->getConstantInt(Val);
7000 BinaryOperator::CreateAnd(LHSI->getOperand(0),
7001 Mask, LHSI->getName()+".mask");
7002 Value *And = InsertNewInstBefore(AndI, ICI);
7003 return new ICmpInst(*Context, ICI.getPredicate(), And,
7004 Context->getConstantExprShl(RHS, ShAmt));
7009 case Instruction::SDiv:
7010 case Instruction::UDiv:
7011 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7012 // Fold this div into the comparison, producing a range check.
7013 // Determine, based on the divide type, what the range is being
7014 // checked. If there is an overflow on the low or high side, remember
7015 // it, otherwise compute the range [low, hi) bounding the new value.
7016 // See: InsertRangeTest above for the kinds of replacements possible.
7017 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7018 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7023 case Instruction::Add:
7024 // Fold: icmp pred (add, X, C1), C2
7026 if (!ICI.isEquality()) {
7027 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7029 const APInt &LHSV = LHSC->getValue();
7031 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7034 if (ICI.isSignedPredicate()) {
7035 if (CR.getLower().isSignBit()) {
7036 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7037 Context->getConstantInt(CR.getUpper()));
7038 } else if (CR.getUpper().isSignBit()) {
7039 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7040 Context->getConstantInt(CR.getLower()));
7043 if (CR.getLower().isMinValue()) {
7044 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7045 Context->getConstantInt(CR.getUpper()));
7046 } else if (CR.getUpper().isMinValue()) {
7047 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7048 Context->getConstantInt(CR.getLower()));
7055 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7056 if (ICI.isEquality()) {
7057 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7059 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7060 // the second operand is a constant, simplify a bit.
7061 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7062 switch (BO->getOpcode()) {
7063 case Instruction::SRem:
7064 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7065 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7066 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7067 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7068 Instruction *NewRem =
7069 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
7071 InsertNewInstBefore(NewRem, ICI);
7072 return new ICmpInst(*Context, ICI.getPredicate(), NewRem,
7073 Context->getNullValue(BO->getType()));
7077 case Instruction::Add:
7078 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7079 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7080 if (BO->hasOneUse())
7081 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7082 Context->getConstantExprSub(RHS, BOp1C));
7083 } else if (RHSV == 0) {
7084 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7085 // efficiently invertible, or if the add has just this one use.
7086 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7088 if (Value *NegVal = dyn_castNegVal(BOp1, Context))
7089 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, NegVal);
7090 else if (Value *NegVal = dyn_castNegVal(BOp0, Context))
7091 return new ICmpInst(*Context, ICI.getPredicate(), NegVal, BOp1);
7092 else if (BO->hasOneUse()) {
7093 Instruction *Neg = BinaryOperator::CreateNeg(*Context, BOp1);
7094 InsertNewInstBefore(Neg, ICI);
7096 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, Neg);
7100 case Instruction::Xor:
7101 // For the xor case, we can xor two constants together, eliminating
7102 // the explicit xor.
7103 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7104 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7105 Context->getConstantExprXor(RHS, BOC));
7108 case Instruction::Sub:
7109 // Replace (([sub|xor] A, B) != 0) with (A != B)
7111 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7115 case Instruction::Or:
7116 // If bits are being or'd in that are not present in the constant we
7117 // are comparing against, then the comparison could never succeed!
7118 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7119 Constant *NotCI = Context->getConstantExprNot(RHS);
7120 if (!Context->getConstantExprAnd(BOC, NotCI)->isNullValue())
7121 return ReplaceInstUsesWith(ICI,
7122 Context->getConstantInt(Type::Int1Ty,
7127 case Instruction::And:
7128 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7129 // If bits are being compared against that are and'd out, then the
7130 // comparison can never succeed!
7131 if ((RHSV & ~BOC->getValue()) != 0)
7132 return ReplaceInstUsesWith(ICI,
7133 Context->getConstantInt(Type::Int1Ty,
7136 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7137 if (RHS == BOC && RHSV.isPowerOf2())
7138 return new ICmpInst(*Context, isICMP_NE ? ICmpInst::ICMP_EQ :
7139 ICmpInst::ICMP_NE, LHSI,
7140 Context->getNullValue(RHS->getType()));
7142 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7143 if (BOC->getValue().isSignBit()) {
7144 Value *X = BO->getOperand(0);
7145 Constant *Zero = Context->getNullValue(X->getType());
7146 ICmpInst::Predicate pred = isICMP_NE ?
7147 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7148 return new ICmpInst(*Context, pred, X, Zero);
7151 // ((X & ~7) == 0) --> X < 8
7152 if (RHSV == 0 && isHighOnes(BOC)) {
7153 Value *X = BO->getOperand(0);
7154 Constant *NegX = Context->getConstantExprNeg(BOC);
7155 ICmpInst::Predicate pred = isICMP_NE ?
7156 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7157 return new ICmpInst(*Context, pred, X, NegX);
7162 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7163 // Handle icmp {eq|ne} <intrinsic>, intcst.
7164 if (II->getIntrinsicID() == Intrinsic::bswap) {
7166 ICI.setOperand(0, II->getOperand(1));
7167 ICI.setOperand(1, Context->getConstantInt(RHSV.byteSwap()));
7175 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7176 /// We only handle extending casts so far.
7178 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7179 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7180 Value *LHSCIOp = LHSCI->getOperand(0);
7181 const Type *SrcTy = LHSCIOp->getType();
7182 const Type *DestTy = LHSCI->getType();
7185 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7186 // integer type is the same size as the pointer type.
7187 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
7188 getTargetData().getPointerSizeInBits() ==
7189 cast<IntegerType>(DestTy)->getBitWidth()) {
7191 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7192 RHSOp = Context->getConstantExprIntToPtr(RHSC, SrcTy);
7193 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7194 RHSOp = RHSC->getOperand(0);
7195 // If the pointer types don't match, insert a bitcast.
7196 if (LHSCIOp->getType() != RHSOp->getType())
7197 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
7201 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSOp);
7204 // The code below only handles extension cast instructions, so far.
7206 if (LHSCI->getOpcode() != Instruction::ZExt &&
7207 LHSCI->getOpcode() != Instruction::SExt)
7210 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7211 bool isSignedCmp = ICI.isSignedPredicate();
7213 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7214 // Not an extension from the same type?
7215 RHSCIOp = CI->getOperand(0);
7216 if (RHSCIOp->getType() != LHSCIOp->getType())
7219 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7220 // and the other is a zext), then we can't handle this.
7221 if (CI->getOpcode() != LHSCI->getOpcode())
7224 // Deal with equality cases early.
7225 if (ICI.isEquality())
7226 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7228 // A signed comparison of sign extended values simplifies into a
7229 // signed comparison.
7230 if (isSignedCmp && isSignedExt)
7231 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7233 // The other three cases all fold into an unsigned comparison.
7234 return new ICmpInst(*Context, ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7237 // If we aren't dealing with a constant on the RHS, exit early
7238 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7242 // Compute the constant that would happen if we truncated to SrcTy then
7243 // reextended to DestTy.
7244 Constant *Res1 = Context->getConstantExprTrunc(CI, SrcTy);
7245 Constant *Res2 = Context->getConstantExprCast(LHSCI->getOpcode(),
7248 // If the re-extended constant didn't change...
7250 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7251 // For example, we might have:
7252 // %A = sext i16 %X to i32
7253 // %B = icmp ugt i32 %A, 1330
7254 // It is incorrect to transform this into
7255 // %B = icmp ugt i16 %X, 1330
7256 // because %A may have negative value.
7258 // However, we allow this when the compare is EQ/NE, because they are
7260 if (isSignedExt == isSignedCmp || ICI.isEquality())
7261 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, Res1);
7265 // The re-extended constant changed so the constant cannot be represented
7266 // in the shorter type. Consequently, we cannot emit a simple comparison.
7268 // First, handle some easy cases. We know the result cannot be equal at this
7269 // point so handle the ICI.isEquality() cases
7270 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7271 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
7272 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7273 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
7275 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7276 // should have been folded away previously and not enter in here.
7279 // We're performing a signed comparison.
7280 if (cast<ConstantInt>(CI)->getValue().isNegative())
7281 Result = Context->getConstantIntFalse(); // X < (small) --> false
7283 Result = Context->getConstantIntTrue(); // X < (large) --> true
7285 // We're performing an unsigned comparison.
7287 // We're performing an unsigned comp with a sign extended value.
7288 // This is true if the input is >= 0. [aka >s -1]
7289 Constant *NegOne = Context->getAllOnesValue(SrcTy);
7290 Result = InsertNewInstBefore(new ICmpInst(*Context, ICmpInst::ICMP_SGT,
7291 LHSCIOp, NegOne, ICI.getName()), ICI);
7293 // Unsigned extend & unsigned compare -> always true.
7294 Result = Context->getConstantIntTrue();
7298 // Finally, return the value computed.
7299 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7300 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7301 return ReplaceInstUsesWith(ICI, Result);
7303 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7304 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7305 "ICmp should be folded!");
7306 if (Constant *CI = dyn_cast<Constant>(Result))
7307 return ReplaceInstUsesWith(ICI, Context->getConstantExprNot(CI));
7308 return BinaryOperator::CreateNot(*Context, Result);
7311 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7312 return commonShiftTransforms(I);
7315 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7316 return commonShiftTransforms(I);
7319 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7320 if (Instruction *R = commonShiftTransforms(I))
7323 Value *Op0 = I.getOperand(0);
7325 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7326 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7327 if (CSI->isAllOnesValue())
7328 return ReplaceInstUsesWith(I, CSI);
7330 // See if we can turn a signed shr into an unsigned shr.
7331 if (MaskedValueIsZero(Op0,
7332 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7333 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7335 // Arithmetic shifting an all-sign-bit value is a no-op.
7336 unsigned NumSignBits = ComputeNumSignBits(Op0);
7337 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7338 return ReplaceInstUsesWith(I, Op0);
7343 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7344 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7345 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7347 // shl X, 0 == X and shr X, 0 == X
7348 // shl 0, X == 0 and shr 0, X == 0
7349 if (Op1 == Context->getNullValue(Op1->getType()) ||
7350 Op0 == Context->getNullValue(Op0->getType()))
7351 return ReplaceInstUsesWith(I, Op0);
7353 if (isa<UndefValue>(Op0)) {
7354 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7355 return ReplaceInstUsesWith(I, Op0);
7356 else // undef << X -> 0, undef >>u X -> 0
7357 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7359 if (isa<UndefValue>(Op1)) {
7360 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7361 return ReplaceInstUsesWith(I, Op0);
7362 else // X << undef, X >>u undef -> 0
7363 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7366 // See if we can fold away this shift.
7367 if (SimplifyDemandedInstructionBits(I))
7370 // Try to fold constant and into select arguments.
7371 if (isa<Constant>(Op0))
7372 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7373 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7376 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7377 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7382 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7383 BinaryOperator &I) {
7384 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7386 // See if we can simplify any instructions used by the instruction whose sole
7387 // purpose is to compute bits we don't care about.
7388 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7390 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7393 if (Op1->uge(TypeBits)) {
7394 if (I.getOpcode() != Instruction::AShr)
7395 return ReplaceInstUsesWith(I, Context->getNullValue(Op0->getType()));
7397 I.setOperand(1, Context->getConstantInt(I.getType(), TypeBits-1));
7402 // ((X*C1) << C2) == (X * (C1 << C2))
7403 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7404 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7405 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7406 return BinaryOperator::CreateMul(BO->getOperand(0),
7407 Context->getConstantExprShl(BOOp, Op1));
7409 // Try to fold constant and into select arguments.
7410 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7411 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7413 if (isa<PHINode>(Op0))
7414 if (Instruction *NV = FoldOpIntoPhi(I))
7417 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7418 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7419 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7420 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7421 // place. Don't try to do this transformation in this case. Also, we
7422 // require that the input operand is a shift-by-constant so that we have
7423 // confidence that the shifts will get folded together. We could do this
7424 // xform in more cases, but it is unlikely to be profitable.
7425 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7426 isa<ConstantInt>(TrOp->getOperand(1))) {
7427 // Okay, we'll do this xform. Make the shift of shift.
7428 Constant *ShAmt = Context->getConstantExprZExt(Op1, TrOp->getType());
7429 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7431 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7433 // For logical shifts, the truncation has the effect of making the high
7434 // part of the register be zeros. Emulate this by inserting an AND to
7435 // clear the top bits as needed. This 'and' will usually be zapped by
7436 // other xforms later if dead.
7437 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7438 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7439 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7441 // The mask we constructed says what the trunc would do if occurring
7442 // between the shifts. We want to know the effect *after* the second
7443 // shift. We know that it is a logical shift by a constant, so adjust the
7444 // mask as appropriate.
7445 if (I.getOpcode() == Instruction::Shl)
7446 MaskV <<= Op1->getZExtValue();
7448 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7449 MaskV = MaskV.lshr(Op1->getZExtValue());
7453 BinaryOperator::CreateAnd(NSh, Context->getConstantInt(MaskV),
7455 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7457 // Return the value truncated to the interesting size.
7458 return new TruncInst(And, I.getType());
7462 if (Op0->hasOneUse()) {
7463 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7464 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7467 switch (Op0BO->getOpcode()) {
7469 case Instruction::Add:
7470 case Instruction::And:
7471 case Instruction::Or:
7472 case Instruction::Xor: {
7473 // These operators commute.
7474 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7475 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7476 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7477 m_Specific(Op1)), *Context)){
7478 Instruction *YS = BinaryOperator::CreateShl(
7479 Op0BO->getOperand(0), Op1,
7481 InsertNewInstBefore(YS, I); // (Y << C)
7483 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7484 Op0BO->getOperand(1)->getName());
7485 InsertNewInstBefore(X, I); // (X + (Y << C))
7486 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7487 return BinaryOperator::CreateAnd(X, Context->getConstantInt(
7488 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7491 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7492 Value *Op0BOOp1 = Op0BO->getOperand(1);
7493 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7495 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7496 m_ConstantInt(CC)), *Context) &&
7497 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7498 Instruction *YS = BinaryOperator::CreateShl(
7499 Op0BO->getOperand(0), Op1,
7501 InsertNewInstBefore(YS, I); // (Y << C)
7503 BinaryOperator::CreateAnd(V1,
7504 Context->getConstantExprShl(CC, Op1),
7505 V1->getName()+".mask");
7506 InsertNewInstBefore(XM, I); // X & (CC << C)
7508 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7513 case Instruction::Sub: {
7514 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7515 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7516 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7517 m_Specific(Op1)), *Context)){
7518 Instruction *YS = BinaryOperator::CreateShl(
7519 Op0BO->getOperand(1), Op1,
7521 InsertNewInstBefore(YS, I); // (Y << C)
7523 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7524 Op0BO->getOperand(0)->getName());
7525 InsertNewInstBefore(X, I); // (X + (Y << C))
7526 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7527 return BinaryOperator::CreateAnd(X, Context->getConstantInt(
7528 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7531 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7532 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7533 match(Op0BO->getOperand(0),
7534 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7535 m_ConstantInt(CC)), *Context) && V2 == Op1 &&
7536 cast<BinaryOperator>(Op0BO->getOperand(0))
7537 ->getOperand(0)->hasOneUse()) {
7538 Instruction *YS = BinaryOperator::CreateShl(
7539 Op0BO->getOperand(1), Op1,
7541 InsertNewInstBefore(YS, I); // (Y << C)
7543 BinaryOperator::CreateAnd(V1,
7544 Context->getConstantExprShl(CC, Op1),
7545 V1->getName()+".mask");
7546 InsertNewInstBefore(XM, I); // X & (CC << C)
7548 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7556 // If the operand is an bitwise operator with a constant RHS, and the
7557 // shift is the only use, we can pull it out of the shift.
7558 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7559 bool isValid = true; // Valid only for And, Or, Xor
7560 bool highBitSet = false; // Transform if high bit of constant set?
7562 switch (Op0BO->getOpcode()) {
7563 default: isValid = false; break; // Do not perform transform!
7564 case Instruction::Add:
7565 isValid = isLeftShift;
7567 case Instruction::Or:
7568 case Instruction::Xor:
7571 case Instruction::And:
7576 // If this is a signed shift right, and the high bit is modified
7577 // by the logical operation, do not perform the transformation.
7578 // The highBitSet boolean indicates the value of the high bit of
7579 // the constant which would cause it to be modified for this
7582 if (isValid && I.getOpcode() == Instruction::AShr)
7583 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7586 Constant *NewRHS = Context->getConstantExpr(I.getOpcode(), Op0C, Op1);
7588 Instruction *NewShift =
7589 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7590 InsertNewInstBefore(NewShift, I);
7591 NewShift->takeName(Op0BO);
7593 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7600 // Find out if this is a shift of a shift by a constant.
7601 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7602 if (ShiftOp && !ShiftOp->isShift())
7605 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7606 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7607 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7608 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7609 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7610 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7611 Value *X = ShiftOp->getOperand(0);
7613 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7615 const IntegerType *Ty = cast<IntegerType>(I.getType());
7617 // Check for (X << c1) << c2 and (X >> c1) >> c2
7618 if (I.getOpcode() == ShiftOp->getOpcode()) {
7619 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7621 if (AmtSum >= TypeBits) {
7622 if (I.getOpcode() != Instruction::AShr)
7623 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7624 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7627 return BinaryOperator::Create(I.getOpcode(), X,
7628 Context->getConstantInt(Ty, AmtSum));
7629 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7630 I.getOpcode() == Instruction::AShr) {
7631 if (AmtSum >= TypeBits)
7632 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7634 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7635 return BinaryOperator::CreateLShr(X, Context->getConstantInt(Ty, AmtSum));
7636 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7637 I.getOpcode() == Instruction::LShr) {
7638 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7639 if (AmtSum >= TypeBits)
7640 AmtSum = TypeBits-1;
7642 Instruction *Shift =
7643 BinaryOperator::CreateAShr(X, Context->getConstantInt(Ty, AmtSum));
7644 InsertNewInstBefore(Shift, I);
7646 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7647 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7650 // Okay, if we get here, one shift must be left, and the other shift must be
7651 // right. See if the amounts are equal.
7652 if (ShiftAmt1 == ShiftAmt2) {
7653 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7654 if (I.getOpcode() == Instruction::Shl) {
7655 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7656 return BinaryOperator::CreateAnd(X, Context->getConstantInt(Mask));
7658 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7659 if (I.getOpcode() == Instruction::LShr) {
7660 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7661 return BinaryOperator::CreateAnd(X, Context->getConstantInt(Mask));
7663 // We can simplify ((X << C) >>s C) into a trunc + sext.
7664 // NOTE: we could do this for any C, but that would make 'unusual' integer
7665 // types. For now, just stick to ones well-supported by the code
7667 const Type *SExtType = 0;
7668 switch (Ty->getBitWidth() - ShiftAmt1) {
7675 SExtType = Context->getIntegerType(Ty->getBitWidth() - ShiftAmt1);
7680 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7681 InsertNewInstBefore(NewTrunc, I);
7682 return new SExtInst(NewTrunc, Ty);
7684 // Otherwise, we can't handle it yet.
7685 } else if (ShiftAmt1 < ShiftAmt2) {
7686 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7688 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7689 if (I.getOpcode() == Instruction::Shl) {
7690 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7691 ShiftOp->getOpcode() == Instruction::AShr);
7692 Instruction *Shift =
7693 BinaryOperator::CreateShl(X, Context->getConstantInt(Ty, ShiftDiff));
7694 InsertNewInstBefore(Shift, I);
7696 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7697 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7700 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7701 if (I.getOpcode() == Instruction::LShr) {
7702 assert(ShiftOp->getOpcode() == Instruction::Shl);
7703 Instruction *Shift =
7704 BinaryOperator::CreateLShr(X, Context->getConstantInt(Ty, ShiftDiff));
7705 InsertNewInstBefore(Shift, I);
7707 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7708 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7711 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7713 assert(ShiftAmt2 < ShiftAmt1);
7714 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7716 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7717 if (I.getOpcode() == Instruction::Shl) {
7718 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7719 ShiftOp->getOpcode() == Instruction::AShr);
7720 Instruction *Shift =
7721 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7722 Context->getConstantInt(Ty, ShiftDiff));
7723 InsertNewInstBefore(Shift, I);
7725 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7726 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7729 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7730 if (I.getOpcode() == Instruction::LShr) {
7731 assert(ShiftOp->getOpcode() == Instruction::Shl);
7732 Instruction *Shift =
7733 BinaryOperator::CreateShl(X, Context->getConstantInt(Ty, ShiftDiff));
7734 InsertNewInstBefore(Shift, I);
7736 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7737 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7740 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7747 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7748 /// expression. If so, decompose it, returning some value X, such that Val is
7751 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7752 int &Offset, LLVMContext *Context) {
7753 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7754 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7755 Offset = CI->getZExtValue();
7757 return Context->getConstantInt(Type::Int32Ty, 0);
7758 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7759 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7760 if (I->getOpcode() == Instruction::Shl) {
7761 // This is a value scaled by '1 << the shift amt'.
7762 Scale = 1U << RHS->getZExtValue();
7764 return I->getOperand(0);
7765 } else if (I->getOpcode() == Instruction::Mul) {
7766 // This value is scaled by 'RHS'.
7767 Scale = RHS->getZExtValue();
7769 return I->getOperand(0);
7770 } else if (I->getOpcode() == Instruction::Add) {
7771 // We have X+C. Check to see if we really have (X*C2)+C1,
7772 // where C1 is divisible by C2.
7775 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7777 Offset += RHS->getZExtValue();
7784 // Otherwise, we can't look past this.
7791 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7792 /// try to eliminate the cast by moving the type information into the alloc.
7793 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7794 AllocationInst &AI) {
7795 const PointerType *PTy = cast<PointerType>(CI.getType());
7797 // Remove any uses of AI that are dead.
7798 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7800 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7801 Instruction *User = cast<Instruction>(*UI++);
7802 if (isInstructionTriviallyDead(User)) {
7803 while (UI != E && *UI == User)
7804 ++UI; // If this instruction uses AI more than once, don't break UI.
7807 DOUT << "IC: DCE: " << *User;
7808 EraseInstFromFunction(*User);
7812 // Get the type really allocated and the type casted to.
7813 const Type *AllocElTy = AI.getAllocatedType();
7814 const Type *CastElTy = PTy->getElementType();
7815 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7817 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7818 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7819 if (CastElTyAlign < AllocElTyAlign) return 0;
7821 // If the allocation has multiple uses, only promote it if we are strictly
7822 // increasing the alignment of the resultant allocation. If we keep it the
7823 // same, we open the door to infinite loops of various kinds. (A reference
7824 // from a dbg.declare doesn't count as a use for this purpose.)
7825 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7826 CastElTyAlign == AllocElTyAlign) return 0;
7828 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7829 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7830 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7832 // See if we can satisfy the modulus by pulling a scale out of the array
7834 unsigned ArraySizeScale;
7836 Value *NumElements = // See if the array size is a decomposable linear expr.
7837 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7838 ArrayOffset, Context);
7840 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7842 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7843 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7845 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7850 // If the allocation size is constant, form a constant mul expression
7851 Amt = Context->getConstantInt(Type::Int32Ty, Scale);
7852 if (isa<ConstantInt>(NumElements))
7853 Amt = Context->getConstantExprMul(cast<ConstantInt>(NumElements),
7854 cast<ConstantInt>(Amt));
7855 // otherwise multiply the amount and the number of elements
7857 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7858 Amt = InsertNewInstBefore(Tmp, AI);
7862 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7863 Value *Off = Context->getConstantInt(Type::Int32Ty, Offset, true);
7864 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7865 Amt = InsertNewInstBefore(Tmp, AI);
7868 AllocationInst *New;
7869 if (isa<MallocInst>(AI))
7870 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7872 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7873 InsertNewInstBefore(New, AI);
7876 // If the allocation has one real use plus a dbg.declare, just remove the
7878 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7879 EraseInstFromFunction(*DI);
7881 // If the allocation has multiple real uses, insert a cast and change all
7882 // things that used it to use the new cast. This will also hack on CI, but it
7884 else if (!AI.hasOneUse()) {
7885 AddUsesToWorkList(AI);
7886 // New is the allocation instruction, pointer typed. AI is the original
7887 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7888 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7889 InsertNewInstBefore(NewCast, AI);
7890 AI.replaceAllUsesWith(NewCast);
7892 return ReplaceInstUsesWith(CI, New);
7895 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7896 /// and return it as type Ty without inserting any new casts and without
7897 /// changing the computed value. This is used by code that tries to decide
7898 /// whether promoting or shrinking integer operations to wider or smaller types
7899 /// will allow us to eliminate a truncate or extend.
7901 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7902 /// extension operation if Ty is larger.
7904 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7905 /// should return true if trunc(V) can be computed by computing V in the smaller
7906 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7907 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7908 /// efficiently truncated.
7910 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7911 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7912 /// the final result.
7913 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7915 int &NumCastsRemoved){
7916 // We can always evaluate constants in another type.
7917 if (isa<Constant>(V))
7920 Instruction *I = dyn_cast<Instruction>(V);
7921 if (!I) return false;
7923 const Type *OrigTy = V->getType();
7925 // If this is an extension or truncate, we can often eliminate it.
7926 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7927 // If this is a cast from the destination type, we can trivially eliminate
7928 // it, and this will remove a cast overall.
7929 if (I->getOperand(0)->getType() == Ty) {
7930 // If the first operand is itself a cast, and is eliminable, do not count
7931 // this as an eliminable cast. We would prefer to eliminate those two
7933 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7939 // We can't extend or shrink something that has multiple uses: doing so would
7940 // require duplicating the instruction in general, which isn't profitable.
7941 if (!I->hasOneUse()) return false;
7943 unsigned Opc = I->getOpcode();
7945 case Instruction::Add:
7946 case Instruction::Sub:
7947 case Instruction::Mul:
7948 case Instruction::And:
7949 case Instruction::Or:
7950 case Instruction::Xor:
7951 // These operators can all arbitrarily be extended or truncated.
7952 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7954 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7957 case Instruction::UDiv:
7958 case Instruction::URem: {
7959 // UDiv and URem can be truncated if all the truncated bits are zero.
7960 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7961 uint32_t BitWidth = Ty->getScalarSizeInBits();
7962 if (BitWidth < OrigBitWidth) {
7963 APInt Mask = APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth);
7964 if (MaskedValueIsZero(I->getOperand(0), Mask) &&
7965 MaskedValueIsZero(I->getOperand(1), Mask)) {
7966 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7968 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7974 case Instruction::Shl:
7975 // If we are truncating the result of this SHL, and if it's a shift of a
7976 // constant amount, we can always perform a SHL in a smaller type.
7977 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7978 uint32_t BitWidth = Ty->getScalarSizeInBits();
7979 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7980 CI->getLimitedValue(BitWidth) < BitWidth)
7981 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7985 case Instruction::LShr:
7986 // If this is a truncate of a logical shr, we can truncate it to a smaller
7987 // lshr iff we know that the bits we would otherwise be shifting in are
7989 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7990 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7991 uint32_t BitWidth = Ty->getScalarSizeInBits();
7992 if (BitWidth < OrigBitWidth &&
7993 MaskedValueIsZero(I->getOperand(0),
7994 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7995 CI->getLimitedValue(BitWidth) < BitWidth) {
7996 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
8001 case Instruction::ZExt:
8002 case Instruction::SExt:
8003 case Instruction::Trunc:
8004 // If this is the same kind of case as our original (e.g. zext+zext), we
8005 // can safely replace it. Note that replacing it does not reduce the number
8006 // of casts in the input.
8010 // sext (zext ty1), ty2 -> zext ty2
8011 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8014 case Instruction::Select: {
8015 SelectInst *SI = cast<SelectInst>(I);
8016 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8018 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8021 case Instruction::PHI: {
8022 // We can change a phi if we can change all operands.
8023 PHINode *PN = cast<PHINode>(I);
8024 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8025 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8031 // TODO: Can handle more cases here.
8038 /// EvaluateInDifferentType - Given an expression that
8039 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8040 /// evaluate the expression.
8041 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8043 if (Constant *C = dyn_cast<Constant>(V))
8044 return Context->getConstantExprIntegerCast(C, Ty,
8045 isSigned /*Sext or ZExt*/);
8047 // Otherwise, it must be an instruction.
8048 Instruction *I = cast<Instruction>(V);
8049 Instruction *Res = 0;
8050 unsigned Opc = I->getOpcode();
8052 case Instruction::Add:
8053 case Instruction::Sub:
8054 case Instruction::Mul:
8055 case Instruction::And:
8056 case Instruction::Or:
8057 case Instruction::Xor:
8058 case Instruction::AShr:
8059 case Instruction::LShr:
8060 case Instruction::Shl:
8061 case Instruction::UDiv:
8062 case Instruction::URem: {
8063 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8064 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8065 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8068 case Instruction::Trunc:
8069 case Instruction::ZExt:
8070 case Instruction::SExt:
8071 // If the source type of the cast is the type we're trying for then we can
8072 // just return the source. There's no need to insert it because it is not
8074 if (I->getOperand(0)->getType() == Ty)
8075 return I->getOperand(0);
8077 // Otherwise, must be the same type of cast, so just reinsert a new one.
8078 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8081 case Instruction::Select: {
8082 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8083 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8084 Res = SelectInst::Create(I->getOperand(0), True, False);
8087 case Instruction::PHI: {
8088 PHINode *OPN = cast<PHINode>(I);
8089 PHINode *NPN = PHINode::Create(Ty);
8090 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8091 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8092 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8098 // TODO: Can handle more cases here.
8099 llvm_unreachable("Unreachable!");
8104 return InsertNewInstBefore(Res, *I);
8107 /// @brief Implement the transforms common to all CastInst visitors.
8108 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8109 Value *Src = CI.getOperand(0);
8111 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8112 // eliminate it now.
8113 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8114 if (Instruction::CastOps opc =
8115 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8116 // The first cast (CSrc) is eliminable so we need to fix up or replace
8117 // the second cast (CI). CSrc will then have a good chance of being dead.
8118 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8122 // If we are casting a select then fold the cast into the select
8123 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8124 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8127 // If we are casting a PHI then fold the cast into the PHI
8128 if (isa<PHINode>(Src))
8129 if (Instruction *NV = FoldOpIntoPhi(CI))
8135 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8136 /// or not there is a sequence of GEP indices into the type that will land us at
8137 /// the specified offset. If so, fill them into NewIndices and return the
8138 /// resultant element type, otherwise return null.
8139 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8140 SmallVectorImpl<Value*> &NewIndices,
8141 const TargetData *TD,
8142 LLVMContext *Context) {
8143 if (!Ty->isSized()) return 0;
8145 // Start with the index over the outer type. Note that the type size
8146 // might be zero (even if the offset isn't zero) if the indexed type
8147 // is something like [0 x {int, int}]
8148 const Type *IntPtrTy = TD->getIntPtrType();
8149 int64_t FirstIdx = 0;
8150 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8151 FirstIdx = Offset/TySize;
8152 Offset -= FirstIdx*TySize;
8154 // Handle hosts where % returns negative instead of values [0..TySize).
8158 assert(Offset >= 0);
8160 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8163 NewIndices.push_back(Context->getConstantInt(IntPtrTy, FirstIdx));
8165 // Index into the types. If we fail, set OrigBase to null.
8167 // Indexing into tail padding between struct/array elements.
8168 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8171 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8172 const StructLayout *SL = TD->getStructLayout(STy);
8173 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8174 "Offset must stay within the indexed type");
8176 unsigned Elt = SL->getElementContainingOffset(Offset);
8177 NewIndices.push_back(Context->getConstantInt(Type::Int32Ty, Elt));
8179 Offset -= SL->getElementOffset(Elt);
8180 Ty = STy->getElementType(Elt);
8181 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8182 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8183 assert(EltSize && "Cannot index into a zero-sized array");
8184 NewIndices.push_back(Context->getConstantInt(IntPtrTy,Offset/EltSize));
8186 Ty = AT->getElementType();
8188 // Otherwise, we can't index into the middle of this atomic type, bail.
8196 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8197 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8198 Value *Src = CI.getOperand(0);
8200 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8201 // If casting the result of a getelementptr instruction with no offset, turn
8202 // this into a cast of the original pointer!
8203 if (GEP->hasAllZeroIndices()) {
8204 // Changing the cast operand is usually not a good idea but it is safe
8205 // here because the pointer operand is being replaced with another
8206 // pointer operand so the opcode doesn't need to change.
8208 CI.setOperand(0, GEP->getOperand(0));
8212 // If the GEP has a single use, and the base pointer is a bitcast, and the
8213 // GEP computes a constant offset, see if we can convert these three
8214 // instructions into fewer. This typically happens with unions and other
8215 // non-type-safe code.
8216 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8217 if (GEP->hasAllConstantIndices()) {
8218 // We are guaranteed to get a constant from EmitGEPOffset.
8219 ConstantInt *OffsetV =
8220 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8221 int64_t Offset = OffsetV->getSExtValue();
8223 // Get the base pointer input of the bitcast, and the type it points to.
8224 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8225 const Type *GEPIdxTy =
8226 cast<PointerType>(OrigBase->getType())->getElementType();
8227 SmallVector<Value*, 8> NewIndices;
8228 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8229 // If we were able to index down into an element, create the GEP
8230 // and bitcast the result. This eliminates one bitcast, potentially
8232 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
8234 NewIndices.end(), "");
8235 InsertNewInstBefore(NGEP, CI);
8236 NGEP->takeName(GEP);
8238 if (isa<BitCastInst>(CI))
8239 return new BitCastInst(NGEP, CI.getType());
8240 assert(isa<PtrToIntInst>(CI));
8241 return new PtrToIntInst(NGEP, CI.getType());
8247 return commonCastTransforms(CI);
8250 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8251 /// type like i42. We don't want to introduce operations on random non-legal
8252 /// integer types where they don't already exist in the code. In the future,
8253 /// we should consider making this based off target-data, so that 32-bit targets
8254 /// won't get i64 operations etc.
8255 static bool isSafeIntegerType(const Type *Ty) {
8256 switch (Ty->getPrimitiveSizeInBits()) {
8267 /// commonIntCastTransforms - This function implements the common transforms
8268 /// for trunc, zext, and sext.
8269 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8270 if (Instruction *Result = commonCastTransforms(CI))
8273 Value *Src = CI.getOperand(0);
8274 const Type *SrcTy = Src->getType();
8275 const Type *DestTy = CI.getType();
8276 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8277 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8279 // See if we can simplify any instructions used by the LHS whose sole
8280 // purpose is to compute bits we don't care about.
8281 if (SimplifyDemandedInstructionBits(CI))
8284 // If the source isn't an instruction or has more than one use then we
8285 // can't do anything more.
8286 Instruction *SrcI = dyn_cast<Instruction>(Src);
8287 if (!SrcI || !Src->hasOneUse())
8290 // Attempt to propagate the cast into the instruction for int->int casts.
8291 int NumCastsRemoved = 0;
8292 // Only do this if the dest type is a simple type, don't convert the
8293 // expression tree to something weird like i93 unless the source is also
8295 if ((isSafeIntegerType(DestTy->getScalarType()) ||
8296 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8297 CanEvaluateInDifferentType(SrcI, DestTy,
8298 CI.getOpcode(), NumCastsRemoved)) {
8299 // If this cast is a truncate, evaluting in a different type always
8300 // eliminates the cast, so it is always a win. If this is a zero-extension,
8301 // we need to do an AND to maintain the clear top-part of the computation,
8302 // so we require that the input have eliminated at least one cast. If this
8303 // is a sign extension, we insert two new casts (to do the extension) so we
8304 // require that two casts have been eliminated.
8305 bool DoXForm = false;
8306 bool JustReplace = false;
8307 switch (CI.getOpcode()) {
8309 // All the others use floating point so we shouldn't actually
8310 // get here because of the check above.
8311 llvm_unreachable("Unknown cast type");
8312 case Instruction::Trunc:
8315 case Instruction::ZExt: {
8316 DoXForm = NumCastsRemoved >= 1;
8317 if (!DoXForm && 0) {
8318 // If it's unnecessary to issue an AND to clear the high bits, it's
8319 // always profitable to do this xform.
8320 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8321 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8322 if (MaskedValueIsZero(TryRes, Mask))
8323 return ReplaceInstUsesWith(CI, TryRes);
8325 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8326 if (TryI->use_empty())
8327 EraseInstFromFunction(*TryI);
8331 case Instruction::SExt: {
8332 DoXForm = NumCastsRemoved >= 2;
8333 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8334 // If we do not have to emit the truncate + sext pair, then it's always
8335 // profitable to do this xform.
8337 // It's not safe to eliminate the trunc + sext pair if one of the
8338 // eliminated cast is a truncate. e.g.
8339 // t2 = trunc i32 t1 to i16
8340 // t3 = sext i16 t2 to i32
8343 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8344 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8345 if (NumSignBits > (DestBitSize - SrcBitSize))
8346 return ReplaceInstUsesWith(CI, TryRes);
8348 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8349 if (TryI->use_empty())
8350 EraseInstFromFunction(*TryI);
8357 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
8359 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8360 CI.getOpcode() == Instruction::SExt);
8362 // Just replace this cast with the result.
8363 return ReplaceInstUsesWith(CI, Res);
8365 assert(Res->getType() == DestTy);
8366 switch (CI.getOpcode()) {
8367 default: llvm_unreachable("Unknown cast type!");
8368 case Instruction::Trunc:
8369 // Just replace this cast with the result.
8370 return ReplaceInstUsesWith(CI, Res);
8371 case Instruction::ZExt: {
8372 assert(SrcBitSize < DestBitSize && "Not a zext?");
8374 // If the high bits are already zero, just replace this cast with the
8376 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8377 if (MaskedValueIsZero(Res, Mask))
8378 return ReplaceInstUsesWith(CI, Res);
8380 // We need to emit an AND to clear the high bits.
8381 Constant *C = Context->getConstantInt(APInt::getLowBitsSet(DestBitSize,
8383 return BinaryOperator::CreateAnd(Res, C);
8385 case Instruction::SExt: {
8386 // If the high bits are already filled with sign bit, just replace this
8387 // cast with the result.
8388 unsigned NumSignBits = ComputeNumSignBits(Res);
8389 if (NumSignBits > (DestBitSize - SrcBitSize))
8390 return ReplaceInstUsesWith(CI, Res);
8392 // We need to emit a cast to truncate, then a cast to sext.
8393 return CastInst::Create(Instruction::SExt,
8394 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8401 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8402 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8404 switch (SrcI->getOpcode()) {
8405 case Instruction::Add:
8406 case Instruction::Mul:
8407 case Instruction::And:
8408 case Instruction::Or:
8409 case Instruction::Xor:
8410 // If we are discarding information, rewrite.
8411 if (DestBitSize < SrcBitSize && DestBitSize != 1) {
8412 // Don't insert two casts unless at least one can be eliminated.
8413 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8414 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8415 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8416 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8417 return BinaryOperator::Create(
8418 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8422 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8423 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8424 SrcI->getOpcode() == Instruction::Xor &&
8425 Op1 == Context->getConstantIntTrue() &&
8426 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8427 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8428 return BinaryOperator::CreateXor(New,
8429 Context->getConstantInt(CI.getType(), 1));
8433 case Instruction::Shl: {
8434 // Canonicalize trunc inside shl, if we can.
8435 ConstantInt *CI = dyn_cast<ConstantInt>(Op1);
8436 if (CI && DestBitSize < SrcBitSize &&
8437 CI->getLimitedValue(DestBitSize) < DestBitSize) {
8438 Value *Op0c = InsertCastBefore(Instruction::Trunc, Op0, DestTy, *SrcI);
8439 Value *Op1c = InsertCastBefore(Instruction::Trunc, Op1, DestTy, *SrcI);
8440 return BinaryOperator::CreateShl(Op0c, Op1c);
8448 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8449 if (Instruction *Result = commonIntCastTransforms(CI))
8452 Value *Src = CI.getOperand(0);
8453 const Type *Ty = CI.getType();
8454 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8455 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8457 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8458 if (DestBitWidth == 1 &&
8459 isa<VectorType>(Ty) == isa<VectorType>(Src->getType())) {
8460 Constant *One = Context->getConstantInt(Src->getType(), 1);
8461 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8462 Value *Zero = Context->getNullValue(Src->getType());
8463 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Src, Zero);
8466 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8467 ConstantInt *ShAmtV = 0;
8469 if (Src->hasOneUse() &&
8470 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)), *Context)) {
8471 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8473 // Get a mask for the bits shifting in.
8474 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8475 if (MaskedValueIsZero(ShiftOp, Mask)) {
8476 if (ShAmt >= DestBitWidth) // All zeros.
8477 return ReplaceInstUsesWith(CI, Context->getNullValue(Ty));
8479 // Okay, we can shrink this. Truncate the input, then return a new
8481 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8482 Value *V2 = Context->getConstantExprTrunc(ShAmtV, Ty);
8483 return BinaryOperator::CreateLShr(V1, V2);
8490 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8491 /// in order to eliminate the icmp.
8492 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8494 // If we are just checking for a icmp eq of a single bit and zext'ing it
8495 // to an integer, then shift the bit to the appropriate place and then
8496 // cast to integer to avoid the comparison.
8497 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8498 const APInt &Op1CV = Op1C->getValue();
8500 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8501 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8502 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8503 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8504 if (!DoXform) return ICI;
8506 Value *In = ICI->getOperand(0);
8507 Value *Sh = Context->getConstantInt(In->getType(),
8508 In->getType()->getScalarSizeInBits()-1);
8509 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8510 In->getName()+".lobit"),
8512 if (In->getType() != CI.getType())
8513 In = CastInst::CreateIntegerCast(In, CI.getType(),
8514 false/*ZExt*/, "tmp", &CI);
8516 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8517 Constant *One = Context->getConstantInt(In->getType(), 1);
8518 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8519 In->getName()+".not"),
8523 return ReplaceInstUsesWith(CI, In);
8528 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8529 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8530 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8531 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8532 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8533 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8534 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8535 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8536 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8537 // This only works for EQ and NE
8538 ICI->isEquality()) {
8539 // If Op1C some other power of two, convert:
8540 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8541 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8542 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8543 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8545 APInt KnownZeroMask(~KnownZero);
8546 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8547 if (!DoXform) return ICI;
8549 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8550 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8551 // (X&4) == 2 --> false
8552 // (X&4) != 2 --> true
8553 Constant *Res = Context->getConstantInt(Type::Int1Ty, isNE);
8554 Res = Context->getConstantExprZExt(Res, CI.getType());
8555 return ReplaceInstUsesWith(CI, Res);
8558 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8559 Value *In = ICI->getOperand(0);
8561 // Perform a logical shr by shiftamt.
8562 // Insert the shift to put the result in the low bit.
8563 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8564 Context->getConstantInt(In->getType(), ShiftAmt),
8565 In->getName()+".lobit"), CI);
8568 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8569 Constant *One = Context->getConstantInt(In->getType(), 1);
8570 In = BinaryOperator::CreateXor(In, One, "tmp");
8571 InsertNewInstBefore(cast<Instruction>(In), CI);
8574 if (CI.getType() == In->getType())
8575 return ReplaceInstUsesWith(CI, In);
8577 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8585 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8586 // If one of the common conversion will work ..
8587 if (Instruction *Result = commonIntCastTransforms(CI))
8590 Value *Src = CI.getOperand(0);
8592 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8593 // types and if the sizes are just right we can convert this into a logical
8594 // 'and' which will be much cheaper than the pair of casts.
8595 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8596 // Get the sizes of the types involved. We know that the intermediate type
8597 // will be smaller than A or C, but don't know the relation between A and C.
8598 Value *A = CSrc->getOperand(0);
8599 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8600 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8601 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8602 // If we're actually extending zero bits, then if
8603 // SrcSize < DstSize: zext(a & mask)
8604 // SrcSize == DstSize: a & mask
8605 // SrcSize > DstSize: trunc(a) & mask
8606 if (SrcSize < DstSize) {
8607 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8608 Constant *AndConst = Context->getConstantInt(A->getType(), AndValue);
8610 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8611 InsertNewInstBefore(And, CI);
8612 return new ZExtInst(And, CI.getType());
8613 } else if (SrcSize == DstSize) {
8614 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8615 return BinaryOperator::CreateAnd(A, Context->getConstantInt(A->getType(),
8617 } else if (SrcSize > DstSize) {
8618 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8619 InsertNewInstBefore(Trunc, CI);
8620 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8621 return BinaryOperator::CreateAnd(Trunc,
8622 Context->getConstantInt(Trunc->getType(),
8627 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8628 return transformZExtICmp(ICI, CI);
8630 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8631 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8632 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8633 // of the (zext icmp) will be transformed.
8634 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8635 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8636 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8637 (transformZExtICmp(LHS, CI, false) ||
8638 transformZExtICmp(RHS, CI, false))) {
8639 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8640 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8641 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8645 // zext(trunc(t) & C) -> (t & zext(C)).
8646 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8647 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8648 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8649 Value *TI0 = TI->getOperand(0);
8650 if (TI0->getType() == CI.getType())
8652 BinaryOperator::CreateAnd(TI0,
8653 Context->getConstantExprZExt(C, CI.getType()));
8656 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8657 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8658 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8659 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8660 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8661 And->getOperand(1) == C)
8662 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8663 Value *TI0 = TI->getOperand(0);
8664 if (TI0->getType() == CI.getType()) {
8665 Constant *ZC = Context->getConstantExprZExt(C, CI.getType());
8666 Instruction *NewAnd = BinaryOperator::CreateAnd(TI0, ZC, "tmp");
8667 InsertNewInstBefore(NewAnd, *And);
8668 return BinaryOperator::CreateXor(NewAnd, ZC);
8675 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8676 if (Instruction *I = commonIntCastTransforms(CI))
8679 Value *Src = CI.getOperand(0);
8681 // Canonicalize sign-extend from i1 to a select.
8682 if (Src->getType() == Type::Int1Ty)
8683 return SelectInst::Create(Src,
8684 Context->getAllOnesValue(CI.getType()),
8685 Context->getNullValue(CI.getType()));
8687 // See if the value being truncated is already sign extended. If so, just
8688 // eliminate the trunc/sext pair.
8689 if (Operator::getOpcode(Src) == Instruction::Trunc) {
8690 Value *Op = cast<User>(Src)->getOperand(0);
8691 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8692 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8693 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8694 unsigned NumSignBits = ComputeNumSignBits(Op);
8696 if (OpBits == DestBits) {
8697 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8698 // bits, it is already ready.
8699 if (NumSignBits > DestBits-MidBits)
8700 return ReplaceInstUsesWith(CI, Op);
8701 } else if (OpBits < DestBits) {
8702 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8703 // bits, just sext from i32.
8704 if (NumSignBits > OpBits-MidBits)
8705 return new SExtInst(Op, CI.getType(), "tmp");
8707 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8708 // bits, just truncate to i32.
8709 if (NumSignBits > OpBits-MidBits)
8710 return new TruncInst(Op, CI.getType(), "tmp");
8714 // If the input is a shl/ashr pair of a same constant, then this is a sign
8715 // extension from a smaller value. If we could trust arbitrary bitwidth
8716 // integers, we could turn this into a truncate to the smaller bit and then
8717 // use a sext for the whole extension. Since we don't, look deeper and check
8718 // for a truncate. If the source and dest are the same type, eliminate the
8719 // trunc and extend and just do shifts. For example, turn:
8720 // %a = trunc i32 %i to i8
8721 // %b = shl i8 %a, 6
8722 // %c = ashr i8 %b, 6
8723 // %d = sext i8 %c to i32
8725 // %a = shl i32 %i, 30
8726 // %d = ashr i32 %a, 30
8728 ConstantInt *BA = 0, *CA = 0;
8729 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8730 m_ConstantInt(CA)), *Context) &&
8731 BA == CA && isa<TruncInst>(A)) {
8732 Value *I = cast<TruncInst>(A)->getOperand(0);
8733 if (I->getType() == CI.getType()) {
8734 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8735 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8736 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8737 Constant *ShAmtV = Context->getConstantInt(CI.getType(), ShAmt);
8738 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8740 return BinaryOperator::CreateAShr(I, ShAmtV);
8747 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8748 /// in the specified FP type without changing its value.
8749 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8750 LLVMContext *Context) {
8752 APFloat F = CFP->getValueAPF();
8753 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8755 return Context->getConstantFP(F);
8759 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8760 /// through it until we get the source value.
8761 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8762 if (Instruction *I = dyn_cast<Instruction>(V))
8763 if (I->getOpcode() == Instruction::FPExt)
8764 return LookThroughFPExtensions(I->getOperand(0), Context);
8766 // If this value is a constant, return the constant in the smallest FP type
8767 // that can accurately represent it. This allows us to turn
8768 // (float)((double)X+2.0) into x+2.0f.
8769 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8770 if (CFP->getType() == Type::PPC_FP128Ty)
8771 return V; // No constant folding of this.
8772 // See if the value can be truncated to float and then reextended.
8773 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8775 if (CFP->getType() == Type::DoubleTy)
8776 return V; // Won't shrink.
8777 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8779 // Don't try to shrink to various long double types.
8785 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8786 if (Instruction *I = commonCastTransforms(CI))
8789 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8790 // smaller than the destination type, we can eliminate the truncate by doing
8791 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8792 // many builtins (sqrt, etc).
8793 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8794 if (OpI && OpI->hasOneUse()) {
8795 switch (OpI->getOpcode()) {
8797 case Instruction::FAdd:
8798 case Instruction::FSub:
8799 case Instruction::FMul:
8800 case Instruction::FDiv:
8801 case Instruction::FRem:
8802 const Type *SrcTy = OpI->getType();
8803 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8804 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8805 if (LHSTrunc->getType() != SrcTy &&
8806 RHSTrunc->getType() != SrcTy) {
8807 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8808 // If the source types were both smaller than the destination type of
8809 // the cast, do this xform.
8810 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8811 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8812 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8814 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8816 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8825 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8826 return commonCastTransforms(CI);
8829 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8830 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8832 return commonCastTransforms(FI);
8834 // fptoui(uitofp(X)) --> X
8835 // fptoui(sitofp(X)) --> X
8836 // This is safe if the intermediate type has enough bits in its mantissa to
8837 // accurately represent all values of X. For example, do not do this with
8838 // i64->float->i64. This is also safe for sitofp case, because any negative
8839 // 'X' value would cause an undefined result for the fptoui.
8840 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8841 OpI->getOperand(0)->getType() == FI.getType() &&
8842 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8843 OpI->getType()->getFPMantissaWidth())
8844 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8846 return commonCastTransforms(FI);
8849 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8850 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8852 return commonCastTransforms(FI);
8854 // fptosi(sitofp(X)) --> X
8855 // fptosi(uitofp(X)) --> X
8856 // This is safe if the intermediate type has enough bits in its mantissa to
8857 // accurately represent all values of X. For example, do not do this with
8858 // i64->float->i64. This is also safe for sitofp case, because any negative
8859 // 'X' value would cause an undefined result for the fptoui.
8860 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8861 OpI->getOperand(0)->getType() == FI.getType() &&
8862 (int)FI.getType()->getScalarSizeInBits() <=
8863 OpI->getType()->getFPMantissaWidth())
8864 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8866 return commonCastTransforms(FI);
8869 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8870 return commonCastTransforms(CI);
8873 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8874 return commonCastTransforms(CI);
8877 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8878 // If the destination integer type is smaller than the intptr_t type for
8879 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8880 // trunc to be exposed to other transforms. Don't do this for extending
8881 // ptrtoint's, because we don't know if the target sign or zero extends its
8883 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8884 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8885 TD->getIntPtrType(),
8887 return new TruncInst(P, CI.getType());
8890 return commonPointerCastTransforms(CI);
8893 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8894 // If the source integer type is larger than the intptr_t type for
8895 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8896 // allows the trunc to be exposed to other transforms. Don't do this for
8897 // extending inttoptr's, because we don't know if the target sign or zero
8898 // extends to pointers.
8899 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
8900 TD->getPointerSizeInBits()) {
8901 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8902 TD->getIntPtrType(),
8904 return new IntToPtrInst(P, CI.getType());
8907 if (Instruction *I = commonCastTransforms(CI))
8910 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8911 if (!DestPointee->isSized()) return 0;
8913 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8916 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8917 m_ConstantInt(Cst)), *Context)) {
8918 // If the source and destination operands have the same type, see if this
8919 // is a single-index GEP.
8920 if (X->getType() == CI.getType()) {
8921 // Get the size of the pointee type.
8922 uint64_t Size = TD->getTypeAllocSize(DestPointee);
8924 // Convert the constant to intptr type.
8925 APInt Offset = Cst->getValue();
8926 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8928 // If Offset is evenly divisible by Size, we can do this xform.
8929 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8930 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8931 GetElementPtrInst *GEP =
8932 GetElementPtrInst::Create(X, Context->getConstantInt(Offset));
8933 // A gep synthesized from inttoptr+add+ptrtoint must be assumed to
8934 // potentially overflow, in the absense of further analysis.
8935 cast<GEPOperator>(GEP)->setHasNoPointerOverflow(false);
8939 // TODO: Could handle other cases, e.g. where add is indexing into field of
8941 } else if (CI.getOperand(0)->hasOneUse() &&
8942 match(CI.getOperand(0), m_Add(m_Value(X),
8943 m_ConstantInt(Cst)), *Context)) {
8944 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8945 // "inttoptr+GEP" instead of "add+intptr".
8947 // Get the size of the pointee type.
8948 uint64_t Size = TD->getTypeAllocSize(DestPointee);
8950 // Convert the constant to intptr type.
8951 APInt Offset = Cst->getValue();
8952 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8954 // If Offset is evenly divisible by Size, we can do this xform.
8955 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8956 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8958 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8960 GetElementPtrInst *GEP =
8961 GetElementPtrInst::Create(P, Context->getConstantInt(Offset), "tmp");
8962 // A gep synthesized from inttoptr+add+ptrtoint must be assumed to
8963 // potentially overflow, in the absense of further analysis.
8964 cast<GEPOperator>(GEP)->setHasNoPointerOverflow(false);
8971 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8972 // If the operands are integer typed then apply the integer transforms,
8973 // otherwise just apply the common ones.
8974 Value *Src = CI.getOperand(0);
8975 const Type *SrcTy = Src->getType();
8976 const Type *DestTy = CI.getType();
8978 if (isa<PointerType>(SrcTy)) {
8979 if (Instruction *I = commonPointerCastTransforms(CI))
8982 if (Instruction *Result = commonCastTransforms(CI))
8987 // Get rid of casts from one type to the same type. These are useless and can
8988 // be replaced by the operand.
8989 if (DestTy == Src->getType())
8990 return ReplaceInstUsesWith(CI, Src);
8992 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
8993 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
8994 const Type *DstElTy = DstPTy->getElementType();
8995 const Type *SrcElTy = SrcPTy->getElementType();
8997 // If the address spaces don't match, don't eliminate the bitcast, which is
8998 // required for changing types.
8999 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
9002 // If we are casting a malloc or alloca to a pointer to a type of the same
9003 // size, rewrite the allocation instruction to allocate the "right" type.
9004 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
9005 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
9008 // If the source and destination are pointers, and this cast is equivalent
9009 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
9010 // This can enhance SROA and other transforms that want type-safe pointers.
9011 Constant *ZeroUInt = Context->getNullValue(Type::Int32Ty);
9012 unsigned NumZeros = 0;
9013 while (SrcElTy != DstElTy &&
9014 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
9015 SrcElTy->getNumContainedTypes() /* not "{}" */) {
9016 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
9020 // If we found a path from the src to dest, create the getelementptr now.
9021 if (SrcElTy == DstElTy) {
9022 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
9023 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
9024 ((Instruction*) NULL));
9028 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9029 if (SVI->hasOneUse()) {
9030 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9031 // a bitconvert to a vector with the same # elts.
9032 if (isa<VectorType>(DestTy) &&
9033 cast<VectorType>(DestTy)->getNumElements() ==
9034 SVI->getType()->getNumElements() &&
9035 SVI->getType()->getNumElements() ==
9036 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9038 // If either of the operands is a cast from CI.getType(), then
9039 // evaluating the shuffle in the casted destination's type will allow
9040 // us to eliminate at least one cast.
9041 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9042 Tmp->getOperand(0)->getType() == DestTy) ||
9043 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9044 Tmp->getOperand(0)->getType() == DestTy)) {
9045 Value *LHS = InsertCastBefore(Instruction::BitCast,
9046 SVI->getOperand(0), DestTy, CI);
9047 Value *RHS = InsertCastBefore(Instruction::BitCast,
9048 SVI->getOperand(1), DestTy, CI);
9049 // Return a new shuffle vector. Use the same element ID's, as we
9050 // know the vector types match #elts.
9051 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9059 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9061 /// %D = select %cond, %C, %A
9063 /// %C = select %cond, %B, 0
9066 /// Assuming that the specified instruction is an operand to the select, return
9067 /// a bitmask indicating which operands of this instruction are foldable if they
9068 /// equal the other incoming value of the select.
9070 static unsigned GetSelectFoldableOperands(Instruction *I) {
9071 switch (I->getOpcode()) {
9072 case Instruction::Add:
9073 case Instruction::Mul:
9074 case Instruction::And:
9075 case Instruction::Or:
9076 case Instruction::Xor:
9077 return 3; // Can fold through either operand.
9078 case Instruction::Sub: // Can only fold on the amount subtracted.
9079 case Instruction::Shl: // Can only fold on the shift amount.
9080 case Instruction::LShr:
9081 case Instruction::AShr:
9084 return 0; // Cannot fold
9088 /// GetSelectFoldableConstant - For the same transformation as the previous
9089 /// function, return the identity constant that goes into the select.
9090 static Constant *GetSelectFoldableConstant(Instruction *I,
9091 LLVMContext *Context) {
9092 switch (I->getOpcode()) {
9093 default: llvm_unreachable("This cannot happen!");
9094 case Instruction::Add:
9095 case Instruction::Sub:
9096 case Instruction::Or:
9097 case Instruction::Xor:
9098 case Instruction::Shl:
9099 case Instruction::LShr:
9100 case Instruction::AShr:
9101 return Context->getNullValue(I->getType());
9102 case Instruction::And:
9103 return Context->getAllOnesValue(I->getType());
9104 case Instruction::Mul:
9105 return Context->getConstantInt(I->getType(), 1);
9109 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9110 /// have the same opcode and only one use each. Try to simplify this.
9111 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9113 if (TI->getNumOperands() == 1) {
9114 // If this is a non-volatile load or a cast from the same type,
9117 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9120 return 0; // unknown unary op.
9123 // Fold this by inserting a select from the input values.
9124 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9125 FI->getOperand(0), SI.getName()+".v");
9126 InsertNewInstBefore(NewSI, SI);
9127 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9131 // Only handle binary operators here.
9132 if (!isa<BinaryOperator>(TI))
9135 // Figure out if the operations have any operands in common.
9136 Value *MatchOp, *OtherOpT, *OtherOpF;
9138 if (TI->getOperand(0) == FI->getOperand(0)) {
9139 MatchOp = TI->getOperand(0);
9140 OtherOpT = TI->getOperand(1);
9141 OtherOpF = FI->getOperand(1);
9142 MatchIsOpZero = true;
9143 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9144 MatchOp = TI->getOperand(1);
9145 OtherOpT = TI->getOperand(0);
9146 OtherOpF = FI->getOperand(0);
9147 MatchIsOpZero = false;
9148 } else if (!TI->isCommutative()) {
9150 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9151 MatchOp = TI->getOperand(0);
9152 OtherOpT = TI->getOperand(1);
9153 OtherOpF = FI->getOperand(0);
9154 MatchIsOpZero = true;
9155 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9156 MatchOp = TI->getOperand(1);
9157 OtherOpT = TI->getOperand(0);
9158 OtherOpF = FI->getOperand(1);
9159 MatchIsOpZero = true;
9164 // If we reach here, they do have operations in common.
9165 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9166 OtherOpF, SI.getName()+".v");
9167 InsertNewInstBefore(NewSI, SI);
9169 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9171 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9173 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9175 llvm_unreachable("Shouldn't get here");
9179 static bool isSelect01(Constant *C1, Constant *C2) {
9180 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9183 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9186 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9189 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9190 /// facilitate further optimization.
9191 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9193 // See the comment above GetSelectFoldableOperands for a description of the
9194 // transformation we are doing here.
9195 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9196 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9197 !isa<Constant>(FalseVal)) {
9198 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9199 unsigned OpToFold = 0;
9200 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9202 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9207 Constant *C = GetSelectFoldableConstant(TVI, Context);
9208 Value *OOp = TVI->getOperand(2-OpToFold);
9209 // Avoid creating select between 2 constants unless it's selecting
9211 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9212 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9213 InsertNewInstBefore(NewSel, SI);
9214 NewSel->takeName(TVI);
9215 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9216 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9217 llvm_unreachable("Unknown instruction!!");
9224 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9225 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9226 !isa<Constant>(TrueVal)) {
9227 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9228 unsigned OpToFold = 0;
9229 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9231 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9236 Constant *C = GetSelectFoldableConstant(FVI, Context);
9237 Value *OOp = FVI->getOperand(2-OpToFold);
9238 // Avoid creating select between 2 constants unless it's selecting
9240 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9241 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9242 InsertNewInstBefore(NewSel, SI);
9243 NewSel->takeName(FVI);
9244 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9245 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9246 llvm_unreachable("Unknown instruction!!");
9256 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9257 /// ICmpInst as its first operand.
9259 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9261 bool Changed = false;
9262 ICmpInst::Predicate Pred = ICI->getPredicate();
9263 Value *CmpLHS = ICI->getOperand(0);
9264 Value *CmpRHS = ICI->getOperand(1);
9265 Value *TrueVal = SI.getTrueValue();
9266 Value *FalseVal = SI.getFalseValue();
9268 // Check cases where the comparison is with a constant that
9269 // can be adjusted to fit the min/max idiom. We may edit ICI in
9270 // place here, so make sure the select is the only user.
9271 if (ICI->hasOneUse())
9272 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9275 case ICmpInst::ICMP_ULT:
9276 case ICmpInst::ICMP_SLT: {
9277 // X < MIN ? T : F --> F
9278 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9279 return ReplaceInstUsesWith(SI, FalseVal);
9280 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9281 Constant *AdjustedRHS = SubOne(CI, Context);
9282 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9283 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9284 Pred = ICmpInst::getSwappedPredicate(Pred);
9285 CmpRHS = AdjustedRHS;
9286 std::swap(FalseVal, TrueVal);
9287 ICI->setPredicate(Pred);
9288 ICI->setOperand(1, CmpRHS);
9289 SI.setOperand(1, TrueVal);
9290 SI.setOperand(2, FalseVal);
9295 case ICmpInst::ICMP_UGT:
9296 case ICmpInst::ICMP_SGT: {
9297 // X > MAX ? T : F --> F
9298 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9299 return ReplaceInstUsesWith(SI, FalseVal);
9300 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9301 Constant *AdjustedRHS = AddOne(CI, Context);
9302 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9303 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9304 Pred = ICmpInst::getSwappedPredicate(Pred);
9305 CmpRHS = AdjustedRHS;
9306 std::swap(FalseVal, TrueVal);
9307 ICI->setPredicate(Pred);
9308 ICI->setOperand(1, CmpRHS);
9309 SI.setOperand(1, TrueVal);
9310 SI.setOperand(2, FalseVal);
9317 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9318 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9319 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9320 if (match(TrueVal, m_ConstantInt<-1>(), *Context) &&
9321 match(FalseVal, m_ConstantInt<0>(), *Context))
9322 Pred = ICI->getPredicate();
9323 else if (match(TrueVal, m_ConstantInt<0>(), *Context) &&
9324 match(FalseVal, m_ConstantInt<-1>(), *Context))
9325 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9327 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9328 // If we are just checking for a icmp eq of a single bit and zext'ing it
9329 // to an integer, then shift the bit to the appropriate place and then
9330 // cast to integer to avoid the comparison.
9331 const APInt &Op1CV = CI->getValue();
9333 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9334 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9335 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9336 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9337 Value *In = ICI->getOperand(0);
9338 Value *Sh = Context->getConstantInt(In->getType(),
9339 In->getType()->getScalarSizeInBits()-1);
9340 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9341 In->getName()+".lobit"),
9343 if (In->getType() != SI.getType())
9344 In = CastInst::CreateIntegerCast(In, SI.getType(),
9345 true/*SExt*/, "tmp", ICI);
9347 if (Pred == ICmpInst::ICMP_SGT)
9348 In = InsertNewInstBefore(BinaryOperator::CreateNot(*Context, In,
9349 In->getName()+".not"), *ICI);
9351 return ReplaceInstUsesWith(SI, In);
9356 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9357 // Transform (X == Y) ? X : Y -> Y
9358 if (Pred == ICmpInst::ICMP_EQ)
9359 return ReplaceInstUsesWith(SI, FalseVal);
9360 // Transform (X != Y) ? X : Y -> X
9361 if (Pred == ICmpInst::ICMP_NE)
9362 return ReplaceInstUsesWith(SI, TrueVal);
9363 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9365 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9366 // Transform (X == Y) ? Y : X -> X
9367 if (Pred == ICmpInst::ICMP_EQ)
9368 return ReplaceInstUsesWith(SI, FalseVal);
9369 // Transform (X != Y) ? Y : X -> Y
9370 if (Pred == ICmpInst::ICMP_NE)
9371 return ReplaceInstUsesWith(SI, TrueVal);
9372 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9375 /// NOTE: if we wanted to, this is where to detect integer ABS
9377 return Changed ? &SI : 0;
9380 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9381 Value *CondVal = SI.getCondition();
9382 Value *TrueVal = SI.getTrueValue();
9383 Value *FalseVal = SI.getFalseValue();
9385 // select true, X, Y -> X
9386 // select false, X, Y -> Y
9387 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9388 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9390 // select C, X, X -> X
9391 if (TrueVal == FalseVal)
9392 return ReplaceInstUsesWith(SI, TrueVal);
9394 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9395 return ReplaceInstUsesWith(SI, FalseVal);
9396 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9397 return ReplaceInstUsesWith(SI, TrueVal);
9398 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9399 if (isa<Constant>(TrueVal))
9400 return ReplaceInstUsesWith(SI, TrueVal);
9402 return ReplaceInstUsesWith(SI, FalseVal);
9405 if (SI.getType() == Type::Int1Ty) {
9406 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9407 if (C->getZExtValue()) {
9408 // Change: A = select B, true, C --> A = or B, C
9409 return BinaryOperator::CreateOr(CondVal, FalseVal);
9411 // Change: A = select B, false, C --> A = and !B, C
9413 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9414 "not."+CondVal->getName()), SI);
9415 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9417 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9418 if (C->getZExtValue() == false) {
9419 // Change: A = select B, C, false --> A = and B, C
9420 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9422 // Change: A = select B, C, true --> A = or !B, C
9424 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9425 "not."+CondVal->getName()), SI);
9426 return BinaryOperator::CreateOr(NotCond, TrueVal);
9430 // select a, b, a -> a&b
9431 // select a, a, b -> a|b
9432 if (CondVal == TrueVal)
9433 return BinaryOperator::CreateOr(CondVal, FalseVal);
9434 else if (CondVal == FalseVal)
9435 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9438 // Selecting between two integer constants?
9439 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9440 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9441 // select C, 1, 0 -> zext C to int
9442 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9443 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9444 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9445 // select C, 0, 1 -> zext !C to int
9447 InsertNewInstBefore(BinaryOperator::CreateNot(*Context, CondVal,
9448 "not."+CondVal->getName()), SI);
9449 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9452 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9453 // If one of the constants is zero (we know they can't both be) and we
9454 // have an icmp instruction with zero, and we have an 'and' with the
9455 // non-constant value, eliminate this whole mess. This corresponds to
9456 // cases like this: ((X & 27) ? 27 : 0)
9457 if (TrueValC->isZero() || FalseValC->isZero())
9458 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9459 cast<Constant>(IC->getOperand(1))->isNullValue())
9460 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9461 if (ICA->getOpcode() == Instruction::And &&
9462 isa<ConstantInt>(ICA->getOperand(1)) &&
9463 (ICA->getOperand(1) == TrueValC ||
9464 ICA->getOperand(1) == FalseValC) &&
9465 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9466 // Okay, now we know that everything is set up, we just don't
9467 // know whether we have a icmp_ne or icmp_eq and whether the
9468 // true or false val is the zero.
9469 bool ShouldNotVal = !TrueValC->isZero();
9470 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9473 V = InsertNewInstBefore(BinaryOperator::Create(
9474 Instruction::Xor, V, ICA->getOperand(1)), SI);
9475 return ReplaceInstUsesWith(SI, V);
9480 // See if we are selecting two values based on a comparison of the two values.
9481 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9482 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9483 // Transform (X == Y) ? X : Y -> Y
9484 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9485 // This is not safe in general for floating point:
9486 // consider X== -0, Y== +0.
9487 // It becomes safe if either operand is a nonzero constant.
9488 ConstantFP *CFPt, *CFPf;
9489 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9490 !CFPt->getValueAPF().isZero()) ||
9491 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9492 !CFPf->getValueAPF().isZero()))
9493 return ReplaceInstUsesWith(SI, FalseVal);
9495 // Transform (X != Y) ? X : Y -> X
9496 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9497 return ReplaceInstUsesWith(SI, TrueVal);
9498 // NOTE: if we wanted to, this is where to detect MIN/MAX
9500 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9501 // Transform (X == Y) ? Y : X -> X
9502 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9503 // This is not safe in general for floating point:
9504 // consider X== -0, Y== +0.
9505 // It becomes safe if either operand is a nonzero constant.
9506 ConstantFP *CFPt, *CFPf;
9507 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9508 !CFPt->getValueAPF().isZero()) ||
9509 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9510 !CFPf->getValueAPF().isZero()))
9511 return ReplaceInstUsesWith(SI, FalseVal);
9513 // Transform (X != Y) ? Y : X -> Y
9514 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9515 return ReplaceInstUsesWith(SI, TrueVal);
9516 // NOTE: if we wanted to, this is where to detect MIN/MAX
9518 // NOTE: if we wanted to, this is where to detect ABS
9521 // See if we are selecting two values based on a comparison of the two values.
9522 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9523 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9526 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9527 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9528 if (TI->hasOneUse() && FI->hasOneUse()) {
9529 Instruction *AddOp = 0, *SubOp = 0;
9531 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9532 if (TI->getOpcode() == FI->getOpcode())
9533 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9536 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9537 // even legal for FP.
9538 if ((TI->getOpcode() == Instruction::Sub &&
9539 FI->getOpcode() == Instruction::Add) ||
9540 (TI->getOpcode() == Instruction::FSub &&
9541 FI->getOpcode() == Instruction::FAdd)) {
9542 AddOp = FI; SubOp = TI;
9543 } else if ((FI->getOpcode() == Instruction::Sub &&
9544 TI->getOpcode() == Instruction::Add) ||
9545 (FI->getOpcode() == Instruction::FSub &&
9546 TI->getOpcode() == Instruction::FAdd)) {
9547 AddOp = TI; SubOp = FI;
9551 Value *OtherAddOp = 0;
9552 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9553 OtherAddOp = AddOp->getOperand(1);
9554 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9555 OtherAddOp = AddOp->getOperand(0);
9559 // So at this point we know we have (Y -> OtherAddOp):
9560 // select C, (add X, Y), (sub X, Z)
9561 Value *NegVal; // Compute -Z
9562 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9563 NegVal = Context->getConstantExprNeg(C);
9565 NegVal = InsertNewInstBefore(
9566 BinaryOperator::CreateNeg(*Context, SubOp->getOperand(1),
9570 Value *NewTrueOp = OtherAddOp;
9571 Value *NewFalseOp = NegVal;
9573 std::swap(NewTrueOp, NewFalseOp);
9574 Instruction *NewSel =
9575 SelectInst::Create(CondVal, NewTrueOp,
9576 NewFalseOp, SI.getName() + ".p");
9578 NewSel = InsertNewInstBefore(NewSel, SI);
9579 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9584 // See if we can fold the select into one of our operands.
9585 if (SI.getType()->isInteger()) {
9586 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9591 if (BinaryOperator::isNot(CondVal)) {
9592 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9593 SI.setOperand(1, FalseVal);
9594 SI.setOperand(2, TrueVal);
9601 /// EnforceKnownAlignment - If the specified pointer points to an object that
9602 /// we control, modify the object's alignment to PrefAlign. This isn't
9603 /// often possible though. If alignment is important, a more reliable approach
9604 /// is to simply align all global variables and allocation instructions to
9605 /// their preferred alignment from the beginning.
9607 static unsigned EnforceKnownAlignment(Value *V,
9608 unsigned Align, unsigned PrefAlign) {
9610 User *U = dyn_cast<User>(V);
9611 if (!U) return Align;
9613 switch (Operator::getOpcode(U)) {
9615 case Instruction::BitCast:
9616 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9617 case Instruction::GetElementPtr: {
9618 // If all indexes are zero, it is just the alignment of the base pointer.
9619 bool AllZeroOperands = true;
9620 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9621 if (!isa<Constant>(*i) ||
9622 !cast<Constant>(*i)->isNullValue()) {
9623 AllZeroOperands = false;
9627 if (AllZeroOperands) {
9628 // Treat this like a bitcast.
9629 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9635 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9636 // If there is a large requested alignment and we can, bump up the alignment
9638 if (!GV->isDeclaration()) {
9639 if (GV->getAlignment() >= PrefAlign)
9640 Align = GV->getAlignment();
9642 GV->setAlignment(PrefAlign);
9646 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9647 // If there is a requested alignment and if this is an alloca, round up. We
9648 // don't do this for malloc, because some systems can't respect the request.
9649 if (isa<AllocaInst>(AI)) {
9650 if (AI->getAlignment() >= PrefAlign)
9651 Align = AI->getAlignment();
9653 AI->setAlignment(PrefAlign);
9662 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9663 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9664 /// and it is more than the alignment of the ultimate object, see if we can
9665 /// increase the alignment of the ultimate object, making this check succeed.
9666 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9667 unsigned PrefAlign) {
9668 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9669 sizeof(PrefAlign) * CHAR_BIT;
9670 APInt Mask = APInt::getAllOnesValue(BitWidth);
9671 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9672 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9673 unsigned TrailZ = KnownZero.countTrailingOnes();
9674 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9676 if (PrefAlign > Align)
9677 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9679 // We don't need to make any adjustment.
9683 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9684 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9685 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9686 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9687 unsigned CopyAlign = MI->getAlignment();
9689 if (CopyAlign < MinAlign) {
9690 MI->setAlignment(Context->getConstantInt(MI->getAlignmentType(),
9695 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9697 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9698 if (MemOpLength == 0) return 0;
9700 // Source and destination pointer types are always "i8*" for intrinsic. See
9701 // if the size is something we can handle with a single primitive load/store.
9702 // A single load+store correctly handles overlapping memory in the memmove
9704 unsigned Size = MemOpLength->getZExtValue();
9705 if (Size == 0) return MI; // Delete this mem transfer.
9707 if (Size > 8 || (Size&(Size-1)))
9708 return 0; // If not 1/2/4/8 bytes, exit.
9710 // Use an integer load+store unless we can find something better.
9712 Context->getPointerTypeUnqual(Context->getIntegerType(Size<<3));
9714 // Memcpy forces the use of i8* for the source and destination. That means
9715 // that if you're using memcpy to move one double around, you'll get a cast
9716 // from double* to i8*. We'd much rather use a double load+store rather than
9717 // an i64 load+store, here because this improves the odds that the source or
9718 // dest address will be promotable. See if we can find a better type than the
9719 // integer datatype.
9720 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9721 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9722 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9723 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9724 // down through these levels if so.
9725 while (!SrcETy->isSingleValueType()) {
9726 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9727 if (STy->getNumElements() == 1)
9728 SrcETy = STy->getElementType(0);
9731 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9732 if (ATy->getNumElements() == 1)
9733 SrcETy = ATy->getElementType();
9740 if (SrcETy->isSingleValueType())
9741 NewPtrTy = Context->getPointerTypeUnqual(SrcETy);
9746 // If the memcpy/memmove provides better alignment info than we can
9748 SrcAlign = std::max(SrcAlign, CopyAlign);
9749 DstAlign = std::max(DstAlign, CopyAlign);
9751 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9752 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9753 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9754 InsertNewInstBefore(L, *MI);
9755 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9757 // Set the size of the copy to 0, it will be deleted on the next iteration.
9758 MI->setOperand(3, Context->getNullValue(MemOpLength->getType()));
9762 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9763 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9764 if (MI->getAlignment() < Alignment) {
9765 MI->setAlignment(Context->getConstantInt(MI->getAlignmentType(),
9770 // Extract the length and alignment and fill if they are constant.
9771 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9772 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9773 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9775 uint64_t Len = LenC->getZExtValue();
9776 Alignment = MI->getAlignment();
9778 // If the length is zero, this is a no-op
9779 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9781 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9782 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9783 const Type *ITy = Context->getIntegerType(Len*8); // n=1 -> i8.
9785 Value *Dest = MI->getDest();
9786 Dest = InsertBitCastBefore(Dest, Context->getPointerTypeUnqual(ITy), *MI);
9788 // Alignment 0 is identity for alignment 1 for memset, but not store.
9789 if (Alignment == 0) Alignment = 1;
9791 // Extract the fill value and store.
9792 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9793 InsertNewInstBefore(new StoreInst(Context->getConstantInt(ITy, Fill),
9794 Dest, false, Alignment), *MI);
9796 // Set the size of the copy to 0, it will be deleted on the next iteration.
9797 MI->setLength(Context->getNullValue(LenC->getType()));
9805 /// visitCallInst - CallInst simplification. This mostly only handles folding
9806 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9807 /// the heavy lifting.
9809 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9810 // If the caller function is nounwind, mark the call as nounwind, even if the
9812 if (CI.getParent()->getParent()->doesNotThrow() &&
9813 !CI.doesNotThrow()) {
9814 CI.setDoesNotThrow();
9820 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9821 if (!II) return visitCallSite(&CI);
9823 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9825 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9826 bool Changed = false;
9828 // memmove/cpy/set of zero bytes is a noop.
9829 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9830 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9832 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9833 if (CI->getZExtValue() == 1) {
9834 // Replace the instruction with just byte operations. We would
9835 // transform other cases to loads/stores, but we don't know if
9836 // alignment is sufficient.
9840 // If we have a memmove and the source operation is a constant global,
9841 // then the source and dest pointers can't alias, so we can change this
9842 // into a call to memcpy.
9843 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9844 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9845 if (GVSrc->isConstant()) {
9846 Module *M = CI.getParent()->getParent()->getParent();
9847 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9849 Tys[0] = CI.getOperand(3)->getType();
9851 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9855 // memmove(x,x,size) -> noop.
9856 if (MMI->getSource() == MMI->getDest())
9857 return EraseInstFromFunction(CI);
9860 // If we can determine a pointer alignment that is bigger than currently
9861 // set, update the alignment.
9862 if (isa<MemTransferInst>(MI)) {
9863 if (Instruction *I = SimplifyMemTransfer(MI))
9865 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9866 if (Instruction *I = SimplifyMemSet(MSI))
9870 if (Changed) return II;
9873 switch (II->getIntrinsicID()) {
9875 case Intrinsic::bswap:
9876 // bswap(bswap(x)) -> x
9877 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9878 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9879 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9881 case Intrinsic::ppc_altivec_lvx:
9882 case Intrinsic::ppc_altivec_lvxl:
9883 case Intrinsic::x86_sse_loadu_ps:
9884 case Intrinsic::x86_sse2_loadu_pd:
9885 case Intrinsic::x86_sse2_loadu_dq:
9886 // Turn PPC lvx -> load if the pointer is known aligned.
9887 // Turn X86 loadups -> load if the pointer is known aligned.
9888 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9889 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9890 Context->getPointerTypeUnqual(II->getType()),
9892 return new LoadInst(Ptr);
9895 case Intrinsic::ppc_altivec_stvx:
9896 case Intrinsic::ppc_altivec_stvxl:
9897 // Turn stvx -> store if the pointer is known aligned.
9898 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9899 const Type *OpPtrTy =
9900 Context->getPointerTypeUnqual(II->getOperand(1)->getType());
9901 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9902 return new StoreInst(II->getOperand(1), Ptr);
9905 case Intrinsic::x86_sse_storeu_ps:
9906 case Intrinsic::x86_sse2_storeu_pd:
9907 case Intrinsic::x86_sse2_storeu_dq:
9908 // Turn X86 storeu -> store if the pointer is known aligned.
9909 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9910 const Type *OpPtrTy =
9911 Context->getPointerTypeUnqual(II->getOperand(2)->getType());
9912 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9913 return new StoreInst(II->getOperand(2), Ptr);
9917 case Intrinsic::x86_sse_cvttss2si: {
9918 // These intrinsics only demands the 0th element of its input vector. If
9919 // we can simplify the input based on that, do so now.
9921 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9922 APInt DemandedElts(VWidth, 1);
9923 APInt UndefElts(VWidth, 0);
9924 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9926 II->setOperand(1, V);
9932 case Intrinsic::ppc_altivec_vperm:
9933 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9934 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9935 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9937 // Check that all of the elements are integer constants or undefs.
9938 bool AllEltsOk = true;
9939 for (unsigned i = 0; i != 16; ++i) {
9940 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9941 !isa<UndefValue>(Mask->getOperand(i))) {
9948 // Cast the input vectors to byte vectors.
9949 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9950 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9951 Value *Result = Context->getUndef(Op0->getType());
9953 // Only extract each element once.
9954 Value *ExtractedElts[32];
9955 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9957 for (unsigned i = 0; i != 16; ++i) {
9958 if (isa<UndefValue>(Mask->getOperand(i)))
9960 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9961 Idx &= 31; // Match the hardware behavior.
9963 if (ExtractedElts[Idx] == 0) {
9965 new ExtractElementInst(Idx < 16 ? Op0 : Op1,
9966 Context->getConstantInt(Type::Int32Ty, Idx&15, false), "tmp");
9967 InsertNewInstBefore(Elt, CI);
9968 ExtractedElts[Idx] = Elt;
9971 // Insert this value into the result vector.
9972 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9973 Context->getConstantInt(Type::Int32Ty, i, false),
9975 InsertNewInstBefore(cast<Instruction>(Result), CI);
9977 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
9982 case Intrinsic::stackrestore: {
9983 // If the save is right next to the restore, remove the restore. This can
9984 // happen when variable allocas are DCE'd.
9985 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
9986 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
9987 BasicBlock::iterator BI = SS;
9989 return EraseInstFromFunction(CI);
9993 // Scan down this block to see if there is another stack restore in the
9994 // same block without an intervening call/alloca.
9995 BasicBlock::iterator BI = II;
9996 TerminatorInst *TI = II->getParent()->getTerminator();
9997 bool CannotRemove = false;
9998 for (++BI; &*BI != TI; ++BI) {
9999 if (isa<AllocaInst>(BI)) {
10000 CannotRemove = true;
10003 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
10004 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
10005 // If there is a stackrestore below this one, remove this one.
10006 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10007 return EraseInstFromFunction(CI);
10008 // Otherwise, ignore the intrinsic.
10010 // If we found a non-intrinsic call, we can't remove the stack
10012 CannotRemove = true;
10018 // If the stack restore is in a return/unwind block and if there are no
10019 // allocas or calls between the restore and the return, nuke the restore.
10020 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10021 return EraseInstFromFunction(CI);
10026 return visitCallSite(II);
10029 // InvokeInst simplification
10031 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10032 return visitCallSite(&II);
10035 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10036 /// passed through the varargs area, we can eliminate the use of the cast.
10037 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10038 const CastInst * const CI,
10039 const TargetData * const TD,
10041 if (!CI->isLosslessCast())
10044 // The size of ByVal arguments is derived from the type, so we
10045 // can't change to a type with a different size. If the size were
10046 // passed explicitly we could avoid this check.
10047 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10050 const Type* SrcTy =
10051 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10052 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10053 if (!SrcTy->isSized() || !DstTy->isSized())
10055 if (TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10060 // visitCallSite - Improvements for call and invoke instructions.
10062 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10063 bool Changed = false;
10065 // If the callee is a constexpr cast of a function, attempt to move the cast
10066 // to the arguments of the call/invoke.
10067 if (transformConstExprCastCall(CS)) return 0;
10069 Value *Callee = CS.getCalledValue();
10071 if (Function *CalleeF = dyn_cast<Function>(Callee))
10072 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10073 Instruction *OldCall = CS.getInstruction();
10074 // If the call and callee calling conventions don't match, this call must
10075 // be unreachable, as the call is undefined.
10076 new StoreInst(Context->getConstantIntTrue(),
10077 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)),
10079 if (!OldCall->use_empty())
10080 OldCall->replaceAllUsesWith(Context->getUndef(OldCall->getType()));
10081 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10082 return EraseInstFromFunction(*OldCall);
10086 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10087 // This instruction is not reachable, just remove it. We insert a store to
10088 // undef so that we know that this code is not reachable, despite the fact
10089 // that we can't modify the CFG here.
10090 new StoreInst(Context->getConstantIntTrue(),
10091 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)),
10092 CS.getInstruction());
10094 if (!CS.getInstruction()->use_empty())
10095 CS.getInstruction()->
10096 replaceAllUsesWith(Context->getUndef(CS.getInstruction()->getType()));
10098 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10099 // Don't break the CFG, insert a dummy cond branch.
10100 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10101 Context->getConstantIntTrue(), II);
10103 return EraseInstFromFunction(*CS.getInstruction());
10106 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10107 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10108 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10109 return transformCallThroughTrampoline(CS);
10111 const PointerType *PTy = cast<PointerType>(Callee->getType());
10112 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10113 if (FTy->isVarArg()) {
10114 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10115 // See if we can optimize any arguments passed through the varargs area of
10117 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10118 E = CS.arg_end(); I != E; ++I, ++ix) {
10119 CastInst *CI = dyn_cast<CastInst>(*I);
10120 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10121 *I = CI->getOperand(0);
10127 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10128 // Inline asm calls cannot throw - mark them 'nounwind'.
10129 CS.setDoesNotThrow();
10133 return Changed ? CS.getInstruction() : 0;
10136 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10137 // attempt to move the cast to the arguments of the call/invoke.
10139 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10140 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10141 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10142 if (CE->getOpcode() != Instruction::BitCast ||
10143 !isa<Function>(CE->getOperand(0)))
10145 Function *Callee = cast<Function>(CE->getOperand(0));
10146 Instruction *Caller = CS.getInstruction();
10147 const AttrListPtr &CallerPAL = CS.getAttributes();
10149 // Okay, this is a cast from a function to a different type. Unless doing so
10150 // would cause a type conversion of one of our arguments, change this call to
10151 // be a direct call with arguments casted to the appropriate types.
10153 const FunctionType *FT = Callee->getFunctionType();
10154 const Type *OldRetTy = Caller->getType();
10155 const Type *NewRetTy = FT->getReturnType();
10157 if (isa<StructType>(NewRetTy))
10158 return false; // TODO: Handle multiple return values.
10160 // Check to see if we are changing the return type...
10161 if (OldRetTy != NewRetTy) {
10162 if (Callee->isDeclaration() &&
10163 // Conversion is ok if changing from one pointer type to another or from
10164 // a pointer to an integer of the same size.
10165 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
10166 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
10167 return false; // Cannot transform this return value.
10169 if (!Caller->use_empty() &&
10170 // void -> non-void is handled specially
10171 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
10172 return false; // Cannot transform this return value.
10174 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10175 Attributes RAttrs = CallerPAL.getRetAttributes();
10176 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10177 return false; // Attribute not compatible with transformed value.
10180 // If the callsite is an invoke instruction, and the return value is used by
10181 // a PHI node in a successor, we cannot change the return type of the call
10182 // because there is no place to put the cast instruction (without breaking
10183 // the critical edge). Bail out in this case.
10184 if (!Caller->use_empty())
10185 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10186 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10188 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10189 if (PN->getParent() == II->getNormalDest() ||
10190 PN->getParent() == II->getUnwindDest())
10194 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10195 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10197 CallSite::arg_iterator AI = CS.arg_begin();
10198 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10199 const Type *ParamTy = FT->getParamType(i);
10200 const Type *ActTy = (*AI)->getType();
10202 if (!CastInst::isCastable(ActTy, ParamTy))
10203 return false; // Cannot transform this parameter value.
10205 if (CallerPAL.getParamAttributes(i + 1)
10206 & Attribute::typeIncompatible(ParamTy))
10207 return false; // Attribute not compatible with transformed value.
10209 // Converting from one pointer type to another or between a pointer and an
10210 // integer of the same size is safe even if we do not have a body.
10211 bool isConvertible = ActTy == ParamTy ||
10212 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
10213 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
10214 if (Callee->isDeclaration() && !isConvertible) return false;
10217 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10218 Callee->isDeclaration())
10219 return false; // Do not delete arguments unless we have a function body.
10221 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10222 !CallerPAL.isEmpty())
10223 // In this case we have more arguments than the new function type, but we
10224 // won't be dropping them. Check that these extra arguments have attributes
10225 // that are compatible with being a vararg call argument.
10226 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10227 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10229 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10230 if (PAttrs & Attribute::VarArgsIncompatible)
10234 // Okay, we decided that this is a safe thing to do: go ahead and start
10235 // inserting cast instructions as necessary...
10236 std::vector<Value*> Args;
10237 Args.reserve(NumActualArgs);
10238 SmallVector<AttributeWithIndex, 8> attrVec;
10239 attrVec.reserve(NumCommonArgs);
10241 // Get any return attributes.
10242 Attributes RAttrs = CallerPAL.getRetAttributes();
10244 // If the return value is not being used, the type may not be compatible
10245 // with the existing attributes. Wipe out any problematic attributes.
10246 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10248 // Add the new return attributes.
10250 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10252 AI = CS.arg_begin();
10253 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10254 const Type *ParamTy = FT->getParamType(i);
10255 if ((*AI)->getType() == ParamTy) {
10256 Args.push_back(*AI);
10258 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10259 false, ParamTy, false);
10260 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
10261 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
10264 // Add any parameter attributes.
10265 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10266 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10269 // If the function takes more arguments than the call was taking, add them
10271 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10272 Args.push_back(Context->getNullValue(FT->getParamType(i)));
10274 // If we are removing arguments to the function, emit an obnoxious warning...
10275 if (FT->getNumParams() < NumActualArgs) {
10276 if (!FT->isVarArg()) {
10277 cerr << "WARNING: While resolving call to function '"
10278 << Callee->getName() << "' arguments were dropped!\n";
10280 // Add all of the arguments in their promoted form to the arg list...
10281 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10282 const Type *PTy = getPromotedType((*AI)->getType());
10283 if (PTy != (*AI)->getType()) {
10284 // Must promote to pass through va_arg area!
10285 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
10287 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
10288 InsertNewInstBefore(Cast, *Caller);
10289 Args.push_back(Cast);
10291 Args.push_back(*AI);
10294 // Add any parameter attributes.
10295 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10296 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10301 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10302 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10304 if (NewRetTy == Type::VoidTy)
10305 Caller->setName(""); // Void type should not have a name.
10307 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
10310 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10311 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10312 Args.begin(), Args.end(),
10313 Caller->getName(), Caller);
10314 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10315 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10317 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10318 Caller->getName(), Caller);
10319 CallInst *CI = cast<CallInst>(Caller);
10320 if (CI->isTailCall())
10321 cast<CallInst>(NC)->setTailCall();
10322 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10323 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10326 // Insert a cast of the return type as necessary.
10328 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10329 if (NV->getType() != Type::VoidTy) {
10330 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10332 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10334 // If this is an invoke instruction, we should insert it after the first
10335 // non-phi, instruction in the normal successor block.
10336 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10337 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10338 InsertNewInstBefore(NC, *I);
10340 // Otherwise, it's a call, just insert cast right after the call instr
10341 InsertNewInstBefore(NC, *Caller);
10343 AddUsersToWorkList(*Caller);
10345 NV = Context->getUndef(Caller->getType());
10349 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10350 Caller->replaceAllUsesWith(NV);
10351 Caller->eraseFromParent();
10352 RemoveFromWorkList(Caller);
10356 // transformCallThroughTrampoline - Turn a call to a function created by the
10357 // init_trampoline intrinsic into a direct call to the underlying function.
10359 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10360 Value *Callee = CS.getCalledValue();
10361 const PointerType *PTy = cast<PointerType>(Callee->getType());
10362 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10363 const AttrListPtr &Attrs = CS.getAttributes();
10365 // If the call already has the 'nest' attribute somewhere then give up -
10366 // otherwise 'nest' would occur twice after splicing in the chain.
10367 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10370 IntrinsicInst *Tramp =
10371 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10373 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10374 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10375 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10377 const AttrListPtr &NestAttrs = NestF->getAttributes();
10378 if (!NestAttrs.isEmpty()) {
10379 unsigned NestIdx = 1;
10380 const Type *NestTy = 0;
10381 Attributes NestAttr = Attribute::None;
10383 // Look for a parameter marked with the 'nest' attribute.
10384 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10385 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10386 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10387 // Record the parameter type and any other attributes.
10389 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10394 Instruction *Caller = CS.getInstruction();
10395 std::vector<Value*> NewArgs;
10396 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10398 SmallVector<AttributeWithIndex, 8> NewAttrs;
10399 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10401 // Insert the nest argument into the call argument list, which may
10402 // mean appending it. Likewise for attributes.
10404 // Add any result attributes.
10405 if (Attributes Attr = Attrs.getRetAttributes())
10406 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10410 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10412 if (Idx == NestIdx) {
10413 // Add the chain argument and attributes.
10414 Value *NestVal = Tramp->getOperand(3);
10415 if (NestVal->getType() != NestTy)
10416 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10417 NewArgs.push_back(NestVal);
10418 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10424 // Add the original argument and attributes.
10425 NewArgs.push_back(*I);
10426 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10428 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10434 // Add any function attributes.
10435 if (Attributes Attr = Attrs.getFnAttributes())
10436 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10438 // The trampoline may have been bitcast to a bogus type (FTy).
10439 // Handle this by synthesizing a new function type, equal to FTy
10440 // with the chain parameter inserted.
10442 std::vector<const Type*> NewTypes;
10443 NewTypes.reserve(FTy->getNumParams()+1);
10445 // Insert the chain's type into the list of parameter types, which may
10446 // mean appending it.
10449 FunctionType::param_iterator I = FTy->param_begin(),
10450 E = FTy->param_end();
10453 if (Idx == NestIdx)
10454 // Add the chain's type.
10455 NewTypes.push_back(NestTy);
10460 // Add the original type.
10461 NewTypes.push_back(*I);
10467 // Replace the trampoline call with a direct call. Let the generic
10468 // code sort out any function type mismatches.
10469 FunctionType *NewFTy =
10470 Context->getFunctionType(FTy->getReturnType(), NewTypes,
10472 Constant *NewCallee =
10473 NestF->getType() == Context->getPointerTypeUnqual(NewFTy) ?
10474 NestF : Context->getConstantExprBitCast(NestF,
10475 Context->getPointerTypeUnqual(NewFTy));
10476 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
10478 Instruction *NewCaller;
10479 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10480 NewCaller = InvokeInst::Create(NewCallee,
10481 II->getNormalDest(), II->getUnwindDest(),
10482 NewArgs.begin(), NewArgs.end(),
10483 Caller->getName(), Caller);
10484 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10485 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10487 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10488 Caller->getName(), Caller);
10489 if (cast<CallInst>(Caller)->isTailCall())
10490 cast<CallInst>(NewCaller)->setTailCall();
10491 cast<CallInst>(NewCaller)->
10492 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10493 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10495 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10496 Caller->replaceAllUsesWith(NewCaller);
10497 Caller->eraseFromParent();
10498 RemoveFromWorkList(Caller);
10503 // Replace the trampoline call with a direct call. Since there is no 'nest'
10504 // parameter, there is no need to adjust the argument list. Let the generic
10505 // code sort out any function type mismatches.
10506 Constant *NewCallee =
10507 NestF->getType() == PTy ? NestF :
10508 Context->getConstantExprBitCast(NestF, PTy);
10509 CS.setCalledFunction(NewCallee);
10510 return CS.getInstruction();
10513 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10514 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10515 /// and a single binop.
10516 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10517 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10518 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10519 unsigned Opc = FirstInst->getOpcode();
10520 Value *LHSVal = FirstInst->getOperand(0);
10521 Value *RHSVal = FirstInst->getOperand(1);
10523 const Type *LHSType = LHSVal->getType();
10524 const Type *RHSType = RHSVal->getType();
10526 // Scan to see if all operands are the same opcode, all have one use, and all
10527 // kill their operands (i.e. the operands have one use).
10528 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10529 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10530 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10531 // Verify type of the LHS matches so we don't fold cmp's of different
10532 // types or GEP's with different index types.
10533 I->getOperand(0)->getType() != LHSType ||
10534 I->getOperand(1)->getType() != RHSType)
10537 // If they are CmpInst instructions, check their predicates
10538 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10539 if (cast<CmpInst>(I)->getPredicate() !=
10540 cast<CmpInst>(FirstInst)->getPredicate())
10543 // Keep track of which operand needs a phi node.
10544 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10545 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10548 // Otherwise, this is safe to transform!
10550 Value *InLHS = FirstInst->getOperand(0);
10551 Value *InRHS = FirstInst->getOperand(1);
10552 PHINode *NewLHS = 0, *NewRHS = 0;
10554 NewLHS = PHINode::Create(LHSType,
10555 FirstInst->getOperand(0)->getName() + ".pn");
10556 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10557 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10558 InsertNewInstBefore(NewLHS, PN);
10563 NewRHS = PHINode::Create(RHSType,
10564 FirstInst->getOperand(1)->getName() + ".pn");
10565 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10566 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10567 InsertNewInstBefore(NewRHS, PN);
10571 // Add all operands to the new PHIs.
10572 if (NewLHS || NewRHS) {
10573 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10574 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10576 Value *NewInLHS = InInst->getOperand(0);
10577 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10580 Value *NewInRHS = InInst->getOperand(1);
10581 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10586 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10587 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10588 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10589 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10593 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10594 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10596 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10597 FirstInst->op_end());
10598 // This is true if all GEP bases are allocas and if all indices into them are
10600 bool AllBasePointersAreAllocas = true;
10602 // Scan to see if all operands are the same opcode, all have one use, and all
10603 // kill their operands (i.e. the operands have one use).
10604 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10605 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10606 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10607 GEP->getNumOperands() != FirstInst->getNumOperands())
10610 // Keep track of whether or not all GEPs are of alloca pointers.
10611 if (AllBasePointersAreAllocas &&
10612 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10613 !GEP->hasAllConstantIndices()))
10614 AllBasePointersAreAllocas = false;
10616 // Compare the operand lists.
10617 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10618 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10621 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10622 // if one of the PHIs has a constant for the index. The index may be
10623 // substantially cheaper to compute for the constants, so making it a
10624 // variable index could pessimize the path. This also handles the case
10625 // for struct indices, which must always be constant.
10626 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10627 isa<ConstantInt>(GEP->getOperand(op)))
10630 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10632 FixedOperands[op] = 0; // Needs a PHI.
10636 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10637 // bother doing this transformation. At best, this will just save a bit of
10638 // offset calculation, but all the predecessors will have to materialize the
10639 // stack address into a register anyway. We'd actually rather *clone* the
10640 // load up into the predecessors so that we have a load of a gep of an alloca,
10641 // which can usually all be folded into the load.
10642 if (AllBasePointersAreAllocas)
10645 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10646 // that is variable.
10647 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10649 bool HasAnyPHIs = false;
10650 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10651 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10652 Value *FirstOp = FirstInst->getOperand(i);
10653 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10654 FirstOp->getName()+".pn");
10655 InsertNewInstBefore(NewPN, PN);
10657 NewPN->reserveOperandSpace(e);
10658 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10659 OperandPhis[i] = NewPN;
10660 FixedOperands[i] = NewPN;
10665 // Add all operands to the new PHIs.
10667 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10668 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10669 BasicBlock *InBB = PN.getIncomingBlock(i);
10671 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10672 if (PHINode *OpPhi = OperandPhis[op])
10673 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10677 Value *Base = FixedOperands[0];
10678 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10679 FixedOperands.end());
10683 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10684 /// sink the load out of the block that defines it. This means that it must be
10685 /// obvious the value of the load is not changed from the point of the load to
10686 /// the end of the block it is in.
10688 /// Finally, it is safe, but not profitable, to sink a load targetting a
10689 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10691 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10692 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10694 for (++BBI; BBI != E; ++BBI)
10695 if (BBI->mayWriteToMemory())
10698 // Check for non-address taken alloca. If not address-taken already, it isn't
10699 // profitable to do this xform.
10700 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10701 bool isAddressTaken = false;
10702 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10704 if (isa<LoadInst>(UI)) continue;
10705 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10706 // If storing TO the alloca, then the address isn't taken.
10707 if (SI->getOperand(1) == AI) continue;
10709 isAddressTaken = true;
10713 if (!isAddressTaken && AI->isStaticAlloca())
10717 // If this load is a load from a GEP with a constant offset from an alloca,
10718 // then we don't want to sink it. In its present form, it will be
10719 // load [constant stack offset]. Sinking it will cause us to have to
10720 // materialize the stack addresses in each predecessor in a register only to
10721 // do a shared load from register in the successor.
10722 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10723 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10724 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10731 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10732 // operator and they all are only used by the PHI, PHI together their
10733 // inputs, and do the operation once, to the result of the PHI.
10734 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10735 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10737 // Scan the instruction, looking for input operations that can be folded away.
10738 // If all input operands to the phi are the same instruction (e.g. a cast from
10739 // the same type or "+42") we can pull the operation through the PHI, reducing
10740 // code size and simplifying code.
10741 Constant *ConstantOp = 0;
10742 const Type *CastSrcTy = 0;
10743 bool isVolatile = false;
10744 if (isa<CastInst>(FirstInst)) {
10745 CastSrcTy = FirstInst->getOperand(0)->getType();
10746 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10747 // Can fold binop, compare or shift here if the RHS is a constant,
10748 // otherwise call FoldPHIArgBinOpIntoPHI.
10749 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10750 if (ConstantOp == 0)
10751 return FoldPHIArgBinOpIntoPHI(PN);
10752 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10753 isVolatile = LI->isVolatile();
10754 // We can't sink the load if the loaded value could be modified between the
10755 // load and the PHI.
10756 if (LI->getParent() != PN.getIncomingBlock(0) ||
10757 !isSafeAndProfitableToSinkLoad(LI))
10760 // If the PHI is of volatile loads and the load block has multiple
10761 // successors, sinking it would remove a load of the volatile value from
10762 // the path through the other successor.
10764 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10767 } else if (isa<GetElementPtrInst>(FirstInst)) {
10768 return FoldPHIArgGEPIntoPHI(PN);
10770 return 0; // Cannot fold this operation.
10773 // Check to see if all arguments are the same operation.
10774 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10775 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10776 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10777 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10780 if (I->getOperand(0)->getType() != CastSrcTy)
10781 return 0; // Cast operation must match.
10782 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10783 // We can't sink the load if the loaded value could be modified between
10784 // the load and the PHI.
10785 if (LI->isVolatile() != isVolatile ||
10786 LI->getParent() != PN.getIncomingBlock(i) ||
10787 !isSafeAndProfitableToSinkLoad(LI))
10790 // If the PHI is of volatile loads and the load block has multiple
10791 // successors, sinking it would remove a load of the volatile value from
10792 // the path through the other successor.
10794 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10797 } else if (I->getOperand(1) != ConstantOp) {
10802 // Okay, they are all the same operation. Create a new PHI node of the
10803 // correct type, and PHI together all of the LHS's of the instructions.
10804 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10805 PN.getName()+".in");
10806 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10808 Value *InVal = FirstInst->getOperand(0);
10809 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10811 // Add all operands to the new PHI.
10812 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10813 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10814 if (NewInVal != InVal)
10816 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10821 // The new PHI unions all of the same values together. This is really
10822 // common, so we handle it intelligently here for compile-time speed.
10826 InsertNewInstBefore(NewPN, PN);
10830 // Insert and return the new operation.
10831 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10832 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10833 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10834 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10835 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10836 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10837 PhiVal, ConstantOp);
10838 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10840 // If this was a volatile load that we are merging, make sure to loop through
10841 // and mark all the input loads as non-volatile. If we don't do this, we will
10842 // insert a new volatile load and the old ones will not be deletable.
10844 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10845 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10847 return new LoadInst(PhiVal, "", isVolatile);
10850 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10852 static bool DeadPHICycle(PHINode *PN,
10853 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10854 if (PN->use_empty()) return true;
10855 if (!PN->hasOneUse()) return false;
10857 // Remember this node, and if we find the cycle, return.
10858 if (!PotentiallyDeadPHIs.insert(PN))
10861 // Don't scan crazily complex things.
10862 if (PotentiallyDeadPHIs.size() == 16)
10865 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10866 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10871 /// PHIsEqualValue - Return true if this phi node is always equal to
10872 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10873 /// z = some value; x = phi (y, z); y = phi (x, z)
10874 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10875 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10876 // See if we already saw this PHI node.
10877 if (!ValueEqualPHIs.insert(PN))
10880 // Don't scan crazily complex things.
10881 if (ValueEqualPHIs.size() == 16)
10884 // Scan the operands to see if they are either phi nodes or are equal to
10886 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10887 Value *Op = PN->getIncomingValue(i);
10888 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10889 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10891 } else if (Op != NonPhiInVal)
10899 // PHINode simplification
10901 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10902 // If LCSSA is around, don't mess with Phi nodes
10903 if (MustPreserveLCSSA) return 0;
10905 if (Value *V = PN.hasConstantValue())
10906 return ReplaceInstUsesWith(PN, V);
10908 // If all PHI operands are the same operation, pull them through the PHI,
10909 // reducing code size.
10910 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10911 isa<Instruction>(PN.getIncomingValue(1)) &&
10912 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10913 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10914 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10915 // than themselves more than once.
10916 PN.getIncomingValue(0)->hasOneUse())
10917 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10920 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10921 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10922 // PHI)... break the cycle.
10923 if (PN.hasOneUse()) {
10924 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10925 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10926 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10927 PotentiallyDeadPHIs.insert(&PN);
10928 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10929 return ReplaceInstUsesWith(PN, Context->getUndef(PN.getType()));
10932 // If this phi has a single use, and if that use just computes a value for
10933 // the next iteration of a loop, delete the phi. This occurs with unused
10934 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10935 // common case here is good because the only other things that catch this
10936 // are induction variable analysis (sometimes) and ADCE, which is only run
10938 if (PHIUser->hasOneUse() &&
10939 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10940 PHIUser->use_back() == &PN) {
10941 return ReplaceInstUsesWith(PN, Context->getUndef(PN.getType()));
10945 // We sometimes end up with phi cycles that non-obviously end up being the
10946 // same value, for example:
10947 // z = some value; x = phi (y, z); y = phi (x, z)
10948 // where the phi nodes don't necessarily need to be in the same block. Do a
10949 // quick check to see if the PHI node only contains a single non-phi value, if
10950 // so, scan to see if the phi cycle is actually equal to that value.
10952 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10953 // Scan for the first non-phi operand.
10954 while (InValNo != NumOperandVals &&
10955 isa<PHINode>(PN.getIncomingValue(InValNo)))
10958 if (InValNo != NumOperandVals) {
10959 Value *NonPhiInVal = PN.getOperand(InValNo);
10961 // Scan the rest of the operands to see if there are any conflicts, if so
10962 // there is no need to recursively scan other phis.
10963 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10964 Value *OpVal = PN.getIncomingValue(InValNo);
10965 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10969 // If we scanned over all operands, then we have one unique value plus
10970 // phi values. Scan PHI nodes to see if they all merge in each other or
10972 if (InValNo == NumOperandVals) {
10973 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10974 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10975 return ReplaceInstUsesWith(PN, NonPhiInVal);
10982 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
10983 Instruction *InsertPoint,
10984 InstCombiner *IC) {
10985 unsigned PtrSize = DTy->getScalarSizeInBits();
10986 unsigned VTySize = V->getType()->getScalarSizeInBits();
10987 // We must cast correctly to the pointer type. Ensure that we
10988 // sign extend the integer value if it is smaller as this is
10989 // used for address computation.
10990 Instruction::CastOps opcode =
10991 (VTySize < PtrSize ? Instruction::SExt :
10992 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
10993 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
10997 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
10998 Value *PtrOp = GEP.getOperand(0);
10999 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
11000 // If so, eliminate the noop.
11001 if (GEP.getNumOperands() == 1)
11002 return ReplaceInstUsesWith(GEP, PtrOp);
11004 if (isa<UndefValue>(GEP.getOperand(0)))
11005 return ReplaceInstUsesWith(GEP, Context->getUndef(GEP.getType()));
11007 bool HasZeroPointerIndex = false;
11008 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
11009 HasZeroPointerIndex = C->isNullValue();
11011 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11012 return ReplaceInstUsesWith(GEP, PtrOp);
11014 // Eliminate unneeded casts for indices.
11015 bool MadeChange = false;
11017 gep_type_iterator GTI = gep_type_begin(GEP);
11018 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
11019 i != e; ++i, ++GTI) {
11020 if (isa<SequentialType>(*GTI)) {
11021 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
11022 if (CI->getOpcode() == Instruction::ZExt ||
11023 CI->getOpcode() == Instruction::SExt) {
11024 const Type *SrcTy = CI->getOperand(0)->getType();
11025 // We can eliminate a cast from i32 to i64 iff the target
11026 // is a 32-bit pointer target.
11027 if (SrcTy->getScalarSizeInBits() >= TD->getPointerSizeInBits()) {
11029 *i = CI->getOperand(0);
11033 // If we are using a wider index than needed for this platform, shrink it
11034 // to what we need. If narrower, sign-extend it to what we need.
11035 // If the incoming value needs a cast instruction,
11036 // insert it. This explicit cast can make subsequent optimizations more
11039 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
11040 if (Constant *C = dyn_cast<Constant>(Op)) {
11041 *i = Context->getConstantExprTrunc(C, TD->getIntPtrType());
11044 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
11049 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
11050 if (Constant *C = dyn_cast<Constant>(Op)) {
11051 *i = Context->getConstantExprSExt(C, TD->getIntPtrType());
11054 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
11062 if (MadeChange) return &GEP;
11064 // Combine Indices - If the source pointer to this getelementptr instruction
11065 // is a getelementptr instruction, combine the indices of the two
11066 // getelementptr instructions into a single instruction.
11068 SmallVector<Value*, 8> SrcGEPOperands;
11069 if (User *Src = dyn_castGetElementPtr(PtrOp))
11070 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
11072 if (!SrcGEPOperands.empty()) {
11073 // Note that if our source is a gep chain itself that we wait for that
11074 // chain to be resolved before we perform this transformation. This
11075 // avoids us creating a TON of code in some cases.
11077 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
11078 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
11079 return 0; // Wait until our source is folded to completion.
11081 SmallVector<Value*, 8> Indices;
11083 // Find out whether the last index in the source GEP is a sequential idx.
11084 bool EndsWithSequential = false;
11085 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
11086 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
11087 EndsWithSequential = !isa<StructType>(*I);
11089 // Can we combine the two pointer arithmetics offsets?
11090 if (EndsWithSequential) {
11091 // Replace: gep (gep %P, long B), long A, ...
11092 // With: T = long A+B; gep %P, T, ...
11094 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
11095 if (SO1 == Context->getNullValue(SO1->getType())) {
11097 } else if (GO1 == Context->getNullValue(GO1->getType())) {
11100 // If they aren't the same type, convert both to an integer of the
11101 // target's pointer size.
11102 if (SO1->getType() != GO1->getType()) {
11103 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
11105 Context->getConstantExprIntegerCast(SO1C, GO1->getType(), true);
11106 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
11108 Context->getConstantExprIntegerCast(GO1C, SO1->getType(), true);
11110 unsigned PS = TD->getPointerSizeInBits();
11111 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
11112 // Convert GO1 to SO1's type.
11113 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
11115 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
11116 // Convert SO1 to GO1's type.
11117 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
11119 const Type *PT = TD->getIntPtrType();
11120 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
11121 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
11125 if (isa<Constant>(SO1) && isa<Constant>(GO1))
11126 Sum = Context->getConstantExprAdd(cast<Constant>(SO1),
11127 cast<Constant>(GO1));
11129 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11130 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
11134 // Recycle the GEP we already have if possible.
11135 if (SrcGEPOperands.size() == 2) {
11136 GEP.setOperand(0, SrcGEPOperands[0]);
11137 GEP.setOperand(1, Sum);
11140 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11141 SrcGEPOperands.end()-1);
11142 Indices.push_back(Sum);
11143 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
11145 } else if (isa<Constant>(*GEP.idx_begin()) &&
11146 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11147 SrcGEPOperands.size() != 1) {
11148 // Otherwise we can do the fold if the first index of the GEP is a zero
11149 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11150 SrcGEPOperands.end());
11151 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
11154 if (!Indices.empty())
11155 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
11156 Indices.end(), GEP.getName());
11158 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
11159 // GEP of global variable. If all of the indices for this GEP are
11160 // constants, we can promote this to a constexpr instead of an instruction.
11162 // Scan for nonconstants...
11163 SmallVector<Constant*, 8> Indices;
11164 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
11165 for (; I != E && isa<Constant>(*I); ++I)
11166 Indices.push_back(cast<Constant>(*I));
11168 if (I == E) { // If they are all constants...
11169 Constant *CE = Context->getConstantExprGetElementPtr(GV,
11170 &Indices[0],Indices.size());
11172 // Replace all uses of the GEP with the new constexpr...
11173 return ReplaceInstUsesWith(GEP, CE);
11175 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
11176 if (!isa<PointerType>(X->getType())) {
11177 // Not interesting. Source pointer must be a cast from pointer.
11178 } else if (HasZeroPointerIndex) {
11179 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11180 // into : GEP [10 x i8]* X, i32 0, ...
11182 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11183 // into : GEP i8* X, ...
11185 // This occurs when the program declares an array extern like "int X[];"
11186 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11187 const PointerType *XTy = cast<PointerType>(X->getType());
11188 if (const ArrayType *CATy =
11189 dyn_cast<ArrayType>(CPTy->getElementType())) {
11190 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11191 if (CATy->getElementType() == XTy->getElementType()) {
11192 // -> GEP i8* X, ...
11193 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11194 return GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11196 } else if (const ArrayType *XATy =
11197 dyn_cast<ArrayType>(XTy->getElementType())) {
11198 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11199 if (CATy->getElementType() == XATy->getElementType()) {
11200 // -> GEP [10 x i8]* X, i32 0, ...
11201 // At this point, we know that the cast source type is a pointer
11202 // to an array of the same type as the destination pointer
11203 // array. Because the array type is never stepped over (there
11204 // is a leading zero) we can fold the cast into this GEP.
11205 GEP.setOperand(0, X);
11210 } else if (GEP.getNumOperands() == 2) {
11211 // Transform things like:
11212 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11213 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11214 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11215 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11216 if (isa<ArrayType>(SrcElTy) &&
11217 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11218 TD->getTypeAllocSize(ResElTy)) {
11220 Idx[0] = Context->getNullValue(Type::Int32Ty);
11221 Idx[1] = GEP.getOperand(1);
11222 Value *V = InsertNewInstBefore(
11223 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
11224 // V and GEP are both pointer types --> BitCast
11225 return new BitCastInst(V, GEP.getType());
11228 // Transform things like:
11229 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11230 // (where tmp = 8*tmp2) into:
11231 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11233 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
11234 uint64_t ArrayEltSize =
11235 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11237 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11238 // allow either a mul, shift, or constant here.
11240 ConstantInt *Scale = 0;
11241 if (ArrayEltSize == 1) {
11242 NewIdx = GEP.getOperand(1);
11244 Context->getConstantInt(cast<IntegerType>(NewIdx->getType()), 1);
11245 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11246 NewIdx = Context->getConstantInt(CI->getType(), 1);
11248 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11249 if (Inst->getOpcode() == Instruction::Shl &&
11250 isa<ConstantInt>(Inst->getOperand(1))) {
11251 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11252 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11253 Scale = Context->getConstantInt(cast<IntegerType>(Inst->getType()),
11255 NewIdx = Inst->getOperand(0);
11256 } else if (Inst->getOpcode() == Instruction::Mul &&
11257 isa<ConstantInt>(Inst->getOperand(1))) {
11258 Scale = cast<ConstantInt>(Inst->getOperand(1));
11259 NewIdx = Inst->getOperand(0);
11263 // If the index will be to exactly the right offset with the scale taken
11264 // out, perform the transformation. Note, we don't know whether Scale is
11265 // signed or not. We'll use unsigned version of division/modulo
11266 // operation after making sure Scale doesn't have the sign bit set.
11267 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11268 Scale->getZExtValue() % ArrayEltSize == 0) {
11269 Scale = Context->getConstantInt(Scale->getType(),
11270 Scale->getZExtValue() / ArrayEltSize);
11271 if (Scale->getZExtValue() != 1) {
11273 Context->getConstantExprIntegerCast(Scale, NewIdx->getType(),
11275 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
11276 NewIdx = InsertNewInstBefore(Sc, GEP);
11279 // Insert the new GEP instruction.
11281 Idx[0] = Context->getNullValue(Type::Int32Ty);
11283 Instruction *NewGEP =
11284 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11285 NewGEP = InsertNewInstBefore(NewGEP, GEP);
11286 // The NewGEP must be pointer typed, so must the old one -> BitCast
11287 return new BitCastInst(NewGEP, GEP.getType());
11293 /// See if we can simplify:
11294 /// X = bitcast A to B*
11295 /// Y = gep X, <...constant indices...>
11296 /// into a gep of the original struct. This is important for SROA and alias
11297 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11298 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11299 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11300 // Determine how much the GEP moves the pointer. We are guaranteed to get
11301 // a constant back from EmitGEPOffset.
11302 ConstantInt *OffsetV =
11303 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11304 int64_t Offset = OffsetV->getSExtValue();
11306 // If this GEP instruction doesn't move the pointer, just replace the GEP
11307 // with a bitcast of the real input to the dest type.
11309 // If the bitcast is of an allocation, and the allocation will be
11310 // converted to match the type of the cast, don't touch this.
11311 if (isa<AllocationInst>(BCI->getOperand(0))) {
11312 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11313 if (Instruction *I = visitBitCast(*BCI)) {
11316 BCI->getParent()->getInstList().insert(BCI, I);
11317 ReplaceInstUsesWith(*BCI, I);
11322 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11325 // Otherwise, if the offset is non-zero, we need to find out if there is a
11326 // field at Offset in 'A's type. If so, we can pull the cast through the
11328 SmallVector<Value*, 8> NewIndices;
11330 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11331 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11332 Instruction *NGEP =
11333 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11335 if (NGEP->getType() == GEP.getType()) return NGEP;
11336 InsertNewInstBefore(NGEP, GEP);
11337 NGEP->takeName(&GEP);
11338 return new BitCastInst(NGEP, GEP.getType());
11346 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11347 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11348 if (AI.isArrayAllocation()) { // Check C != 1
11349 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11350 const Type *NewTy =
11351 Context->getArrayType(AI.getAllocatedType(), C->getZExtValue());
11352 AllocationInst *New = 0;
11354 // Create and insert the replacement instruction...
11355 if (isa<MallocInst>(AI))
11356 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11358 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11359 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11362 InsertNewInstBefore(New, AI);
11364 // Scan to the end of the allocation instructions, to skip over a block of
11365 // allocas if possible...also skip interleaved debug info
11367 BasicBlock::iterator It = New;
11368 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11370 // Now that I is pointing to the first non-allocation-inst in the block,
11371 // insert our getelementptr instruction...
11373 Value *NullIdx = Context->getNullValue(Type::Int32Ty);
11377 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11378 New->getName()+".sub", It);
11380 // Now make everything use the getelementptr instead of the original
11382 return ReplaceInstUsesWith(AI, V);
11383 } else if (isa<UndefValue>(AI.getArraySize())) {
11384 return ReplaceInstUsesWith(AI, Context->getNullValue(AI.getType()));
11388 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11389 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11390 // Note that we only do this for alloca's, because malloc should allocate
11391 // and return a unique pointer, even for a zero byte allocation.
11392 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11393 return ReplaceInstUsesWith(AI, Context->getNullValue(AI.getType()));
11395 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11396 if (AI.getAlignment() == 0)
11397 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11403 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11404 Value *Op = FI.getOperand(0);
11406 // free undef -> unreachable.
11407 if (isa<UndefValue>(Op)) {
11408 // Insert a new store to null because we cannot modify the CFG here.
11409 new StoreInst(Context->getConstantIntTrue(),
11410 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)), &FI);
11411 return EraseInstFromFunction(FI);
11414 // If we have 'free null' delete the instruction. This can happen in stl code
11415 // when lots of inlining happens.
11416 if (isa<ConstantPointerNull>(Op))
11417 return EraseInstFromFunction(FI);
11419 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11420 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11421 FI.setOperand(0, CI->getOperand(0));
11425 // Change free (gep X, 0,0,0,0) into free(X)
11426 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11427 if (GEPI->hasAllZeroIndices()) {
11428 AddToWorkList(GEPI);
11429 FI.setOperand(0, GEPI->getOperand(0));
11434 // Change free(malloc) into nothing, if the malloc has a single use.
11435 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11436 if (MI->hasOneUse()) {
11437 EraseInstFromFunction(FI);
11438 return EraseInstFromFunction(*MI);
11445 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11446 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11447 const TargetData *TD) {
11448 User *CI = cast<User>(LI.getOperand(0));
11449 Value *CastOp = CI->getOperand(0);
11450 LLVMContext *Context = IC.getContext();
11453 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11454 // Instead of loading constant c string, use corresponding integer value
11455 // directly if string length is small enough.
11457 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11458 unsigned len = Str.length();
11459 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11460 unsigned numBits = Ty->getPrimitiveSizeInBits();
11461 // Replace LI with immediate integer store.
11462 if ((numBits >> 3) == len + 1) {
11463 APInt StrVal(numBits, 0);
11464 APInt SingleChar(numBits, 0);
11465 if (TD->isLittleEndian()) {
11466 for (signed i = len-1; i >= 0; i--) {
11467 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11468 StrVal = (StrVal << 8) | SingleChar;
11471 for (unsigned i = 0; i < len; i++) {
11472 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11473 StrVal = (StrVal << 8) | SingleChar;
11475 // Append NULL at the end.
11477 StrVal = (StrVal << 8) | SingleChar;
11479 Value *NL = Context->getConstantInt(StrVal);
11480 return IC.ReplaceInstUsesWith(LI, NL);
11486 const PointerType *DestTy = cast<PointerType>(CI->getType());
11487 const Type *DestPTy = DestTy->getElementType();
11488 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11490 // If the address spaces don't match, don't eliminate the cast.
11491 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11494 const Type *SrcPTy = SrcTy->getElementType();
11496 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11497 isa<VectorType>(DestPTy)) {
11498 // If the source is an array, the code below will not succeed. Check to
11499 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11501 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11502 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11503 if (ASrcTy->getNumElements() != 0) {
11505 Idxs[0] = Idxs[1] = Context->getNullValue(Type::Int32Ty);
11506 CastOp = Context->getConstantExprGetElementPtr(CSrc, Idxs, 2);
11507 SrcTy = cast<PointerType>(CastOp->getType());
11508 SrcPTy = SrcTy->getElementType();
11511 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11512 isa<VectorType>(SrcPTy)) &&
11513 // Do not allow turning this into a load of an integer, which is then
11514 // casted to a pointer, this pessimizes pointer analysis a lot.
11515 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11516 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11517 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11519 // Okay, we are casting from one integer or pointer type to another of
11520 // the same size. Instead of casting the pointer before the load, cast
11521 // the result of the loaded value.
11522 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11524 LI.isVolatile()),LI);
11525 // Now cast the result of the load.
11526 return new BitCastInst(NewLoad, LI.getType());
11533 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11534 Value *Op = LI.getOperand(0);
11536 // Attempt to improve the alignment.
11537 unsigned KnownAlign =
11538 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11540 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11541 LI.getAlignment()))
11542 LI.setAlignment(KnownAlign);
11544 // load (cast X) --> cast (load X) iff safe
11545 if (isa<CastInst>(Op))
11546 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11549 // None of the following transforms are legal for volatile loads.
11550 if (LI.isVolatile()) return 0;
11552 // Do really simple store-to-load forwarding and load CSE, to catch cases
11553 // where there are several consequtive memory accesses to the same location,
11554 // separated by a few arithmetic operations.
11555 BasicBlock::iterator BBI = &LI;
11556 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11557 return ReplaceInstUsesWith(LI, AvailableVal);
11559 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11560 const Value *GEPI0 = GEPI->getOperand(0);
11561 // TODO: Consider a target hook for valid address spaces for this xform.
11562 if (isa<ConstantPointerNull>(GEPI0) &&
11563 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11564 // Insert a new store to null instruction before the load to indicate
11565 // that this code is not reachable. We do this instead of inserting
11566 // an unreachable instruction directly because we cannot modify the
11568 new StoreInst(Context->getUndef(LI.getType()),
11569 Context->getNullValue(Op->getType()), &LI);
11570 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11574 if (Constant *C = dyn_cast<Constant>(Op)) {
11575 // load null/undef -> undef
11576 // TODO: Consider a target hook for valid address spaces for this xform.
11577 if (isa<UndefValue>(C) || (C->isNullValue() &&
11578 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11579 // Insert a new store to null instruction before the load to indicate that
11580 // this code is not reachable. We do this instead of inserting an
11581 // unreachable instruction directly because we cannot modify the CFG.
11582 new StoreInst(Context->getUndef(LI.getType()),
11583 Context->getNullValue(Op->getType()), &LI);
11584 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11587 // Instcombine load (constant global) into the value loaded.
11588 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11589 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11590 return ReplaceInstUsesWith(LI, GV->getInitializer());
11592 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11593 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11594 if (CE->getOpcode() == Instruction::GetElementPtr) {
11595 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11596 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11598 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11600 return ReplaceInstUsesWith(LI, V);
11601 if (CE->getOperand(0)->isNullValue()) {
11602 // Insert a new store to null instruction before the load to indicate
11603 // that this code is not reachable. We do this instead of inserting
11604 // an unreachable instruction directly because we cannot modify the
11606 new StoreInst(Context->getUndef(LI.getType()),
11607 Context->getNullValue(Op->getType()), &LI);
11608 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11611 } else if (CE->isCast()) {
11612 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11618 // If this load comes from anywhere in a constant global, and if the global
11619 // is all undef or zero, we know what it loads.
11620 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11621 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11622 if (GV->getInitializer()->isNullValue())
11623 return ReplaceInstUsesWith(LI, Context->getNullValue(LI.getType()));
11624 else if (isa<UndefValue>(GV->getInitializer()))
11625 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11629 if (Op->hasOneUse()) {
11630 // Change select and PHI nodes to select values instead of addresses: this
11631 // helps alias analysis out a lot, allows many others simplifications, and
11632 // exposes redundancy in the code.
11634 // Note that we cannot do the transformation unless we know that the
11635 // introduced loads cannot trap! Something like this is valid as long as
11636 // the condition is always false: load (select bool %C, int* null, int* %G),
11637 // but it would not be valid if we transformed it to load from null
11638 // unconditionally.
11640 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11641 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11642 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11643 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11644 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11645 SI->getOperand(1)->getName()+".val"), LI);
11646 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11647 SI->getOperand(2)->getName()+".val"), LI);
11648 return SelectInst::Create(SI->getCondition(), V1, V2);
11651 // load (select (cond, null, P)) -> load P
11652 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11653 if (C->isNullValue()) {
11654 LI.setOperand(0, SI->getOperand(2));
11658 // load (select (cond, P, null)) -> load P
11659 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11660 if (C->isNullValue()) {
11661 LI.setOperand(0, SI->getOperand(1));
11669 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11670 /// when possible. This makes it generally easy to do alias analysis and/or
11671 /// SROA/mem2reg of the memory object.
11672 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11673 User *CI = cast<User>(SI.getOperand(1));
11674 Value *CastOp = CI->getOperand(0);
11675 LLVMContext *Context = IC.getContext();
11677 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11678 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11679 if (SrcTy == 0) return 0;
11681 const Type *SrcPTy = SrcTy->getElementType();
11683 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11686 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11687 /// to its first element. This allows us to handle things like:
11688 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11689 /// on 32-bit hosts.
11690 SmallVector<Value*, 4> NewGEPIndices;
11692 // If the source is an array, the code below will not succeed. Check to
11693 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11695 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11696 // Index through pointer.
11697 Constant *Zero = Context->getNullValue(Type::Int32Ty);
11698 NewGEPIndices.push_back(Zero);
11701 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11702 if (!STy->getNumElements()) /* Struct can be empty {} */
11704 NewGEPIndices.push_back(Zero);
11705 SrcPTy = STy->getElementType(0);
11706 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11707 NewGEPIndices.push_back(Zero);
11708 SrcPTy = ATy->getElementType();
11714 SrcTy = Context->getPointerType(SrcPTy, SrcTy->getAddressSpace());
11717 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11720 // If the pointers point into different address spaces or if they point to
11721 // values with different sizes, we can't do the transformation.
11722 if (SrcTy->getAddressSpace() !=
11723 cast<PointerType>(CI->getType())->getAddressSpace() ||
11724 IC.getTargetData().getTypeSizeInBits(SrcPTy) !=
11725 IC.getTargetData().getTypeSizeInBits(DestPTy))
11728 // Okay, we are casting from one integer or pointer type to another of
11729 // the same size. Instead of casting the pointer before
11730 // the store, cast the value to be stored.
11732 Value *SIOp0 = SI.getOperand(0);
11733 Instruction::CastOps opcode = Instruction::BitCast;
11734 const Type* CastSrcTy = SIOp0->getType();
11735 const Type* CastDstTy = SrcPTy;
11736 if (isa<PointerType>(CastDstTy)) {
11737 if (CastSrcTy->isInteger())
11738 opcode = Instruction::IntToPtr;
11739 } else if (isa<IntegerType>(CastDstTy)) {
11740 if (isa<PointerType>(SIOp0->getType()))
11741 opcode = Instruction::PtrToInt;
11744 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11745 // emit a GEP to index into its first field.
11746 if (!NewGEPIndices.empty()) {
11747 if (Constant *C = dyn_cast<Constant>(CastOp))
11748 CastOp = Context->getConstantExprGetElementPtr(C, &NewGEPIndices[0],
11749 NewGEPIndices.size());
11751 CastOp = IC.InsertNewInstBefore(
11752 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11753 NewGEPIndices.end()), SI);
11756 if (Constant *C = dyn_cast<Constant>(SIOp0))
11757 NewCast = Context->getConstantExprCast(opcode, C, CastDstTy);
11759 NewCast = IC.InsertNewInstBefore(
11760 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11762 return new StoreInst(NewCast, CastOp);
11765 /// equivalentAddressValues - Test if A and B will obviously have the same
11766 /// value. This includes recognizing that %t0 and %t1 will have the same
11767 /// value in code like this:
11768 /// %t0 = getelementptr \@a, 0, 3
11769 /// store i32 0, i32* %t0
11770 /// %t1 = getelementptr \@a, 0, 3
11771 /// %t2 = load i32* %t1
11773 static bool equivalentAddressValues(Value *A, Value *B) {
11774 // Test if the values are trivially equivalent.
11775 if (A == B) return true;
11777 // Test if the values come form identical arithmetic instructions.
11778 if (isa<BinaryOperator>(A) ||
11779 isa<CastInst>(A) ||
11781 isa<GetElementPtrInst>(A))
11782 if (Instruction *BI = dyn_cast<Instruction>(B))
11783 if (cast<Instruction>(A)->isIdenticalTo(BI))
11786 // Otherwise they may not be equivalent.
11790 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11791 // return the llvm.dbg.declare.
11792 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11793 if (!V->hasNUses(2))
11795 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11797 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11799 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11800 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11807 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11808 Value *Val = SI.getOperand(0);
11809 Value *Ptr = SI.getOperand(1);
11811 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11812 EraseInstFromFunction(SI);
11817 // If the RHS is an alloca with a single use, zapify the store, making the
11819 // If the RHS is an alloca with a two uses, the other one being a
11820 // llvm.dbg.declare, zapify the store and the declare, making the
11821 // alloca dead. We must do this to prevent declare's from affecting
11823 if (!SI.isVolatile()) {
11824 if (Ptr->hasOneUse()) {
11825 if (isa<AllocaInst>(Ptr)) {
11826 EraseInstFromFunction(SI);
11830 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11831 if (isa<AllocaInst>(GEP->getOperand(0))) {
11832 if (GEP->getOperand(0)->hasOneUse()) {
11833 EraseInstFromFunction(SI);
11837 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11838 EraseInstFromFunction(*DI);
11839 EraseInstFromFunction(SI);
11846 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11847 EraseInstFromFunction(*DI);
11848 EraseInstFromFunction(SI);
11854 // Attempt to improve the alignment.
11855 unsigned KnownAlign =
11856 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11858 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11859 SI.getAlignment()))
11860 SI.setAlignment(KnownAlign);
11862 // Do really simple DSE, to catch cases where there are several consecutive
11863 // stores to the same location, separated by a few arithmetic operations. This
11864 // situation often occurs with bitfield accesses.
11865 BasicBlock::iterator BBI = &SI;
11866 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11869 // Don't count debug info directives, lest they affect codegen,
11870 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11871 // It is necessary for correctness to skip those that feed into a
11872 // llvm.dbg.declare, as these are not present when debugging is off.
11873 if (isa<DbgInfoIntrinsic>(BBI) ||
11874 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11879 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11880 // Prev store isn't volatile, and stores to the same location?
11881 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11882 SI.getOperand(1))) {
11885 EraseInstFromFunction(*PrevSI);
11891 // If this is a load, we have to stop. However, if the loaded value is from
11892 // the pointer we're loading and is producing the pointer we're storing,
11893 // then *this* store is dead (X = load P; store X -> P).
11894 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11895 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11896 !SI.isVolatile()) {
11897 EraseInstFromFunction(SI);
11901 // Otherwise, this is a load from some other location. Stores before it
11902 // may not be dead.
11906 // Don't skip over loads or things that can modify memory.
11907 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11912 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11914 // store X, null -> turns into 'unreachable' in SimplifyCFG
11915 if (isa<ConstantPointerNull>(Ptr) &&
11916 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11917 if (!isa<UndefValue>(Val)) {
11918 SI.setOperand(0, Context->getUndef(Val->getType()));
11919 if (Instruction *U = dyn_cast<Instruction>(Val))
11920 AddToWorkList(U); // Dropped a use.
11923 return 0; // Do not modify these!
11926 // store undef, Ptr -> noop
11927 if (isa<UndefValue>(Val)) {
11928 EraseInstFromFunction(SI);
11933 // If the pointer destination is a cast, see if we can fold the cast into the
11935 if (isa<CastInst>(Ptr))
11936 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11938 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11940 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11944 // If this store is the last instruction in the basic block (possibly
11945 // excepting debug info instructions and the pointer bitcasts that feed
11946 // into them), and if the block ends with an unconditional branch, try
11947 // to move it to the successor block.
11951 } while (isa<DbgInfoIntrinsic>(BBI) ||
11952 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11953 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11954 if (BI->isUnconditional())
11955 if (SimplifyStoreAtEndOfBlock(SI))
11956 return 0; // xform done!
11961 /// SimplifyStoreAtEndOfBlock - Turn things like:
11962 /// if () { *P = v1; } else { *P = v2 }
11963 /// into a phi node with a store in the successor.
11965 /// Simplify things like:
11966 /// *P = v1; if () { *P = v2; }
11967 /// into a phi node with a store in the successor.
11969 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11970 BasicBlock *StoreBB = SI.getParent();
11972 // Check to see if the successor block has exactly two incoming edges. If
11973 // so, see if the other predecessor contains a store to the same location.
11974 // if so, insert a PHI node (if needed) and move the stores down.
11975 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11977 // Determine whether Dest has exactly two predecessors and, if so, compute
11978 // the other predecessor.
11979 pred_iterator PI = pred_begin(DestBB);
11980 BasicBlock *OtherBB = 0;
11981 if (*PI != StoreBB)
11984 if (PI == pred_end(DestBB))
11987 if (*PI != StoreBB) {
11992 if (++PI != pred_end(DestBB))
11995 // Bail out if all the relevant blocks aren't distinct (this can happen,
11996 // for example, if SI is in an infinite loop)
11997 if (StoreBB == DestBB || OtherBB == DestBB)
12000 // Verify that the other block ends in a branch and is not otherwise empty.
12001 BasicBlock::iterator BBI = OtherBB->getTerminator();
12002 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
12003 if (!OtherBr || BBI == OtherBB->begin())
12006 // If the other block ends in an unconditional branch, check for the 'if then
12007 // else' case. there is an instruction before the branch.
12008 StoreInst *OtherStore = 0;
12009 if (OtherBr->isUnconditional()) {
12011 // Skip over debugging info.
12012 while (isa<DbgInfoIntrinsic>(BBI) ||
12013 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12014 if (BBI==OtherBB->begin())
12018 // If this isn't a store, or isn't a store to the same location, bail out.
12019 OtherStore = dyn_cast<StoreInst>(BBI);
12020 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
12023 // Otherwise, the other block ended with a conditional branch. If one of the
12024 // destinations is StoreBB, then we have the if/then case.
12025 if (OtherBr->getSuccessor(0) != StoreBB &&
12026 OtherBr->getSuccessor(1) != StoreBB)
12029 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12030 // if/then triangle. See if there is a store to the same ptr as SI that
12031 // lives in OtherBB.
12033 // Check to see if we find the matching store.
12034 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12035 if (OtherStore->getOperand(1) != SI.getOperand(1))
12039 // If we find something that may be using or overwriting the stored
12040 // value, or if we run out of instructions, we can't do the xform.
12041 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12042 BBI == OtherBB->begin())
12046 // In order to eliminate the store in OtherBr, we have to
12047 // make sure nothing reads or overwrites the stored value in
12049 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12050 // FIXME: This should really be AA driven.
12051 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12056 // Insert a PHI node now if we need it.
12057 Value *MergedVal = OtherStore->getOperand(0);
12058 if (MergedVal != SI.getOperand(0)) {
12059 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12060 PN->reserveOperandSpace(2);
12061 PN->addIncoming(SI.getOperand(0), SI.getParent());
12062 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12063 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12066 // Advance to a place where it is safe to insert the new store and
12068 BBI = DestBB->getFirstNonPHI();
12069 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12070 OtherStore->isVolatile()), *BBI);
12072 // Nuke the old stores.
12073 EraseInstFromFunction(SI);
12074 EraseInstFromFunction(*OtherStore);
12080 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12081 // Change br (not X), label True, label False to: br X, label False, True
12083 BasicBlock *TrueDest;
12084 BasicBlock *FalseDest;
12085 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest), *Context) &&
12086 !isa<Constant>(X)) {
12087 // Swap Destinations and condition...
12088 BI.setCondition(X);
12089 BI.setSuccessor(0, FalseDest);
12090 BI.setSuccessor(1, TrueDest);
12094 // Cannonicalize fcmp_one -> fcmp_oeq
12095 FCmpInst::Predicate FPred; Value *Y;
12096 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12097 TrueDest, FalseDest), *Context))
12098 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12099 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
12100 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
12101 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
12102 Instruction *NewSCC = new FCmpInst(I, NewPred, X, Y, "");
12103 NewSCC->takeName(I);
12104 // Swap Destinations and condition...
12105 BI.setCondition(NewSCC);
12106 BI.setSuccessor(0, FalseDest);
12107 BI.setSuccessor(1, TrueDest);
12108 RemoveFromWorkList(I);
12109 I->eraseFromParent();
12110 AddToWorkList(NewSCC);
12114 // Cannonicalize icmp_ne -> icmp_eq
12115 ICmpInst::Predicate IPred;
12116 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12117 TrueDest, FalseDest), *Context))
12118 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12119 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12120 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
12121 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
12122 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
12123 Instruction *NewSCC = new ICmpInst(I, NewPred, X, Y, "");
12124 NewSCC->takeName(I);
12125 // Swap Destinations and condition...
12126 BI.setCondition(NewSCC);
12127 BI.setSuccessor(0, FalseDest);
12128 BI.setSuccessor(1, TrueDest);
12129 RemoveFromWorkList(I);
12130 I->eraseFromParent();;
12131 AddToWorkList(NewSCC);
12138 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12139 Value *Cond = SI.getCondition();
12140 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12141 if (I->getOpcode() == Instruction::Add)
12142 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12143 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12144 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12146 Context->getConstantExprSub(cast<Constant>(SI.getOperand(i)),
12148 SI.setOperand(0, I->getOperand(0));
12156 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12157 Value *Agg = EV.getAggregateOperand();
12159 if (!EV.hasIndices())
12160 return ReplaceInstUsesWith(EV, Agg);
12162 if (Constant *C = dyn_cast<Constant>(Agg)) {
12163 if (isa<UndefValue>(C))
12164 return ReplaceInstUsesWith(EV, Context->getUndef(EV.getType()));
12166 if (isa<ConstantAggregateZero>(C))
12167 return ReplaceInstUsesWith(EV, Context->getNullValue(EV.getType()));
12169 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12170 // Extract the element indexed by the first index out of the constant
12171 Value *V = C->getOperand(*EV.idx_begin());
12172 if (EV.getNumIndices() > 1)
12173 // Extract the remaining indices out of the constant indexed by the
12175 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12177 return ReplaceInstUsesWith(EV, V);
12179 return 0; // Can't handle other constants
12181 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12182 // We're extracting from an insertvalue instruction, compare the indices
12183 const unsigned *exti, *exte, *insi, *inse;
12184 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12185 exte = EV.idx_end(), inse = IV->idx_end();
12186 exti != exte && insi != inse;
12188 if (*insi != *exti)
12189 // The insert and extract both reference distinctly different elements.
12190 // This means the extract is not influenced by the insert, and we can
12191 // replace the aggregate operand of the extract with the aggregate
12192 // operand of the insert. i.e., replace
12193 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12194 // %E = extractvalue { i32, { i32 } } %I, 0
12196 // %E = extractvalue { i32, { i32 } } %A, 0
12197 return ExtractValueInst::Create(IV->getAggregateOperand(),
12198 EV.idx_begin(), EV.idx_end());
12200 if (exti == exte && insi == inse)
12201 // Both iterators are at the end: Index lists are identical. Replace
12202 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12203 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12205 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12206 if (exti == exte) {
12207 // The extract list is a prefix of the insert list. i.e. replace
12208 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12209 // %E = extractvalue { i32, { i32 } } %I, 1
12211 // %X = extractvalue { i32, { i32 } } %A, 1
12212 // %E = insertvalue { i32 } %X, i32 42, 0
12213 // by switching the order of the insert and extract (though the
12214 // insertvalue should be left in, since it may have other uses).
12215 Value *NewEV = InsertNewInstBefore(
12216 ExtractValueInst::Create(IV->getAggregateOperand(),
12217 EV.idx_begin(), EV.idx_end()),
12219 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12223 // The insert list is a prefix of the extract list
12224 // We can simply remove the common indices from the extract and make it
12225 // operate on the inserted value instead of the insertvalue result.
12227 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12228 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12230 // %E extractvalue { i32 } { i32 42 }, 0
12231 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12234 // Can't simplify extracts from other values. Note that nested extracts are
12235 // already simplified implicitely by the above (extract ( extract (insert) )
12236 // will be translated into extract ( insert ( extract ) ) first and then just
12237 // the value inserted, if appropriate).
12241 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12242 /// is to leave as a vector operation.
12243 static bool CheapToScalarize(Value *V, bool isConstant) {
12244 if (isa<ConstantAggregateZero>(V))
12246 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12247 if (isConstant) return true;
12248 // If all elts are the same, we can extract.
12249 Constant *Op0 = C->getOperand(0);
12250 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12251 if (C->getOperand(i) != Op0)
12255 Instruction *I = dyn_cast<Instruction>(V);
12256 if (!I) return false;
12258 // Insert element gets simplified to the inserted element or is deleted if
12259 // this is constant idx extract element and its a constant idx insertelt.
12260 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12261 isa<ConstantInt>(I->getOperand(2)))
12263 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12265 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12266 if (BO->hasOneUse() &&
12267 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12268 CheapToScalarize(BO->getOperand(1), isConstant)))
12270 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12271 if (CI->hasOneUse() &&
12272 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12273 CheapToScalarize(CI->getOperand(1), isConstant)))
12279 /// Read and decode a shufflevector mask.
12281 /// It turns undef elements into values that are larger than the number of
12282 /// elements in the input.
12283 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12284 unsigned NElts = SVI->getType()->getNumElements();
12285 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12286 return std::vector<unsigned>(NElts, 0);
12287 if (isa<UndefValue>(SVI->getOperand(2)))
12288 return std::vector<unsigned>(NElts, 2*NElts);
12290 std::vector<unsigned> Result;
12291 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12292 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12293 if (isa<UndefValue>(*i))
12294 Result.push_back(NElts*2); // undef -> 8
12296 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12300 /// FindScalarElement - Given a vector and an element number, see if the scalar
12301 /// value is already around as a register, for example if it were inserted then
12302 /// extracted from the vector.
12303 static Value *FindScalarElement(Value *V, unsigned EltNo,
12304 LLVMContext *Context) {
12305 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12306 const VectorType *PTy = cast<VectorType>(V->getType());
12307 unsigned Width = PTy->getNumElements();
12308 if (EltNo >= Width) // Out of range access.
12309 return Context->getUndef(PTy->getElementType());
12311 if (isa<UndefValue>(V))
12312 return Context->getUndef(PTy->getElementType());
12313 else if (isa<ConstantAggregateZero>(V))
12314 return Context->getNullValue(PTy->getElementType());
12315 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12316 return CP->getOperand(EltNo);
12317 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12318 // If this is an insert to a variable element, we don't know what it is.
12319 if (!isa<ConstantInt>(III->getOperand(2)))
12321 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12323 // If this is an insert to the element we are looking for, return the
12325 if (EltNo == IIElt)
12326 return III->getOperand(1);
12328 // Otherwise, the insertelement doesn't modify the value, recurse on its
12330 return FindScalarElement(III->getOperand(0), EltNo, Context);
12331 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12332 unsigned LHSWidth =
12333 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12334 unsigned InEl = getShuffleMask(SVI)[EltNo];
12335 if (InEl < LHSWidth)
12336 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12337 else if (InEl < LHSWidth*2)
12338 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12340 return Context->getUndef(PTy->getElementType());
12343 // Otherwise, we don't know.
12347 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12348 // If vector val is undef, replace extract with scalar undef.
12349 if (isa<UndefValue>(EI.getOperand(0)))
12350 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12352 // If vector val is constant 0, replace extract with scalar 0.
12353 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12354 return ReplaceInstUsesWith(EI, Context->getNullValue(EI.getType()));
12356 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12357 // If vector val is constant with all elements the same, replace EI with
12358 // that element. When the elements are not identical, we cannot replace yet
12359 // (we do that below, but only when the index is constant).
12360 Constant *op0 = C->getOperand(0);
12361 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12362 if (C->getOperand(i) != op0) {
12367 return ReplaceInstUsesWith(EI, op0);
12370 // If extracting a specified index from the vector, see if we can recursively
12371 // find a previously computed scalar that was inserted into the vector.
12372 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12373 unsigned IndexVal = IdxC->getZExtValue();
12374 unsigned VectorWidth =
12375 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12377 // If this is extracting an invalid index, turn this into undef, to avoid
12378 // crashing the code below.
12379 if (IndexVal >= VectorWidth)
12380 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12382 // This instruction only demands the single element from the input vector.
12383 // If the input vector has a single use, simplify it based on this use
12385 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12386 APInt UndefElts(VectorWidth, 0);
12387 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12388 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12389 DemandedMask, UndefElts)) {
12390 EI.setOperand(0, V);
12395 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12396 return ReplaceInstUsesWith(EI, Elt);
12398 // If the this extractelement is directly using a bitcast from a vector of
12399 // the same number of elements, see if we can find the source element from
12400 // it. In this case, we will end up needing to bitcast the scalars.
12401 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12402 if (const VectorType *VT =
12403 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12404 if (VT->getNumElements() == VectorWidth)
12405 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12406 IndexVal, Context))
12407 return new BitCastInst(Elt, EI.getType());
12411 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12412 if (I->hasOneUse()) {
12413 // Push extractelement into predecessor operation if legal and
12414 // profitable to do so
12415 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12416 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12417 if (CheapToScalarize(BO, isConstantElt)) {
12418 ExtractElementInst *newEI0 =
12419 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
12420 EI.getName()+".lhs");
12421 ExtractElementInst *newEI1 =
12422 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
12423 EI.getName()+".rhs");
12424 InsertNewInstBefore(newEI0, EI);
12425 InsertNewInstBefore(newEI1, EI);
12426 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12428 } else if (isa<LoadInst>(I)) {
12430 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12431 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12432 Context->getPointerType(EI.getType(), AS),EI);
12433 GetElementPtrInst *GEP =
12434 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12435 InsertNewInstBefore(GEP, EI);
12436 return new LoadInst(GEP);
12439 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12440 // Extracting the inserted element?
12441 if (IE->getOperand(2) == EI.getOperand(1))
12442 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12443 // If the inserted and extracted elements are constants, they must not
12444 // be the same value, extract from the pre-inserted value instead.
12445 if (isa<Constant>(IE->getOperand(2)) &&
12446 isa<Constant>(EI.getOperand(1))) {
12447 AddUsesToWorkList(EI);
12448 EI.setOperand(0, IE->getOperand(0));
12451 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12452 // If this is extracting an element from a shufflevector, figure out where
12453 // it came from and extract from the appropriate input element instead.
12454 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12455 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12457 unsigned LHSWidth =
12458 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12460 if (SrcIdx < LHSWidth)
12461 Src = SVI->getOperand(0);
12462 else if (SrcIdx < LHSWidth*2) {
12463 SrcIdx -= LHSWidth;
12464 Src = SVI->getOperand(1);
12466 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12468 return new ExtractElementInst(Src,
12469 Context->getConstantInt(Type::Int32Ty, SrcIdx, false));
12476 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12477 /// elements from either LHS or RHS, return the shuffle mask and true.
12478 /// Otherwise, return false.
12479 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12480 std::vector<Constant*> &Mask,
12481 LLVMContext *Context) {
12482 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12483 "Invalid CollectSingleShuffleElements");
12484 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12486 if (isa<UndefValue>(V)) {
12487 Mask.assign(NumElts, Context->getUndef(Type::Int32Ty));
12489 } else if (V == LHS) {
12490 for (unsigned i = 0; i != NumElts; ++i)
12491 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i));
12493 } else if (V == RHS) {
12494 for (unsigned i = 0; i != NumElts; ++i)
12495 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i+NumElts));
12497 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12498 // If this is an insert of an extract from some other vector, include it.
12499 Value *VecOp = IEI->getOperand(0);
12500 Value *ScalarOp = IEI->getOperand(1);
12501 Value *IdxOp = IEI->getOperand(2);
12503 if (!isa<ConstantInt>(IdxOp))
12505 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12507 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12508 // Okay, we can handle this if the vector we are insertinting into is
12509 // transitively ok.
12510 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12511 // If so, update the mask to reflect the inserted undef.
12512 Mask[InsertedIdx] = Context->getUndef(Type::Int32Ty);
12515 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12516 if (isa<ConstantInt>(EI->getOperand(1)) &&
12517 EI->getOperand(0)->getType() == V->getType()) {
12518 unsigned ExtractedIdx =
12519 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12521 // This must be extracting from either LHS or RHS.
12522 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12523 // Okay, we can handle this if the vector we are insertinting into is
12524 // transitively ok.
12525 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12526 // If so, update the mask to reflect the inserted value.
12527 if (EI->getOperand(0) == LHS) {
12528 Mask[InsertedIdx % NumElts] =
12529 Context->getConstantInt(Type::Int32Ty, ExtractedIdx);
12531 assert(EI->getOperand(0) == RHS);
12532 Mask[InsertedIdx % NumElts] =
12533 Context->getConstantInt(Type::Int32Ty, ExtractedIdx+NumElts);
12542 // TODO: Handle shufflevector here!
12547 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12548 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12549 /// that computes V and the LHS value of the shuffle.
12550 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12551 Value *&RHS, LLVMContext *Context) {
12552 assert(isa<VectorType>(V->getType()) &&
12553 (RHS == 0 || V->getType() == RHS->getType()) &&
12554 "Invalid shuffle!");
12555 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12557 if (isa<UndefValue>(V)) {
12558 Mask.assign(NumElts, Context->getUndef(Type::Int32Ty));
12560 } else if (isa<ConstantAggregateZero>(V)) {
12561 Mask.assign(NumElts, Context->getConstantInt(Type::Int32Ty, 0));
12563 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12564 // If this is an insert of an extract from some other vector, include it.
12565 Value *VecOp = IEI->getOperand(0);
12566 Value *ScalarOp = IEI->getOperand(1);
12567 Value *IdxOp = IEI->getOperand(2);
12569 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12570 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12571 EI->getOperand(0)->getType() == V->getType()) {
12572 unsigned ExtractedIdx =
12573 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12574 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12576 // Either the extracted from or inserted into vector must be RHSVec,
12577 // otherwise we'd end up with a shuffle of three inputs.
12578 if (EI->getOperand(0) == RHS || RHS == 0) {
12579 RHS = EI->getOperand(0);
12580 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12581 Mask[InsertedIdx % NumElts] =
12582 Context->getConstantInt(Type::Int32Ty, NumElts+ExtractedIdx);
12586 if (VecOp == RHS) {
12587 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12589 // Everything but the extracted element is replaced with the RHS.
12590 for (unsigned i = 0; i != NumElts; ++i) {
12591 if (i != InsertedIdx)
12592 Mask[i] = Context->getConstantInt(Type::Int32Ty, NumElts+i);
12597 // If this insertelement is a chain that comes from exactly these two
12598 // vectors, return the vector and the effective shuffle.
12599 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12601 return EI->getOperand(0);
12606 // TODO: Handle shufflevector here!
12608 // Otherwise, can't do anything fancy. Return an identity vector.
12609 for (unsigned i = 0; i != NumElts; ++i)
12610 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i));
12614 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12615 Value *VecOp = IE.getOperand(0);
12616 Value *ScalarOp = IE.getOperand(1);
12617 Value *IdxOp = IE.getOperand(2);
12619 // Inserting an undef or into an undefined place, remove this.
12620 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12621 ReplaceInstUsesWith(IE, VecOp);
12623 // If the inserted element was extracted from some other vector, and if the
12624 // indexes are constant, try to turn this into a shufflevector operation.
12625 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12626 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12627 EI->getOperand(0)->getType() == IE.getType()) {
12628 unsigned NumVectorElts = IE.getType()->getNumElements();
12629 unsigned ExtractedIdx =
12630 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12631 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12633 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12634 return ReplaceInstUsesWith(IE, VecOp);
12636 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12637 return ReplaceInstUsesWith(IE, Context->getUndef(IE.getType()));
12639 // If we are extracting a value from a vector, then inserting it right
12640 // back into the same place, just use the input vector.
12641 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12642 return ReplaceInstUsesWith(IE, VecOp);
12644 // We could theoretically do this for ANY input. However, doing so could
12645 // turn chains of insertelement instructions into a chain of shufflevector
12646 // instructions, and right now we do not merge shufflevectors. As such,
12647 // only do this in a situation where it is clear that there is benefit.
12648 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12649 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12650 // the values of VecOp, except then one read from EIOp0.
12651 // Build a new shuffle mask.
12652 std::vector<Constant*> Mask;
12653 if (isa<UndefValue>(VecOp))
12654 Mask.assign(NumVectorElts, Context->getUndef(Type::Int32Ty));
12656 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12657 Mask.assign(NumVectorElts, Context->getConstantInt(Type::Int32Ty,
12660 Mask[InsertedIdx] =
12661 Context->getConstantInt(Type::Int32Ty, ExtractedIdx);
12662 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12663 Context->getConstantVector(Mask));
12666 // If this insertelement isn't used by some other insertelement, turn it
12667 // (and any insertelements it points to), into one big shuffle.
12668 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12669 std::vector<Constant*> Mask;
12671 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12672 if (RHS == 0) RHS = Context->getUndef(LHS->getType());
12673 // We now have a shuffle of LHS, RHS, Mask.
12674 return new ShuffleVectorInst(LHS, RHS,
12675 Context->getConstantVector(Mask));
12680 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12681 APInt UndefElts(VWidth, 0);
12682 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12683 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12690 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12691 Value *LHS = SVI.getOperand(0);
12692 Value *RHS = SVI.getOperand(1);
12693 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12695 bool MadeChange = false;
12697 // Undefined shuffle mask -> undefined value.
12698 if (isa<UndefValue>(SVI.getOperand(2)))
12699 return ReplaceInstUsesWith(SVI, Context->getUndef(SVI.getType()));
12701 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12703 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12706 APInt UndefElts(VWidth, 0);
12707 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12708 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12709 LHS = SVI.getOperand(0);
12710 RHS = SVI.getOperand(1);
12714 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12715 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12716 if (LHS == RHS || isa<UndefValue>(LHS)) {
12717 if (isa<UndefValue>(LHS) && LHS == RHS) {
12718 // shuffle(undef,undef,mask) -> undef.
12719 return ReplaceInstUsesWith(SVI, LHS);
12722 // Remap any references to RHS to use LHS.
12723 std::vector<Constant*> Elts;
12724 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12725 if (Mask[i] >= 2*e)
12726 Elts.push_back(Context->getUndef(Type::Int32Ty));
12728 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12729 (Mask[i] < e && isa<UndefValue>(LHS))) {
12730 Mask[i] = 2*e; // Turn into undef.
12731 Elts.push_back(Context->getUndef(Type::Int32Ty));
12733 Mask[i] = Mask[i] % e; // Force to LHS.
12734 Elts.push_back(Context->getConstantInt(Type::Int32Ty, Mask[i]));
12738 SVI.setOperand(0, SVI.getOperand(1));
12739 SVI.setOperand(1, Context->getUndef(RHS->getType()));
12740 SVI.setOperand(2, Context->getConstantVector(Elts));
12741 LHS = SVI.getOperand(0);
12742 RHS = SVI.getOperand(1);
12746 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12747 bool isLHSID = true, isRHSID = true;
12749 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12750 if (Mask[i] >= e*2) continue; // Ignore undef values.
12751 // Is this an identity shuffle of the LHS value?
12752 isLHSID &= (Mask[i] == i);
12754 // Is this an identity shuffle of the RHS value?
12755 isRHSID &= (Mask[i]-e == i);
12758 // Eliminate identity shuffles.
12759 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12760 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12762 // If the LHS is a shufflevector itself, see if we can combine it with this
12763 // one without producing an unusual shuffle. Here we are really conservative:
12764 // we are absolutely afraid of producing a shuffle mask not in the input
12765 // program, because the code gen may not be smart enough to turn a merged
12766 // shuffle into two specific shuffles: it may produce worse code. As such,
12767 // we only merge two shuffles if the result is one of the two input shuffle
12768 // masks. In this case, merging the shuffles just removes one instruction,
12769 // which we know is safe. This is good for things like turning:
12770 // (splat(splat)) -> splat.
12771 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12772 if (isa<UndefValue>(RHS)) {
12773 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12775 std::vector<unsigned> NewMask;
12776 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12777 if (Mask[i] >= 2*e)
12778 NewMask.push_back(2*e);
12780 NewMask.push_back(LHSMask[Mask[i]]);
12782 // If the result mask is equal to the src shuffle or this shuffle mask, do
12783 // the replacement.
12784 if (NewMask == LHSMask || NewMask == Mask) {
12785 unsigned LHSInNElts =
12786 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12787 std::vector<Constant*> Elts;
12788 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12789 if (NewMask[i] >= LHSInNElts*2) {
12790 Elts.push_back(Context->getUndef(Type::Int32Ty));
12792 Elts.push_back(Context->getConstantInt(Type::Int32Ty, NewMask[i]));
12795 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12796 LHSSVI->getOperand(1),
12797 Context->getConstantVector(Elts));
12802 return MadeChange ? &SVI : 0;
12808 /// TryToSinkInstruction - Try to move the specified instruction from its
12809 /// current block into the beginning of DestBlock, which can only happen if it's
12810 /// safe to move the instruction past all of the instructions between it and the
12811 /// end of its block.
12812 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12813 assert(I->hasOneUse() && "Invariants didn't hold!");
12815 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12816 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12819 // Do not sink alloca instructions out of the entry block.
12820 if (isa<AllocaInst>(I) && I->getParent() ==
12821 &DestBlock->getParent()->getEntryBlock())
12824 // We can only sink load instructions if there is nothing between the load and
12825 // the end of block that could change the value.
12826 if (I->mayReadFromMemory()) {
12827 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12829 if (Scan->mayWriteToMemory())
12833 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12835 CopyPrecedingStopPoint(I, InsertPos);
12836 I->moveBefore(InsertPos);
12842 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12843 /// all reachable code to the worklist.
12845 /// This has a couple of tricks to make the code faster and more powerful. In
12846 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12847 /// them to the worklist (this significantly speeds up instcombine on code where
12848 /// many instructions are dead or constant). Additionally, if we find a branch
12849 /// whose condition is a known constant, we only visit the reachable successors.
12851 static void AddReachableCodeToWorklist(BasicBlock *BB,
12852 SmallPtrSet<BasicBlock*, 64> &Visited,
12854 const TargetData *TD) {
12855 SmallVector<BasicBlock*, 256> Worklist;
12856 Worklist.push_back(BB);
12858 while (!Worklist.empty()) {
12859 BB = Worklist.back();
12860 Worklist.pop_back();
12862 // We have now visited this block! If we've already been here, ignore it.
12863 if (!Visited.insert(BB)) continue;
12865 DbgInfoIntrinsic *DBI_Prev = NULL;
12866 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12867 Instruction *Inst = BBI++;
12869 // DCE instruction if trivially dead.
12870 if (isInstructionTriviallyDead(Inst)) {
12872 DOUT << "IC: DCE: " << *Inst;
12873 Inst->eraseFromParent();
12877 // ConstantProp instruction if trivially constant.
12878 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12879 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12880 Inst->replaceAllUsesWith(C);
12882 Inst->eraseFromParent();
12886 // If there are two consecutive llvm.dbg.stoppoint calls then
12887 // it is likely that the optimizer deleted code in between these
12889 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12892 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12893 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12894 IC.RemoveFromWorkList(DBI_Prev);
12895 DBI_Prev->eraseFromParent();
12897 DBI_Prev = DBI_Next;
12902 IC.AddToWorkList(Inst);
12905 // Recursively visit successors. If this is a branch or switch on a
12906 // constant, only visit the reachable successor.
12907 TerminatorInst *TI = BB->getTerminator();
12908 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12909 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12910 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12911 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12912 Worklist.push_back(ReachableBB);
12915 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12916 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12917 // See if this is an explicit destination.
12918 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12919 if (SI->getCaseValue(i) == Cond) {
12920 BasicBlock *ReachableBB = SI->getSuccessor(i);
12921 Worklist.push_back(ReachableBB);
12925 // Otherwise it is the default destination.
12926 Worklist.push_back(SI->getSuccessor(0));
12931 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12932 Worklist.push_back(TI->getSuccessor(i));
12936 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12937 bool Changed = false;
12938 TD = &getAnalysis<TargetData>();
12940 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12941 << F.getNameStr() << "\n");
12944 // Do a depth-first traversal of the function, populate the worklist with
12945 // the reachable instructions. Ignore blocks that are not reachable. Keep
12946 // track of which blocks we visit.
12947 SmallPtrSet<BasicBlock*, 64> Visited;
12948 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12950 // Do a quick scan over the function. If we find any blocks that are
12951 // unreachable, remove any instructions inside of them. This prevents
12952 // the instcombine code from having to deal with some bad special cases.
12953 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12954 if (!Visited.count(BB)) {
12955 Instruction *Term = BB->getTerminator();
12956 while (Term != BB->begin()) { // Remove instrs bottom-up
12957 BasicBlock::iterator I = Term; --I;
12959 DOUT << "IC: DCE: " << *I;
12960 // A debug intrinsic shouldn't force another iteration if we weren't
12961 // going to do one without it.
12962 if (!isa<DbgInfoIntrinsic>(I)) {
12966 if (!I->use_empty())
12967 I->replaceAllUsesWith(Context->getUndef(I->getType()));
12968 I->eraseFromParent();
12973 while (!Worklist.empty()) {
12974 Instruction *I = RemoveOneFromWorkList();
12975 if (I == 0) continue; // skip null values.
12977 // Check to see if we can DCE the instruction.
12978 if (isInstructionTriviallyDead(I)) {
12979 // Add operands to the worklist.
12980 if (I->getNumOperands() < 4)
12981 AddUsesToWorkList(*I);
12984 DOUT << "IC: DCE: " << *I;
12986 I->eraseFromParent();
12987 RemoveFromWorkList(I);
12992 // Instruction isn't dead, see if we can constant propagate it.
12993 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
12994 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
12996 // Add operands to the worklist.
12997 AddUsesToWorkList(*I);
12998 ReplaceInstUsesWith(*I, C);
13001 I->eraseFromParent();
13002 RemoveFromWorkList(I);
13008 // See if we can constant fold its operands.
13009 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
13010 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
13011 if (Constant *NewC = ConstantFoldConstantExpression(CE,
13012 F.getContext(), TD))
13019 // See if we can trivially sink this instruction to a successor basic block.
13020 if (I->hasOneUse()) {
13021 BasicBlock *BB = I->getParent();
13022 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
13023 if (UserParent != BB) {
13024 bool UserIsSuccessor = false;
13025 // See if the user is one of our successors.
13026 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13027 if (*SI == UserParent) {
13028 UserIsSuccessor = true;
13032 // If the user is one of our immediate successors, and if that successor
13033 // only has us as a predecessors (we'd have to split the critical edge
13034 // otherwise), we can keep going.
13035 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
13036 next(pred_begin(UserParent)) == pred_end(UserParent))
13037 // Okay, the CFG is simple enough, try to sink this instruction.
13038 Changed |= TryToSinkInstruction(I, UserParent);
13042 // Now that we have an instruction, try combining it to simplify it...
13046 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
13047 if (Instruction *Result = visit(*I)) {
13049 // Should we replace the old instruction with a new one?
13051 DOUT << "IC: Old = " << *I
13052 << " New = " << *Result;
13054 // Everything uses the new instruction now.
13055 I->replaceAllUsesWith(Result);
13057 // Push the new instruction and any users onto the worklist.
13058 AddToWorkList(Result);
13059 AddUsersToWorkList(*Result);
13061 // Move the name to the new instruction first.
13062 Result->takeName(I);
13064 // Insert the new instruction into the basic block...
13065 BasicBlock *InstParent = I->getParent();
13066 BasicBlock::iterator InsertPos = I;
13068 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13069 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13072 InstParent->getInstList().insert(InsertPos, Result);
13074 // Make sure that we reprocess all operands now that we reduced their
13076 AddUsesToWorkList(*I);
13078 // Instructions can end up on the worklist more than once. Make sure
13079 // we do not process an instruction that has been deleted.
13080 RemoveFromWorkList(I);
13082 // Erase the old instruction.
13083 InstParent->getInstList().erase(I);
13086 DOUT << "IC: Mod = " << OrigI
13087 << " New = " << *I;
13090 // If the instruction was modified, it's possible that it is now dead.
13091 // if so, remove it.
13092 if (isInstructionTriviallyDead(I)) {
13093 // Make sure we process all operands now that we are reducing their
13095 AddUsesToWorkList(*I);
13097 // Instructions may end up in the worklist more than once. Erase all
13098 // occurrences of this instruction.
13099 RemoveFromWorkList(I);
13100 I->eraseFromParent();
13103 AddUsersToWorkList(*I);
13110 assert(WorklistMap.empty() && "Worklist empty, but map not?");
13112 // Do an explicit clear, this shrinks the map if needed.
13113 WorklistMap.clear();
13118 bool InstCombiner::runOnFunction(Function &F) {
13119 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13121 bool EverMadeChange = false;
13123 // Iterate while there is work to do.
13124 unsigned Iteration = 0;
13125 while (DoOneIteration(F, Iteration++))
13126 EverMadeChange = true;
13127 return EverMadeChange;
13130 FunctionPass *llvm::createInstructionCombiningPass() {
13131 return new InstCombiner();