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/Analysis/ConstantFolding.h"
44 #include "llvm/Analysis/ValueTracking.h"
45 #include "llvm/Target/TargetData.h"
46 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
47 #include "llvm/Transforms/Utils/Local.h"
48 #include "llvm/Support/CallSite.h"
49 #include "llvm/Support/ConstantRange.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/GetElementPtrTypeIterator.h"
53 #include "llvm/Support/InstVisitor.h"
54 #include "llvm/Support/MathExtras.h"
55 #include "llvm/Support/PatternMatch.h"
56 #include "llvm/Support/Compiler.h"
57 #include "llvm/ADT/DenseMap.h"
58 #include "llvm/ADT/SmallVector.h"
59 #include "llvm/ADT/SmallPtrSet.h"
60 #include "llvm/ADT/Statistic.h"
61 #include "llvm/ADT/STLExtras.h"
66 using namespace llvm::PatternMatch;
68 STATISTIC(NumCombined , "Number of insts combined");
69 STATISTIC(NumConstProp, "Number of constant folds");
70 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
71 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
72 STATISTIC(NumSunkInst , "Number of instructions sunk");
75 class VISIBILITY_HIDDEN InstCombiner
76 : public FunctionPass,
77 public InstVisitor<InstCombiner, Instruction*> {
78 // Worklist of all of the instructions that need to be simplified.
79 SmallVector<Instruction*, 256> Worklist;
80 DenseMap<Instruction*, unsigned> WorklistMap;
82 bool MustPreserveLCSSA;
84 static char ID; // Pass identification, replacement for typeid
85 InstCombiner() : FunctionPass(&ID) {}
87 LLVMContext *getContext() { return Context; }
89 /// AddToWorkList - Add the specified instruction to the worklist if it
90 /// isn't already in it.
91 void AddToWorkList(Instruction *I) {
92 if (WorklistMap.insert(std::make_pair(I, Worklist.size())).second)
93 Worklist.push_back(I);
96 // RemoveFromWorkList - remove I from the worklist if it exists.
97 void RemoveFromWorkList(Instruction *I) {
98 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
99 if (It == WorklistMap.end()) return; // Not in worklist.
101 // Don't bother moving everything down, just null out the slot.
102 Worklist[It->second] = 0;
104 WorklistMap.erase(It);
107 Instruction *RemoveOneFromWorkList() {
108 Instruction *I = Worklist.back();
110 WorklistMap.erase(I);
115 /// AddUsersToWorkList - When an instruction is simplified, add all users of
116 /// the instruction to the work lists because they might get more simplified
119 void AddUsersToWorkList(Value &I) {
120 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
122 AddToWorkList(cast<Instruction>(*UI));
125 /// AddUsesToWorkList - When an instruction is simplified, add operands to
126 /// the work lists because they might get more simplified now.
128 void AddUsesToWorkList(Instruction &I) {
129 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
130 if (Instruction *Op = dyn_cast<Instruction>(*i))
134 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
135 /// dead. Add all of its operands to the worklist, turning them into
136 /// undef's to reduce the number of uses of those instructions.
138 /// Return the specified operand before it is turned into an undef.
140 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
141 Value *R = I.getOperand(op);
143 for (User::op_iterator i = I.op_begin(), e = I.op_end(); i != e; ++i)
144 if (Instruction *Op = dyn_cast<Instruction>(*i)) {
146 // Set the operand to undef to drop the use.
147 *i = Context->getUndef(Op->getType());
154 virtual bool runOnFunction(Function &F);
156 bool DoOneIteration(Function &F, unsigned ItNum);
158 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
159 AU.addRequired<TargetData>();
160 AU.addPreservedID(LCSSAID);
161 AU.setPreservesCFG();
164 TargetData &getTargetData() const { return *TD; }
166 // Visitation implementation - Implement instruction combining for different
167 // instruction types. The semantics are as follows:
169 // null - No change was made
170 // I - Change was made, I is still valid, I may be dead though
171 // otherwise - Change was made, replace I with returned instruction
173 Instruction *visitAdd(BinaryOperator &I);
174 Instruction *visitFAdd(BinaryOperator &I);
175 Instruction *visitSub(BinaryOperator &I);
176 Instruction *visitFSub(BinaryOperator &I);
177 Instruction *visitMul(BinaryOperator &I);
178 Instruction *visitFMul(BinaryOperator &I);
179 Instruction *visitURem(BinaryOperator &I);
180 Instruction *visitSRem(BinaryOperator &I);
181 Instruction *visitFRem(BinaryOperator &I);
182 bool SimplifyDivRemOfSelect(BinaryOperator &I);
183 Instruction *commonRemTransforms(BinaryOperator &I);
184 Instruction *commonIRemTransforms(BinaryOperator &I);
185 Instruction *commonDivTransforms(BinaryOperator &I);
186 Instruction *commonIDivTransforms(BinaryOperator &I);
187 Instruction *visitUDiv(BinaryOperator &I);
188 Instruction *visitSDiv(BinaryOperator &I);
189 Instruction *visitFDiv(BinaryOperator &I);
190 Instruction *FoldAndOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
191 Instruction *visitAnd(BinaryOperator &I);
192 Instruction *FoldOrOfICmps(Instruction &I, ICmpInst *LHS, ICmpInst *RHS);
193 Instruction *FoldOrWithConstants(BinaryOperator &I, Value *Op,
194 Value *A, Value *B, Value *C);
195 Instruction *visitOr (BinaryOperator &I);
196 Instruction *visitXor(BinaryOperator &I);
197 Instruction *visitShl(BinaryOperator &I);
198 Instruction *visitAShr(BinaryOperator &I);
199 Instruction *visitLShr(BinaryOperator &I);
200 Instruction *commonShiftTransforms(BinaryOperator &I);
201 Instruction *FoldFCmp_IntToFP_Cst(FCmpInst &I, Instruction *LHSI,
203 Instruction *visitFCmpInst(FCmpInst &I);
204 Instruction *visitICmpInst(ICmpInst &I);
205 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
206 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
209 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
210 ConstantInt *DivRHS);
212 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
213 ICmpInst::Predicate Cond, Instruction &I);
214 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
216 Instruction *commonCastTransforms(CastInst &CI);
217 Instruction *commonIntCastTransforms(CastInst &CI);
218 Instruction *commonPointerCastTransforms(CastInst &CI);
219 Instruction *visitTrunc(TruncInst &CI);
220 Instruction *visitZExt(ZExtInst &CI);
221 Instruction *visitSExt(SExtInst &CI);
222 Instruction *visitFPTrunc(FPTruncInst &CI);
223 Instruction *visitFPExt(CastInst &CI);
224 Instruction *visitFPToUI(FPToUIInst &FI);
225 Instruction *visitFPToSI(FPToSIInst &FI);
226 Instruction *visitUIToFP(CastInst &CI);
227 Instruction *visitSIToFP(CastInst &CI);
228 Instruction *visitPtrToInt(PtrToIntInst &CI);
229 Instruction *visitIntToPtr(IntToPtrInst &CI);
230 Instruction *visitBitCast(BitCastInst &CI);
231 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
233 Instruction *FoldSelectIntoOp(SelectInst &SI, Value*, Value*);
234 Instruction *visitSelectInst(SelectInst &SI);
235 Instruction *visitSelectInstWithICmp(SelectInst &SI, ICmpInst *ICI);
236 Instruction *visitCallInst(CallInst &CI);
237 Instruction *visitInvokeInst(InvokeInst &II);
238 Instruction *visitPHINode(PHINode &PN);
239 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
240 Instruction *visitAllocationInst(AllocationInst &AI);
241 Instruction *visitFreeInst(FreeInst &FI);
242 Instruction *visitLoadInst(LoadInst &LI);
243 Instruction *visitStoreInst(StoreInst &SI);
244 Instruction *visitBranchInst(BranchInst &BI);
245 Instruction *visitSwitchInst(SwitchInst &SI);
246 Instruction *visitInsertElementInst(InsertElementInst &IE);
247 Instruction *visitExtractElementInst(ExtractElementInst &EI);
248 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
249 Instruction *visitExtractValueInst(ExtractValueInst &EV);
251 // visitInstruction - Specify what to return for unhandled instructions...
252 Instruction *visitInstruction(Instruction &I) { return 0; }
255 Instruction *visitCallSite(CallSite CS);
256 bool transformConstExprCastCall(CallSite CS);
257 Instruction *transformCallThroughTrampoline(CallSite CS);
258 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
259 bool DoXform = true);
260 bool WillNotOverflowSignedAdd(Value *LHS, Value *RHS);
261 DbgDeclareInst *hasOneUsePlusDeclare(Value *V);
265 // InsertNewInstBefore - insert an instruction New before instruction Old
266 // in the program. Add the new instruction to the worklist.
268 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
269 assert(New && New->getParent() == 0 &&
270 "New instruction already inserted into a basic block!");
271 BasicBlock *BB = Old.getParent();
272 BB->getInstList().insert(&Old, New); // Insert inst
277 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
278 /// This also adds the cast to the worklist. Finally, this returns the
280 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
282 if (V->getType() == Ty) return V;
284 if (Constant *CV = dyn_cast<Constant>(V))
285 return Context->getConstantExprCast(opc, CV, Ty);
287 Instruction *C = CastInst::Create(opc, V, Ty, V->getName(), &Pos);
292 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
293 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
297 // ReplaceInstUsesWith - This method is to be used when an instruction is
298 // found to be dead, replacable with another preexisting expression. Here
299 // we add all uses of I to the worklist, replace all uses of I with the new
300 // value, then return I, so that the inst combiner will know that I was
303 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
304 AddUsersToWorkList(I); // Add all modified instrs to worklist
306 I.replaceAllUsesWith(V);
309 // If we are replacing the instruction with itself, this must be in a
310 // segment of unreachable code, so just clobber the instruction.
311 I.replaceAllUsesWith(Context->getUndef(I.getType()));
316 // EraseInstFromFunction - When dealing with an instruction that has side
317 // effects or produces a void value, we can't rely on DCE to delete the
318 // instruction. Instead, visit methods should return the value returned by
320 Instruction *EraseInstFromFunction(Instruction &I) {
321 assert(I.use_empty() && "Cannot erase instruction that is used!");
322 AddUsesToWorkList(I);
323 RemoveFromWorkList(&I);
325 return 0; // Don't do anything with FI
328 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt &KnownZero,
329 APInt &KnownOne, unsigned Depth = 0) const {
330 return llvm::ComputeMaskedBits(V, Mask, KnownZero, KnownOne, TD, Depth);
333 bool MaskedValueIsZero(Value *V, const APInt &Mask,
334 unsigned Depth = 0) const {
335 return llvm::MaskedValueIsZero(V, Mask, TD, Depth);
337 unsigned ComputeNumSignBits(Value *Op, unsigned Depth = 0) const {
338 return llvm::ComputeNumSignBits(Op, TD, Depth);
343 /// SimplifyCommutative - This performs a few simplifications for
344 /// commutative operators.
345 bool SimplifyCommutative(BinaryOperator &I);
347 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
348 /// most-complex to least-complex order.
349 bool SimplifyCompare(CmpInst &I);
351 /// SimplifyDemandedUseBits - Attempts to replace V with a simpler value
352 /// based on the demanded bits.
353 Value *SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
354 APInt& KnownZero, APInt& KnownOne,
356 bool SimplifyDemandedBits(Use &U, APInt DemandedMask,
357 APInt& KnownZero, APInt& KnownOne,
360 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
361 /// SimplifyDemandedBits knows about. See if the instruction has any
362 /// properties that allow us to simplify its operands.
363 bool SimplifyDemandedInstructionBits(Instruction &Inst);
365 Value *SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
366 APInt& UndefElts, unsigned Depth = 0);
368 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
369 // PHI node as operand #0, see if we can fold the instruction into the PHI
370 // (which is only possible if all operands to the PHI are constants).
371 Instruction *FoldOpIntoPhi(Instruction &I);
373 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
374 // operator and they all are only used by the PHI, PHI together their
375 // inputs, and do the operation once, to the result of the PHI.
376 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
377 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
378 Instruction *FoldPHIArgGEPIntoPHI(PHINode &PN);
381 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
382 ConstantInt *AndRHS, BinaryOperator &TheAnd);
384 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
385 bool isSub, Instruction &I);
386 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
387 bool isSigned, bool Inside, Instruction &IB);
388 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
389 Instruction *MatchBSwap(BinaryOperator &I);
390 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
391 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
392 Instruction *SimplifyMemSet(MemSetInst *MI);
395 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
397 bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
398 unsigned CastOpc, int &NumCastsRemoved);
399 unsigned GetOrEnforceKnownAlignment(Value *V,
400 unsigned PrefAlign = 0);
405 char InstCombiner::ID = 0;
406 static RegisterPass<InstCombiner>
407 X("instcombine", "Combine redundant instructions");
409 // getComplexity: Assign a complexity or rank value to LLVM Values...
410 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
411 static unsigned getComplexity(Value *V) {
412 if (isa<Instruction>(V)) {
413 if (BinaryOperator::isNeg(V) || BinaryOperator::isFNeg(V) ||
414 BinaryOperator::isNot(V))
418 if (isa<Argument>(V)) return 3;
419 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
422 // isOnlyUse - Return true if this instruction will be deleted if we stop using
424 static bool isOnlyUse(Value *V) {
425 return V->hasOneUse() || isa<Constant>(V);
428 // getPromotedType - Return the specified type promoted as it would be to pass
429 // though a va_arg area...
430 static const Type *getPromotedType(const Type *Ty) {
431 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
432 if (ITy->getBitWidth() < 32)
433 return Type::Int32Ty;
438 /// getBitCastOperand - If the specified operand is a CastInst, a constant
439 /// expression bitcast, or a GetElementPtrInst with all zero indices, return the
440 /// operand value, otherwise return null.
441 static Value *getBitCastOperand(Value *V) {
442 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
444 return I->getOperand(0);
445 else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
446 // GetElementPtrInst?
447 if (GEP->hasAllZeroIndices())
448 return GEP->getOperand(0);
449 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V)) {
450 if (CE->getOpcode() == Instruction::BitCast)
451 // BitCast ConstantExp?
452 return CE->getOperand(0);
453 else if (CE->getOpcode() == Instruction::GetElementPtr) {
454 // GetElementPtr ConstantExp?
455 for (User::op_iterator I = CE->op_begin() + 1, E = CE->op_end();
457 ConstantInt *CI = dyn_cast<ConstantInt>(I);
458 if (!CI || !CI->isZero())
459 // Any non-zero indices? Not cast-like.
462 // All-zero indices? This is just like casting.
463 return CE->getOperand(0);
469 /// This function is a wrapper around CastInst::isEliminableCastPair. It
470 /// simply extracts arguments and returns what that function returns.
471 static Instruction::CastOps
472 isEliminableCastPair(
473 const CastInst *CI, ///< The first cast instruction
474 unsigned opcode, ///< The opcode of the second cast instruction
475 const Type *DstTy, ///< The target type for the second cast instruction
476 TargetData *TD ///< The target data for pointer size
479 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
480 const Type *MidTy = CI->getType(); // B from above
482 // Get the opcodes of the two Cast instructions
483 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
484 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
486 unsigned Res = CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
487 DstTy, TD->getIntPtrType());
489 // We don't want to form an inttoptr or ptrtoint that converts to an integer
490 // type that differs from the pointer size.
491 if ((Res == Instruction::IntToPtr && SrcTy != TD->getIntPtrType()) ||
492 (Res == Instruction::PtrToInt && DstTy != TD->getIntPtrType()))
495 return Instruction::CastOps(Res);
498 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
499 /// in any code being generated. It does not require codegen if V is simple
500 /// enough or if the cast can be folded into other casts.
501 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
502 const Type *Ty, TargetData *TD) {
503 if (V->getType() == Ty || isa<Constant>(V)) return false;
505 // If this is another cast that can be eliminated, it isn't codegen either.
506 if (const CastInst *CI = dyn_cast<CastInst>(V))
507 if (isEliminableCastPair(CI, opcode, Ty, TD))
512 // SimplifyCommutative - This performs a few simplifications for commutative
515 // 1. Order operands such that they are listed from right (least complex) to
516 // left (most complex). This puts constants before unary operators before
519 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
520 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
522 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
523 bool Changed = false;
524 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
525 Changed = !I.swapOperands();
527 if (!I.isAssociative()) return Changed;
528 Instruction::BinaryOps Opcode = I.getOpcode();
529 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
530 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
531 if (isa<Constant>(I.getOperand(1))) {
532 Constant *Folded = Context->getConstantExpr(I.getOpcode(),
533 cast<Constant>(I.getOperand(1)),
534 cast<Constant>(Op->getOperand(1)));
535 I.setOperand(0, Op->getOperand(0));
536 I.setOperand(1, Folded);
538 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
539 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
540 isOnlyUse(Op) && isOnlyUse(Op1)) {
541 Constant *C1 = cast<Constant>(Op->getOperand(1));
542 Constant *C2 = cast<Constant>(Op1->getOperand(1));
544 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
545 Constant *Folded = Context->getConstantExpr(I.getOpcode(), C1, C2);
546 Instruction *New = BinaryOperator::Create(Opcode, Op->getOperand(0),
550 I.setOperand(0, New);
551 I.setOperand(1, Folded);
558 /// SimplifyCompare - For a CmpInst this function just orders the operands
559 /// so that theyare listed from right (least complex) to left (most complex).
560 /// This puts constants before unary operators before binary operators.
561 bool InstCombiner::SimplifyCompare(CmpInst &I) {
562 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
565 // Compare instructions are not associative so there's nothing else we can do.
569 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
570 // if the LHS is a constant zero (which is the 'negate' form).
572 static inline Value *dyn_castNegVal(Value *V, LLVMContext *Context) {
573 if (BinaryOperator::isNeg(V))
574 return BinaryOperator::getNegArgument(V);
576 // Constants can be considered to be negated values if they can be folded.
577 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
578 return Context->getConstantExprNeg(C);
580 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
581 if (C->getType()->getElementType()->isInteger())
582 return Context->getConstantExprNeg(C);
587 // dyn_castFNegVal - Given a 'fsub' instruction, return the RHS of the
588 // instruction if the LHS is a constant negative zero (which is the 'negate'
591 static inline Value *dyn_castFNegVal(Value *V, LLVMContext *Context) {
592 if (BinaryOperator::isFNeg(V))
593 return BinaryOperator::getFNegArgument(V);
595 // Constants can be considered to be negated values if they can be folded.
596 if (ConstantFP *C = dyn_cast<ConstantFP>(V))
597 return Context->getConstantExprFNeg(C);
599 if (ConstantVector *C = dyn_cast<ConstantVector>(V))
600 if (C->getType()->getElementType()->isFloatingPoint())
601 return Context->getConstantExprFNeg(C);
606 static inline Value *dyn_castNotVal(Value *V, LLVMContext *Context) {
607 if (BinaryOperator::isNot(V))
608 return BinaryOperator::getNotArgument(V);
610 // Constants can be considered to be not'ed values...
611 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
612 return Context->getConstantInt(~C->getValue());
616 // dyn_castFoldableMul - If this value is a multiply that can be folded into
617 // other computations (because it has a constant operand), return the
618 // non-constant operand of the multiply, and set CST to point to the multiplier.
619 // Otherwise, return null.
621 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST,
622 LLVMContext *Context) {
623 if (V->hasOneUse() && V->getType()->isInteger())
624 if (Instruction *I = dyn_cast<Instruction>(V)) {
625 if (I->getOpcode() == Instruction::Mul)
626 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
627 return I->getOperand(0);
628 if (I->getOpcode() == Instruction::Shl)
629 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
630 // The multiplier is really 1 << CST.
631 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
632 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
633 CST = Context->getConstantInt(APInt(BitWidth, 1).shl(CSTVal));
634 return I->getOperand(0);
640 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
641 /// expression, return it.
642 static User *dyn_castGetElementPtr(Value *V) {
643 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
644 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
645 if (CE->getOpcode() == Instruction::GetElementPtr)
646 return cast<User>(V);
650 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
651 /// opcode value. Otherwise return UserOp1.
652 static unsigned getOpcode(const Value *V) {
653 if (const Instruction *I = dyn_cast<Instruction>(V))
654 return I->getOpcode();
655 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
656 return CE->getOpcode();
657 // Use UserOp1 to mean there's no opcode.
658 return Instruction::UserOp1;
661 /// AddOne - Add one to a ConstantInt
662 static Constant *AddOne(Constant *C, LLVMContext *Context) {
663 return Context->getConstantExprAdd(C,
664 Context->getConstantInt(C->getType(), 1));
666 /// SubOne - Subtract one from a ConstantInt
667 static Constant *SubOne(ConstantInt *C, LLVMContext *Context) {
668 return Context->getConstantExprSub(C,
669 Context->getConstantInt(C->getType(), 1));
671 /// MultiplyOverflows - True if the multiply can not be expressed in an int
673 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign,
674 LLVMContext *Context) {
675 uint32_t W = C1->getBitWidth();
676 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
685 APInt MulExt = LHSExt * RHSExt;
688 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
689 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
690 return MulExt.slt(Min) || MulExt.sgt(Max);
692 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
696 /// ShrinkDemandedConstant - Check to see if the specified operand of the
697 /// specified instruction is a constant integer. If so, check to see if there
698 /// are any bits set in the constant that are not demanded. If so, shrink the
699 /// constant and return true.
700 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
701 APInt Demanded, LLVMContext *Context) {
702 assert(I && "No instruction?");
703 assert(OpNo < I->getNumOperands() && "Operand index too large");
705 // If the operand is not a constant integer, nothing to do.
706 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
707 if (!OpC) return false;
709 // If there are no bits set that aren't demanded, nothing to do.
710 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
711 if ((~Demanded & OpC->getValue()) == 0)
714 // This instruction is producing bits that are not demanded. Shrink the RHS.
715 Demanded &= OpC->getValue();
716 I->setOperand(OpNo, Context->getConstantInt(Demanded));
720 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
721 // 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 ComputeSignedMinMaxValuesFromKnownBits(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 "KnownZero, KnownOne and Min, Max must have equal bitwidth.");
731 APInt UnknownBits = ~(KnownZero|KnownOne);
733 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
734 // bit if it is unknown.
736 Max = KnownOne|UnknownBits;
738 if (UnknownBits.isNegative()) { // Sign bit is unknown
739 Min.set(Min.getBitWidth()-1);
740 Max.clear(Max.getBitWidth()-1);
744 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
745 // a set of known zero and one bits, compute the maximum and minimum values that
746 // could have the specified known zero and known one bits, returning them in
748 static void ComputeUnsignedMinMaxValuesFromKnownBits(const APInt &KnownZero,
749 const APInt &KnownOne,
750 APInt &Min, APInt &Max) {
751 assert(KnownZero.getBitWidth() == KnownOne.getBitWidth() &&
752 KnownZero.getBitWidth() == Min.getBitWidth() &&
753 KnownZero.getBitWidth() == Max.getBitWidth() &&
754 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
755 APInt UnknownBits = ~(KnownZero|KnownOne);
757 // The minimum value is when the unknown bits are all zeros.
759 // The maximum value is when the unknown bits are all ones.
760 Max = KnownOne|UnknownBits;
763 /// SimplifyDemandedInstructionBits - Inst is an integer instruction that
764 /// SimplifyDemandedBits knows about. See if the instruction has any
765 /// properties that allow us to simplify its operands.
766 bool InstCombiner::SimplifyDemandedInstructionBits(Instruction &Inst) {
767 unsigned BitWidth = Inst.getType()->getScalarSizeInBits();
768 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
769 APInt DemandedMask(APInt::getAllOnesValue(BitWidth));
771 Value *V = SimplifyDemandedUseBits(&Inst, DemandedMask,
772 KnownZero, KnownOne, 0);
773 if (V == 0) return false;
774 if (V == &Inst) return true;
775 ReplaceInstUsesWith(Inst, V);
779 /// SimplifyDemandedBits - This form of SimplifyDemandedBits simplifies the
780 /// specified instruction operand if possible, updating it in place. It returns
781 /// true if it made any change and false otherwise.
782 bool InstCombiner::SimplifyDemandedBits(Use &U, APInt DemandedMask,
783 APInt &KnownZero, APInt &KnownOne,
785 Value *NewVal = SimplifyDemandedUseBits(U.get(), DemandedMask,
786 KnownZero, KnownOne, Depth);
787 if (NewVal == 0) return false;
793 /// SimplifyDemandedUseBits - This function attempts to replace V with a simpler
794 /// value based on the demanded bits. When this function is called, it is known
795 /// that only the bits set in DemandedMask of the result of V are ever used
796 /// downstream. Consequently, depending on the mask and V, it may be possible
797 /// to replace V with a constant or one of its operands. In such cases, this
798 /// function does the replacement and returns true. In all other cases, it
799 /// returns false after analyzing the expression and setting KnownOne and known
800 /// to be one in the expression. KnownZero contains all the bits that are known
801 /// to be zero in the expression. These are provided to potentially allow the
802 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
803 /// the expression. KnownOne and KnownZero always follow the invariant that
804 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
805 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
806 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
807 /// and KnownOne must all be the same.
809 /// This returns null if it did not change anything and it permits no
810 /// simplification. This returns V itself if it did some simplification of V's
811 /// operands based on the information about what bits are demanded. This returns
812 /// some other non-null value if it found out that V is equal to another value
813 /// in the context where the specified bits are demanded, but not for all users.
814 Value *InstCombiner::SimplifyDemandedUseBits(Value *V, APInt DemandedMask,
815 APInt &KnownZero, APInt &KnownOne,
817 assert(V != 0 && "Null pointer of Value???");
818 assert(Depth <= 6 && "Limit Search Depth");
819 uint32_t BitWidth = DemandedMask.getBitWidth();
820 const Type *VTy = V->getType();
821 assert((TD || !isa<PointerType>(VTy)) &&
822 "SimplifyDemandedBits needs to know bit widths!");
823 assert((!TD || TD->getTypeSizeInBits(VTy->getScalarType()) == BitWidth) &&
824 (!VTy->isIntOrIntVector() ||
825 VTy->getScalarSizeInBits() == BitWidth) &&
826 KnownZero.getBitWidth() == BitWidth &&
827 KnownOne.getBitWidth() == BitWidth &&
828 "Value *V, DemandedMask, KnownZero and KnownOne "
829 "must have same BitWidth");
830 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
831 // We know all of the bits for a constant!
832 KnownOne = CI->getValue() & DemandedMask;
833 KnownZero = ~KnownOne & DemandedMask;
836 if (isa<ConstantPointerNull>(V)) {
837 // We know all of the bits for a constant!
839 KnownZero = DemandedMask;
845 if (DemandedMask == 0) { // Not demanding any bits from V.
846 if (isa<UndefValue>(V))
848 return Context->getUndef(VTy);
851 if (Depth == 6) // Limit search depth.
854 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
855 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
857 Instruction *I = dyn_cast<Instruction>(V);
859 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
860 return 0; // Only analyze instructions.
863 // If there are multiple uses of this value and we aren't at the root, then
864 // we can't do any simplifications of the operands, because DemandedMask
865 // only reflects the bits demanded by *one* of the users.
866 if (Depth != 0 && !I->hasOneUse()) {
867 // Despite the fact that we can't simplify this instruction in all User's
868 // context, we can at least compute the knownzero/knownone bits, and we can
869 // do simplifications that apply to *just* the one user if we know that
870 // this instruction has a simpler value in that context.
871 if (I->getOpcode() == Instruction::And) {
872 // If either the LHS or the RHS are Zero, the result is zero.
873 ComputeMaskedBits(I->getOperand(1), DemandedMask,
874 RHSKnownZero, RHSKnownOne, Depth+1);
875 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
876 LHSKnownZero, LHSKnownOne, Depth+1);
878 // If all of the demanded bits are known 1 on one side, return the other.
879 // These bits cannot contribute to the result of the 'and' in this
881 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
882 (DemandedMask & ~LHSKnownZero))
883 return I->getOperand(0);
884 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
885 (DemandedMask & ~RHSKnownZero))
886 return I->getOperand(1);
888 // If all of the demanded bits in the inputs are known zeros, return zero.
889 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
890 return Context->getNullValue(VTy);
892 } else if (I->getOpcode() == Instruction::Or) {
893 // We can simplify (X|Y) -> X or Y in the user's context if we know that
894 // only bits from X or Y are demanded.
896 // If either the LHS or the RHS are One, the result is One.
897 ComputeMaskedBits(I->getOperand(1), DemandedMask,
898 RHSKnownZero, RHSKnownOne, Depth+1);
899 ComputeMaskedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
900 LHSKnownZero, LHSKnownOne, Depth+1);
902 // If all of the demanded bits are known zero on one side, return the
903 // other. These bits cannot contribute to the result of the 'or' in this
905 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
906 (DemandedMask & ~LHSKnownOne))
907 return I->getOperand(0);
908 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
909 (DemandedMask & ~RHSKnownOne))
910 return I->getOperand(1);
912 // If all of the potentially set bits on one side are known to be set on
913 // the other side, just use the 'other' side.
914 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
915 (DemandedMask & (~RHSKnownZero)))
916 return I->getOperand(0);
917 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
918 (DemandedMask & (~LHSKnownZero)))
919 return I->getOperand(1);
922 // Compute the KnownZero/KnownOne bits to simplify things downstream.
923 ComputeMaskedBits(I, DemandedMask, KnownZero, KnownOne, Depth);
927 // If this is the root being simplified, allow it to have multiple uses,
928 // just set the DemandedMask to all bits so that we can try to simplify the
929 // operands. This allows visitTruncInst (for example) to simplify the
930 // operand of a trunc without duplicating all the logic below.
931 if (Depth == 0 && !V->hasOneUse())
932 DemandedMask = APInt::getAllOnesValue(BitWidth);
934 switch (I->getOpcode()) {
936 ComputeMaskedBits(I, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
938 case Instruction::And:
939 // If either the LHS or the RHS are Zero, the result is zero.
940 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
941 RHSKnownZero, RHSKnownOne, Depth+1) ||
942 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownZero,
943 LHSKnownZero, LHSKnownOne, Depth+1))
945 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
946 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
948 // If all of the demanded bits are known 1 on one side, return the other.
949 // These bits cannot contribute to the result of the 'and'.
950 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
951 (DemandedMask & ~LHSKnownZero))
952 return I->getOperand(0);
953 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
954 (DemandedMask & ~RHSKnownZero))
955 return I->getOperand(1);
957 // If all of the demanded bits in the inputs are known zeros, return zero.
958 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
959 return Context->getNullValue(VTy);
961 // If the RHS is a constant, see if we can simplify it.
962 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero, Context))
965 // Output known-1 bits are only known if set in both the LHS & RHS.
966 RHSKnownOne &= LHSKnownOne;
967 // Output known-0 are known to be clear if zero in either the LHS | RHS.
968 RHSKnownZero |= LHSKnownZero;
970 case Instruction::Or:
971 // If either the LHS or the RHS are One, the result is One.
972 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
973 RHSKnownZero, RHSKnownOne, Depth+1) ||
974 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask & ~RHSKnownOne,
975 LHSKnownZero, LHSKnownOne, Depth+1))
977 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
978 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
980 // If all of the demanded bits are known zero on one side, return the other.
981 // These bits cannot contribute to the result of the 'or'.
982 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
983 (DemandedMask & ~LHSKnownOne))
984 return I->getOperand(0);
985 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
986 (DemandedMask & ~RHSKnownOne))
987 return I->getOperand(1);
989 // If all of the potentially set bits on one side are known to be set on
990 // the other side, just use the 'other' side.
991 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
992 (DemandedMask & (~RHSKnownZero)))
993 return I->getOperand(0);
994 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
995 (DemandedMask & (~LHSKnownZero)))
996 return I->getOperand(1);
998 // If the RHS is a constant, see if we can simplify it.
999 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context))
1002 // Output known-0 bits are only known if clear in both the LHS & RHS.
1003 RHSKnownZero &= LHSKnownZero;
1004 // Output known-1 are known to be set if set in either the LHS | RHS.
1005 RHSKnownOne |= LHSKnownOne;
1007 case Instruction::Xor: {
1008 if (SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1009 RHSKnownZero, RHSKnownOne, Depth+1) ||
1010 SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1011 LHSKnownZero, LHSKnownOne, Depth+1))
1013 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1014 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1016 // If all of the demanded bits are known zero on one side, return the other.
1017 // These bits cannot contribute to the result of the 'xor'.
1018 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1019 return I->getOperand(0);
1020 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1021 return I->getOperand(1);
1023 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1024 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1025 (RHSKnownOne & LHSKnownOne);
1026 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1027 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1028 (RHSKnownOne & LHSKnownZero);
1030 // If all of the demanded bits are known to be zero on one side or the
1031 // other, turn this into an *inclusive* or.
1032 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1033 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1035 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1037 return InsertNewInstBefore(Or, *I);
1040 // If all of the demanded bits on one side are known, and all of the set
1041 // bits on that side are also known to be set on the other side, turn this
1042 // into an AND, as we know the bits will be cleared.
1043 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1044 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1046 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1047 Constant *AndC = Context->getConstantInt(~RHSKnownOne & DemandedMask);
1049 BinaryOperator::CreateAnd(I->getOperand(0), AndC, "tmp");
1050 return InsertNewInstBefore(And, *I);
1054 // If the RHS is a constant, see if we can simplify it.
1055 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1056 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context))
1059 RHSKnownZero = KnownZeroOut;
1060 RHSKnownOne = KnownOneOut;
1063 case Instruction::Select:
1064 if (SimplifyDemandedBits(I->getOperandUse(2), DemandedMask,
1065 RHSKnownZero, RHSKnownOne, Depth+1) ||
1066 SimplifyDemandedBits(I->getOperandUse(1), DemandedMask,
1067 LHSKnownZero, LHSKnownOne, Depth+1))
1069 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1070 assert(!(LHSKnownZero & LHSKnownOne) && "Bits known to be one AND zero?");
1072 // If the operands are constants, see if we can simplify them.
1073 if (ShrinkDemandedConstant(I, 1, DemandedMask, Context) ||
1074 ShrinkDemandedConstant(I, 2, DemandedMask, Context))
1077 // Only known if known in both the LHS and RHS.
1078 RHSKnownOne &= LHSKnownOne;
1079 RHSKnownZero &= LHSKnownZero;
1081 case Instruction::Trunc: {
1082 unsigned truncBf = I->getOperand(0)->getType()->getScalarSizeInBits();
1083 DemandedMask.zext(truncBf);
1084 RHSKnownZero.zext(truncBf);
1085 RHSKnownOne.zext(truncBf);
1086 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1087 RHSKnownZero, RHSKnownOne, Depth+1))
1089 DemandedMask.trunc(BitWidth);
1090 RHSKnownZero.trunc(BitWidth);
1091 RHSKnownOne.trunc(BitWidth);
1092 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1095 case Instruction::BitCast:
1096 if (!I->getOperand(0)->getType()->isIntOrIntVector())
1097 return false; // vector->int or fp->int?
1099 if (const VectorType *DstVTy = dyn_cast<VectorType>(I->getType())) {
1100 if (const VectorType *SrcVTy =
1101 dyn_cast<VectorType>(I->getOperand(0)->getType())) {
1102 if (DstVTy->getNumElements() != SrcVTy->getNumElements())
1103 // Don't touch a bitcast between vectors of different element counts.
1106 // Don't touch a scalar-to-vector bitcast.
1108 } else if (isa<VectorType>(I->getOperand(0)->getType()))
1109 // Don't touch a vector-to-scalar bitcast.
1112 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1113 RHSKnownZero, RHSKnownOne, Depth+1))
1115 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1117 case Instruction::ZExt: {
1118 // Compute the bits in the result that are not present in the input.
1119 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1121 DemandedMask.trunc(SrcBitWidth);
1122 RHSKnownZero.trunc(SrcBitWidth);
1123 RHSKnownOne.trunc(SrcBitWidth);
1124 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMask,
1125 RHSKnownZero, RHSKnownOne, Depth+1))
1127 DemandedMask.zext(BitWidth);
1128 RHSKnownZero.zext(BitWidth);
1129 RHSKnownOne.zext(BitWidth);
1130 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1131 // The top bits are known to be zero.
1132 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1135 case Instruction::SExt: {
1136 // Compute the bits in the result that are not present in the input.
1137 unsigned SrcBitWidth =I->getOperand(0)->getType()->getScalarSizeInBits();
1139 APInt InputDemandedBits = DemandedMask &
1140 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1142 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1143 // If any of the sign extended bits are demanded, we know that the sign
1145 if ((NewBits & DemandedMask) != 0)
1146 InputDemandedBits.set(SrcBitWidth-1);
1148 InputDemandedBits.trunc(SrcBitWidth);
1149 RHSKnownZero.trunc(SrcBitWidth);
1150 RHSKnownOne.trunc(SrcBitWidth);
1151 if (SimplifyDemandedBits(I->getOperandUse(0), InputDemandedBits,
1152 RHSKnownZero, RHSKnownOne, Depth+1))
1154 InputDemandedBits.zext(BitWidth);
1155 RHSKnownZero.zext(BitWidth);
1156 RHSKnownOne.zext(BitWidth);
1157 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1159 // If the sign bit of the input is known set or clear, then we know the
1160 // top bits of the result.
1162 // If the input sign bit is known zero, or if the NewBits are not demanded
1163 // convert this into a zero extension.
1164 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits) {
1165 // Convert to ZExt cast
1166 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName());
1167 return InsertNewInstBefore(NewCast, *I);
1168 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1169 RHSKnownOne |= NewBits;
1173 case Instruction::Add: {
1174 // Figure out what the input bits are. If the top bits of the and result
1175 // are not demanded, then the add doesn't demand them from its input
1177 unsigned NLZ = DemandedMask.countLeadingZeros();
1179 // If there is a constant on the RHS, there are a variety of xformations
1181 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1182 // If null, this should be simplified elsewhere. Some of the xforms here
1183 // won't work if the RHS is zero.
1187 // If the top bit of the output is demanded, demand everything from the
1188 // input. Otherwise, we demand all the input bits except NLZ top bits.
1189 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1191 // Find information about known zero/one bits in the input.
1192 if (SimplifyDemandedBits(I->getOperandUse(0), InDemandedBits,
1193 LHSKnownZero, LHSKnownOne, Depth+1))
1196 // If the RHS of the add has bits set that can't affect the input, reduce
1198 if (ShrinkDemandedConstant(I, 1, InDemandedBits, Context))
1201 // Avoid excess work.
1202 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1205 // Turn it into OR if input bits are zero.
1206 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1208 BinaryOperator::CreateOr(I->getOperand(0), I->getOperand(1),
1210 return InsertNewInstBefore(Or, *I);
1213 // We can say something about the output known-zero and known-one bits,
1214 // depending on potential carries from the input constant and the
1215 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1216 // bits set and the RHS constant is 0x01001, then we know we have a known
1217 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1219 // To compute this, we first compute the potential carry bits. These are
1220 // the bits which may be modified. I'm not aware of a better way to do
1222 const APInt &RHSVal = RHS->getValue();
1223 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1225 // Now that we know which bits have carries, compute the known-1/0 sets.
1227 // Bits are known one if they are known zero in one operand and one in the
1228 // other, and there is no input carry.
1229 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1230 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1232 // Bits are known zero if they are known zero in both operands and there
1233 // is no input carry.
1234 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1236 // If the high-bits of this ADD are not demanded, then it does not demand
1237 // the high bits of its LHS or RHS.
1238 if (DemandedMask[BitWidth-1] == 0) {
1239 // Right fill the mask of bits for this ADD to demand the most
1240 // significant bit and all those below it.
1241 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1242 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1243 LHSKnownZero, LHSKnownOne, Depth+1) ||
1244 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1245 LHSKnownZero, LHSKnownOne, Depth+1))
1251 case Instruction::Sub:
1252 // If the high-bits of this SUB are not demanded, then it does not demand
1253 // the high bits of its LHS or RHS.
1254 if (DemandedMask[BitWidth-1] == 0) {
1255 // Right fill the mask of bits for this SUB to demand the most
1256 // significant bit and all those below it.
1257 uint32_t NLZ = DemandedMask.countLeadingZeros();
1258 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1259 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedFromOps,
1260 LHSKnownZero, LHSKnownOne, Depth+1) ||
1261 SimplifyDemandedBits(I->getOperandUse(1), DemandedFromOps,
1262 LHSKnownZero, LHSKnownOne, Depth+1))
1265 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1266 // the known zeros and ones.
1267 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1269 case Instruction::Shl:
1270 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1271 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1272 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1273 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1274 RHSKnownZero, RHSKnownOne, Depth+1))
1276 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1277 RHSKnownZero <<= ShiftAmt;
1278 RHSKnownOne <<= ShiftAmt;
1279 // low bits known zero.
1281 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1284 case Instruction::LShr:
1285 // For a logical shift right
1286 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1287 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1289 // Unsigned shift right.
1290 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1291 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1292 RHSKnownZero, RHSKnownOne, Depth+1))
1294 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1295 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1296 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1298 // Compute the new bits that are at the top now.
1299 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1300 RHSKnownZero |= HighBits; // high bits known zero.
1304 case Instruction::AShr:
1305 // If this is an arithmetic shift right and only the low-bit is set, we can
1306 // always convert this into a logical shr, even if the shift amount is
1307 // variable. The low bit of the shift cannot be an input sign bit unless
1308 // the shift amount is >= the size of the datatype, which is undefined.
1309 if (DemandedMask == 1) {
1310 // Perform the logical shift right.
1311 Instruction *NewVal = BinaryOperator::CreateLShr(
1312 I->getOperand(0), I->getOperand(1), I->getName());
1313 return InsertNewInstBefore(NewVal, *I);
1316 // If the sign bit is the only bit demanded by this ashr, then there is no
1317 // need to do it, the shift doesn't change the high bit.
1318 if (DemandedMask.isSignBit())
1319 return I->getOperand(0);
1321 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1322 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1324 // Signed shift right.
1325 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1326 // If any of the "high bits" are demanded, we should set the sign bit as
1328 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1329 DemandedMaskIn.set(BitWidth-1);
1330 if (SimplifyDemandedBits(I->getOperandUse(0), DemandedMaskIn,
1331 RHSKnownZero, RHSKnownOne, Depth+1))
1333 assert(!(RHSKnownZero & RHSKnownOne) && "Bits known to be one AND zero?");
1334 // Compute the new bits that are at the top now.
1335 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1336 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1337 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1339 // Handle the sign bits.
1340 APInt SignBit(APInt::getSignBit(BitWidth));
1341 // Adjust to where it is now in the mask.
1342 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1344 // If the input sign bit is known to be zero, or if none of the top bits
1345 // are demanded, turn this into an unsigned shift right.
1346 if (BitWidth <= ShiftAmt || RHSKnownZero[BitWidth-ShiftAmt-1] ||
1347 (HighBits & ~DemandedMask) == HighBits) {
1348 // Perform the logical shift right.
1349 Instruction *NewVal = BinaryOperator::CreateLShr(
1350 I->getOperand(0), SA, I->getName());
1351 return InsertNewInstBefore(NewVal, *I);
1352 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1353 RHSKnownOne |= HighBits;
1357 case Instruction::SRem:
1358 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1359 APInt RA = Rem->getValue().abs();
1360 if (RA.isPowerOf2()) {
1361 if (DemandedMask.ult(RA)) // srem won't affect demanded bits
1362 return I->getOperand(0);
1364 APInt LowBits = RA - 1;
1365 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1366 if (SimplifyDemandedBits(I->getOperandUse(0), Mask2,
1367 LHSKnownZero, LHSKnownOne, Depth+1))
1370 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1371 LHSKnownZero |= ~LowBits;
1373 KnownZero |= LHSKnownZero & DemandedMask;
1375 assert(!(KnownZero & KnownOne) && "Bits known to be one AND zero?");
1379 case Instruction::URem: {
1380 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1381 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1382 if (SimplifyDemandedBits(I->getOperandUse(0), AllOnes,
1383 KnownZero2, KnownOne2, Depth+1) ||
1384 SimplifyDemandedBits(I->getOperandUse(1), AllOnes,
1385 KnownZero2, KnownOne2, Depth+1))
1388 unsigned Leaders = KnownZero2.countLeadingOnes();
1389 Leaders = std::max(Leaders,
1390 KnownZero2.countLeadingOnes());
1391 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1394 case Instruction::Call:
1395 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1396 switch (II->getIntrinsicID()) {
1398 case Intrinsic::bswap: {
1399 // If the only bits demanded come from one byte of the bswap result,
1400 // just shift the input byte into position to eliminate the bswap.
1401 unsigned NLZ = DemandedMask.countLeadingZeros();
1402 unsigned NTZ = DemandedMask.countTrailingZeros();
1404 // Round NTZ down to the next byte. If we have 11 trailing zeros, then
1405 // we need all the bits down to bit 8. Likewise, round NLZ. If we
1406 // have 14 leading zeros, round to 8.
1409 // If we need exactly one byte, we can do this transformation.
1410 if (BitWidth-NLZ-NTZ == 8) {
1411 unsigned ResultBit = NTZ;
1412 unsigned InputBit = BitWidth-NTZ-8;
1414 // Replace this with either a left or right shift to get the byte into
1416 Instruction *NewVal;
1417 if (InputBit > ResultBit)
1418 NewVal = BinaryOperator::CreateLShr(I->getOperand(1),
1419 Context->getConstantInt(I->getType(), InputBit-ResultBit));
1421 NewVal = BinaryOperator::CreateShl(I->getOperand(1),
1422 Context->getConstantInt(I->getType(), ResultBit-InputBit));
1423 NewVal->takeName(I);
1424 return InsertNewInstBefore(NewVal, *I);
1427 // TODO: Could compute known zero/one bits based on the input.
1432 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1436 // If the client is only demanding bits that we know, return the known
1438 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1439 Constant *C = Context->getConstantInt(RHSKnownOne);
1440 if (isa<PointerType>(V->getType()))
1441 C = Context->getConstantExprIntToPtr(C, V->getType());
1448 /// SimplifyDemandedVectorElts - The specified value produces a vector with
1449 /// any number of elements. DemandedElts contains the set of elements that are
1450 /// actually used by the caller. This method analyzes which elements of the
1451 /// operand are undef and returns that information in UndefElts.
1453 /// If the information about demanded elements can be used to simplify the
1454 /// operation, the operation is simplified, then the resultant value is
1455 /// returned. This returns null if no change was made.
1456 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, APInt DemandedElts,
1459 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1460 APInt EltMask(APInt::getAllOnesValue(VWidth));
1461 assert((DemandedElts & ~EltMask) == 0 && "Invalid DemandedElts!");
1463 if (isa<UndefValue>(V)) {
1464 // If the entire vector is undefined, just return this info.
1465 UndefElts = EltMask;
1467 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1468 UndefElts = EltMask;
1469 return Context->getUndef(V->getType());
1473 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1474 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1475 Constant *Undef = Context->getUndef(EltTy);
1477 std::vector<Constant*> Elts;
1478 for (unsigned i = 0; i != VWidth; ++i)
1479 if (!DemandedElts[i]) { // If not demanded, set to undef.
1480 Elts.push_back(Undef);
1482 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1483 Elts.push_back(Undef);
1485 } else { // Otherwise, defined.
1486 Elts.push_back(CP->getOperand(i));
1489 // If we changed the constant, return it.
1490 Constant *NewCP = Context->getConstantVector(Elts);
1491 return NewCP != CP ? NewCP : 0;
1492 } else if (isa<ConstantAggregateZero>(V)) {
1493 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1496 // Check if this is identity. If so, return 0 since we are not simplifying
1498 if (DemandedElts == ((1ULL << VWidth) -1))
1501 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1502 Constant *Zero = Context->getNullValue(EltTy);
1503 Constant *Undef = Context->getUndef(EltTy);
1504 std::vector<Constant*> Elts;
1505 for (unsigned i = 0; i != VWidth; ++i) {
1506 Constant *Elt = DemandedElts[i] ? Zero : Undef;
1507 Elts.push_back(Elt);
1509 UndefElts = DemandedElts ^ EltMask;
1510 return Context->getConstantVector(Elts);
1513 // Limit search depth.
1517 // If multiple users are using the root value, procede with
1518 // simplification conservatively assuming that all elements
1520 if (!V->hasOneUse()) {
1521 // Quit if we find multiple users of a non-root value though.
1522 // They'll be handled when it's their turn to be visited by
1523 // the main instcombine process.
1525 // TODO: Just compute the UndefElts information recursively.
1528 // Conservatively assume that all elements are needed.
1529 DemandedElts = EltMask;
1532 Instruction *I = dyn_cast<Instruction>(V);
1533 if (!I) return 0; // Only analyze instructions.
1535 bool MadeChange = false;
1536 APInt UndefElts2(VWidth, 0);
1538 switch (I->getOpcode()) {
1541 case Instruction::InsertElement: {
1542 // If this is a variable index, we don't know which element it overwrites.
1543 // demand exactly the same input as we produce.
1544 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1546 // Note that we can't propagate undef elt info, because we don't know
1547 // which elt is getting updated.
1548 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1549 UndefElts2, Depth+1);
1550 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1554 // If this is inserting an element that isn't demanded, remove this
1556 unsigned IdxNo = Idx->getZExtValue();
1557 if (IdxNo >= VWidth || !DemandedElts[IdxNo])
1558 return AddSoonDeadInstToWorklist(*I, 0);
1560 // Otherwise, the element inserted overwrites whatever was there, so the
1561 // input demanded set is simpler than the output set.
1562 APInt DemandedElts2 = DemandedElts;
1563 DemandedElts2.clear(IdxNo);
1564 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts2,
1565 UndefElts, Depth+1);
1566 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1568 // The inserted element is defined.
1569 UndefElts.clear(IdxNo);
1572 case Instruction::ShuffleVector: {
1573 ShuffleVectorInst *Shuffle = cast<ShuffleVectorInst>(I);
1574 uint64_t LHSVWidth =
1575 cast<VectorType>(Shuffle->getOperand(0)->getType())->getNumElements();
1576 APInt LeftDemanded(LHSVWidth, 0), RightDemanded(LHSVWidth, 0);
1577 for (unsigned i = 0; i < VWidth; i++) {
1578 if (DemandedElts[i]) {
1579 unsigned MaskVal = Shuffle->getMaskValue(i);
1580 if (MaskVal != -1u) {
1581 assert(MaskVal < LHSVWidth * 2 &&
1582 "shufflevector mask index out of range!");
1583 if (MaskVal < LHSVWidth)
1584 LeftDemanded.set(MaskVal);
1586 RightDemanded.set(MaskVal - LHSVWidth);
1591 APInt UndefElts4(LHSVWidth, 0);
1592 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), LeftDemanded,
1593 UndefElts4, Depth+1);
1594 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1596 APInt UndefElts3(LHSVWidth, 0);
1597 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), RightDemanded,
1598 UndefElts3, Depth+1);
1599 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1601 bool NewUndefElts = false;
1602 for (unsigned i = 0; i < VWidth; i++) {
1603 unsigned MaskVal = Shuffle->getMaskValue(i);
1604 if (MaskVal == -1u) {
1606 } else if (MaskVal < LHSVWidth) {
1607 if (UndefElts4[MaskVal]) {
1608 NewUndefElts = true;
1612 if (UndefElts3[MaskVal - LHSVWidth]) {
1613 NewUndefElts = true;
1620 // Add additional discovered undefs.
1621 std::vector<Constant*> Elts;
1622 for (unsigned i = 0; i < VWidth; ++i) {
1624 Elts.push_back(Context->getUndef(Type::Int32Ty));
1626 Elts.push_back(Context->getConstantInt(Type::Int32Ty,
1627 Shuffle->getMaskValue(i)));
1629 I->setOperand(2, Context->getConstantVector(Elts));
1634 case Instruction::BitCast: {
1635 // Vector->vector casts only.
1636 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1638 unsigned InVWidth = VTy->getNumElements();
1639 APInt InputDemandedElts(InVWidth, 0);
1642 if (VWidth == InVWidth) {
1643 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1644 // elements as are demanded of us.
1646 InputDemandedElts = DemandedElts;
1647 } else if (VWidth > InVWidth) {
1651 // If there are more elements in the result than there are in the source,
1652 // then an input element is live if any of the corresponding output
1653 // elements are live.
1654 Ratio = VWidth/InVWidth;
1655 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1656 if (DemandedElts[OutIdx])
1657 InputDemandedElts.set(OutIdx/Ratio);
1663 // If there are more elements in the source than there are in the result,
1664 // then an input element is live if the corresponding output element is
1666 Ratio = InVWidth/VWidth;
1667 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1668 if (DemandedElts[InIdx/Ratio])
1669 InputDemandedElts.set(InIdx);
1672 // div/rem demand all inputs, because they don't want divide by zero.
1673 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1674 UndefElts2, Depth+1);
1676 I->setOperand(0, TmpV);
1680 UndefElts = UndefElts2;
1681 if (VWidth > InVWidth) {
1682 LLVM_UNREACHABLE("Unimp");
1683 // If there are more elements in the result than there are in the source,
1684 // then an output element is undef if the corresponding input element is
1686 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1687 if (UndefElts2[OutIdx/Ratio])
1688 UndefElts.set(OutIdx);
1689 } else if (VWidth < InVWidth) {
1690 LLVM_UNREACHABLE("Unimp");
1691 // If there are more elements in the source than there are in the result,
1692 // then a result element is undef if all of the corresponding input
1693 // elements are undef.
1694 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1695 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1696 if (!UndefElts2[InIdx]) // Not undef?
1697 UndefElts.clear(InIdx/Ratio); // Clear undef bit.
1701 case Instruction::And:
1702 case Instruction::Or:
1703 case Instruction::Xor:
1704 case Instruction::Add:
1705 case Instruction::Sub:
1706 case Instruction::Mul:
1707 // div/rem demand all inputs, because they don't want divide by zero.
1708 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1709 UndefElts, Depth+1);
1710 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1711 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1712 UndefElts2, Depth+1);
1713 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1715 // Output elements are undefined if both are undefined. Consider things
1716 // like undef&0. The result is known zero, not undef.
1717 UndefElts &= UndefElts2;
1720 case Instruction::Call: {
1721 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1723 switch (II->getIntrinsicID()) {
1726 // Binary vector operations that work column-wise. A dest element is a
1727 // function of the corresponding input elements from the two inputs.
1728 case Intrinsic::x86_sse_sub_ss:
1729 case Intrinsic::x86_sse_mul_ss:
1730 case Intrinsic::x86_sse_min_ss:
1731 case Intrinsic::x86_sse_max_ss:
1732 case Intrinsic::x86_sse2_sub_sd:
1733 case Intrinsic::x86_sse2_mul_sd:
1734 case Intrinsic::x86_sse2_min_sd:
1735 case Intrinsic::x86_sse2_max_sd:
1736 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1737 UndefElts, Depth+1);
1738 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1739 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1740 UndefElts2, Depth+1);
1741 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1743 // If only the low elt is demanded and this is a scalarizable intrinsic,
1744 // scalarize it now.
1745 if (DemandedElts == 1) {
1746 switch (II->getIntrinsicID()) {
1748 case Intrinsic::x86_sse_sub_ss:
1749 case Intrinsic::x86_sse_mul_ss:
1750 case Intrinsic::x86_sse2_sub_sd:
1751 case Intrinsic::x86_sse2_mul_sd:
1752 // TODO: Lower MIN/MAX/ABS/etc
1753 Value *LHS = II->getOperand(1);
1754 Value *RHS = II->getOperand(2);
1755 // Extract the element as scalars.
1756 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1757 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1759 switch (II->getIntrinsicID()) {
1760 default: LLVM_UNREACHABLE("Case stmts out of sync!");
1761 case Intrinsic::x86_sse_sub_ss:
1762 case Intrinsic::x86_sse2_sub_sd:
1763 TmpV = InsertNewInstBefore(BinaryOperator::CreateFSub(LHS, RHS,
1764 II->getName()), *II);
1766 case Intrinsic::x86_sse_mul_ss:
1767 case Intrinsic::x86_sse2_mul_sd:
1768 TmpV = InsertNewInstBefore(BinaryOperator::CreateFMul(LHS, RHS,
1769 II->getName()), *II);
1774 InsertElementInst::Create(
1775 Context->getUndef(II->getType()), TmpV, 0U, II->getName());
1776 InsertNewInstBefore(New, *II);
1777 AddSoonDeadInstToWorklist(*II, 0);
1782 // Output elements are undefined if both are undefined. Consider things
1783 // like undef&0. The result is known zero, not undef.
1784 UndefElts &= UndefElts2;
1790 return MadeChange ? I : 0;
1794 /// AssociativeOpt - Perform an optimization on an associative operator. This
1795 /// function is designed to check a chain of associative operators for a
1796 /// potential to apply a certain optimization. Since the optimization may be
1797 /// applicable if the expression was reassociated, this checks the chain, then
1798 /// reassociates the expression as necessary to expose the optimization
1799 /// opportunity. This makes use of a special Functor, which must define
1800 /// 'shouldApply' and 'apply' methods.
1802 template<typename Functor>
1803 static Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F,
1804 LLVMContext *Context) {
1805 unsigned Opcode = Root.getOpcode();
1806 Value *LHS = Root.getOperand(0);
1808 // Quick check, see if the immediate LHS matches...
1809 if (F.shouldApply(LHS))
1810 return F.apply(Root);
1812 // Otherwise, if the LHS is not of the same opcode as the root, return.
1813 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1814 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1815 // Should we apply this transform to the RHS?
1816 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1818 // If not to the RHS, check to see if we should apply to the LHS...
1819 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1820 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1824 // If the functor wants to apply the optimization to the RHS of LHSI,
1825 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1827 // Now all of the instructions are in the current basic block, go ahead
1828 // and perform the reassociation.
1829 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1831 // First move the selected RHS to the LHS of the root...
1832 Root.setOperand(0, LHSI->getOperand(1));
1834 // Make what used to be the LHS of the root be the user of the root...
1835 Value *ExtraOperand = TmpLHSI->getOperand(1);
1836 if (&Root == TmpLHSI) {
1837 Root.replaceAllUsesWith(Context->getNullValue(TmpLHSI->getType()));
1840 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1841 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1842 BasicBlock::iterator ARI = &Root; ++ARI;
1843 TmpLHSI->moveBefore(ARI); // Move TmpLHSI to after Root
1846 // Now propagate the ExtraOperand down the chain of instructions until we
1848 while (TmpLHSI != LHSI) {
1849 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1850 // Move the instruction to immediately before the chain we are
1851 // constructing to avoid breaking dominance properties.
1852 NextLHSI->moveBefore(ARI);
1855 Value *NextOp = NextLHSI->getOperand(1);
1856 NextLHSI->setOperand(1, ExtraOperand);
1858 ExtraOperand = NextOp;
1861 // Now that the instructions are reassociated, have the functor perform
1862 // the transformation...
1863 return F.apply(Root);
1866 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1873 // AddRHS - Implements: X + X --> X << 1
1876 LLVMContext *Context;
1877 AddRHS(Value *rhs, LLVMContext *C) : RHS(rhs), Context(C) {}
1878 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1879 Instruction *apply(BinaryOperator &Add) const {
1880 return BinaryOperator::CreateShl(Add.getOperand(0),
1881 Context->getConstantInt(Add.getType(), 1));
1885 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1887 struct AddMaskingAnd {
1889 LLVMContext *Context;
1890 AddMaskingAnd(Constant *c, LLVMContext *C) : C2(c), Context(C) {}
1891 bool shouldApply(Value *LHS) const {
1893 return match(LHS, m_And(m_Value(), m_ConstantInt(C1)), *Context) &&
1894 Context->getConstantExprAnd(C1, C2)->isNullValue();
1896 Instruction *apply(BinaryOperator &Add) const {
1897 return BinaryOperator::CreateOr(Add.getOperand(0), Add.getOperand(1));
1903 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1905 LLVMContext *Context = IC->getContext();
1907 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1908 return IC->InsertCastBefore(CI->getOpcode(), SO, I.getType(), I);
1911 // Figure out if the constant is the left or the right argument.
1912 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1913 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1915 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1917 return Context->getConstantExpr(I.getOpcode(), SOC, ConstOperand);
1918 return Context->getConstantExpr(I.getOpcode(), ConstOperand, SOC);
1921 Value *Op0 = SO, *Op1 = ConstOperand;
1923 std::swap(Op0, Op1);
1925 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1926 New = BinaryOperator::Create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1927 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1928 New = CmpInst::Create(*Context, CI->getOpcode(), CI->getPredicate(),
1929 Op0, Op1, SO->getName()+".cmp");
1931 LLVM_UNREACHABLE("Unknown binary instruction type!");
1933 return IC->InsertNewInstBefore(New, I);
1936 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1937 // constant as the other operand, try to fold the binary operator into the
1938 // select arguments. This also works for Cast instructions, which obviously do
1939 // not have a second operand.
1940 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1942 // Don't modify shared select instructions
1943 if (!SI->hasOneUse()) return 0;
1944 Value *TV = SI->getOperand(1);
1945 Value *FV = SI->getOperand(2);
1947 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1948 // Bool selects with constant operands can be folded to logical ops.
1949 if (SI->getType() == Type::Int1Ty) return 0;
1951 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1952 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1954 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
1961 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1962 /// node as operand #0, see if we can fold the instruction into the PHI (which
1963 /// is only possible if all operands to the PHI are constants).
1964 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1965 PHINode *PN = cast<PHINode>(I.getOperand(0));
1966 unsigned NumPHIValues = PN->getNumIncomingValues();
1967 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1969 // Check to see if all of the operands of the PHI are constants. If there is
1970 // one non-constant value, remember the BB it is. If there is more than one
1971 // or if *it* is a PHI, bail out.
1972 BasicBlock *NonConstBB = 0;
1973 for (unsigned i = 0; i != NumPHIValues; ++i)
1974 if (!isa<Constant>(PN->getIncomingValue(i))) {
1975 if (NonConstBB) return 0; // More than one non-const value.
1976 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1977 NonConstBB = PN->getIncomingBlock(i);
1979 // If the incoming non-constant value is in I's block, we have an infinite
1981 if (NonConstBB == I.getParent())
1985 // If there is exactly one non-constant value, we can insert a copy of the
1986 // operation in that block. However, if this is a critical edge, we would be
1987 // inserting the computation one some other paths (e.g. inside a loop). Only
1988 // do this if the pred block is unconditionally branching into the phi block.
1990 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1991 if (!BI || !BI->isUnconditional()) return 0;
1994 // Okay, we can do the transformation: create the new PHI node.
1995 PHINode *NewPN = PHINode::Create(I.getType(), "");
1996 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1997 InsertNewInstBefore(NewPN, *PN);
1998 NewPN->takeName(PN);
2000 // Next, add all of the operands to the PHI.
2001 if (I.getNumOperands() == 2) {
2002 Constant *C = cast<Constant>(I.getOperand(1));
2003 for (unsigned i = 0; i != NumPHIValues; ++i) {
2005 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2006 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2007 InV = Context->getConstantExprCompare(CI->getPredicate(), InC, C);
2009 InV = Context->getConstantExpr(I.getOpcode(), InC, C);
2011 assert(PN->getIncomingBlock(i) == NonConstBB);
2012 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2013 InV = BinaryOperator::Create(BO->getOpcode(),
2014 PN->getIncomingValue(i), C, "phitmp",
2015 NonConstBB->getTerminator());
2016 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2017 InV = CmpInst::Create(*Context, CI->getOpcode(),
2019 PN->getIncomingValue(i), C, "phitmp",
2020 NonConstBB->getTerminator());
2022 LLVM_UNREACHABLE("Unknown binop!");
2024 AddToWorkList(cast<Instruction>(InV));
2026 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2029 CastInst *CI = cast<CastInst>(&I);
2030 const Type *RetTy = CI->getType();
2031 for (unsigned i = 0; i != NumPHIValues; ++i) {
2033 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2034 InV = Context->getConstantExprCast(CI->getOpcode(), InC, RetTy);
2036 assert(PN->getIncomingBlock(i) == NonConstBB);
2037 InV = CastInst::Create(CI->getOpcode(), PN->getIncomingValue(i),
2038 I.getType(), "phitmp",
2039 NonConstBB->getTerminator());
2040 AddToWorkList(cast<Instruction>(InV));
2042 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2045 return ReplaceInstUsesWith(I, NewPN);
2049 /// WillNotOverflowSignedAdd - Return true if we can prove that:
2050 /// (sext (add LHS, RHS)) === (add (sext LHS), (sext RHS))
2051 /// This basically requires proving that the add in the original type would not
2052 /// overflow to change the sign bit or have a carry out.
2053 bool InstCombiner::WillNotOverflowSignedAdd(Value *LHS, Value *RHS) {
2054 // There are different heuristics we can use for this. Here are some simple
2057 // Add has the property that adding any two 2's complement numbers can only
2058 // have one carry bit which can change a sign. As such, if LHS and RHS each
2059 // have at least two sign bits, we know that the addition of the two values will
2060 // sign extend fine.
2061 if (ComputeNumSignBits(LHS) > 1 && ComputeNumSignBits(RHS) > 1)
2065 // If one of the operands only has one non-zero bit, and if the other operand
2066 // has a known-zero bit in a more significant place than it (not including the
2067 // sign bit) the ripple may go up to and fill the zero, but won't change the
2068 // sign. For example, (X & ~4) + 1.
2076 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2077 bool Changed = SimplifyCommutative(I);
2078 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2080 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2081 // X + undef -> undef
2082 if (isa<UndefValue>(RHS))
2083 return ReplaceInstUsesWith(I, RHS);
2086 if (RHSC->isNullValue())
2087 return ReplaceInstUsesWith(I, LHS);
2089 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2090 // X + (signbit) --> X ^ signbit
2091 const APInt& Val = CI->getValue();
2092 uint32_t BitWidth = Val.getBitWidth();
2093 if (Val == APInt::getSignBit(BitWidth))
2094 return BinaryOperator::CreateXor(LHS, RHS);
2096 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2097 // (X & 254)+1 -> (X&254)|1
2098 if (SimplifyDemandedInstructionBits(I))
2101 // zext(i1) - 1 -> select i1, 0, -1
2102 if (ZExtInst *ZI = dyn_cast<ZExtInst>(LHS))
2103 if (CI->isAllOnesValue() &&
2104 ZI->getOperand(0)->getType() == Type::Int1Ty)
2105 return SelectInst::Create(ZI->getOperand(0),
2106 Context->getNullValue(I.getType()),
2107 Context->getConstantIntAllOnesValue(I.getType()));
2110 if (isa<PHINode>(LHS))
2111 if (Instruction *NV = FoldOpIntoPhi(I))
2114 ConstantInt *XorRHS = 0;
2116 if (isa<ConstantInt>(RHSC) &&
2117 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)), *Context)) {
2118 uint32_t TySizeBits = I.getType()->getScalarSizeInBits();
2119 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2121 uint32_t Size = TySizeBits / 2;
2122 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2123 APInt CFF80Val(-C0080Val);
2125 if (TySizeBits > Size) {
2126 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2127 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2128 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2129 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2130 // This is a sign extend if the top bits are known zero.
2131 if (!MaskedValueIsZero(XorLHS,
2132 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2133 Size = 0; // Not a sign ext, but can't be any others either.
2138 C0080Val = APIntOps::lshr(C0080Val, Size);
2139 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2140 } while (Size >= 1);
2142 // FIXME: This shouldn't be necessary. When the backends can handle types
2143 // with funny bit widths then this switch statement should be removed. It
2144 // is just here to get the size of the "middle" type back up to something
2145 // that the back ends can handle.
2146 const Type *MiddleType = 0;
2149 case 32: MiddleType = Type::Int32Ty; break;
2150 case 16: MiddleType = Type::Int16Ty; break;
2151 case 8: MiddleType = Type::Int8Ty; break;
2154 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2155 InsertNewInstBefore(NewTrunc, I);
2156 return new SExtInst(NewTrunc, I.getType(), I.getName());
2161 if (I.getType() == Type::Int1Ty)
2162 return BinaryOperator::CreateXor(LHS, RHS);
2165 if (I.getType()->isInteger()) {
2166 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS, Context), Context))
2169 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2170 if (RHSI->getOpcode() == Instruction::Sub)
2171 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2172 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2174 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2175 if (LHSI->getOpcode() == Instruction::Sub)
2176 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2177 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2182 // -A + -B --> -(A + B)
2183 if (Value *LHSV = dyn_castNegVal(LHS, Context)) {
2184 if (LHS->getType()->isIntOrIntVector()) {
2185 if (Value *RHSV = dyn_castNegVal(RHS, Context)) {
2186 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSV, RHSV, "sum");
2187 InsertNewInstBefore(NewAdd, I);
2188 return BinaryOperator::CreateNeg(NewAdd);
2192 return BinaryOperator::CreateSub(RHS, LHSV);
2196 if (!isa<Constant>(RHS))
2197 if (Value *V = dyn_castNegVal(RHS, Context))
2198 return BinaryOperator::CreateSub(LHS, V);
2202 if (Value *X = dyn_castFoldableMul(LHS, C2, Context)) {
2203 if (X == RHS) // X*C + X --> X * (C+1)
2204 return BinaryOperator::CreateMul(RHS, AddOne(C2, Context));
2206 // X*C1 + X*C2 --> X * (C1+C2)
2208 if (X == dyn_castFoldableMul(RHS, C1, Context))
2209 return BinaryOperator::CreateMul(X, Context->getConstantExprAdd(C1, C2));
2212 // X + X*C --> X * (C+1)
2213 if (dyn_castFoldableMul(RHS, C2, Context) == LHS)
2214 return BinaryOperator::CreateMul(LHS, AddOne(C2, Context));
2216 // X + ~X --> -1 since ~X = -X-1
2217 if (dyn_castNotVal(LHS, Context) == RHS ||
2218 dyn_castNotVal(RHS, Context) == LHS)
2219 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
2222 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2223 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2)), *Context))
2224 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2, Context), Context))
2227 // A+B --> A|B iff A and B have no bits set in common.
2228 if (const IntegerType *IT = dyn_cast<IntegerType>(I.getType())) {
2229 APInt Mask = APInt::getAllOnesValue(IT->getBitWidth());
2230 APInt LHSKnownOne(IT->getBitWidth(), 0);
2231 APInt LHSKnownZero(IT->getBitWidth(), 0);
2232 ComputeMaskedBits(LHS, Mask, LHSKnownZero, LHSKnownOne);
2233 if (LHSKnownZero != 0) {
2234 APInt RHSKnownOne(IT->getBitWidth(), 0);
2235 APInt RHSKnownZero(IT->getBitWidth(), 0);
2236 ComputeMaskedBits(RHS, Mask, RHSKnownZero, RHSKnownOne);
2238 // No bits in common -> bitwise or.
2239 if ((LHSKnownZero|RHSKnownZero).isAllOnesValue())
2240 return BinaryOperator::CreateOr(LHS, RHS);
2244 // W*X + Y*Z --> W * (X+Z) iff W == Y
2245 if (I.getType()->isIntOrIntVector()) {
2246 Value *W, *X, *Y, *Z;
2247 if (match(LHS, m_Mul(m_Value(W), m_Value(X)), *Context) &&
2248 match(RHS, m_Mul(m_Value(Y), m_Value(Z)), *Context)) {
2252 } else if (Y == X) {
2254 } else if (X == Z) {
2261 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, Z,
2262 LHS->getName()), I);
2263 return BinaryOperator::CreateMul(W, NewAdd);
2268 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2270 if (match(LHS, m_Not(m_Value(X)), *Context)) // ~X + C --> (C-1) - X
2271 return BinaryOperator::CreateSub(SubOne(CRHS, Context), X);
2273 // (X & FF00) + xx00 -> (X+xx00) & FF00
2274 if (LHS->hasOneUse() &&
2275 match(LHS, m_And(m_Value(X), m_ConstantInt(C2)), *Context)) {
2276 Constant *Anded = Context->getConstantExprAnd(CRHS, C2);
2277 if (Anded == CRHS) {
2278 // See if all bits from the first bit set in the Add RHS up are included
2279 // in the mask. First, get the rightmost bit.
2280 const APInt& AddRHSV = CRHS->getValue();
2282 // Form a mask of all bits from the lowest bit added through the top.
2283 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2285 // See if the and mask includes all of these bits.
2286 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2288 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2289 // Okay, the xform is safe. Insert the new add pronto.
2290 Value *NewAdd = InsertNewInstBefore(BinaryOperator::CreateAdd(X, CRHS,
2291 LHS->getName()), I);
2292 return BinaryOperator::CreateAnd(NewAdd, C2);
2297 // Try to fold constant add into select arguments.
2298 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2299 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2303 // add (cast *A to intptrtype) B ->
2304 // cast (GEP (cast *A to i8*) B) --> intptrtype
2306 CastInst *CI = dyn_cast<CastInst>(LHS);
2309 CI = dyn_cast<CastInst>(RHS);
2312 if (CI && CI->getType()->isSized() &&
2313 (CI->getType()->getScalarSizeInBits() ==
2314 TD->getIntPtrType()->getPrimitiveSizeInBits())
2315 && isa<PointerType>(CI->getOperand(0)->getType())) {
2317 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2318 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2319 Context->getPointerType(Type::Int8Ty, AS), I);
2320 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2321 return new PtrToIntInst(I2, CI->getType());
2325 // add (select X 0 (sub n A)) A --> select X A n
2327 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2330 SI = dyn_cast<SelectInst>(RHS);
2333 if (SI && SI->hasOneUse()) {
2334 Value *TV = SI->getTrueValue();
2335 Value *FV = SI->getFalseValue();
2338 // Can we fold the add into the argument of the select?
2339 // We check both true and false select arguments for a matching subtract.
2340 if (match(FV, m_Zero(), *Context) &&
2341 match(TV, m_Sub(m_Value(N), m_Specific(A)), *Context))
2342 // Fold the add into the true select value.
2343 return SelectInst::Create(SI->getCondition(), N, A);
2344 if (match(TV, m_Zero(), *Context) &&
2345 match(FV, m_Sub(m_Value(N), m_Specific(A)), *Context))
2346 // Fold the add into the false select value.
2347 return SelectInst::Create(SI->getCondition(), A, N);
2351 // Check for (add (sext x), y), see if we can merge this into an
2352 // integer add followed by a sext.
2353 if (SExtInst *LHSConv = dyn_cast<SExtInst>(LHS)) {
2354 // (add (sext x), cst) --> (sext (add x, cst'))
2355 if (ConstantInt *RHSC = dyn_cast<ConstantInt>(RHS)) {
2357 Context->getConstantExprTrunc(RHSC, LHSConv->getOperand(0)->getType());
2358 if (LHSConv->hasOneUse() &&
2359 Context->getConstantExprSExt(CI, I.getType()) == RHSC &&
2360 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2361 // Insert the new, smaller add.
2362 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2364 InsertNewInstBefore(NewAdd, I);
2365 return new SExtInst(NewAdd, I.getType());
2369 // (add (sext x), (sext y)) --> (sext (add int x, y))
2370 if (SExtInst *RHSConv = dyn_cast<SExtInst>(RHS)) {
2371 // Only do this if x/y have the same type, if at last one of them has a
2372 // single use (so we don't increase the number of sexts), and if the
2373 // integer add will not overflow.
2374 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2375 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2376 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2377 RHSConv->getOperand(0))) {
2378 // Insert the new integer add.
2379 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2380 RHSConv->getOperand(0),
2382 InsertNewInstBefore(NewAdd, I);
2383 return new SExtInst(NewAdd, I.getType());
2388 return Changed ? &I : 0;
2391 Instruction *InstCombiner::visitFAdd(BinaryOperator &I) {
2392 bool Changed = SimplifyCommutative(I);
2393 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2395 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2397 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2398 if (CFP->isExactlyValue(Context->getConstantFPNegativeZero
2399 (I.getType())->getValueAPF()))
2400 return ReplaceInstUsesWith(I, LHS);
2403 if (isa<PHINode>(LHS))
2404 if (Instruction *NV = FoldOpIntoPhi(I))
2409 // -A + -B --> -(A + B)
2410 if (Value *LHSV = dyn_castFNegVal(LHS, Context))
2411 return BinaryOperator::CreateFSub(RHS, LHSV);
2414 if (!isa<Constant>(RHS))
2415 if (Value *V = dyn_castFNegVal(RHS, Context))
2416 return BinaryOperator::CreateFSub(LHS, V);
2418 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2419 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2420 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2421 return ReplaceInstUsesWith(I, LHS);
2423 // Check for (add double (sitofp x), y), see if we can merge this into an
2424 // integer add followed by a promotion.
2425 if (SIToFPInst *LHSConv = dyn_cast<SIToFPInst>(LHS)) {
2426 // (add double (sitofp x), fpcst) --> (sitofp (add int x, intcst))
2427 // ... if the constant fits in the integer value. This is useful for things
2428 // like (double)(x & 1234) + 4.0 -> (double)((X & 1234)+4) which no longer
2429 // requires a constant pool load, and generally allows the add to be better
2431 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS)) {
2433 Context->getConstantExprFPToSI(CFP, LHSConv->getOperand(0)->getType());
2434 if (LHSConv->hasOneUse() &&
2435 Context->getConstantExprSIToFP(CI, I.getType()) == CFP &&
2436 WillNotOverflowSignedAdd(LHSConv->getOperand(0), CI)) {
2437 // Insert the new integer add.
2438 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2440 InsertNewInstBefore(NewAdd, I);
2441 return new SIToFPInst(NewAdd, I.getType());
2445 // (add double (sitofp x), (sitofp y)) --> (sitofp (add int x, y))
2446 if (SIToFPInst *RHSConv = dyn_cast<SIToFPInst>(RHS)) {
2447 // Only do this if x/y have the same type, if at last one of them has a
2448 // single use (so we don't increase the number of int->fp conversions),
2449 // and if the integer add will not overflow.
2450 if (LHSConv->getOperand(0)->getType()==RHSConv->getOperand(0)->getType()&&
2451 (LHSConv->hasOneUse() || RHSConv->hasOneUse()) &&
2452 WillNotOverflowSignedAdd(LHSConv->getOperand(0),
2453 RHSConv->getOperand(0))) {
2454 // Insert the new integer add.
2455 Instruction *NewAdd = BinaryOperator::CreateAdd(LHSConv->getOperand(0),
2456 RHSConv->getOperand(0),
2458 InsertNewInstBefore(NewAdd, I);
2459 return new SIToFPInst(NewAdd, I.getType());
2464 return Changed ? &I : 0;
2467 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2468 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2470 if (Op0 == Op1) // sub X, X -> 0
2471 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2473 // If this is a 'B = x-(-A)', change to B = x+A...
2474 if (Value *V = dyn_castNegVal(Op1, Context))
2475 return BinaryOperator::CreateAdd(Op0, V);
2477 if (isa<UndefValue>(Op0))
2478 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2479 if (isa<UndefValue>(Op1))
2480 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2482 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2483 // Replace (-1 - A) with (~A)...
2484 if (C->isAllOnesValue())
2485 return BinaryOperator::CreateNot(Op1);
2487 // C - ~X == X + (1+C)
2489 if (match(Op1, m_Not(m_Value(X)), *Context))
2490 return BinaryOperator::CreateAdd(X, AddOne(C, Context));
2492 // -(X >>u 31) -> (X >>s 31)
2493 // -(X >>s 31) -> (X >>u 31)
2495 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2496 if (SI->getOpcode() == Instruction::LShr) {
2497 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2498 // Check to see if we are shifting out everything but the sign bit.
2499 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2500 SI->getType()->getPrimitiveSizeInBits()-1) {
2501 // Ok, the transformation is safe. Insert AShr.
2502 return BinaryOperator::Create(Instruction::AShr,
2503 SI->getOperand(0), CU, SI->getName());
2507 else if (SI->getOpcode() == Instruction::AShr) {
2508 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2509 // Check to see if we are shifting out everything but the sign bit.
2510 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2511 SI->getType()->getPrimitiveSizeInBits()-1) {
2512 // Ok, the transformation is safe. Insert LShr.
2513 return BinaryOperator::CreateLShr(
2514 SI->getOperand(0), CU, SI->getName());
2521 // Try to fold constant sub into select arguments.
2522 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2523 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2527 if (I.getType() == Type::Int1Ty)
2528 return BinaryOperator::CreateXor(Op0, Op1);
2530 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2531 if (Op1I->getOpcode() == Instruction::Add) {
2532 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2533 return BinaryOperator::CreateNeg(Op1I->getOperand(1), I.getName());
2534 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2535 return BinaryOperator::CreateNeg(Op1I->getOperand(0), I.getName());
2536 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2537 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2538 // C1-(X+C2) --> (C1-C2)-X
2539 return BinaryOperator::CreateSub(
2540 Context->getConstantExprSub(CI1, CI2), Op1I->getOperand(0));
2544 if (Op1I->hasOneUse()) {
2545 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2546 // is not used by anyone else...
2548 if (Op1I->getOpcode() == Instruction::Sub) {
2549 // Swap the two operands of the subexpr...
2550 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2551 Op1I->setOperand(0, IIOp1);
2552 Op1I->setOperand(1, IIOp0);
2554 // Create the new top level add instruction...
2555 return BinaryOperator::CreateAdd(Op0, Op1);
2558 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2560 if (Op1I->getOpcode() == Instruction::And &&
2561 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2562 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2565 InsertNewInstBefore(BinaryOperator::CreateNot(OtherOp, "B.not"), I);
2566 return BinaryOperator::CreateAnd(Op0, NewNot);
2569 // 0 - (X sdiv C) -> (X sdiv -C)
2570 if (Op1I->getOpcode() == Instruction::SDiv)
2571 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2573 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2574 return BinaryOperator::CreateSDiv(Op1I->getOperand(0),
2575 Context->getConstantExprNeg(DivRHS));
2577 // X - X*C --> X * (1-C)
2578 ConstantInt *C2 = 0;
2579 if (dyn_castFoldableMul(Op1I, C2, Context) == Op0) {
2581 Context->getConstantExprSub(Context->getConstantInt(I.getType(), 1),
2583 return BinaryOperator::CreateMul(Op0, CP1);
2588 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2589 if (Op0I->getOpcode() == Instruction::Add) {
2590 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2591 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2592 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2593 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2594 } else if (Op0I->getOpcode() == Instruction::Sub) {
2595 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2596 return BinaryOperator::CreateNeg(Op0I->getOperand(1), I.getName());
2601 if (Value *X = dyn_castFoldableMul(Op0, C1, Context)) {
2602 if (X == Op1) // X*C - X --> X * (C-1)
2603 return BinaryOperator::CreateMul(Op1, SubOne(C1, Context));
2605 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2606 if (X == dyn_castFoldableMul(Op1, C2, Context))
2607 return BinaryOperator::CreateMul(X, Context->getConstantExprSub(C1, C2));
2612 Instruction *InstCombiner::visitFSub(BinaryOperator &I) {
2613 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2615 // If this is a 'B = x-(-A)', change to B = x+A...
2616 if (Value *V = dyn_castFNegVal(Op1, Context))
2617 return BinaryOperator::CreateFAdd(Op0, V);
2619 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2620 if (Op1I->getOpcode() == Instruction::FAdd) {
2621 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2622 return BinaryOperator::CreateFNeg(Op1I->getOperand(1), I.getName());
2623 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2624 return BinaryOperator::CreateFNeg(Op1I->getOperand(0), I.getName());
2631 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2632 /// comparison only checks the sign bit. If it only checks the sign bit, set
2633 /// TrueIfSigned if the result of the comparison is true when the input value is
2635 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2636 bool &TrueIfSigned) {
2638 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2639 TrueIfSigned = true;
2640 return RHS->isZero();
2641 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2642 TrueIfSigned = true;
2643 return RHS->isAllOnesValue();
2644 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2645 TrueIfSigned = false;
2646 return RHS->isAllOnesValue();
2647 case ICmpInst::ICMP_UGT:
2648 // True if LHS u> RHS and RHS == high-bit-mask - 1
2649 TrueIfSigned = true;
2650 return RHS->getValue() ==
2651 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2652 case ICmpInst::ICMP_UGE:
2653 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2654 TrueIfSigned = true;
2655 return RHS->getValue().isSignBit();
2661 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2662 bool Changed = SimplifyCommutative(I);
2663 Value *Op0 = I.getOperand(0);
2665 // TODO: If Op1 is undef and Op0 is finite, return zero.
2666 if (!I.getType()->isFPOrFPVector() &&
2667 isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2668 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2670 // Simplify mul instructions with a constant RHS...
2671 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2672 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2674 // ((X << C1)*C2) == (X * (C2 << C1))
2675 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2676 if (SI->getOpcode() == Instruction::Shl)
2677 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2678 return BinaryOperator::CreateMul(SI->getOperand(0),
2679 Context->getConstantExprShl(CI, ShOp));
2682 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2683 if (CI->equalsInt(1)) // X * 1 == X
2684 return ReplaceInstUsesWith(I, Op0);
2685 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2686 return BinaryOperator::CreateNeg(Op0, I.getName());
2688 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2689 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2690 return BinaryOperator::CreateShl(Op0,
2691 Context->getConstantInt(Op0->getType(), Val.logBase2()));
2693 } else if (isa<VectorType>(Op1->getType())) {
2694 // TODO: If Op1 is all zeros and Op0 is all finite, return all zeros.
2696 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2697 if (Op1V->isAllOnesValue()) // X * -1 == 0 - X
2698 return BinaryOperator::CreateNeg(Op0, I.getName());
2700 // As above, vector X*splat(1.0) -> X in all defined cases.
2701 if (Constant *Splat = Op1V->getSplatValue()) {
2702 if (ConstantInt *CI = dyn_cast<ConstantInt>(Splat))
2703 if (CI->equalsInt(1))
2704 return ReplaceInstUsesWith(I, Op0);
2709 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2710 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2711 isa<ConstantInt>(Op0I->getOperand(1)) && isa<ConstantInt>(Op1)) {
2712 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2713 Instruction *Add = BinaryOperator::CreateMul(Op0I->getOperand(0),
2715 InsertNewInstBefore(Add, I);
2716 Value *C1C2 = Context->getConstantExprMul(Op1,
2717 cast<Constant>(Op0I->getOperand(1)));
2718 return BinaryOperator::CreateAdd(Add, C1C2);
2722 // Try to fold constant mul into select arguments.
2723 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2724 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2727 if (isa<PHINode>(Op0))
2728 if (Instruction *NV = FoldOpIntoPhi(I))
2732 if (Value *Op0v = dyn_castNegVal(Op0, Context)) // -X * -Y = X*Y
2733 if (Value *Op1v = dyn_castNegVal(I.getOperand(1), Context))
2734 return BinaryOperator::CreateMul(Op0v, Op1v);
2736 // (X / Y) * Y = X - (X % Y)
2737 // (X / Y) * -Y = (X % Y) - X
2739 Value *Op1 = I.getOperand(1);
2740 BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0);
2742 (BO->getOpcode() != Instruction::UDiv &&
2743 BO->getOpcode() != Instruction::SDiv)) {
2745 BO = dyn_cast<BinaryOperator>(I.getOperand(1));
2747 Value *Neg = dyn_castNegVal(Op1, Context);
2748 if (BO && BO->hasOneUse() &&
2749 (BO->getOperand(1) == Op1 || BO->getOperand(1) == Neg) &&
2750 (BO->getOpcode() == Instruction::UDiv ||
2751 BO->getOpcode() == Instruction::SDiv)) {
2752 Value *Op0BO = BO->getOperand(0), *Op1BO = BO->getOperand(1);
2755 if (BO->getOpcode() == Instruction::UDiv)
2756 Rem = BinaryOperator::CreateURem(Op0BO, Op1BO);
2758 Rem = BinaryOperator::CreateSRem(Op0BO, Op1BO);
2760 InsertNewInstBefore(Rem, I);
2764 return BinaryOperator::CreateSub(Op0BO, Rem);
2766 return BinaryOperator::CreateSub(Rem, Op0BO);
2770 if (I.getType() == Type::Int1Ty)
2771 return BinaryOperator::CreateAnd(Op0, I.getOperand(1));
2773 // If one of the operands of the multiply is a cast from a boolean value, then
2774 // we know the bool is either zero or one, so this is a 'masking' multiply.
2775 // See if we can simplify things based on how the boolean was originally
2777 CastInst *BoolCast = 0;
2778 if (ZExtInst *CI = dyn_cast<ZExtInst>(Op0))
2779 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2782 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2783 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2786 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2787 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2788 const Type *SCOpTy = SCIOp0->getType();
2791 // If the icmp is true iff the sign bit of X is set, then convert this
2792 // multiply into a shift/and combination.
2793 if (isa<ConstantInt>(SCIOp1) &&
2794 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2796 // Shift the X value right to turn it into "all signbits".
2797 Constant *Amt = Context->getConstantInt(SCIOp0->getType(),
2798 SCOpTy->getPrimitiveSizeInBits()-1);
2800 InsertNewInstBefore(
2801 BinaryOperator::Create(Instruction::AShr, SCIOp0, Amt,
2802 BoolCast->getOperand(0)->getName()+
2805 // If the multiply type is not the same as the source type, sign extend
2806 // or truncate to the multiply type.
2807 if (I.getType() != V->getType()) {
2808 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2809 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2810 Instruction::CastOps opcode =
2811 (SrcBits == DstBits ? Instruction::BitCast :
2812 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2813 V = InsertCastBefore(opcode, V, I.getType(), I);
2816 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2817 return BinaryOperator::CreateAnd(V, OtherOp);
2822 return Changed ? &I : 0;
2825 Instruction *InstCombiner::visitFMul(BinaryOperator &I) {
2826 bool Changed = SimplifyCommutative(I);
2827 Value *Op0 = I.getOperand(0);
2829 // Simplify mul instructions with a constant RHS...
2830 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2831 if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2832 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2833 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2834 if (Op1F->isExactlyValue(1.0))
2835 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2836 } else if (isa<VectorType>(Op1->getType())) {
2837 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
2838 // As above, vector X*splat(1.0) -> X in all defined cases.
2839 if (Constant *Splat = Op1V->getSplatValue()) {
2840 if (ConstantFP *F = dyn_cast<ConstantFP>(Splat))
2841 if (F->isExactlyValue(1.0))
2842 return ReplaceInstUsesWith(I, Op0);
2847 // Try to fold constant mul into select arguments.
2848 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2849 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2852 if (isa<PHINode>(Op0))
2853 if (Instruction *NV = FoldOpIntoPhi(I))
2857 if (Value *Op0v = dyn_castFNegVal(Op0, Context)) // -X * -Y = X*Y
2858 if (Value *Op1v = dyn_castFNegVal(I.getOperand(1), Context))
2859 return BinaryOperator::CreateFMul(Op0v, Op1v);
2861 return Changed ? &I : 0;
2864 /// SimplifyDivRemOfSelect - Try to fold a divide or remainder of a select
2866 bool InstCombiner::SimplifyDivRemOfSelect(BinaryOperator &I) {
2867 SelectInst *SI = cast<SelectInst>(I.getOperand(1));
2869 // div/rem X, (Cond ? 0 : Y) -> div/rem X, Y
2870 int NonNullOperand = -1;
2871 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2872 if (ST->isNullValue())
2874 // div/rem X, (Cond ? Y : 0) -> div/rem X, Y
2875 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2876 if (ST->isNullValue())
2879 if (NonNullOperand == -1)
2882 Value *SelectCond = SI->getOperand(0);
2884 // Change the div/rem to use 'Y' instead of the select.
2885 I.setOperand(1, SI->getOperand(NonNullOperand));
2887 // Okay, we know we replace the operand of the div/rem with 'Y' with no
2888 // problem. However, the select, or the condition of the select may have
2889 // multiple uses. Based on our knowledge that the operand must be non-zero,
2890 // propagate the known value for the select into other uses of it, and
2891 // propagate a known value of the condition into its other users.
2893 // If the select and condition only have a single use, don't bother with this,
2895 if (SI->use_empty() && SelectCond->hasOneUse())
2898 // Scan the current block backward, looking for other uses of SI.
2899 BasicBlock::iterator BBI = &I, BBFront = I.getParent()->begin();
2901 while (BBI != BBFront) {
2903 // If we found a call to a function, we can't assume it will return, so
2904 // information from below it cannot be propagated above it.
2905 if (isa<CallInst>(BBI) && !isa<IntrinsicInst>(BBI))
2908 // Replace uses of the select or its condition with the known values.
2909 for (Instruction::op_iterator I = BBI->op_begin(), E = BBI->op_end();
2912 *I = SI->getOperand(NonNullOperand);
2914 } else if (*I == SelectCond) {
2915 *I = NonNullOperand == 1 ? Context->getConstantIntTrue() :
2916 Context->getConstantIntFalse();
2921 // If we past the instruction, quit looking for it.
2924 if (&*BBI == SelectCond)
2927 // If we ran out of things to eliminate, break out of the loop.
2928 if (SelectCond == 0 && SI == 0)
2936 /// This function implements the transforms on div instructions that work
2937 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2938 /// used by the visitors to those instructions.
2939 /// @brief Transforms common to all three div instructions
2940 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2941 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2943 // undef / X -> 0 for integer.
2944 // undef / X -> undef for FP (the undef could be a snan).
2945 if (isa<UndefValue>(Op0)) {
2946 if (Op0->getType()->isFPOrFPVector())
2947 return ReplaceInstUsesWith(I, Op0);
2948 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2951 // X / undef -> undef
2952 if (isa<UndefValue>(Op1))
2953 return ReplaceInstUsesWith(I, Op1);
2958 /// This function implements the transforms common to both integer division
2959 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2960 /// division instructions.
2961 /// @brief Common integer divide transforms
2962 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2963 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2965 // (sdiv X, X) --> 1 (udiv X, X) --> 1
2967 if (const VectorType *Ty = dyn_cast<VectorType>(I.getType())) {
2968 Constant *CI = Context->getConstantInt(Ty->getElementType(), 1);
2969 std::vector<Constant*> Elts(Ty->getNumElements(), CI);
2970 return ReplaceInstUsesWith(I, Context->getConstantVector(Elts));
2973 Constant *CI = Context->getConstantInt(I.getType(), 1);
2974 return ReplaceInstUsesWith(I, CI);
2977 if (Instruction *Common = commonDivTransforms(I))
2980 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2981 // This does not apply for fdiv.
2982 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
2985 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2987 if (RHS->equalsInt(1))
2988 return ReplaceInstUsesWith(I, Op0);
2990 // (X / C1) / C2 -> X / (C1*C2)
2991 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2992 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2993 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2994 if (MultiplyOverflows(RHS, LHSRHS,
2995 I.getOpcode()==Instruction::SDiv, Context))
2996 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
2998 return BinaryOperator::Create(I.getOpcode(), LHS->getOperand(0),
2999 Context->getConstantExprMul(RHS, LHSRHS));
3002 if (!RHS->isZero()) { // avoid X udiv 0
3003 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3004 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3006 if (isa<PHINode>(Op0))
3007 if (Instruction *NV = FoldOpIntoPhi(I))
3012 // 0 / X == 0, we don't need to preserve faults!
3013 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3014 if (LHS->equalsInt(0))
3015 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3017 // It can't be division by zero, hence it must be division by one.
3018 if (I.getType() == Type::Int1Ty)
3019 return ReplaceInstUsesWith(I, Op0);
3021 if (ConstantVector *Op1V = dyn_cast<ConstantVector>(Op1)) {
3022 if (ConstantInt *X = cast_or_null<ConstantInt>(Op1V->getSplatValue()))
3025 return ReplaceInstUsesWith(I, Op0);
3031 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3032 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3034 // Handle the integer div common cases
3035 if (Instruction *Common = commonIDivTransforms(I))
3038 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3039 // X udiv C^2 -> X >> C
3040 // Check to see if this is an unsigned division with an exact power of 2,
3041 // if so, convert to a right shift.
3042 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3043 return BinaryOperator::CreateLShr(Op0,
3044 Context->getConstantInt(Op0->getType(), C->getValue().logBase2()));
3046 // X udiv C, where C >= signbit
3047 if (C->getValue().isNegative()) {
3048 Value *IC = InsertNewInstBefore(new ICmpInst(*Context,
3049 ICmpInst::ICMP_ULT, Op0, C),
3051 return SelectInst::Create(IC, Context->getNullValue(I.getType()),
3052 Context->getConstantInt(I.getType(), 1));
3056 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3057 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3058 if (RHSI->getOpcode() == Instruction::Shl &&
3059 isa<ConstantInt>(RHSI->getOperand(0))) {
3060 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3061 if (C1.isPowerOf2()) {
3062 Value *N = RHSI->getOperand(1);
3063 const Type *NTy = N->getType();
3064 if (uint32_t C2 = C1.logBase2()) {
3065 Constant *C2V = Context->getConstantInt(NTy, C2);
3066 N = InsertNewInstBefore(BinaryOperator::CreateAdd(N, C2V, "tmp"), I);
3068 return BinaryOperator::CreateLShr(Op0, N);
3073 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3074 // where C1&C2 are powers of two.
3075 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3076 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3077 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3078 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3079 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3080 // Compute the shift amounts
3081 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3082 // Construct the "on true" case of the select
3083 Constant *TC = Context->getConstantInt(Op0->getType(), TSA);
3084 Instruction *TSI = BinaryOperator::CreateLShr(
3085 Op0, TC, SI->getName()+".t");
3086 TSI = InsertNewInstBefore(TSI, I);
3088 // Construct the "on false" case of the select
3089 Constant *FC = Context->getConstantInt(Op0->getType(), FSA);
3090 Instruction *FSI = BinaryOperator::CreateLShr(
3091 Op0, FC, SI->getName()+".f");
3092 FSI = InsertNewInstBefore(FSI, I);
3094 // construct the select instruction and return it.
3095 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3101 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3102 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3104 // Handle the integer div common cases
3105 if (Instruction *Common = commonIDivTransforms(I))
3108 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3110 if (RHS->isAllOnesValue())
3111 return BinaryOperator::CreateNeg(Op0);
3114 // If the sign bits of both operands are zero (i.e. we can prove they are
3115 // unsigned inputs), turn this into a udiv.
3116 if (I.getType()->isInteger()) {
3117 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3118 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3119 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3120 return BinaryOperator::CreateUDiv(Op0, Op1, I.getName());
3127 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3128 return commonDivTransforms(I);
3131 /// This function implements the transforms on rem instructions that work
3132 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3133 /// is used by the visitors to those instructions.
3134 /// @brief Transforms common to all three rem instructions
3135 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3136 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3138 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3139 if (I.getType()->isFPOrFPVector())
3140 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3141 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3143 if (isa<UndefValue>(Op1))
3144 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3146 // Handle cases involving: rem X, (select Cond, Y, Z)
3147 if (isa<SelectInst>(Op1) && SimplifyDivRemOfSelect(I))
3153 /// This function implements the transforms common to both integer remainder
3154 /// instructions (urem and srem). It is called by the visitors to those integer
3155 /// remainder instructions.
3156 /// @brief Common integer remainder transforms
3157 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3158 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3160 if (Instruction *common = commonRemTransforms(I))
3163 // 0 % X == 0 for integer, we don't need to preserve faults!
3164 if (Constant *LHS = dyn_cast<Constant>(Op0))
3165 if (LHS->isNullValue())
3166 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3168 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3169 // X % 0 == undef, we don't need to preserve faults!
3170 if (RHS->equalsInt(0))
3171 return ReplaceInstUsesWith(I, Context->getUndef(I.getType()));
3173 if (RHS->equalsInt(1)) // X % 1 == 0
3174 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3176 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3177 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3178 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3180 } else if (isa<PHINode>(Op0I)) {
3181 if (Instruction *NV = FoldOpIntoPhi(I))
3185 // See if we can fold away this rem instruction.
3186 if (SimplifyDemandedInstructionBits(I))
3194 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3195 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3197 if (Instruction *common = commonIRemTransforms(I))
3200 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3201 // X urem C^2 -> X and C
3202 // Check to see if this is an unsigned remainder with an exact power of 2,
3203 // if so, convert to a bitwise and.
3204 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3205 if (C->getValue().isPowerOf2())
3206 return BinaryOperator::CreateAnd(Op0, SubOne(C, Context));
3209 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3210 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3211 if (RHSI->getOpcode() == Instruction::Shl &&
3212 isa<ConstantInt>(RHSI->getOperand(0))) {
3213 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3214 Constant *N1 = Context->getConstantIntAllOnesValue(I.getType());
3215 Value *Add = InsertNewInstBefore(BinaryOperator::CreateAdd(RHSI, N1,
3217 return BinaryOperator::CreateAnd(Op0, Add);
3222 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3223 // where C1&C2 are powers of two.
3224 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3225 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3226 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3227 // STO == 0 and SFO == 0 handled above.
3228 if ((STO->getValue().isPowerOf2()) &&
3229 (SFO->getValue().isPowerOf2())) {
3230 Value *TrueAnd = InsertNewInstBefore(
3231 BinaryOperator::CreateAnd(Op0, SubOne(STO, Context),
3232 SI->getName()+".t"), I);
3233 Value *FalseAnd = InsertNewInstBefore(
3234 BinaryOperator::CreateAnd(Op0, SubOne(SFO, Context),
3235 SI->getName()+".f"), I);
3236 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3244 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3245 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3247 // Handle the integer rem common cases
3248 if (Instruction *common = commonIRemTransforms(I))
3251 if (Value *RHSNeg = dyn_castNegVal(Op1, Context))
3252 if (!isa<Constant>(RHSNeg) ||
3253 (isa<ConstantInt>(RHSNeg) &&
3254 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive())) {
3256 AddUsesToWorkList(I);
3257 I.setOperand(1, RHSNeg);
3261 // If the sign bits of both operands are zero (i.e. we can prove they are
3262 // unsigned inputs), turn this into a urem.
3263 if (I.getType()->isInteger()) {
3264 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3265 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3266 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3267 return BinaryOperator::CreateURem(Op0, Op1, I.getName());
3271 // If it's a constant vector, flip any negative values positive.
3272 if (ConstantVector *RHSV = dyn_cast<ConstantVector>(Op1)) {
3273 unsigned VWidth = RHSV->getNumOperands();
3275 bool hasNegative = false;
3276 for (unsigned i = 0; !hasNegative && i != VWidth; ++i)
3277 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i)))
3278 if (RHS->getValue().isNegative())
3282 std::vector<Constant *> Elts(VWidth);
3283 for (unsigned i = 0; i != VWidth; ++i) {
3284 if (ConstantInt *RHS = dyn_cast<ConstantInt>(RHSV->getOperand(i))) {
3285 if (RHS->getValue().isNegative())
3286 Elts[i] = cast<ConstantInt>(Context->getConstantExprNeg(RHS));
3292 Constant *NewRHSV = Context->getConstantVector(Elts);
3293 if (NewRHSV != RHSV) {
3294 AddUsesToWorkList(I);
3295 I.setOperand(1, NewRHSV);
3304 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3305 return commonRemTransforms(I);
3308 // isOneBitSet - Return true if there is exactly one bit set in the specified
3310 static bool isOneBitSet(const ConstantInt *CI) {
3311 return CI->getValue().isPowerOf2();
3314 // isHighOnes - Return true if the constant is of the form 1+0+.
3315 // This is the same as lowones(~X).
3316 static bool isHighOnes(const ConstantInt *CI) {
3317 return (~CI->getValue() + 1).isPowerOf2();
3320 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3321 /// are carefully arranged to allow folding of expressions such as:
3323 /// (A < B) | (A > B) --> (A != B)
3325 /// Note that this is only valid if the first and second predicates have the
3326 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3328 /// Three bits are used to represent the condition, as follows:
3333 /// <=> Value Definition
3334 /// 000 0 Always false
3341 /// 111 7 Always true
3343 static unsigned getICmpCode(const ICmpInst *ICI) {
3344 switch (ICI->getPredicate()) {
3346 case ICmpInst::ICMP_UGT: return 1; // 001
3347 case ICmpInst::ICMP_SGT: return 1; // 001
3348 case ICmpInst::ICMP_EQ: return 2; // 010
3349 case ICmpInst::ICMP_UGE: return 3; // 011
3350 case ICmpInst::ICMP_SGE: return 3; // 011
3351 case ICmpInst::ICMP_ULT: return 4; // 100
3352 case ICmpInst::ICMP_SLT: return 4; // 100
3353 case ICmpInst::ICMP_NE: return 5; // 101
3354 case ICmpInst::ICMP_ULE: return 6; // 110
3355 case ICmpInst::ICMP_SLE: return 6; // 110
3358 LLVM_UNREACHABLE("Invalid ICmp predicate!");
3363 /// getFCmpCode - Similar to getICmpCode but for FCmpInst. This encodes a fcmp
3364 /// predicate into a three bit mask. It also returns whether it is an ordered
3365 /// predicate by reference.
3366 static unsigned getFCmpCode(FCmpInst::Predicate CC, bool &isOrdered) {
3369 case FCmpInst::FCMP_ORD: isOrdered = true; return 0; // 000
3370 case FCmpInst::FCMP_UNO: return 0; // 000
3371 case FCmpInst::FCMP_OGT: isOrdered = true; return 1; // 001
3372 case FCmpInst::FCMP_UGT: return 1; // 001
3373 case FCmpInst::FCMP_OEQ: isOrdered = true; return 2; // 010
3374 case FCmpInst::FCMP_UEQ: return 2; // 010
3375 case FCmpInst::FCMP_OGE: isOrdered = true; return 3; // 011
3376 case FCmpInst::FCMP_UGE: return 3; // 011
3377 case FCmpInst::FCMP_OLT: isOrdered = true; return 4; // 100
3378 case FCmpInst::FCMP_ULT: return 4; // 100
3379 case FCmpInst::FCMP_ONE: isOrdered = true; return 5; // 101
3380 case FCmpInst::FCMP_UNE: return 5; // 101
3381 case FCmpInst::FCMP_OLE: isOrdered = true; return 6; // 110
3382 case FCmpInst::FCMP_ULE: return 6; // 110
3385 // Not expecting FCMP_FALSE and FCMP_TRUE;
3386 LLVM_UNREACHABLE("Unexpected FCmp predicate!");
3391 /// getICmpValue - This is the complement of getICmpCode, which turns an
3392 /// opcode and two operands into either a constant true or false, or a brand
3393 /// new ICmp instruction. The sign is passed in to determine which kind
3394 /// of predicate to use in the new icmp instruction.
3395 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS,
3396 LLVMContext *Context) {
3398 default: LLVM_UNREACHABLE("Illegal ICmp code!");
3399 case 0: return Context->getConstantIntFalse();
3402 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, LHS, RHS);
3404 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, LHS, RHS);
3405 case 2: return new ICmpInst(*Context, ICmpInst::ICMP_EQ, LHS, RHS);
3408 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHS, RHS);
3410 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHS, RHS);
3413 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHS, RHS);
3415 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHS, RHS);
3416 case 5: return new ICmpInst(*Context, ICmpInst::ICMP_NE, LHS, RHS);
3419 return new ICmpInst(*Context, ICmpInst::ICMP_SLE, LHS, RHS);
3421 return new ICmpInst(*Context, ICmpInst::ICMP_ULE, LHS, RHS);
3422 case 7: return Context->getConstantIntTrue();
3426 /// getFCmpValue - This is the complement of getFCmpCode, which turns an
3427 /// opcode and two operands into either a FCmp instruction. isordered is passed
3428 /// in to determine which kind of predicate to use in the new fcmp instruction.
3429 static Value *getFCmpValue(bool isordered, unsigned code,
3430 Value *LHS, Value *RHS, LLVMContext *Context) {
3432 default: LLVM_UNREACHABLE("Illegal FCmp code!");
3435 return new FCmpInst(*Context, FCmpInst::FCMP_ORD, LHS, RHS);
3437 return new FCmpInst(*Context, FCmpInst::FCMP_UNO, LHS, RHS);
3440 return new FCmpInst(*Context, FCmpInst::FCMP_OGT, LHS, RHS);
3442 return new FCmpInst(*Context, FCmpInst::FCMP_UGT, LHS, RHS);
3445 return new FCmpInst(*Context, FCmpInst::FCMP_OEQ, LHS, RHS);
3447 return new FCmpInst(*Context, FCmpInst::FCMP_UEQ, LHS, RHS);
3450 return new FCmpInst(*Context, FCmpInst::FCMP_OGE, LHS, RHS);
3452 return new FCmpInst(*Context, FCmpInst::FCMP_UGE, LHS, RHS);
3455 return new FCmpInst(*Context, FCmpInst::FCMP_OLT, LHS, RHS);
3457 return new FCmpInst(*Context, FCmpInst::FCMP_ULT, LHS, RHS);
3460 return new FCmpInst(*Context, FCmpInst::FCMP_ONE, LHS, RHS);
3462 return new FCmpInst(*Context, FCmpInst::FCMP_UNE, LHS, RHS);
3465 return new FCmpInst(*Context, FCmpInst::FCMP_OLE, LHS, RHS);
3467 return new FCmpInst(*Context, FCmpInst::FCMP_ULE, LHS, RHS);
3468 case 7: return Context->getConstantIntTrue();
3472 /// PredicatesFoldable - Return true if both predicates match sign or if at
3473 /// least one of them is an equality comparison (which is signless).
3474 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3475 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3476 (ICmpInst::isSignedPredicate(p1) && ICmpInst::isEquality(p2)) ||
3477 (ICmpInst::isSignedPredicate(p2) && ICmpInst::isEquality(p1));
3481 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3482 struct FoldICmpLogical {
3485 ICmpInst::Predicate pred;
3486 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3487 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3488 pred(ICI->getPredicate()) {}
3489 bool shouldApply(Value *V) const {
3490 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3491 if (PredicatesFoldable(pred, ICI->getPredicate()))
3492 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3493 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3496 Instruction *apply(Instruction &Log) const {
3497 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3498 if (ICI->getOperand(0) != LHS) {
3499 assert(ICI->getOperand(1) == LHS);
3500 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3503 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3504 unsigned LHSCode = getICmpCode(ICI);
3505 unsigned RHSCode = getICmpCode(RHSICI);
3507 switch (Log.getOpcode()) {
3508 case Instruction::And: Code = LHSCode & RHSCode; break;
3509 case Instruction::Or: Code = LHSCode | RHSCode; break;
3510 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3511 default: LLVM_UNREACHABLE("Illegal logical opcode!"); return 0;
3514 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3515 ICmpInst::isSignedPredicate(ICI->getPredicate());
3517 Value *RV = getICmpValue(isSigned, Code, LHS, RHS, IC.getContext());
3518 if (Instruction *I = dyn_cast<Instruction>(RV))
3520 // Otherwise, it's a constant boolean value...
3521 return IC.ReplaceInstUsesWith(Log, RV);
3524 } // end anonymous namespace
3526 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3527 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3528 // guaranteed to be a binary operator.
3529 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3531 ConstantInt *AndRHS,
3532 BinaryOperator &TheAnd) {
3533 Value *X = Op->getOperand(0);
3534 Constant *Together = 0;
3536 Together = Context->getConstantExprAnd(AndRHS, OpRHS);
3538 switch (Op->getOpcode()) {
3539 case Instruction::Xor:
3540 if (Op->hasOneUse()) {
3541 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3542 Instruction *And = BinaryOperator::CreateAnd(X, AndRHS);
3543 InsertNewInstBefore(And, TheAnd);
3545 return BinaryOperator::CreateXor(And, Together);
3548 case Instruction::Or:
3549 if (Together == AndRHS) // (X | C) & C --> C
3550 return ReplaceInstUsesWith(TheAnd, AndRHS);
3552 if (Op->hasOneUse() && Together != OpRHS) {
3553 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3554 Instruction *Or = BinaryOperator::CreateOr(X, Together);
3555 InsertNewInstBefore(Or, TheAnd);
3557 return BinaryOperator::CreateAnd(Or, AndRHS);
3560 case Instruction::Add:
3561 if (Op->hasOneUse()) {
3562 // Adding a one to a single bit bit-field should be turned into an XOR
3563 // of the bit. First thing to check is to see if this AND is with a
3564 // single bit constant.
3565 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3567 // If there is only one bit set...
3568 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3569 // Ok, at this point, we know that we are masking the result of the
3570 // ADD down to exactly one bit. If the constant we are adding has
3571 // no bits set below this bit, then we can eliminate the ADD.
3572 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3574 // Check to see if any bits below the one bit set in AndRHSV are set.
3575 if ((AddRHS & (AndRHSV-1)) == 0) {
3576 // If not, the only thing that can effect the output of the AND is
3577 // the bit specified by AndRHSV. If that bit is set, the effect of
3578 // the XOR is to toggle the bit. If it is clear, then the ADD has
3580 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3581 TheAnd.setOperand(0, X);
3584 // Pull the XOR out of the AND.
3585 Instruction *NewAnd = BinaryOperator::CreateAnd(X, AndRHS);
3586 InsertNewInstBefore(NewAnd, TheAnd);
3587 NewAnd->takeName(Op);
3588 return BinaryOperator::CreateXor(NewAnd, AndRHS);
3595 case Instruction::Shl: {
3596 // We know that the AND will not produce any of the bits shifted in, so if
3597 // the anded constant includes them, clear them now!
3599 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3600 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3601 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3602 ConstantInt *CI = Context->getConstantInt(AndRHS->getValue() & ShlMask);
3604 if (CI->getValue() == ShlMask) {
3605 // Masking out bits that the shift already masks
3606 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3607 } else if (CI != AndRHS) { // Reducing bits set in and.
3608 TheAnd.setOperand(1, CI);
3613 case Instruction::LShr:
3615 // We know that the AND will not produce any of the bits shifted in, so if
3616 // the anded constant includes them, clear them now! This only applies to
3617 // unsigned shifts, because a signed shr may bring in set bits!
3619 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3620 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3621 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3622 ConstantInt *CI = Context->getConstantInt(AndRHS->getValue() & ShrMask);
3624 if (CI->getValue() == ShrMask) {
3625 // Masking out bits that the shift already masks.
3626 return ReplaceInstUsesWith(TheAnd, Op);
3627 } else if (CI != AndRHS) {
3628 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3633 case Instruction::AShr:
3635 // See if this is shifting in some sign extension, then masking it out
3637 if (Op->hasOneUse()) {
3638 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3639 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3640 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3641 Constant *C = Context->getConstantInt(AndRHS->getValue() & ShrMask);
3642 if (C == AndRHS) { // Masking out bits shifted in.
3643 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3644 // Make the argument unsigned.
3645 Value *ShVal = Op->getOperand(0);
3646 ShVal = InsertNewInstBefore(
3647 BinaryOperator::CreateLShr(ShVal, OpRHS,
3648 Op->getName()), TheAnd);
3649 return BinaryOperator::CreateAnd(ShVal, AndRHS, TheAnd.getName());
3658 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3659 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3660 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3661 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3662 /// insert new instructions.
3663 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3664 bool isSigned, bool Inside,
3666 assert(cast<ConstantInt>(Context->getConstantExprICmp((isSigned ?
3667 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3668 "Lo is not <= Hi in range emission code!");
3671 if (Lo == Hi) // Trivially false.
3672 return new ICmpInst(*Context, ICmpInst::ICMP_NE, V, V);
3674 // V >= Min && V < Hi --> V < Hi
3675 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3676 ICmpInst::Predicate pred = (isSigned ?
3677 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3678 return new ICmpInst(*Context, pred, V, Hi);
3681 // Emit V-Lo <u Hi-Lo
3682 Constant *NegLo = Context->getConstantExprNeg(Lo);
3683 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3684 InsertNewInstBefore(Add, IB);
3685 Constant *UpperBound = Context->getConstantExprAdd(NegLo, Hi);
3686 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, UpperBound);
3689 if (Lo == Hi) // Trivially true.
3690 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, V, V);
3692 // V < Min || V >= Hi -> V > Hi-1
3693 Hi = SubOne(cast<ConstantInt>(Hi), Context);
3694 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3695 ICmpInst::Predicate pred = (isSigned ?
3696 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3697 return new ICmpInst(*Context, pred, V, Hi);
3700 // Emit V-Lo >u Hi-1-Lo
3701 // Note that Hi has already had one subtracted from it, above.
3702 ConstantInt *NegLo = cast<ConstantInt>(Context->getConstantExprNeg(Lo));
3703 Instruction *Add = BinaryOperator::CreateAdd(V, NegLo, V->getName()+".off");
3704 InsertNewInstBefore(Add, IB);
3705 Constant *LowerBound = Context->getConstantExprAdd(NegLo, Hi);
3706 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add, LowerBound);
3709 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3710 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3711 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3712 // not, since all 1s are not contiguous.
3713 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3714 const APInt& V = Val->getValue();
3715 uint32_t BitWidth = Val->getType()->getBitWidth();
3716 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3718 // look for the first zero bit after the run of ones
3719 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3720 // look for the first non-zero bit
3721 ME = V.getActiveBits();
3725 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3726 /// where isSub determines whether the operator is a sub. If we can fold one of
3727 /// the following xforms:
3729 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3730 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3731 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3733 /// return (A +/- B).
3735 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3736 ConstantInt *Mask, bool isSub,
3738 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3739 if (!LHSI || LHSI->getNumOperands() != 2 ||
3740 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3742 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3744 switch (LHSI->getOpcode()) {
3746 case Instruction::And:
3747 if (Context->getConstantExprAnd(N, Mask) == Mask) {
3748 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3749 if ((Mask->getValue().countLeadingZeros() +
3750 Mask->getValue().countPopulation()) ==
3751 Mask->getValue().getBitWidth())
3754 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3755 // part, we don't need any explicit masks to take them out of A. If that
3756 // is all N is, ignore it.
3757 uint32_t MB = 0, ME = 0;
3758 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3759 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3760 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3761 if (MaskedValueIsZero(RHS, Mask))
3766 case Instruction::Or:
3767 case Instruction::Xor:
3768 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3769 if ((Mask->getValue().countLeadingZeros() +
3770 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3771 && Context->getConstantExprAnd(N, Mask)->isNullValue())
3778 New = BinaryOperator::CreateSub(LHSI->getOperand(0), RHS, "fold");
3780 New = BinaryOperator::CreateAdd(LHSI->getOperand(0), RHS, "fold");
3781 return InsertNewInstBefore(New, I);
3784 /// FoldAndOfICmps - Fold (icmp)&(icmp) if possible.
3785 Instruction *InstCombiner::FoldAndOfICmps(Instruction &I,
3786 ICmpInst *LHS, ICmpInst *RHS) {
3788 ConstantInt *LHSCst, *RHSCst;
3789 ICmpInst::Predicate LHSCC, RHSCC;
3791 // This only handles icmp of constants: (icmp1 A, C1) & (icmp2 B, C2).
3792 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
3793 m_ConstantInt(LHSCst)), *Context) ||
3794 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
3795 m_ConstantInt(RHSCst)), *Context))
3798 // (icmp ult A, C) & (icmp ult B, C) --> (icmp ult (A|B), C)
3799 // where C is a power of 2
3800 if (LHSCst == RHSCst && LHSCC == RHSCC && LHSCC == ICmpInst::ICMP_ULT &&
3801 LHSCst->getValue().isPowerOf2()) {
3802 Instruction *NewOr = BinaryOperator::CreateOr(Val, Val2);
3803 InsertNewInstBefore(NewOr, I);
3804 return new ICmpInst(*Context, LHSCC, NewOr, LHSCst);
3807 // From here on, we only handle:
3808 // (icmp1 A, C1) & (icmp2 A, C2) --> something simpler.
3809 if (Val != Val2) return 0;
3811 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
3812 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
3813 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
3814 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
3815 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
3818 // We can't fold (ugt x, C) & (sgt x, C2).
3819 if (!PredicatesFoldable(LHSCC, RHSCC))
3822 // Ensure that the larger constant is on the RHS.
3824 if (ICmpInst::isSignedPredicate(LHSCC) ||
3825 (ICmpInst::isEquality(LHSCC) &&
3826 ICmpInst::isSignedPredicate(RHSCC)))
3827 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3829 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3832 std::swap(LHS, RHS);
3833 std::swap(LHSCst, RHSCst);
3834 std::swap(LHSCC, RHSCC);
3837 // At this point, we know we have have two icmp instructions
3838 // comparing a value against two constants and and'ing the result
3839 // together. Because of the above check, we know that we only have
3840 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3841 // (from the FoldICmpLogical check above), that the two constants
3842 // are not equal and that the larger constant is on the RHS
3843 assert(LHSCst != RHSCst && "Compares not folded above?");
3846 default: LLVM_UNREACHABLE("Unknown integer condition code!");
3847 case ICmpInst::ICMP_EQ:
3849 default: LLVM_UNREACHABLE("Unknown integer condition code!");
3850 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3851 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3852 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3853 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3854 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3855 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3856 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3857 return ReplaceInstUsesWith(I, LHS);
3859 case ICmpInst::ICMP_NE:
3861 default: LLVM_UNREACHABLE("Unknown integer condition code!");
3862 case ICmpInst::ICMP_ULT:
3863 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X u< 14) -> X < 13
3864 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Val, LHSCst);
3865 break; // (X != 13 & X u< 15) -> no change
3866 case ICmpInst::ICMP_SLT:
3867 if (LHSCst == SubOne(RHSCst, Context)) // (X != 13 & X s< 14) -> X < 13
3868 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Val, LHSCst);
3869 break; // (X != 13 & X s< 15) -> no change
3870 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3871 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3872 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3873 return ReplaceInstUsesWith(I, RHS);
3874 case ICmpInst::ICMP_NE:
3875 if (LHSCst == SubOne(RHSCst, Context)){// (X != 13 & X != 14) -> X-13 >u 1
3876 Constant *AddCST = Context->getConstantExprNeg(LHSCst);
3877 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
3878 Val->getName()+".off");
3879 InsertNewInstBefore(Add, I);
3880 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Add,
3881 Context->getConstantInt(Add->getType(), 1));
3883 break; // (X != 13 & X != 15) -> no change
3886 case ICmpInst::ICMP_ULT:
3888 default: LLVM_UNREACHABLE("Unknown integer condition code!");
3889 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3890 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3891 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3892 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3894 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3895 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3896 return ReplaceInstUsesWith(I, LHS);
3897 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3901 case ICmpInst::ICMP_SLT:
3903 default: LLVM_UNREACHABLE("Unknown integer condition code!");
3904 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3905 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3906 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
3907 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3909 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3910 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3911 return ReplaceInstUsesWith(I, LHS);
3912 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3916 case ICmpInst::ICMP_UGT:
3918 default: LLVM_UNREACHABLE("Unknown integer condition code!");
3919 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X == 15
3920 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3921 return ReplaceInstUsesWith(I, RHS);
3922 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3924 case ICmpInst::ICMP_NE:
3925 if (RHSCst == AddOne(LHSCst, Context)) // (X u> 13 & X != 14) -> X u> 14
3926 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3927 break; // (X u> 13 & X != 15) -> no change
3928 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) -> (X-14) <u 1
3929 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3930 RHSCst, false, true, I);
3931 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3935 case ICmpInst::ICMP_SGT:
3937 default: LLVM_UNREACHABLE("Unknown integer condition code!");
3938 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
3939 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3940 return ReplaceInstUsesWith(I, RHS);
3941 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3943 case ICmpInst::ICMP_NE:
3944 if (RHSCst == AddOne(LHSCst, Context)) // (X s> 13 & X != 14) -> X s> 14
3945 return new ICmpInst(*Context, LHSCC, Val, RHSCst);
3946 break; // (X s> 13 & X != 15) -> no change
3947 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) -> (X-14) s< 1
3948 return InsertRangeTest(Val, AddOne(LHSCst, Context),
3949 RHSCst, true, true, I);
3950 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3960 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3961 bool Changed = SimplifyCommutative(I);
3962 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3964 if (isa<UndefValue>(Op1)) // X & undef -> 0
3965 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
3969 return ReplaceInstUsesWith(I, Op1);
3971 // See if we can simplify any instructions used by the instruction whose sole
3972 // purpose is to compute bits we don't care about.
3973 if (SimplifyDemandedInstructionBits(I))
3975 if (isa<VectorType>(I.getType())) {
3976 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3977 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3978 return ReplaceInstUsesWith(I, I.getOperand(0));
3979 } else if (isa<ConstantAggregateZero>(Op1)) {
3980 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3984 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3985 const APInt& AndRHSMask = AndRHS->getValue();
3986 APInt NotAndRHS(~AndRHSMask);
3988 // Optimize a variety of ((val OP C1) & C2) combinations...
3989 if (isa<BinaryOperator>(Op0)) {
3990 Instruction *Op0I = cast<Instruction>(Op0);
3991 Value *Op0LHS = Op0I->getOperand(0);
3992 Value *Op0RHS = Op0I->getOperand(1);
3993 switch (Op0I->getOpcode()) {
3994 case Instruction::Xor:
3995 case Instruction::Or:
3996 // If the mask is only needed on one incoming arm, push it up.
3997 if (Op0I->hasOneUse()) {
3998 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3999 // Not masking anything out for the LHS, move to RHS.
4000 Instruction *NewRHS = BinaryOperator::CreateAnd(Op0RHS, AndRHS,
4001 Op0RHS->getName()+".masked");
4002 InsertNewInstBefore(NewRHS, I);
4003 return BinaryOperator::Create(
4004 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
4006 if (!isa<Constant>(Op0RHS) &&
4007 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
4008 // Not masking anything out for the RHS, move to LHS.
4009 Instruction *NewLHS = BinaryOperator::CreateAnd(Op0LHS, AndRHS,
4010 Op0LHS->getName()+".masked");
4011 InsertNewInstBefore(NewLHS, I);
4012 return BinaryOperator::Create(
4013 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
4018 case Instruction::Add:
4019 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
4020 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4021 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
4022 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
4023 return BinaryOperator::CreateAnd(V, AndRHS);
4024 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
4025 return BinaryOperator::CreateAnd(V, AndRHS); // Add commutes
4028 case Instruction::Sub:
4029 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
4030 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4031 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
4032 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
4033 return BinaryOperator::CreateAnd(V, AndRHS);
4035 // (A - N) & AndRHS -> -N & AndRHS iff A&AndRHS==0 and AndRHS
4036 // has 1's for all bits that the subtraction with A might affect.
4037 if (Op0I->hasOneUse()) {
4038 uint32_t BitWidth = AndRHSMask.getBitWidth();
4039 uint32_t Zeros = AndRHSMask.countLeadingZeros();
4040 APInt Mask = APInt::getLowBitsSet(BitWidth, BitWidth - Zeros);
4042 ConstantInt *A = dyn_cast<ConstantInt>(Op0LHS);
4043 if (!(A && A->isZero()) && // avoid infinite recursion.
4044 MaskedValueIsZero(Op0LHS, Mask)) {
4045 Instruction *NewNeg = BinaryOperator::CreateNeg(Op0RHS);
4046 InsertNewInstBefore(NewNeg, I);
4047 return BinaryOperator::CreateAnd(NewNeg, AndRHS);
4052 case Instruction::Shl:
4053 case Instruction::LShr:
4054 // (1 << x) & 1 --> zext(x == 0)
4055 // (1 >> x) & 1 --> zext(x == 0)
4056 if (AndRHSMask == 1 && Op0LHS == AndRHS) {
4057 Instruction *NewICmp = new ICmpInst(*Context, ICmpInst::ICMP_EQ,
4058 Op0RHS, Context->getNullValue(I.getType()));
4059 InsertNewInstBefore(NewICmp, I);
4060 return new ZExtInst(NewICmp, I.getType());
4065 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4066 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
4068 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4069 // If this is an integer truncation or change from signed-to-unsigned, and
4070 // if the source is an and/or with immediate, transform it. This
4071 // frequently occurs for bitfield accesses.
4072 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
4073 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
4074 CastOp->getNumOperands() == 2)
4075 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
4076 if (CastOp->getOpcode() == Instruction::And) {
4077 // Change: and (cast (and X, C1) to T), C2
4078 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
4079 // This will fold the two constants together, which may allow
4080 // other simplifications.
4081 Instruction *NewCast = CastInst::CreateTruncOrBitCast(
4082 CastOp->getOperand(0), I.getType(),
4083 CastOp->getName()+".shrunk");
4084 NewCast = InsertNewInstBefore(NewCast, I);
4085 // trunc_or_bitcast(C1)&C2
4087 Context->getConstantExprTruncOrBitCast(AndCI,I.getType());
4088 C3 = Context->getConstantExprAnd(C3, AndRHS);
4089 return BinaryOperator::CreateAnd(NewCast, C3);
4090 } else if (CastOp->getOpcode() == Instruction::Or) {
4091 // Change: and (cast (or X, C1) to T), C2
4092 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
4094 Context->getConstantExprTruncOrBitCast(AndCI,I.getType());
4095 if (Context->getConstantExprAnd(C3, AndRHS) == AndRHS)
4097 return ReplaceInstUsesWith(I, AndRHS);
4103 // Try to fold constant and into select arguments.
4104 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4105 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4107 if (isa<PHINode>(Op0))
4108 if (Instruction *NV = FoldOpIntoPhi(I))
4112 Value *Op0NotVal = dyn_castNotVal(Op0, Context);
4113 Value *Op1NotVal = dyn_castNotVal(Op1, Context);
4115 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
4116 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
4118 // (~A & ~B) == (~(A | B)) - De Morgan's Law
4119 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4120 Instruction *Or = BinaryOperator::CreateOr(Op0NotVal, Op1NotVal,
4121 I.getName()+".demorgan");
4122 InsertNewInstBefore(Or, I);
4123 return BinaryOperator::CreateNot(Or);
4127 Value *A = 0, *B = 0, *C = 0, *D = 0;
4128 if (match(Op0, m_Or(m_Value(A), m_Value(B)), *Context)) {
4129 if (A == Op1 || B == Op1) // (A | ?) & A --> A
4130 return ReplaceInstUsesWith(I, Op1);
4132 // (A|B) & ~(A&B) -> A^B
4133 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4134 if ((A == C && B == D) || (A == D && B == C))
4135 return BinaryOperator::CreateXor(A, B);
4139 if (match(Op1, m_Or(m_Value(A), m_Value(B)), *Context)) {
4140 if (A == Op0 || B == Op0) // A & (A | ?) --> A
4141 return ReplaceInstUsesWith(I, Op0);
4143 // ~(A&B) & (A|B) -> A^B
4144 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))), *Context)) {
4145 if ((A == C && B == D) || (A == D && B == C))
4146 return BinaryOperator::CreateXor(A, B);
4150 if (Op0->hasOneUse() &&
4151 match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4152 if (A == Op1) { // (A^B)&A -> A&(A^B)
4153 I.swapOperands(); // Simplify below
4154 std::swap(Op0, Op1);
4155 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
4156 cast<BinaryOperator>(Op0)->swapOperands();
4157 I.swapOperands(); // Simplify below
4158 std::swap(Op0, Op1);
4162 if (Op1->hasOneUse() &&
4163 match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context)) {
4164 if (B == Op0) { // B&(A^B) -> B&(B^A)
4165 cast<BinaryOperator>(Op1)->swapOperands();
4168 if (A == Op0) { // A&(A^B) -> A & ~B
4169 Instruction *NotB = BinaryOperator::CreateNot(B, "tmp");
4170 InsertNewInstBefore(NotB, I);
4171 return BinaryOperator::CreateAnd(A, NotB);
4175 // (A&((~A)|B)) -> A&B
4176 if (match(Op0, m_Or(m_Not(m_Specific(Op1)), m_Value(A)), *Context) ||
4177 match(Op0, m_Or(m_Value(A), m_Not(m_Specific(Op1))), *Context))
4178 return BinaryOperator::CreateAnd(A, Op1);
4179 if (match(Op1, m_Or(m_Not(m_Specific(Op0)), m_Value(A)), *Context) ||
4180 match(Op1, m_Or(m_Value(A), m_Not(m_Specific(Op0))), *Context))
4181 return BinaryOperator::CreateAnd(A, Op0);
4184 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
4185 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
4186 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4189 if (ICmpInst *LHS = dyn_cast<ICmpInst>(Op0))
4190 if (Instruction *Res = FoldAndOfICmps(I, LHS, RHS))
4194 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4195 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4196 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4197 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4198 const Type *SrcTy = Op0C->getOperand(0)->getType();
4199 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4200 // Only do this if the casts both really cause code to be generated.
4201 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4203 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4205 Instruction *NewOp = BinaryOperator::CreateAnd(Op0C->getOperand(0),
4206 Op1C->getOperand(0),
4208 InsertNewInstBefore(NewOp, I);
4209 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4213 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4214 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4215 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4216 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4217 SI0->getOperand(1) == SI1->getOperand(1) &&
4218 (SI0->hasOneUse() || SI1->hasOneUse())) {
4219 Instruction *NewOp =
4220 InsertNewInstBefore(BinaryOperator::CreateAnd(SI0->getOperand(0),
4222 SI0->getName()), I);
4223 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4224 SI1->getOperand(1));
4228 // If and'ing two fcmp, try combine them into one.
4229 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4230 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4231 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4232 RHS->getPredicate() == FCmpInst::FCMP_ORD) {
4233 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4234 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4235 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4236 // If either of the constants are nans, then the whole thing returns
4238 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4239 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4240 return new FCmpInst(*Context, FCmpInst::FCMP_ORD,
4241 LHS->getOperand(0), RHS->getOperand(0));
4244 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4245 FCmpInst::Predicate Op0CC, Op1CC;
4246 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS),
4247 m_Value(Op0RHS)), *Context) &&
4248 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS),
4249 m_Value(Op1RHS)), *Context)) {
4250 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4251 // Swap RHS operands to match LHS.
4252 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4253 std::swap(Op1LHS, Op1RHS);
4255 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4256 // Simplify (fcmp cc0 x, y) & (fcmp cc1 x, y).
4258 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4260 else if (Op0CC == FCmpInst::FCMP_FALSE ||
4261 Op1CC == FCmpInst::FCMP_FALSE)
4262 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4263 else if (Op0CC == FCmpInst::FCMP_TRUE)
4264 return ReplaceInstUsesWith(I, Op1);
4265 else if (Op1CC == FCmpInst::FCMP_TRUE)
4266 return ReplaceInstUsesWith(I, Op0);
4269 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4270 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4272 std::swap(Op0, Op1);
4273 std::swap(Op0Pred, Op1Pred);
4274 std::swap(Op0Ordered, Op1Ordered);
4277 // uno && ueq -> uno && (uno || eq) -> ueq
4278 // ord && olt -> ord && (ord && lt) -> olt
4279 if (Op0Ordered == Op1Ordered)
4280 return ReplaceInstUsesWith(I, Op1);
4281 // uno && oeq -> uno && (ord && eq) -> false
4282 // uno && ord -> false
4284 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
4285 // ord && ueq -> ord && (uno || eq) -> oeq
4286 return cast<Instruction>(getFCmpValue(true, Op1Pred,
4287 Op0LHS, Op0RHS, Context));
4295 return Changed ? &I : 0;
4298 /// CollectBSwapParts - Analyze the specified subexpression and see if it is
4299 /// capable of providing pieces of a bswap. The subexpression provides pieces
4300 /// of a bswap if it is proven that each of the non-zero bytes in the output of
4301 /// the expression came from the corresponding "byte swapped" byte in some other
4302 /// value. For example, if the current subexpression is "(shl i32 %X, 24)" then
4303 /// we know that the expression deposits the low byte of %X into the high byte
4304 /// of the bswap result and that all other bytes are zero. This expression is
4305 /// accepted, the high byte of ByteValues is set to X to indicate a correct
4308 /// This function returns true if the match was unsuccessful and false if so.
4309 /// On entry to the function the "OverallLeftShift" is a signed integer value
4310 /// indicating the number of bytes that the subexpression is later shifted. For
4311 /// example, if the expression is later right shifted by 16 bits, the
4312 /// OverallLeftShift value would be -2 on entry. This is used to specify which
4313 /// byte of ByteValues is actually being set.
4315 /// Similarly, ByteMask is a bitmask where a bit is clear if its corresponding
4316 /// byte is masked to zero by a user. For example, in (X & 255), X will be
4317 /// processed with a bytemask of 1. Because bytemask is 32-bits, this limits
4318 /// this function to working on up to 32-byte (256 bit) values. ByteMask is
4319 /// always in the local (OverallLeftShift) coordinate space.
4321 static bool CollectBSwapParts(Value *V, int OverallLeftShift, uint32_t ByteMask,
4322 SmallVector<Value*, 8> &ByteValues) {
4323 if (Instruction *I = dyn_cast<Instruction>(V)) {
4324 // If this is an or instruction, it may be an inner node of the bswap.
4325 if (I->getOpcode() == Instruction::Or) {
4326 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4328 CollectBSwapParts(I->getOperand(1), OverallLeftShift, ByteMask,
4332 // If this is a logical shift by a constant multiple of 8, recurse with
4333 // OverallLeftShift and ByteMask adjusted.
4334 if (I->isLogicalShift() && isa<ConstantInt>(I->getOperand(1))) {
4336 cast<ConstantInt>(I->getOperand(1))->getLimitedValue(~0U);
4337 // Ensure the shift amount is defined and of a byte value.
4338 if ((ShAmt & 7) || (ShAmt > 8*ByteValues.size()))
4341 unsigned ByteShift = ShAmt >> 3;
4342 if (I->getOpcode() == Instruction::Shl) {
4343 // X << 2 -> collect(X, +2)
4344 OverallLeftShift += ByteShift;
4345 ByteMask >>= ByteShift;
4347 // X >>u 2 -> collect(X, -2)
4348 OverallLeftShift -= ByteShift;
4349 ByteMask <<= ByteShift;
4350 ByteMask &= (~0U >> (32-ByteValues.size()));
4353 if (OverallLeftShift >= (int)ByteValues.size()) return true;
4354 if (OverallLeftShift <= -(int)ByteValues.size()) return true;
4356 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4360 // If this is a logical 'and' with a mask that clears bytes, clear the
4361 // corresponding bytes in ByteMask.
4362 if (I->getOpcode() == Instruction::And &&
4363 isa<ConstantInt>(I->getOperand(1))) {
4364 // Scan every byte of the and mask, seeing if the byte is either 0 or 255.
4365 unsigned NumBytes = ByteValues.size();
4366 APInt Byte(I->getType()->getPrimitiveSizeInBits(), 255);
4367 const APInt &AndMask = cast<ConstantInt>(I->getOperand(1))->getValue();
4369 for (unsigned i = 0; i != NumBytes; ++i, Byte <<= 8) {
4370 // If this byte is masked out by a later operation, we don't care what
4372 if ((ByteMask & (1 << i)) == 0)
4375 // If the AndMask is all zeros for this byte, clear the bit.
4376 APInt MaskB = AndMask & Byte;
4378 ByteMask &= ~(1U << i);
4382 // If the AndMask is not all ones for this byte, it's not a bytezap.
4386 // Otherwise, this byte is kept.
4389 return CollectBSwapParts(I->getOperand(0), OverallLeftShift, ByteMask,
4394 // Okay, we got to something that isn't a shift, 'or' or 'and'. This must be
4395 // the input value to the bswap. Some observations: 1) if more than one byte
4396 // is demanded from this input, then it could not be successfully assembled
4397 // into a byteswap. At least one of the two bytes would not be aligned with
4398 // their ultimate destination.
4399 if (!isPowerOf2_32(ByteMask)) return true;
4400 unsigned InputByteNo = CountTrailingZeros_32(ByteMask);
4402 // 2) The input and ultimate destinations must line up: if byte 3 of an i32
4403 // is demanded, it needs to go into byte 0 of the result. This means that the
4404 // byte needs to be shifted until it lands in the right byte bucket. The
4405 // shift amount depends on the position: if the byte is coming from the high
4406 // part of the value (e.g. byte 3) then it must be shifted right. If from the
4407 // low part, it must be shifted left.
4408 unsigned DestByteNo = InputByteNo + OverallLeftShift;
4409 if (InputByteNo < ByteValues.size()/2) {
4410 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4413 if (ByteValues.size()-1-DestByteNo != InputByteNo)
4417 // If the destination byte value is already defined, the values are or'd
4418 // together, which isn't a bswap (unless it's an or of the same bits).
4419 if (ByteValues[DestByteNo] && ByteValues[DestByteNo] != V)
4421 ByteValues[DestByteNo] = V;
4425 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4426 /// If so, insert the new bswap intrinsic and return it.
4427 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4428 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4429 if (!ITy || ITy->getBitWidth() % 16 ||
4430 // ByteMask only allows up to 32-byte values.
4431 ITy->getBitWidth() > 32*8)
4432 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4434 /// ByteValues - For each byte of the result, we keep track of which value
4435 /// defines each byte.
4436 SmallVector<Value*, 8> ByteValues;
4437 ByteValues.resize(ITy->getBitWidth()/8);
4439 // Try to find all the pieces corresponding to the bswap.
4440 uint32_t ByteMask = ~0U >> (32-ByteValues.size());
4441 if (CollectBSwapParts(&I, 0, ByteMask, ByteValues))
4444 // Check to see if all of the bytes come from the same value.
4445 Value *V = ByteValues[0];
4446 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4448 // Check to make sure that all of the bytes come from the same value.
4449 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4450 if (ByteValues[i] != V)
4452 const Type *Tys[] = { ITy };
4453 Module *M = I.getParent()->getParent()->getParent();
4454 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4455 return CallInst::Create(F, V);
4458 /// MatchSelectFromAndOr - We have an expression of the form (A&C)|(B&D). Check
4459 /// If A is (cond?-1:0) and either B or D is ~(cond?-1,0) or (cond?0,-1), then
4460 /// we can simplify this expression to "cond ? C : D or B".
4461 static Instruction *MatchSelectFromAndOr(Value *A, Value *B,
4463 LLVMContext *Context) {
4464 // If A is not a select of -1/0, this cannot match.
4466 if (!match(A, m_SelectCst<-1, 0>(m_Value(Cond)), *Context))
4469 // ((cond?-1:0)&C) | (B&(cond?0:-1)) -> cond ? C : B.
4470 if (match(D, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4471 return SelectInst::Create(Cond, C, B);
4472 if (match(D, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4473 return SelectInst::Create(Cond, C, B);
4474 // ((cond?-1:0)&C) | ((cond?0:-1)&D) -> cond ? C : D.
4475 if (match(B, m_SelectCst<0, -1>(m_Specific(Cond)), *Context))
4476 return SelectInst::Create(Cond, C, D);
4477 if (match(B, m_Not(m_SelectCst<-1, 0>(m_Specific(Cond))), *Context))
4478 return SelectInst::Create(Cond, C, D);
4482 /// FoldOrOfICmps - Fold (icmp)|(icmp) if possible.
4483 Instruction *InstCombiner::FoldOrOfICmps(Instruction &I,
4484 ICmpInst *LHS, ICmpInst *RHS) {
4486 ConstantInt *LHSCst, *RHSCst;
4487 ICmpInst::Predicate LHSCC, RHSCC;
4489 // This only handles icmp of constants: (icmp1 A, C1) | (icmp2 B, C2).
4490 if (!match(LHS, m_ICmp(LHSCC, m_Value(Val),
4491 m_ConstantInt(LHSCst)), *Context) ||
4492 !match(RHS, m_ICmp(RHSCC, m_Value(Val2),
4493 m_ConstantInt(RHSCst)), *Context))
4496 // From here on, we only handle:
4497 // (icmp1 A, C1) | (icmp2 A, C2) --> something simpler.
4498 if (Val != Val2) return 0;
4500 // ICMP_[US][GL]E X, CST is folded to ICMP_[US][GL]T elsewhere.
4501 if (LHSCC == ICmpInst::ICMP_UGE || LHSCC == ICmpInst::ICMP_ULE ||
4502 RHSCC == ICmpInst::ICMP_UGE || RHSCC == ICmpInst::ICMP_ULE ||
4503 LHSCC == ICmpInst::ICMP_SGE || LHSCC == ICmpInst::ICMP_SLE ||
4504 RHSCC == ICmpInst::ICMP_SGE || RHSCC == ICmpInst::ICMP_SLE)
4507 // We can't fold (ugt x, C) | (sgt x, C2).
4508 if (!PredicatesFoldable(LHSCC, RHSCC))
4511 // Ensure that the larger constant is on the RHS.
4513 if (ICmpInst::isSignedPredicate(LHSCC) ||
4514 (ICmpInst::isEquality(LHSCC) &&
4515 ICmpInst::isSignedPredicate(RHSCC)))
4516 ShouldSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4518 ShouldSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4521 std::swap(LHS, RHS);
4522 std::swap(LHSCst, RHSCst);
4523 std::swap(LHSCC, RHSCC);
4526 // At this point, we know we have have two icmp instructions
4527 // comparing a value against two constants and or'ing the result
4528 // together. Because of the above check, we know that we only have
4529 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4530 // FoldICmpLogical check above), that the two constants are not
4532 assert(LHSCst != RHSCst && "Compares not folded above?");
4535 default: LLVM_UNREACHABLE("Unknown integer condition code!");
4536 case ICmpInst::ICMP_EQ:
4538 default: LLVM_UNREACHABLE("Unknown integer condition code!");
4539 case ICmpInst::ICMP_EQ:
4540 if (LHSCst == SubOne(RHSCst, Context)) {
4541 // (X == 13 | X == 14) -> X-13 <u 2
4542 Constant *AddCST = Context->getConstantExprNeg(LHSCst);
4543 Instruction *Add = BinaryOperator::CreateAdd(Val, AddCST,
4544 Val->getName()+".off");
4545 InsertNewInstBefore(Add, I);
4546 AddCST = Context->getConstantExprSub(AddOne(RHSCst, Context), LHSCst);
4547 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Add, AddCST);
4549 break; // (X == 13 | X == 15) -> no change
4550 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4551 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4553 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4554 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4555 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4556 return ReplaceInstUsesWith(I, RHS);
4559 case ICmpInst::ICMP_NE:
4561 default: LLVM_UNREACHABLE("Unknown integer condition code!");
4562 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4563 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4564 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4565 return ReplaceInstUsesWith(I, LHS);
4566 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4567 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4568 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4569 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4572 case ICmpInst::ICMP_ULT:
4574 default: LLVM_UNREACHABLE("Unknown integer condition code!");
4575 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4577 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) -> (X-13) u> 2
4578 // If RHSCst is [us]MAXINT, it is always false. Not handling
4579 // this can cause overflow.
4580 if (RHSCst->isMaxValue(false))
4581 return ReplaceInstUsesWith(I, LHS);
4582 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4584 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4586 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4587 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4588 return ReplaceInstUsesWith(I, RHS);
4589 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4593 case ICmpInst::ICMP_SLT:
4595 default: LLVM_UNREACHABLE("Unknown integer condition code!");
4596 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4598 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) -> (X-13) s> 2
4599 // If RHSCst is [us]MAXINT, it is always false. Not handling
4600 // this can cause overflow.
4601 if (RHSCst->isMaxValue(true))
4602 return ReplaceInstUsesWith(I, LHS);
4603 return InsertRangeTest(Val, LHSCst, AddOne(RHSCst, Context),
4605 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4607 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4608 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4609 return ReplaceInstUsesWith(I, RHS);
4610 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4614 case ICmpInst::ICMP_UGT:
4616 default: LLVM_UNREACHABLE("Unknown integer condition code!");
4617 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4618 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4619 return ReplaceInstUsesWith(I, LHS);
4620 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4622 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4623 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4624 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4625 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4629 case ICmpInst::ICMP_SGT:
4631 default: LLVM_UNREACHABLE("Unknown integer condition code!");
4632 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4633 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4634 return ReplaceInstUsesWith(I, LHS);
4635 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4637 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4638 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4639 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4640 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4648 /// FoldOrWithConstants - This helper function folds:
4650 /// ((A | B) & C1) | (B & C2)
4656 /// when the XOR of the two constants is "all ones" (-1).
4657 Instruction *InstCombiner::FoldOrWithConstants(BinaryOperator &I, Value *Op,
4658 Value *A, Value *B, Value *C) {
4659 ConstantInt *CI1 = dyn_cast<ConstantInt>(C);
4663 ConstantInt *CI2 = 0;
4664 if (!match(Op, m_And(m_Value(V1), m_ConstantInt(CI2)), *Context)) return 0;
4666 APInt Xor = CI1->getValue() ^ CI2->getValue();
4667 if (!Xor.isAllOnesValue()) return 0;
4669 if (V1 == A || V1 == B) {
4670 Instruction *NewOp =
4671 InsertNewInstBefore(BinaryOperator::CreateAnd((V1 == A) ? B : A, CI1), I);
4672 return BinaryOperator::CreateOr(NewOp, V1);
4678 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4679 bool Changed = SimplifyCommutative(I);
4680 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4682 if (isa<UndefValue>(Op1)) // X | undef -> -1
4683 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4687 return ReplaceInstUsesWith(I, Op0);
4689 // See if we can simplify any instructions used by the instruction whose sole
4690 // purpose is to compute bits we don't care about.
4691 if (SimplifyDemandedInstructionBits(I))
4693 if (isa<VectorType>(I.getType())) {
4694 if (isa<ConstantAggregateZero>(Op1)) {
4695 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4696 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4697 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4698 return ReplaceInstUsesWith(I, I.getOperand(1));
4703 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4704 ConstantInt *C1 = 0; Value *X = 0;
4705 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4706 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1)), *Context) &&
4708 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4709 InsertNewInstBefore(Or, I);
4711 return BinaryOperator::CreateAnd(Or,
4712 Context->getConstantInt(RHS->getValue() | C1->getValue()));
4715 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4716 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1)), *Context) &&
4718 Instruction *Or = BinaryOperator::CreateOr(X, RHS);
4719 InsertNewInstBefore(Or, I);
4721 return BinaryOperator::CreateXor(Or,
4722 Context->getConstantInt(C1->getValue() & ~RHS->getValue()));
4725 // Try to fold constant and into select arguments.
4726 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4727 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4729 if (isa<PHINode>(Op0))
4730 if (Instruction *NV = FoldOpIntoPhi(I))
4734 Value *A = 0, *B = 0;
4735 ConstantInt *C1 = 0, *C2 = 0;
4737 if (match(Op0, m_And(m_Value(A), m_Value(B)), *Context))
4738 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4739 return ReplaceInstUsesWith(I, Op1);
4740 if (match(Op1, m_And(m_Value(A), m_Value(B)), *Context))
4741 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4742 return ReplaceInstUsesWith(I, Op0);
4744 // (A | B) | C and A | (B | C) -> bswap if possible.
4745 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4746 if (match(Op0, m_Or(m_Value(), m_Value()), *Context) ||
4747 match(Op1, m_Or(m_Value(), m_Value()), *Context) ||
4748 (match(Op0, m_Shift(m_Value(), m_Value()), *Context) &&
4749 match(Op1, m_Shift(m_Value(), m_Value()), *Context))) {
4750 if (Instruction *BSwap = MatchBSwap(I))
4754 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4755 if (Op0->hasOneUse() &&
4756 match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4757 MaskedValueIsZero(Op1, C1->getValue())) {
4758 Instruction *NOr = BinaryOperator::CreateOr(A, Op1);
4759 InsertNewInstBefore(NOr, I);
4761 return BinaryOperator::CreateXor(NOr, C1);
4764 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4765 if (Op1->hasOneUse() &&
4766 match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1)), *Context) &&
4767 MaskedValueIsZero(Op0, C1->getValue())) {
4768 Instruction *NOr = BinaryOperator::CreateOr(A, Op0);
4769 InsertNewInstBefore(NOr, I);
4771 return BinaryOperator::CreateXor(NOr, C1);
4775 Value *C = 0, *D = 0;
4776 if (match(Op0, m_And(m_Value(A), m_Value(C)), *Context) &&
4777 match(Op1, m_And(m_Value(B), m_Value(D)), *Context)) {
4778 Value *V1 = 0, *V2 = 0, *V3 = 0;
4779 C1 = dyn_cast<ConstantInt>(C);
4780 C2 = dyn_cast<ConstantInt>(D);
4781 if (C1 && C2) { // (A & C1)|(B & C2)
4782 // If we have: ((V + N) & C1) | (V & C2)
4783 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4784 // replace with V+N.
4785 if (C1->getValue() == ~C2->getValue()) {
4786 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4787 match(A, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4788 // Add commutes, try both ways.
4789 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4790 return ReplaceInstUsesWith(I, A);
4791 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4792 return ReplaceInstUsesWith(I, A);
4794 // Or commutes, try both ways.
4795 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4796 match(B, m_Add(m_Value(V1), m_Value(V2)), *Context)) {
4797 // Add commutes, try both ways.
4798 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4799 return ReplaceInstUsesWith(I, B);
4800 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4801 return ReplaceInstUsesWith(I, B);
4804 V1 = 0; V2 = 0; V3 = 0;
4807 // Check to see if we have any common things being and'ed. If so, find the
4808 // terms for V1 & (V2|V3).
4809 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4810 if (A == B) // (A & C)|(A & D) == A & (C|D)
4811 V1 = A, V2 = C, V3 = D;
4812 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4813 V1 = A, V2 = B, V3 = C;
4814 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4815 V1 = C, V2 = A, V3 = D;
4816 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4817 V1 = C, V2 = A, V3 = B;
4821 InsertNewInstBefore(BinaryOperator::CreateOr(V2, V3, "tmp"), I);
4822 return BinaryOperator::CreateAnd(V1, Or);
4826 // (A & (C0?-1:0)) | (B & ~(C0?-1:0)) -> C0 ? A : B, and commuted variants
4827 if (Instruction *Match = MatchSelectFromAndOr(A, B, C, D, Context))
4829 if (Instruction *Match = MatchSelectFromAndOr(B, A, D, C, Context))
4831 if (Instruction *Match = MatchSelectFromAndOr(C, B, A, D, Context))
4833 if (Instruction *Match = MatchSelectFromAndOr(D, A, B, C, Context))
4836 // ((A&~B)|(~A&B)) -> A^B
4837 if ((match(C, m_Not(m_Specific(D)), *Context) &&
4838 match(B, m_Not(m_Specific(A)), *Context)))
4839 return BinaryOperator::CreateXor(A, D);
4840 // ((~B&A)|(~A&B)) -> A^B
4841 if ((match(A, m_Not(m_Specific(D)), *Context) &&
4842 match(B, m_Not(m_Specific(C)), *Context)))
4843 return BinaryOperator::CreateXor(C, D);
4844 // ((A&~B)|(B&~A)) -> A^B
4845 if ((match(C, m_Not(m_Specific(B)), *Context) &&
4846 match(D, m_Not(m_Specific(A)), *Context)))
4847 return BinaryOperator::CreateXor(A, B);
4848 // ((~B&A)|(B&~A)) -> A^B
4849 if ((match(A, m_Not(m_Specific(B)), *Context) &&
4850 match(D, m_Not(m_Specific(C)), *Context)))
4851 return BinaryOperator::CreateXor(C, B);
4854 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4855 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4856 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4857 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4858 SI0->getOperand(1) == SI1->getOperand(1) &&
4859 (SI0->hasOneUse() || SI1->hasOneUse())) {
4860 Instruction *NewOp =
4861 InsertNewInstBefore(BinaryOperator::CreateOr(SI0->getOperand(0),
4863 SI0->getName()), I);
4864 return BinaryOperator::Create(SI1->getOpcode(), NewOp,
4865 SI1->getOperand(1));
4869 // ((A|B)&1)|(B&-2) -> (A&1) | B
4870 if (match(Op0, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4871 match(Op0, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4872 Instruction *Ret = FoldOrWithConstants(I, Op1, A, B, C);
4873 if (Ret) return Ret;
4875 // (B&-2)|((A|B)&1) -> (A&1) | B
4876 if (match(Op1, m_And(m_Or(m_Value(A), m_Value(B)), m_Value(C)), *Context) ||
4877 match(Op1, m_And(m_Value(C), m_Or(m_Value(A), m_Value(B))), *Context)) {
4878 Instruction *Ret = FoldOrWithConstants(I, Op0, A, B, C);
4879 if (Ret) return Ret;
4882 if (match(Op0, m_Not(m_Value(A)), *Context)) { // ~A | Op1
4883 if (A == Op1) // ~A | A == -1
4884 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4888 // Note, A is still live here!
4889 if (match(Op1, m_Not(m_Value(B)), *Context)) { // Op0 | ~B
4891 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
4893 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4894 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4895 Value *And = InsertNewInstBefore(BinaryOperator::CreateAnd(A, B,
4896 I.getName()+".demorgan"), I);
4897 return BinaryOperator::CreateNot(And);
4901 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4902 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4903 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
4906 if (ICmpInst *LHS = dyn_cast<ICmpInst>(I.getOperand(0)))
4907 if (Instruction *Res = FoldOrOfICmps(I, LHS, RHS))
4911 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4912 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4913 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4914 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4915 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4916 !isa<ICmpInst>(Op1C->getOperand(0))) {
4917 const Type *SrcTy = Op0C->getOperand(0)->getType();
4918 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4919 // Only do this if the casts both really cause code to be
4921 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4923 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4925 Instruction *NewOp = BinaryOperator::CreateOr(Op0C->getOperand(0),
4926 Op1C->getOperand(0),
4928 InsertNewInstBefore(NewOp, I);
4929 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
4936 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4937 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4938 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4939 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4940 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4941 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType()) {
4942 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4943 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4944 // If either of the constants are nans, then the whole thing returns
4946 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4947 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4949 // Otherwise, no need to compare the two constants, compare the
4951 return new FCmpInst(*Context, FCmpInst::FCMP_UNO,
4952 LHS->getOperand(0), RHS->getOperand(0));
4955 Value *Op0LHS, *Op0RHS, *Op1LHS, *Op1RHS;
4956 FCmpInst::Predicate Op0CC, Op1CC;
4957 if (match(Op0, m_FCmp(Op0CC, m_Value(Op0LHS),
4958 m_Value(Op0RHS)), *Context) &&
4959 match(Op1, m_FCmp(Op1CC, m_Value(Op1LHS),
4960 m_Value(Op1RHS)), *Context)) {
4961 if (Op0LHS == Op1RHS && Op0RHS == Op1LHS) {
4962 // Swap RHS operands to match LHS.
4963 Op1CC = FCmpInst::getSwappedPredicate(Op1CC);
4964 std::swap(Op1LHS, Op1RHS);
4966 if (Op0LHS == Op1LHS && Op0RHS == Op1RHS) {
4967 // Simplify (fcmp cc0 x, y) | (fcmp cc1 x, y).
4969 return new FCmpInst(*Context, (FCmpInst::Predicate)Op0CC,
4971 else if (Op0CC == FCmpInst::FCMP_TRUE ||
4972 Op1CC == FCmpInst::FCMP_TRUE)
4973 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
4974 else if (Op0CC == FCmpInst::FCMP_FALSE)
4975 return ReplaceInstUsesWith(I, Op1);
4976 else if (Op1CC == FCmpInst::FCMP_FALSE)
4977 return ReplaceInstUsesWith(I, Op0);
4980 unsigned Op0Pred = getFCmpCode(Op0CC, Op0Ordered);
4981 unsigned Op1Pred = getFCmpCode(Op1CC, Op1Ordered);
4982 if (Op0Ordered == Op1Ordered) {
4983 // If both are ordered or unordered, return a new fcmp with
4984 // or'ed predicates.
4985 Value *RV = getFCmpValue(Op0Ordered, Op0Pred|Op1Pred,
4986 Op0LHS, Op0RHS, Context);
4987 if (Instruction *I = dyn_cast<Instruction>(RV))
4989 // Otherwise, it's a constant boolean value...
4990 return ReplaceInstUsesWith(I, RV);
4998 return Changed ? &I : 0;
5003 // XorSelf - Implements: X ^ X --> 0
5006 XorSelf(Value *rhs) : RHS(rhs) {}
5007 bool shouldApply(Value *LHS) const { return LHS == RHS; }
5008 Instruction *apply(BinaryOperator &Xor) const {
5015 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
5016 bool Changed = SimplifyCommutative(I);
5017 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5019 if (isa<UndefValue>(Op1)) {
5020 if (isa<UndefValue>(Op0))
5021 // Handle undef ^ undef -> 0 special case. This is a common
5023 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
5024 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
5027 // xor X, X = 0, even if X is nested in a sequence of Xor's.
5028 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1), Context)) {
5029 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
5030 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
5033 // See if we can simplify any instructions used by the instruction whose sole
5034 // purpose is to compute bits we don't care about.
5035 if (SimplifyDemandedInstructionBits(I))
5037 if (isa<VectorType>(I.getType()))
5038 if (isa<ConstantAggregateZero>(Op1))
5039 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
5041 // Is this a ~ operation?
5042 if (Value *NotOp = dyn_castNotVal(&I, Context)) {
5043 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
5044 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
5045 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
5046 if (Op0I->getOpcode() == Instruction::And ||
5047 Op0I->getOpcode() == Instruction::Or) {
5048 if (dyn_castNotVal(Op0I->getOperand(1), Context)) Op0I->swapOperands();
5049 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0), Context)) {
5051 BinaryOperator::CreateNot(Op0I->getOperand(1),
5052 Op0I->getOperand(1)->getName()+".not");
5053 InsertNewInstBefore(NotY, I);
5054 if (Op0I->getOpcode() == Instruction::And)
5055 return BinaryOperator::CreateOr(Op0NotVal, NotY);
5057 return BinaryOperator::CreateAnd(Op0NotVal, NotY);
5064 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
5065 if (RHS == Context->getConstantIntTrue() && Op0->hasOneUse()) {
5066 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
5067 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
5068 return new ICmpInst(*Context, ICI->getInversePredicate(),
5069 ICI->getOperand(0), ICI->getOperand(1));
5071 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
5072 return new FCmpInst(*Context, FCI->getInversePredicate(),
5073 FCI->getOperand(0), FCI->getOperand(1));
5076 // fold (xor(zext(cmp)), 1) and (xor(sext(cmp)), -1) to ext(!cmp).
5077 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5078 if (CmpInst *CI = dyn_cast<CmpInst>(Op0C->getOperand(0))) {
5079 if (CI->hasOneUse() && Op0C->hasOneUse()) {
5080 Instruction::CastOps Opcode = Op0C->getOpcode();
5081 if (Opcode == Instruction::ZExt || Opcode == Instruction::SExt) {
5082 if (RHS == Context->getConstantExprCast(Opcode,
5083 Context->getConstantIntTrue(),
5084 Op0C->getDestTy())) {
5085 Instruction *NewCI = InsertNewInstBefore(CmpInst::Create(
5087 CI->getOpcode(), CI->getInversePredicate(),
5088 CI->getOperand(0), CI->getOperand(1)), I);
5089 NewCI->takeName(CI);
5090 return CastInst::Create(Opcode, NewCI, Op0C->getType());
5097 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
5098 // ~(c-X) == X-c-1 == X+(-c-1)
5099 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
5100 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
5101 Constant *NegOp0I0C = Context->getConstantExprNeg(Op0I0C);
5102 Constant *ConstantRHS = Context->getConstantExprSub(NegOp0I0C,
5103 Context->getConstantInt(I.getType(), 1));
5104 return BinaryOperator::CreateAdd(Op0I->getOperand(1), ConstantRHS);
5107 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
5108 if (Op0I->getOpcode() == Instruction::Add) {
5109 // ~(X-c) --> (-c-1)-X
5110 if (RHS->isAllOnesValue()) {
5111 Constant *NegOp0CI = Context->getConstantExprNeg(Op0CI);
5112 return BinaryOperator::CreateSub(
5113 Context->getConstantExprSub(NegOp0CI,
5114 Context->getConstantInt(I.getType(), 1)),
5115 Op0I->getOperand(0));
5116 } else if (RHS->getValue().isSignBit()) {
5117 // (X + C) ^ signbit -> (X + C + signbit)
5119 Context->getConstantInt(RHS->getValue() + Op0CI->getValue());
5120 return BinaryOperator::CreateAdd(Op0I->getOperand(0), C);
5123 } else if (Op0I->getOpcode() == Instruction::Or) {
5124 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
5125 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
5126 Constant *NewRHS = Context->getConstantExprOr(Op0CI, RHS);
5127 // Anything in both C1 and C2 is known to be zero, remove it from
5129 Constant *CommonBits = Context->getConstantExprAnd(Op0CI, RHS);
5130 NewRHS = Context->getConstantExprAnd(NewRHS,
5131 Context->getConstantExprNot(CommonBits));
5132 AddToWorkList(Op0I);
5133 I.setOperand(0, Op0I->getOperand(0));
5134 I.setOperand(1, NewRHS);
5141 // Try to fold constant and into select arguments.
5142 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5143 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5145 if (isa<PHINode>(Op0))
5146 if (Instruction *NV = FoldOpIntoPhi(I))
5150 if (Value *X = dyn_castNotVal(Op0, Context)) // ~A ^ A == -1
5152 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
5154 if (Value *X = dyn_castNotVal(Op1, Context)) // A ^ ~A == -1
5156 return ReplaceInstUsesWith(I, Context->getAllOnesValue(I.getType()));
5159 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
5162 if (match(Op1I, m_Or(m_Value(A), m_Value(B)), *Context)) {
5163 if (A == Op0) { // B^(B|A) == (A|B)^B
5164 Op1I->swapOperands();
5166 std::swap(Op0, Op1);
5167 } else if (B == Op0) { // B^(A|B) == (A|B)^B
5168 I.swapOperands(); // Simplified below.
5169 std::swap(Op0, Op1);
5171 } else if (match(Op1I, m_Xor(m_Specific(Op0), m_Value(B)), *Context)) {
5172 return ReplaceInstUsesWith(I, B); // A^(A^B) == B
5173 } else if (match(Op1I, m_Xor(m_Value(A), m_Specific(Op0)), *Context)) {
5174 return ReplaceInstUsesWith(I, A); // A^(B^A) == B
5175 } else if (match(Op1I, m_And(m_Value(A), m_Value(B)), *Context) &&
5177 if (A == Op0) { // A^(A&B) -> A^(B&A)
5178 Op1I->swapOperands();
5181 if (B == Op0) { // A^(B&A) -> (B&A)^A
5182 I.swapOperands(); // Simplified below.
5183 std::swap(Op0, Op1);
5188 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
5191 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5192 Op0I->hasOneUse()) {
5193 if (A == Op1) // (B|A)^B == (A|B)^B
5195 if (B == Op1) { // (A|B)^B == A & ~B
5197 InsertNewInstBefore(BinaryOperator::CreateNot(Op1, "tmp"), I);
5198 return BinaryOperator::CreateAnd(A, NotB);
5200 } else if (match(Op0I, m_Xor(m_Specific(Op1), m_Value(B)), *Context)) {
5201 return ReplaceInstUsesWith(I, B); // (A^B)^A == B
5202 } else if (match(Op0I, m_Xor(m_Value(A), m_Specific(Op1)), *Context)) {
5203 return ReplaceInstUsesWith(I, A); // (B^A)^A == B
5204 } else if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5206 if (A == Op1) // (A&B)^A -> (B&A)^A
5208 if (B == Op1 && // (B&A)^A == ~B & A
5209 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
5211 InsertNewInstBefore(BinaryOperator::CreateNot(A, "tmp"), I);
5212 return BinaryOperator::CreateAnd(N, Op1);
5217 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
5218 if (Op0I && Op1I && Op0I->isShift() &&
5219 Op0I->getOpcode() == Op1I->getOpcode() &&
5220 Op0I->getOperand(1) == Op1I->getOperand(1) &&
5221 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
5222 Instruction *NewOp =
5223 InsertNewInstBefore(BinaryOperator::CreateXor(Op0I->getOperand(0),
5224 Op1I->getOperand(0),
5225 Op0I->getName()), I);
5226 return BinaryOperator::Create(Op1I->getOpcode(), NewOp,
5227 Op1I->getOperand(1));
5231 Value *A, *B, *C, *D;
5232 // (A & B)^(A | B) -> A ^ B
5233 if (match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5234 match(Op1I, m_Or(m_Value(C), m_Value(D)), *Context)) {
5235 if ((A == C && B == D) || (A == D && B == C))
5236 return BinaryOperator::CreateXor(A, B);
5238 // (A | B)^(A & B) -> A ^ B
5239 if (match(Op0I, m_Or(m_Value(A), m_Value(B)), *Context) &&
5240 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5241 if ((A == C && B == D) || (A == D && B == C))
5242 return BinaryOperator::CreateXor(A, B);
5246 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
5247 match(Op0I, m_And(m_Value(A), m_Value(B)), *Context) &&
5248 match(Op1I, m_And(m_Value(C), m_Value(D)), *Context)) {
5249 // (X & Y)^(X & Y) -> (Y^Z) & X
5250 Value *X = 0, *Y = 0, *Z = 0;
5252 X = A, Y = B, Z = D;
5254 X = A, Y = B, Z = C;
5256 X = B, Y = A, Z = D;
5258 X = B, Y = A, Z = C;
5261 Instruction *NewOp =
5262 InsertNewInstBefore(BinaryOperator::CreateXor(Y, Z, Op0->getName()), I);
5263 return BinaryOperator::CreateAnd(NewOp, X);
5268 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
5269 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
5270 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS),Context))
5273 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
5274 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
5275 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
5276 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
5277 const Type *SrcTy = Op0C->getOperand(0)->getType();
5278 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
5279 // Only do this if the casts both really cause code to be generated.
5280 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
5282 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
5284 Instruction *NewOp = BinaryOperator::CreateXor(Op0C->getOperand(0),
5285 Op1C->getOperand(0),
5287 InsertNewInstBefore(NewOp, I);
5288 return CastInst::Create(Op0C->getOpcode(), NewOp, I.getType());
5293 return Changed ? &I : 0;
5296 static ConstantInt *ExtractElement(Constant *V, Constant *Idx,
5297 LLVMContext *Context) {
5298 return cast<ConstantInt>(Context->getConstantExprExtractElement(V, Idx));
5301 static bool HasAddOverflow(ConstantInt *Result,
5302 ConstantInt *In1, ConstantInt *In2,
5305 if (In2->getValue().isNegative())
5306 return Result->getValue().sgt(In1->getValue());
5308 return Result->getValue().slt(In1->getValue());
5310 return Result->getValue().ult(In1->getValue());
5313 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
5314 /// overflowed for this type.
5315 static bool AddWithOverflow(Constant *&Result, Constant *In1,
5316 Constant *In2, LLVMContext *Context,
5317 bool IsSigned = false) {
5318 Result = Context->getConstantExprAdd(In1, In2);
5320 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5321 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5322 Constant *Idx = Context->getConstantInt(Type::Int32Ty, i);
5323 if (HasAddOverflow(ExtractElement(Result, Idx, Context),
5324 ExtractElement(In1, Idx, Context),
5325 ExtractElement(In2, Idx, Context),
5332 return HasAddOverflow(cast<ConstantInt>(Result),
5333 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5337 static bool HasSubOverflow(ConstantInt *Result,
5338 ConstantInt *In1, ConstantInt *In2,
5341 if (In2->getValue().isNegative())
5342 return Result->getValue().slt(In1->getValue());
5344 return Result->getValue().sgt(In1->getValue());
5346 return Result->getValue().ugt(In1->getValue());
5349 /// SubWithOverflow - Compute Result = In1-In2, returning true if the result
5350 /// overflowed for this type.
5351 static bool SubWithOverflow(Constant *&Result, Constant *In1,
5352 Constant *In2, LLVMContext *Context,
5353 bool IsSigned = false) {
5354 Result = Context->getConstantExprSub(In1, In2);
5356 if (const VectorType *VTy = dyn_cast<VectorType>(In1->getType())) {
5357 for (unsigned i = 0, e = VTy->getNumElements(); i != e; ++i) {
5358 Constant *Idx = Context->getConstantInt(Type::Int32Ty, i);
5359 if (HasSubOverflow(ExtractElement(Result, Idx, Context),
5360 ExtractElement(In1, Idx, Context),
5361 ExtractElement(In2, Idx, Context),
5368 return HasSubOverflow(cast<ConstantInt>(Result),
5369 cast<ConstantInt>(In1), cast<ConstantInt>(In2),
5373 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
5374 /// code necessary to compute the offset from the base pointer (without adding
5375 /// in the base pointer). Return the result as a signed integer of intptr size.
5376 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
5377 TargetData &TD = IC.getTargetData();
5378 gep_type_iterator GTI = gep_type_begin(GEP);
5379 const Type *IntPtrTy = TD.getIntPtrType();
5380 LLVMContext *Context = IC.getContext();
5381 Value *Result = Context->getNullValue(IntPtrTy);
5383 // Build a mask for high order bits.
5384 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5385 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5387 for (User::op_iterator i = GEP->op_begin() + 1, e = GEP->op_end(); i != e;
5390 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType()) & PtrSizeMask;
5391 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
5392 if (OpC->isZero()) continue;
5394 // Handle a struct index, which adds its field offset to the pointer.
5395 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5396 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
5398 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
5400 Context->getConstantInt(RC->getValue() + APInt(IntPtrWidth, Size));
5402 Result = IC.InsertNewInstBefore(
5403 BinaryOperator::CreateAdd(Result,
5404 Context->getConstantInt(IntPtrTy, Size),
5405 GEP->getName()+".offs"), I);
5409 Constant *Scale = Context->getConstantInt(IntPtrTy, Size);
5411 Context->getConstantExprIntegerCast(OpC, IntPtrTy, true /*SExt*/);
5412 Scale = Context->getConstantExprMul(OC, Scale);
5413 if (Constant *RC = dyn_cast<Constant>(Result))
5414 Result = Context->getConstantExprAdd(RC, Scale);
5416 // Emit an add instruction.
5417 Result = IC.InsertNewInstBefore(
5418 BinaryOperator::CreateAdd(Result, Scale,
5419 GEP->getName()+".offs"), I);
5423 // Convert to correct type.
5424 if (Op->getType() != IntPtrTy) {
5425 if (Constant *OpC = dyn_cast<Constant>(Op))
5426 Op = Context->getConstantExprIntegerCast(OpC, IntPtrTy, true);
5428 Op = IC.InsertNewInstBefore(CastInst::CreateIntegerCast(Op, IntPtrTy,
5430 Op->getName()+".c"), I);
5433 Constant *Scale = Context->getConstantInt(IntPtrTy, Size);
5434 if (Constant *OpC = dyn_cast<Constant>(Op))
5435 Op = Context->getConstantExprMul(OpC, Scale);
5436 else // We'll let instcombine(mul) convert this to a shl if possible.
5437 Op = IC.InsertNewInstBefore(BinaryOperator::CreateMul(Op, Scale,
5438 GEP->getName()+".idx"), I);
5441 // Emit an add instruction.
5442 if (isa<Constant>(Op) && isa<Constant>(Result))
5443 Result = Context->getConstantExprAdd(cast<Constant>(Op),
5444 cast<Constant>(Result));
5446 Result = IC.InsertNewInstBefore(BinaryOperator::CreateAdd(Op, Result,
5447 GEP->getName()+".offs"), I);
5453 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5454 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5455 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5456 /// complex, and scales are involved. The above expression would also be legal
5457 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5458 /// later form is less amenable to optimization though, and we are allowed to
5459 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5461 /// If we can't emit an optimized form for this expression, this returns null.
5463 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5465 TargetData &TD = IC.getTargetData();
5466 gep_type_iterator GTI = gep_type_begin(GEP);
5468 // Check to see if this gep only has a single variable index. If so, and if
5469 // any constant indices are a multiple of its scale, then we can compute this
5470 // in terms of the scale of the variable index. For example, if the GEP
5471 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5472 // because the expression will cross zero at the same point.
5473 unsigned i, e = GEP->getNumOperands();
5475 for (i = 1; i != e; ++i, ++GTI) {
5476 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5477 // Compute the aggregate offset of constant indices.
5478 if (CI->isZero()) continue;
5480 // Handle a struct index, which adds its field offset to the pointer.
5481 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5482 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5484 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5485 Offset += Size*CI->getSExtValue();
5488 // Found our variable index.
5493 // If there are no variable indices, we must have a constant offset, just
5494 // evaluate it the general way.
5495 if (i == e) return 0;
5497 Value *VariableIdx = GEP->getOperand(i);
5498 // Determine the scale factor of the variable element. For example, this is
5499 // 4 if the variable index is into an array of i32.
5500 uint64_t VariableScale = TD.getTypeAllocSize(GTI.getIndexedType());
5502 // Verify that there are no other variable indices. If so, emit the hard way.
5503 for (++i, ++GTI; i != e; ++i, ++GTI) {
5504 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5507 // Compute the aggregate offset of constant indices.
5508 if (CI->isZero()) continue;
5510 // Handle a struct index, which adds its field offset to the pointer.
5511 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5512 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5514 uint64_t Size = TD.getTypeAllocSize(GTI.getIndexedType());
5515 Offset += Size*CI->getSExtValue();
5519 // Okay, we know we have a single variable index, which must be a
5520 // pointer/array/vector index. If there is no offset, life is simple, return
5522 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5524 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5525 // we don't need to bother extending: the extension won't affect where the
5526 // computation crosses zero.
5527 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5528 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5529 VariableIdx->getNameStart(), &I);
5533 // Otherwise, there is an index. The computation we will do will be modulo
5534 // the pointer size, so get it.
5535 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5537 Offset &= PtrSizeMask;
5538 VariableScale &= PtrSizeMask;
5540 // To do this transformation, any constant index must be a multiple of the
5541 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5542 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5543 // multiple of the variable scale.
5544 int64_t NewOffs = Offset / (int64_t)VariableScale;
5545 if (Offset != NewOffs*(int64_t)VariableScale)
5548 // Okay, we can do this evaluation. Start by converting the index to intptr.
5549 const Type *IntPtrTy = TD.getIntPtrType();
5550 if (VariableIdx->getType() != IntPtrTy)
5551 VariableIdx = CastInst::CreateIntegerCast(VariableIdx, IntPtrTy,
5553 VariableIdx->getNameStart(), &I);
5554 Constant *OffsetVal = IC.getContext()->getConstantInt(IntPtrTy, NewOffs);
5555 return BinaryOperator::CreateAdd(VariableIdx, OffsetVal, "offset", &I);
5559 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5560 /// else. At this point we know that the GEP is on the LHS of the comparison.
5561 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5562 ICmpInst::Predicate Cond,
5564 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5566 // Look through bitcasts.
5567 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5568 RHS = BCI->getOperand(0);
5570 Value *PtrBase = GEPLHS->getOperand(0);
5571 if (PtrBase == RHS) {
5572 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5573 // This transformation (ignoring the base and scales) is valid because we
5574 // know pointers can't overflow. See if we can output an optimized form.
5575 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5577 // If not, synthesize the offset the hard way.
5579 Offset = EmitGEPOffset(GEPLHS, I, *this);
5580 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), Offset,
5581 Context->getNullValue(Offset->getType()));
5582 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5583 // If the base pointers are different, but the indices are the same, just
5584 // compare the base pointer.
5585 if (PtrBase != GEPRHS->getOperand(0)) {
5586 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5587 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5588 GEPRHS->getOperand(0)->getType();
5590 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5591 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5592 IndicesTheSame = false;
5596 // If all indices are the same, just compare the base pointers.
5598 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond),
5599 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5601 // Otherwise, the base pointers are different and the indices are
5602 // different, bail out.
5606 // If one of the GEPs has all zero indices, recurse.
5607 bool AllZeros = true;
5608 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5609 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5610 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5615 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5616 ICmpInst::getSwappedPredicate(Cond), I);
5618 // If the other GEP has all zero indices, recurse.
5620 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5621 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5622 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5627 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5629 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5630 // If the GEPs only differ by one index, compare it.
5631 unsigned NumDifferences = 0; // Keep track of # differences.
5632 unsigned DiffOperand = 0; // The operand that differs.
5633 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5634 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5635 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5636 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5637 // Irreconcilable differences.
5641 if (NumDifferences++) break;
5646 if (NumDifferences == 0) // SAME GEP?
5647 return ReplaceInstUsesWith(I, // No comparison is needed here.
5648 Context->getConstantInt(Type::Int1Ty,
5649 ICmpInst::isTrueWhenEqual(Cond)));
5651 else if (NumDifferences == 1) {
5652 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5653 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5654 // Make sure we do a signed comparison here.
5655 return new ICmpInst(*Context,
5656 ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5660 // Only lower this if the icmp is the only user of the GEP or if we expect
5661 // the result to fold to a constant!
5662 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5663 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5664 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5665 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5666 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5667 return new ICmpInst(*Context, ICmpInst::getSignedPredicate(Cond), L, R);
5673 /// FoldFCmp_IntToFP_Cst - Fold fcmp ([us]itofp x, cst) if possible.
5675 Instruction *InstCombiner::FoldFCmp_IntToFP_Cst(FCmpInst &I,
5678 if (!isa<ConstantFP>(RHSC)) return 0;
5679 const APFloat &RHS = cast<ConstantFP>(RHSC)->getValueAPF();
5681 // Get the width of the mantissa. We don't want to hack on conversions that
5682 // might lose information from the integer, e.g. "i64 -> float"
5683 int MantissaWidth = LHSI->getType()->getFPMantissaWidth();
5684 if (MantissaWidth == -1) return 0; // Unknown.
5686 // Check to see that the input is converted from an integer type that is small
5687 // enough that preserves all bits. TODO: check here for "known" sign bits.
5688 // This would allow us to handle (fptosi (x >>s 62) to float) if x is i64 f.e.
5689 unsigned InputSize = LHSI->getOperand(0)->getType()->getScalarSizeInBits();
5691 // If this is a uitofp instruction, we need an extra bit to hold the sign.
5692 bool LHSUnsigned = isa<UIToFPInst>(LHSI);
5696 // If the conversion would lose info, don't hack on this.
5697 if ((int)InputSize > MantissaWidth)
5700 // Otherwise, we can potentially simplify the comparison. We know that it
5701 // will always come through as an integer value and we know the constant is
5702 // not a NAN (it would have been previously simplified).
5703 assert(!RHS.isNaN() && "NaN comparison not already folded!");
5705 ICmpInst::Predicate Pred;
5706 switch (I.getPredicate()) {
5707 default: LLVM_UNREACHABLE("Unexpected predicate!");
5708 case FCmpInst::FCMP_UEQ:
5709 case FCmpInst::FCMP_OEQ:
5710 Pred = ICmpInst::ICMP_EQ;
5712 case FCmpInst::FCMP_UGT:
5713 case FCmpInst::FCMP_OGT:
5714 Pred = LHSUnsigned ? ICmpInst::ICMP_UGT : ICmpInst::ICMP_SGT;
5716 case FCmpInst::FCMP_UGE:
5717 case FCmpInst::FCMP_OGE:
5718 Pred = LHSUnsigned ? ICmpInst::ICMP_UGE : ICmpInst::ICMP_SGE;
5720 case FCmpInst::FCMP_ULT:
5721 case FCmpInst::FCMP_OLT:
5722 Pred = LHSUnsigned ? ICmpInst::ICMP_ULT : ICmpInst::ICMP_SLT;
5724 case FCmpInst::FCMP_ULE:
5725 case FCmpInst::FCMP_OLE:
5726 Pred = LHSUnsigned ? ICmpInst::ICMP_ULE : ICmpInst::ICMP_SLE;
5728 case FCmpInst::FCMP_UNE:
5729 case FCmpInst::FCMP_ONE:
5730 Pred = ICmpInst::ICMP_NE;
5732 case FCmpInst::FCMP_ORD:
5733 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5734 case FCmpInst::FCMP_UNO:
5735 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5738 const IntegerType *IntTy = cast<IntegerType>(LHSI->getOperand(0)->getType());
5740 // Now we know that the APFloat is a normal number, zero or inf.
5742 // See if the FP constant is too large for the integer. For example,
5743 // comparing an i8 to 300.0.
5744 unsigned IntWidth = IntTy->getScalarSizeInBits();
5747 // If the RHS value is > SignedMax, fold the comparison. This handles +INF
5748 // and large values.
5749 APFloat SMax(RHS.getSemantics(), APFloat::fcZero, false);
5750 SMax.convertFromAPInt(APInt::getSignedMaxValue(IntWidth), true,
5751 APFloat::rmNearestTiesToEven);
5752 if (SMax.compare(RHS) == APFloat::cmpLessThan) { // smax < 13123.0
5753 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SLT ||
5754 Pred == ICmpInst::ICMP_SLE)
5755 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5756 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5759 // If the RHS value is > UnsignedMax, fold the comparison. This handles
5760 // +INF and large values.
5761 APFloat UMax(RHS.getSemantics(), APFloat::fcZero, false);
5762 UMax.convertFromAPInt(APInt::getMaxValue(IntWidth), false,
5763 APFloat::rmNearestTiesToEven);
5764 if (UMax.compare(RHS) == APFloat::cmpLessThan) { // umax < 13123.0
5765 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_ULT ||
5766 Pred == ICmpInst::ICMP_ULE)
5767 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5768 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5773 // See if the RHS value is < SignedMin.
5774 APFloat SMin(RHS.getSemantics(), APFloat::fcZero, false);
5775 SMin.convertFromAPInt(APInt::getSignedMinValue(IntWidth), true,
5776 APFloat::rmNearestTiesToEven);
5777 if (SMin.compare(RHS) == APFloat::cmpGreaterThan) { // smin > 12312.0
5778 if (Pred == ICmpInst::ICMP_NE || Pred == ICmpInst::ICMP_SGT ||
5779 Pred == ICmpInst::ICMP_SGE)
5780 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5781 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5785 // Okay, now we know that the FP constant fits in the range [SMIN, SMAX] or
5786 // [0, UMAX], but it may still be fractional. See if it is fractional by
5787 // casting the FP value to the integer value and back, checking for equality.
5788 // Don't do this for zero, because -0.0 is not fractional.
5789 Constant *RHSInt = LHSUnsigned
5790 ? Context->getConstantExprFPToUI(RHSC, IntTy)
5791 : Context->getConstantExprFPToSI(RHSC, IntTy);
5792 if (!RHS.isZero()) {
5793 bool Equal = LHSUnsigned
5794 ? Context->getConstantExprUIToFP(RHSInt, RHSC->getType()) == RHSC
5795 : Context->getConstantExprSIToFP(RHSInt, RHSC->getType()) == RHSC;
5797 // If we had a comparison against a fractional value, we have to adjust
5798 // the compare predicate and sometimes the value. RHSC is rounded towards
5799 // zero at this point.
5801 default: LLVM_UNREACHABLE("Unexpected integer comparison!");
5802 case ICmpInst::ICMP_NE: // (float)int != 4.4 --> true
5803 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5804 case ICmpInst::ICMP_EQ: // (float)int == 4.4 --> false
5805 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5806 case ICmpInst::ICMP_ULE:
5807 // (float)int <= 4.4 --> int <= 4
5808 // (float)int <= -4.4 --> false
5809 if (RHS.isNegative())
5810 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5812 case ICmpInst::ICMP_SLE:
5813 // (float)int <= 4.4 --> int <= 4
5814 // (float)int <= -4.4 --> int < -4
5815 if (RHS.isNegative())
5816 Pred = ICmpInst::ICMP_SLT;
5818 case ICmpInst::ICMP_ULT:
5819 // (float)int < -4.4 --> false
5820 // (float)int < 4.4 --> int <= 4
5821 if (RHS.isNegative())
5822 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5823 Pred = ICmpInst::ICMP_ULE;
5825 case ICmpInst::ICMP_SLT:
5826 // (float)int < -4.4 --> int < -4
5827 // (float)int < 4.4 --> int <= 4
5828 if (!RHS.isNegative())
5829 Pred = ICmpInst::ICMP_SLE;
5831 case ICmpInst::ICMP_UGT:
5832 // (float)int > 4.4 --> int > 4
5833 // (float)int > -4.4 --> true
5834 if (RHS.isNegative())
5835 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5837 case ICmpInst::ICMP_SGT:
5838 // (float)int > 4.4 --> int > 4
5839 // (float)int > -4.4 --> int >= -4
5840 if (RHS.isNegative())
5841 Pred = ICmpInst::ICMP_SGE;
5843 case ICmpInst::ICMP_UGE:
5844 // (float)int >= -4.4 --> true
5845 // (float)int >= 4.4 --> int > 4
5846 if (!RHS.isNegative())
5847 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5848 Pred = ICmpInst::ICMP_UGT;
5850 case ICmpInst::ICMP_SGE:
5851 // (float)int >= -4.4 --> int >= -4
5852 // (float)int >= 4.4 --> int > 4
5853 if (!RHS.isNegative())
5854 Pred = ICmpInst::ICMP_SGT;
5860 // Lower this FP comparison into an appropriate integer version of the
5862 return new ICmpInst(*Context, Pred, LHSI->getOperand(0), RHSInt);
5865 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5866 bool Changed = SimplifyCompare(I);
5867 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5869 // Fold trivial predicates.
5870 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5871 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5872 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5873 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5875 // Simplify 'fcmp pred X, X'
5877 switch (I.getPredicate()) {
5878 default: LLVM_UNREACHABLE("Unknown predicate!");
5879 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5880 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5881 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5882 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5883 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5884 case FCmpInst::FCMP_OLT: // True if ordered and less than
5885 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5886 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5888 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5889 case FCmpInst::FCMP_ULT: // True if unordered or less than
5890 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5891 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5892 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5893 I.setPredicate(FCmpInst::FCMP_UNO);
5894 I.setOperand(1, Context->getNullValue(Op0->getType()));
5897 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5898 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5899 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5900 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5901 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5902 I.setPredicate(FCmpInst::FCMP_ORD);
5903 I.setOperand(1, Context->getNullValue(Op0->getType()));
5908 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5909 return ReplaceInstUsesWith(I, Context->getUndef(Type::Int1Ty));
5911 // Handle fcmp with constant RHS
5912 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5913 // If the constant is a nan, see if we can fold the comparison based on it.
5914 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
5915 if (CFP->getValueAPF().isNaN()) {
5916 if (FCmpInst::isOrdered(I.getPredicate())) // True if ordered and...
5917 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
5918 assert(FCmpInst::isUnordered(I.getPredicate()) &&
5919 "Comparison must be either ordered or unordered!");
5920 // True if unordered.
5921 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
5925 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5926 switch (LHSI->getOpcode()) {
5927 case Instruction::PHI:
5928 // Only fold fcmp into the PHI if the phi and fcmp are in the same
5929 // block. If in the same block, we're encouraging jump threading. If
5930 // not, we are just pessimizing the code by making an i1 phi.
5931 if (LHSI->getParent() == I.getParent())
5932 if (Instruction *NV = FoldOpIntoPhi(I))
5935 case Instruction::SIToFP:
5936 case Instruction::UIToFP:
5937 if (Instruction *NV = FoldFCmp_IntToFP_Cst(I, LHSI, RHSC))
5940 case Instruction::Select:
5941 // If either operand of the select is a constant, we can fold the
5942 // comparison into the select arms, which will cause one to be
5943 // constant folded and the select turned into a bitwise or.
5944 Value *Op1 = 0, *Op2 = 0;
5945 if (LHSI->hasOneUse()) {
5946 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5947 // Fold the known value into the constant operand.
5948 Op1 = Context->getConstantExprCompare(I.getPredicate(), C, RHSC);
5949 // Insert a new FCmp of the other select operand.
5950 Op2 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5951 LHSI->getOperand(2), RHSC,
5953 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5954 // Fold the known value into the constant operand.
5955 Op2 = Context->getConstantExprCompare(I.getPredicate(), C, RHSC);
5956 // Insert a new FCmp of the other select operand.
5957 Op1 = InsertNewInstBefore(new FCmpInst(*Context, I.getPredicate(),
5958 LHSI->getOperand(1), RHSC,
5964 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5969 return Changed ? &I : 0;
5972 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5973 bool Changed = SimplifyCompare(I);
5974 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5975 const Type *Ty = Op0->getType();
5979 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
5980 I.isTrueWhenEqual()));
5982 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5983 return ReplaceInstUsesWith(I, Context->getUndef(Type::Int1Ty));
5985 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5986 // addresses never equal each other! We already know that Op0 != Op1.
5987 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5988 isa<ConstantPointerNull>(Op0)) &&
5989 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5990 isa<ConstantPointerNull>(Op1)))
5991 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
5992 !I.isTrueWhenEqual()));
5994 // icmp's with boolean values can always be turned into bitwise operations
5995 if (Ty == Type::Int1Ty) {
5996 switch (I.getPredicate()) {
5997 default: LLVM_UNREACHABLE("Invalid icmp instruction!");
5998 case ICmpInst::ICMP_EQ: { // icmp eq i1 A, B -> ~(A^B)
5999 Instruction *Xor = BinaryOperator::CreateXor(Op0, Op1, I.getName()+"tmp");
6000 InsertNewInstBefore(Xor, I);
6001 return BinaryOperator::CreateNot(Xor);
6003 case ICmpInst::ICMP_NE: // icmp eq i1 A, B -> A^B
6004 return BinaryOperator::CreateXor(Op0, Op1);
6006 case ICmpInst::ICMP_UGT:
6007 std::swap(Op0, Op1); // Change icmp ugt -> icmp ult
6009 case ICmpInst::ICMP_ULT:{ // icmp ult i1 A, B -> ~A & B
6010 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
6011 InsertNewInstBefore(Not, I);
6012 return BinaryOperator::CreateAnd(Not, Op1);
6014 case ICmpInst::ICMP_SGT:
6015 std::swap(Op0, Op1); // Change icmp sgt -> icmp slt
6017 case ICmpInst::ICMP_SLT: { // icmp slt i1 A, B -> A & ~B
6018 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
6019 InsertNewInstBefore(Not, I);
6020 return BinaryOperator::CreateAnd(Not, Op0);
6022 case ICmpInst::ICMP_UGE:
6023 std::swap(Op0, Op1); // Change icmp uge -> icmp ule
6025 case ICmpInst::ICMP_ULE: { // icmp ule i1 A, B -> ~A | B
6026 Instruction *Not = BinaryOperator::CreateNot(Op0, I.getName()+"tmp");
6027 InsertNewInstBefore(Not, I);
6028 return BinaryOperator::CreateOr(Not, Op1);
6030 case ICmpInst::ICMP_SGE:
6031 std::swap(Op0, Op1); // Change icmp sge -> icmp sle
6033 case ICmpInst::ICMP_SLE: { // icmp sle i1 A, B -> A | ~B
6034 Instruction *Not = BinaryOperator::CreateNot(Op1, I.getName()+"tmp");
6035 InsertNewInstBefore(Not, I);
6036 return BinaryOperator::CreateOr(Not, Op0);
6041 unsigned BitWidth = 0;
6043 BitWidth = TD->getTypeSizeInBits(Ty->getScalarType());
6044 else if (Ty->isIntOrIntVector())
6045 BitWidth = Ty->getScalarSizeInBits();
6047 bool isSignBit = false;
6049 // See if we are doing a comparison with a constant.
6050 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6051 Value *A = 0, *B = 0;
6053 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
6054 if (I.isEquality() && CI->isNullValue() &&
6055 match(Op0, m_Sub(m_Value(A), m_Value(B)), *Context)) {
6056 // (icmp cond A B) if cond is equality
6057 return new ICmpInst(*Context, I.getPredicate(), A, B);
6060 // If we have an icmp le or icmp ge instruction, turn it into the
6061 // appropriate icmp lt or icmp gt instruction. This allows us to rely on
6062 // them being folded in the code below.
6063 switch (I.getPredicate()) {
6065 case ICmpInst::ICMP_ULE:
6066 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
6067 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6068 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, Op0,
6069 AddOne(CI, Context));
6070 case ICmpInst::ICMP_SLE:
6071 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
6072 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6073 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6074 AddOne(CI, Context));
6075 case ICmpInst::ICMP_UGE:
6076 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
6077 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6078 return new ICmpInst(*Context, ICmpInst::ICMP_UGT, Op0,
6079 SubOne(CI, Context));
6080 case ICmpInst::ICMP_SGE:
6081 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
6082 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6083 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6084 SubOne(CI, Context));
6087 // If this comparison is a normal comparison, it demands all
6088 // bits, if it is a sign bit comparison, it only demands the sign bit.
6090 isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
6093 // See if we can fold the comparison based on range information we can get
6094 // by checking whether bits are known to be zero or one in the input.
6095 if (BitWidth != 0) {
6096 APInt Op0KnownZero(BitWidth, 0), Op0KnownOne(BitWidth, 0);
6097 APInt Op1KnownZero(BitWidth, 0), Op1KnownOne(BitWidth, 0);
6099 if (SimplifyDemandedBits(I.getOperandUse(0),
6100 isSignBit ? APInt::getSignBit(BitWidth)
6101 : APInt::getAllOnesValue(BitWidth),
6102 Op0KnownZero, Op0KnownOne, 0))
6104 if (SimplifyDemandedBits(I.getOperandUse(1),
6105 APInt::getAllOnesValue(BitWidth),
6106 Op1KnownZero, Op1KnownOne, 0))
6109 // Given the known and unknown bits, compute a range that the LHS could be
6110 // in. Compute the Min, Max and RHS values based on the known bits. For the
6111 // EQ and NE we use unsigned values.
6112 APInt Op0Min(BitWidth, 0), Op0Max(BitWidth, 0);
6113 APInt Op1Min(BitWidth, 0), Op1Max(BitWidth, 0);
6114 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
6115 ComputeSignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6117 ComputeSignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6120 ComputeUnsignedMinMaxValuesFromKnownBits(Op0KnownZero, Op0KnownOne,
6122 ComputeUnsignedMinMaxValuesFromKnownBits(Op1KnownZero, Op1KnownOne,
6126 // If Min and Max are known to be the same, then SimplifyDemandedBits
6127 // figured out that the LHS is a constant. Just constant fold this now so
6128 // that code below can assume that Min != Max.
6129 if (!isa<Constant>(Op0) && Op0Min == Op0Max)
6130 return new ICmpInst(*Context, I.getPredicate(),
6131 Context->getConstantInt(Op0Min), Op1);
6132 if (!isa<Constant>(Op1) && Op1Min == Op1Max)
6133 return new ICmpInst(*Context, I.getPredicate(), Op0,
6134 Context->getConstantInt(Op1Min));
6136 // Based on the range information we know about the LHS, see if we can
6137 // simplify this comparison. For example, (x&4) < 8 is always true.
6138 switch (I.getPredicate()) {
6139 default: LLVM_UNREACHABLE("Unknown icmp opcode!");
6140 case ICmpInst::ICMP_EQ:
6141 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6142 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6144 case ICmpInst::ICMP_NE:
6145 if (Op0Max.ult(Op1Min) || Op0Min.ugt(Op1Max))
6146 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6148 case ICmpInst::ICMP_ULT:
6149 if (Op0Max.ult(Op1Min)) // A <u B -> true if max(A) < min(B)
6150 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6151 if (Op0Min.uge(Op1Max)) // A <u B -> false if min(A) >= max(B)
6152 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6153 if (Op1Min == Op0Max) // A <u B -> A != B if max(A) == min(B)
6154 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6155 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6156 if (Op1Max == Op0Min+1) // A <u C -> A == C-1 if min(A)+1 == C
6157 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6158 SubOne(CI, Context));
6160 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
6161 if (CI->isMinValue(true))
6162 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, Op0,
6163 Context->getConstantIntAllOnesValue(Op0->getType()));
6166 case ICmpInst::ICMP_UGT:
6167 if (Op0Min.ugt(Op1Max)) // A >u B -> true if min(A) > max(B)
6168 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6169 if (Op0Max.ule(Op1Min)) // A >u B -> false if max(A) <= max(B)
6170 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6172 if (Op1Max == Op0Min) // A >u B -> A != B if min(A) == max(B)
6173 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6174 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6175 if (Op1Min == Op0Max-1) // A >u C -> A == C+1 if max(a)-1 == C
6176 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6177 AddOne(CI, Context));
6179 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
6180 if (CI->isMaxValue(true))
6181 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, Op0,
6182 Context->getNullValue(Op0->getType()));
6185 case ICmpInst::ICMP_SLT:
6186 if (Op0Max.slt(Op1Min)) // A <s B -> true if max(A) < min(C)
6187 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6188 if (Op0Min.sge(Op1Max)) // A <s B -> false if min(A) >= max(C)
6189 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6190 if (Op1Min == Op0Max) // A <s B -> A != B if max(A) == min(B)
6191 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6192 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6193 if (Op1Max == Op0Min+1) // A <s C -> A == C-1 if min(A)+1 == C
6194 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6195 SubOne(CI, Context));
6198 case ICmpInst::ICMP_SGT:
6199 if (Op0Min.sgt(Op1Max)) // A >s B -> true if min(A) > max(B)
6200 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6201 if (Op0Max.sle(Op1Min)) // A >s B -> false if max(A) <= min(B)
6202 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6204 if (Op1Max == Op0Min) // A >s B -> A != B if min(A) == max(B)
6205 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Op0, Op1);
6206 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6207 if (Op1Min == Op0Max-1) // A >s C -> A == C+1 if max(A)-1 == C
6208 return new ICmpInst(*Context, ICmpInst::ICMP_EQ, Op0,
6209 AddOne(CI, Context));
6212 case ICmpInst::ICMP_SGE:
6213 assert(!isa<ConstantInt>(Op1) && "ICMP_SGE with ConstantInt not folded!");
6214 if (Op0Min.sge(Op1Max)) // A >=s B -> true if min(A) >= max(B)
6215 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6216 if (Op0Max.slt(Op1Min)) // A >=s B -> false if max(A) < min(B)
6217 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6219 case ICmpInst::ICMP_SLE:
6220 assert(!isa<ConstantInt>(Op1) && "ICMP_SLE with ConstantInt not folded!");
6221 if (Op0Max.sle(Op1Min)) // A <=s B -> true if max(A) <= min(B)
6222 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6223 if (Op0Min.sgt(Op1Max)) // A <=s B -> false if min(A) > max(B)
6224 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6226 case ICmpInst::ICMP_UGE:
6227 assert(!isa<ConstantInt>(Op1) && "ICMP_UGE with ConstantInt not folded!");
6228 if (Op0Min.uge(Op1Max)) // A >=u B -> true if min(A) >= max(B)
6229 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6230 if (Op0Max.ult(Op1Min)) // A >=u B -> false if max(A) < min(B)
6231 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6233 case ICmpInst::ICMP_ULE:
6234 assert(!isa<ConstantInt>(Op1) && "ICMP_ULE with ConstantInt not folded!");
6235 if (Op0Max.ule(Op1Min)) // A <=u B -> true if max(A) <= min(B)
6236 return ReplaceInstUsesWith(I, Context->getConstantIntTrue());
6237 if (Op0Min.ugt(Op1Max)) // A <=u B -> false if min(A) > max(B)
6238 return ReplaceInstUsesWith(I, Context->getConstantIntFalse());
6242 // Turn a signed comparison into an unsigned one if both operands
6243 // are known to have the same sign.
6244 if (I.isSignedPredicate() &&
6245 ((Op0KnownZero.isNegative() && Op1KnownZero.isNegative()) ||
6246 (Op0KnownOne.isNegative() && Op1KnownOne.isNegative())))
6247 return new ICmpInst(*Context, I.getUnsignedPredicate(), Op0, Op1);
6250 // Test if the ICmpInst instruction is used exclusively by a select as
6251 // part of a minimum or maximum operation. If so, refrain from doing
6252 // any other folding. This helps out other analyses which understand
6253 // non-obfuscated minimum and maximum idioms, such as ScalarEvolution
6254 // and CodeGen. And in this case, at least one of the comparison
6255 // operands has at least one user besides the compare (the select),
6256 // which would often largely negate the benefit of folding anyway.
6258 if (SelectInst *SI = dyn_cast<SelectInst>(*I.use_begin()))
6259 if ((SI->getOperand(1) == Op0 && SI->getOperand(2) == Op1) ||
6260 (SI->getOperand(2) == Op0 && SI->getOperand(1) == Op1))
6263 // See if we are doing a comparison between a constant and an instruction that
6264 // can be folded into the comparison.
6265 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
6266 // Since the RHS is a ConstantInt (CI), if the left hand side is an
6267 // instruction, see if that instruction also has constants so that the
6268 // instruction can be folded into the icmp
6269 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6270 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
6274 // Handle icmp with constant (but not simple integer constant) RHS
6275 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
6276 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
6277 switch (LHSI->getOpcode()) {
6278 case Instruction::GetElementPtr:
6279 if (RHSC->isNullValue()) {
6280 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
6281 bool isAllZeros = true;
6282 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
6283 if (!isa<Constant>(LHSI->getOperand(i)) ||
6284 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
6289 return new ICmpInst(*Context, I.getPredicate(), LHSI->getOperand(0),
6290 Context->getNullValue(LHSI->getOperand(0)->getType()));
6294 case Instruction::PHI:
6295 // Only fold icmp into the PHI if the phi and fcmp are in the same
6296 // block. If in the same block, we're encouraging jump threading. If
6297 // not, we are just pessimizing the code by making an i1 phi.
6298 if (LHSI->getParent() == I.getParent())
6299 if (Instruction *NV = FoldOpIntoPhi(I))
6302 case Instruction::Select: {
6303 // If either operand of the select is a constant, we can fold the
6304 // comparison into the select arms, which will cause one to be
6305 // constant folded and the select turned into a bitwise or.
6306 Value *Op1 = 0, *Op2 = 0;
6307 if (LHSI->hasOneUse()) {
6308 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
6309 // Fold the known value into the constant operand.
6310 Op1 = Context->getConstantExprICmp(I.getPredicate(), C, RHSC);
6311 // Insert a new ICmp of the other select operand.
6312 Op2 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6313 LHSI->getOperand(2), RHSC,
6315 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
6316 // Fold the known value into the constant operand.
6317 Op2 = Context->getConstantExprICmp(I.getPredicate(), C, RHSC);
6318 // Insert a new ICmp of the other select operand.
6319 Op1 = InsertNewInstBefore(new ICmpInst(*Context, I.getPredicate(),
6320 LHSI->getOperand(1), RHSC,
6326 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
6329 case Instruction::Malloc:
6330 // If we have (malloc != null), and if the malloc has a single use, we
6331 // can assume it is successful and remove the malloc.
6332 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
6333 AddToWorkList(LHSI);
6334 return ReplaceInstUsesWith(I, Context->getConstantInt(Type::Int1Ty,
6335 !I.isTrueWhenEqual()));
6341 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
6342 if (User *GEP = dyn_castGetElementPtr(Op0))
6343 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
6345 if (User *GEP = dyn_castGetElementPtr(Op1))
6346 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
6347 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
6350 // Test to see if the operands of the icmp are casted versions of other
6351 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
6353 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
6354 if (isa<PointerType>(Op0->getType()) &&
6355 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
6356 // We keep moving the cast from the left operand over to the right
6357 // operand, where it can often be eliminated completely.
6358 Op0 = CI->getOperand(0);
6360 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
6361 // so eliminate it as well.
6362 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
6363 Op1 = CI2->getOperand(0);
6365 // If Op1 is a constant, we can fold the cast into the constant.
6366 if (Op0->getType() != Op1->getType()) {
6367 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
6368 Op1 = Context->getConstantExprBitCast(Op1C, Op0->getType());
6370 // Otherwise, cast the RHS right before the icmp
6371 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
6374 return new ICmpInst(*Context, I.getPredicate(), Op0, Op1);
6378 if (isa<CastInst>(Op0)) {
6379 // Handle the special case of: icmp (cast bool to X), <cst>
6380 // This comes up when you have code like
6383 // For generality, we handle any zero-extension of any operand comparison
6384 // with a constant or another cast from the same type.
6385 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
6386 if (Instruction *R = visitICmpInstWithCastAndCast(I))
6390 // See if it's the same type of instruction on the left and right.
6391 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
6392 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
6393 if (Op0I->getOpcode() == Op1I->getOpcode() && Op0I->hasOneUse() &&
6394 Op1I->hasOneUse() && Op0I->getOperand(1) == Op1I->getOperand(1)) {
6395 switch (Op0I->getOpcode()) {
6397 case Instruction::Add:
6398 case Instruction::Sub:
6399 case Instruction::Xor:
6400 if (I.isEquality()) // a+x icmp eq/ne b+x --> a icmp b
6401 return new ICmpInst(*Context, I.getPredicate(), Op0I->getOperand(0),
6402 Op1I->getOperand(0));
6403 // icmp u/s (a ^ signbit), (b ^ signbit) --> icmp s/u a, b
6404 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6405 if (CI->getValue().isSignBit()) {
6406 ICmpInst::Predicate Pred = I.isSignedPredicate()
6407 ? I.getUnsignedPredicate()
6408 : I.getSignedPredicate();
6409 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6410 Op1I->getOperand(0));
6413 if (CI->getValue().isMaxSignedValue()) {
6414 ICmpInst::Predicate Pred = I.isSignedPredicate()
6415 ? I.getUnsignedPredicate()
6416 : I.getSignedPredicate();
6417 Pred = I.getSwappedPredicate(Pred);
6418 return new ICmpInst(*Context, Pred, Op0I->getOperand(0),
6419 Op1I->getOperand(0));
6423 case Instruction::Mul:
6424 if (!I.isEquality())
6427 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
6428 // a * Cst icmp eq/ne b * Cst --> a & Mask icmp b & Mask
6429 // Mask = -1 >> count-trailing-zeros(Cst).
6430 if (!CI->isZero() && !CI->isOne()) {
6431 const APInt &AP = CI->getValue();
6432 ConstantInt *Mask = Context->getConstantInt(
6433 APInt::getLowBitsSet(AP.getBitWidth(),
6435 AP.countTrailingZeros()));
6436 Instruction *And1 = BinaryOperator::CreateAnd(Op0I->getOperand(0),
6438 Instruction *And2 = BinaryOperator::CreateAnd(Op1I->getOperand(0),
6440 InsertNewInstBefore(And1, I);
6441 InsertNewInstBefore(And2, I);
6442 return new ICmpInst(*Context, I.getPredicate(), And1, And2);
6451 // ~x < ~y --> y < x
6453 if (match(Op0, m_Not(m_Value(A)), *Context) &&
6454 match(Op1, m_Not(m_Value(B)), *Context))
6455 return new ICmpInst(*Context, I.getPredicate(), B, A);
6458 if (I.isEquality()) {
6459 Value *A, *B, *C, *D;
6461 // -x == -y --> x == y
6462 if (match(Op0, m_Neg(m_Value(A)), *Context) &&
6463 match(Op1, m_Neg(m_Value(B)), *Context))
6464 return new ICmpInst(*Context, I.getPredicate(), A, B);
6466 if (match(Op0, m_Xor(m_Value(A), m_Value(B)), *Context)) {
6467 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
6468 Value *OtherVal = A == Op1 ? B : A;
6469 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6470 Context->getNullValue(A->getType()));
6473 if (match(Op1, m_Xor(m_Value(C), m_Value(D)), *Context)) {
6474 // A^c1 == C^c2 --> A == C^(c1^c2)
6475 ConstantInt *C1, *C2;
6476 if (match(B, m_ConstantInt(C1), *Context) &&
6477 match(D, m_ConstantInt(C2), *Context) && Op1->hasOneUse()) {
6479 Context->getConstantInt(C1->getValue() ^ C2->getValue());
6480 Instruction *Xor = BinaryOperator::CreateXor(C, NC, "tmp");
6481 return new ICmpInst(*Context, I.getPredicate(), A,
6482 InsertNewInstBefore(Xor, I));
6485 // A^B == A^D -> B == D
6486 if (A == C) return new ICmpInst(*Context, I.getPredicate(), B, D);
6487 if (A == D) return new ICmpInst(*Context, I.getPredicate(), B, C);
6488 if (B == C) return new ICmpInst(*Context, I.getPredicate(), A, D);
6489 if (B == D) return new ICmpInst(*Context, I.getPredicate(), A, C);
6493 if (match(Op1, m_Xor(m_Value(A), m_Value(B)), *Context) &&
6494 (A == Op0 || B == Op0)) {
6495 // A == (A^B) -> B == 0
6496 Value *OtherVal = A == Op0 ? B : A;
6497 return new ICmpInst(*Context, I.getPredicate(), OtherVal,
6498 Context->getNullValue(A->getType()));
6501 // (A-B) == A -> B == 0
6502 if (match(Op0, m_Sub(m_Specific(Op1), m_Value(B)), *Context))
6503 return new ICmpInst(*Context, I.getPredicate(), B,
6504 Context->getNullValue(B->getType()));
6506 // A == (A-B) -> B == 0
6507 if (match(Op1, m_Sub(m_Specific(Op0), m_Value(B)), *Context))
6508 return new ICmpInst(*Context, I.getPredicate(), B,
6509 Context->getNullValue(B->getType()));
6511 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
6512 if (Op0->hasOneUse() && Op1->hasOneUse() &&
6513 match(Op0, m_And(m_Value(A), m_Value(B)), *Context) &&
6514 match(Op1, m_And(m_Value(C), m_Value(D)), *Context)) {
6515 Value *X = 0, *Y = 0, *Z = 0;
6518 X = B; Y = D; Z = A;
6519 } else if (A == D) {
6520 X = B; Y = C; Z = A;
6521 } else if (B == C) {
6522 X = A; Y = D; Z = B;
6523 } else if (B == D) {
6524 X = A; Y = C; Z = B;
6527 if (X) { // Build (X^Y) & Z
6528 Op1 = InsertNewInstBefore(BinaryOperator::CreateXor(X, Y, "tmp"), I);
6529 Op1 = InsertNewInstBefore(BinaryOperator::CreateAnd(Op1, Z, "tmp"), I);
6530 I.setOperand(0, Op1);
6531 I.setOperand(1, Context->getNullValue(Op1->getType()));
6536 return Changed ? &I : 0;
6540 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
6541 /// and CmpRHS are both known to be integer constants.
6542 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
6543 ConstantInt *DivRHS) {
6544 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
6545 const APInt &CmpRHSV = CmpRHS->getValue();
6547 // FIXME: If the operand types don't match the type of the divide
6548 // then don't attempt this transform. The code below doesn't have the
6549 // logic to deal with a signed divide and an unsigned compare (and
6550 // vice versa). This is because (x /s C1) <s C2 produces different
6551 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
6552 // (x /u C1) <u C2. Simply casting the operands and result won't
6553 // work. :( The if statement below tests that condition and bails
6555 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
6556 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
6558 if (DivRHS->isZero())
6559 return 0; // The ProdOV computation fails on divide by zero.
6560 if (DivIsSigned && DivRHS->isAllOnesValue())
6561 return 0; // The overflow computation also screws up here
6562 if (DivRHS->isOne())
6563 return 0; // Not worth bothering, and eliminates some funny cases
6566 // Compute Prod = CI * DivRHS. We are essentially solving an equation
6567 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
6568 // C2 (CI). By solving for X we can turn this into a range check
6569 // instead of computing a divide.
6570 Constant *Prod = Context->getConstantExprMul(CmpRHS, DivRHS);
6572 // Determine if the product overflows by seeing if the product is
6573 // not equal to the divide. Make sure we do the same kind of divide
6574 // as in the LHS instruction that we're folding.
6575 bool ProdOV = (DivIsSigned ? Context->getConstantExprSDiv(Prod, DivRHS) :
6576 Context->getConstantExprUDiv(Prod, DivRHS)) != CmpRHS;
6578 // Get the ICmp opcode
6579 ICmpInst::Predicate Pred = ICI.getPredicate();
6581 // Figure out the interval that is being checked. For example, a comparison
6582 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
6583 // Compute this interval based on the constants involved and the signedness of
6584 // the compare/divide. This computes a half-open interval, keeping track of
6585 // whether either value in the interval overflows. After analysis each
6586 // overflow variable is set to 0 if it's corresponding bound variable is valid
6587 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
6588 int LoOverflow = 0, HiOverflow = 0;
6589 Constant *LoBound = 0, *HiBound = 0;
6591 if (!DivIsSigned) { // udiv
6592 // e.g. X/5 op 3 --> [15, 20)
6594 HiOverflow = LoOverflow = ProdOV;
6596 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, Context, false);
6597 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
6598 if (CmpRHSV == 0) { // (X / pos) op 0
6599 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
6600 LoBound = cast<ConstantInt>(Context->getConstantExprNeg(SubOne(DivRHS,
6603 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
6604 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
6605 HiOverflow = LoOverflow = ProdOV;
6607 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, Context, true);
6608 } else { // (X / pos) op neg
6609 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
6610 HiBound = AddOne(Prod, Context);
6611 LoOverflow = HiOverflow = ProdOV ? -1 : 0;
6613 ConstantInt* DivNeg =
6614 cast<ConstantInt>(Context->getConstantExprNeg(DivRHS));
6615 LoOverflow = AddWithOverflow(LoBound, HiBound, DivNeg, Context,
6619 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
6620 if (CmpRHSV == 0) { // (X / neg) op 0
6621 // e.g. X/-5 op 0 --> [-4, 5)
6622 LoBound = AddOne(DivRHS, Context);
6623 HiBound = cast<ConstantInt>(Context->getConstantExprNeg(DivRHS));
6624 if (HiBound == DivRHS) { // -INTMIN = INTMIN
6625 HiOverflow = 1; // [INTMIN+1, overflow)
6626 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
6628 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
6629 // e.g. X/-5 op 3 --> [-19, -14)
6630 HiBound = AddOne(Prod, Context);
6631 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
6633 LoOverflow = AddWithOverflow(LoBound, HiBound,
6634 DivRHS, Context, true) ? -1 : 0;
6635 } else { // (X / neg) op neg
6636 LoBound = Prod; // e.g. X/-5 op -3 --> [15, 20)
6637 LoOverflow = HiOverflow = ProdOV;
6639 HiOverflow = SubWithOverflow(HiBound, Prod, DivRHS, Context, true);
6642 // Dividing by a negative swaps the condition. LT <-> GT
6643 Pred = ICmpInst::getSwappedPredicate(Pred);
6646 Value *X = DivI->getOperand(0);
6648 default: LLVM_UNREACHABLE("Unhandled icmp opcode!");
6649 case ICmpInst::ICMP_EQ:
6650 if (LoOverflow && HiOverflow)
6651 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6652 else if (HiOverflow)
6653 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6654 ICmpInst::ICMP_UGE, X, LoBound);
6655 else if (LoOverflow)
6656 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6657 ICmpInst::ICMP_ULT, X, HiBound);
6659 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
6660 case ICmpInst::ICMP_NE:
6661 if (LoOverflow && HiOverflow)
6662 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6663 else if (HiOverflow)
6664 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SLT :
6665 ICmpInst::ICMP_ULT, X, LoBound);
6666 else if (LoOverflow)
6667 return new ICmpInst(*Context, DivIsSigned ? ICmpInst::ICMP_SGE :
6668 ICmpInst::ICMP_UGE, X, HiBound);
6670 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
6671 case ICmpInst::ICMP_ULT:
6672 case ICmpInst::ICMP_SLT:
6673 if (LoOverflow == +1) // Low bound is greater than input range.
6674 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6675 if (LoOverflow == -1) // Low bound is less than input range.
6676 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6677 return new ICmpInst(*Context, Pred, X, LoBound);
6678 case ICmpInst::ICMP_UGT:
6679 case ICmpInst::ICMP_SGT:
6680 if (HiOverflow == +1) // High bound greater than input range.
6681 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6682 else if (HiOverflow == -1) // High bound less than input range.
6683 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6684 if (Pred == ICmpInst::ICMP_UGT)
6685 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, X, HiBound);
6687 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, X, HiBound);
6692 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
6694 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
6697 const APInt &RHSV = RHS->getValue();
6699 switch (LHSI->getOpcode()) {
6700 case Instruction::Trunc:
6701 if (ICI.isEquality() && LHSI->hasOneUse()) {
6702 // Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
6703 // of the high bits truncated out of x are known.
6704 unsigned DstBits = LHSI->getType()->getPrimitiveSizeInBits(),
6705 SrcBits = LHSI->getOperand(0)->getType()->getPrimitiveSizeInBits();
6706 APInt Mask(APInt::getHighBitsSet(SrcBits, SrcBits-DstBits));
6707 APInt KnownZero(SrcBits, 0), KnownOne(SrcBits, 0);
6708 ComputeMaskedBits(LHSI->getOperand(0), Mask, KnownZero, KnownOne);
6710 // If all the high bits are known, we can do this xform.
6711 if ((KnownZero|KnownOne).countLeadingOnes() >= SrcBits-DstBits) {
6712 // Pull in the high bits from known-ones set.
6713 APInt NewRHS(RHS->getValue());
6714 NewRHS.zext(SrcBits);
6716 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6717 Context->getConstantInt(NewRHS));
6722 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
6723 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
6724 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
6726 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
6727 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
6728 Value *CompareVal = LHSI->getOperand(0);
6730 // If the sign bit of the XorCST is not set, there is no change to
6731 // the operation, just stop using the Xor.
6732 if (!XorCST->getValue().isNegative()) {
6733 ICI.setOperand(0, CompareVal);
6734 AddToWorkList(LHSI);
6738 // Was the old condition true if the operand is positive?
6739 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
6741 // If so, the new one isn't.
6742 isTrueIfPositive ^= true;
6744 if (isTrueIfPositive)
6745 return new ICmpInst(*Context, ICmpInst::ICMP_SGT, CompareVal,
6746 SubOne(RHS, Context));
6748 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, CompareVal,
6749 AddOne(RHS, Context));
6752 if (LHSI->hasOneUse()) {
6753 // (icmp u/s (xor A SignBit), C) -> (icmp s/u A, (xor C SignBit))
6754 if (!ICI.isEquality() && XorCST->getValue().isSignBit()) {
6755 const APInt &SignBit = XorCST->getValue();
6756 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6757 ? ICI.getUnsignedPredicate()
6758 : ICI.getSignedPredicate();
6759 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6760 Context->getConstantInt(RHSV ^ SignBit));
6763 // (icmp u/s (xor A ~SignBit), C) -> (icmp s/u (xor C ~SignBit), A)
6764 if (!ICI.isEquality() && XorCST->getValue().isMaxSignedValue()) {
6765 const APInt &NotSignBit = XorCST->getValue();
6766 ICmpInst::Predicate Pred = ICI.isSignedPredicate()
6767 ? ICI.getUnsignedPredicate()
6768 : ICI.getSignedPredicate();
6769 Pred = ICI.getSwappedPredicate(Pred);
6770 return new ICmpInst(*Context, Pred, LHSI->getOperand(0),
6771 Context->getConstantInt(RHSV ^ NotSignBit));
6776 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
6777 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
6778 LHSI->getOperand(0)->hasOneUse()) {
6779 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
6781 // If the LHS is an AND of a truncating cast, we can widen the
6782 // and/compare to be the input width without changing the value
6783 // produced, eliminating a cast.
6784 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
6785 // We can do this transformation if either the AND constant does not
6786 // have its sign bit set or if it is an equality comparison.
6787 // Extending a relational comparison when we're checking the sign
6788 // bit would not work.
6789 if (Cast->hasOneUse() &&
6790 (ICI.isEquality() ||
6791 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
6793 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
6794 APInt NewCST = AndCST->getValue();
6795 NewCST.zext(BitWidth);
6797 NewCI.zext(BitWidth);
6798 Instruction *NewAnd =
6799 BinaryOperator::CreateAnd(Cast->getOperand(0),
6800 Context->getConstantInt(NewCST),LHSI->getName());
6801 InsertNewInstBefore(NewAnd, ICI);
6802 return new ICmpInst(*Context, ICI.getPredicate(), NewAnd,
6803 Context->getConstantInt(NewCI));
6807 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
6808 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
6809 // happens a LOT in code produced by the C front-end, for bitfield
6811 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
6812 if (Shift && !Shift->isShift())
6816 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
6817 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
6818 const Type *AndTy = AndCST->getType(); // Type of the and.
6820 // We can fold this as long as we can't shift unknown bits
6821 // into the mask. This can only happen with signed shift
6822 // rights, as they sign-extend.
6824 bool CanFold = Shift->isLogicalShift();
6826 // To test for the bad case of the signed shr, see if any
6827 // of the bits shifted in could be tested after the mask.
6828 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
6829 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
6831 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
6832 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
6833 AndCST->getValue()) == 0)
6839 if (Shift->getOpcode() == Instruction::Shl)
6840 NewCst = Context->getConstantExprLShr(RHS, ShAmt);
6842 NewCst = Context->getConstantExprShl(RHS, ShAmt);
6844 // Check to see if we are shifting out any of the bits being
6846 if (Context->getConstantExpr(Shift->getOpcode(),
6847 NewCst, ShAmt) != RHS) {
6848 // If we shifted bits out, the fold is not going to work out.
6849 // As a special case, check to see if this means that the
6850 // result is always true or false now.
6851 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6852 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
6853 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6854 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
6856 ICI.setOperand(1, NewCst);
6857 Constant *NewAndCST;
6858 if (Shift->getOpcode() == Instruction::Shl)
6859 NewAndCST = Context->getConstantExprLShr(AndCST, ShAmt);
6861 NewAndCST = Context->getConstantExprShl(AndCST, ShAmt);
6862 LHSI->setOperand(1, NewAndCST);
6863 LHSI->setOperand(0, Shift->getOperand(0));
6864 AddToWorkList(Shift); // Shift is dead.
6865 AddUsesToWorkList(ICI);
6871 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6872 // preferable because it allows the C<<Y expression to be hoisted out
6873 // of a loop if Y is invariant and X is not.
6874 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6875 ICI.isEquality() && !Shift->isArithmeticShift() &&
6876 !isa<Constant>(Shift->getOperand(0))) {
6879 if (Shift->getOpcode() == Instruction::LShr) {
6880 NS = BinaryOperator::CreateShl(AndCST,
6881 Shift->getOperand(1), "tmp");
6883 // Insert a logical shift.
6884 NS = BinaryOperator::CreateLShr(AndCST,
6885 Shift->getOperand(1), "tmp");
6887 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6889 // Compute X & (C << Y).
6890 Instruction *NewAnd =
6891 BinaryOperator::CreateAnd(Shift->getOperand(0), NS, LHSI->getName());
6892 InsertNewInstBefore(NewAnd, ICI);
6894 ICI.setOperand(0, NewAnd);
6900 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6901 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6904 uint32_t TypeBits = RHSV.getBitWidth();
6906 // Check that the shift amount is in range. If not, don't perform
6907 // undefined shifts. When the shift is visited it will be
6909 if (ShAmt->uge(TypeBits))
6912 if (ICI.isEquality()) {
6913 // If we are comparing against bits always shifted out, the
6914 // comparison cannot succeed.
6916 Context->getConstantExprShl(Context->getConstantExprLShr(RHS, ShAmt),
6918 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6919 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6920 Constant *Cst = Context->getConstantInt(Type::Int1Ty, IsICMP_NE);
6921 return ReplaceInstUsesWith(ICI, Cst);
6924 if (LHSI->hasOneUse()) {
6925 // Otherwise strength reduce the shift into an and.
6926 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6928 Context->getConstantInt(APInt::getLowBitsSet(TypeBits,
6929 TypeBits-ShAmtVal));
6932 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6933 Mask, LHSI->getName()+".mask");
6934 Value *And = InsertNewInstBefore(AndI, ICI);
6935 return new ICmpInst(*Context, ICI.getPredicate(), And,
6936 Context->getConstantInt(RHSV.lshr(ShAmtVal)));
6940 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6941 bool TrueIfSigned = false;
6942 if (LHSI->hasOneUse() &&
6943 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6944 // (X << 31) <s 0 --> (X&1) != 0
6945 Constant *Mask = Context->getConstantInt(APInt(TypeBits, 1) <<
6946 (TypeBits-ShAmt->getZExtValue()-1));
6948 BinaryOperator::CreateAnd(LHSI->getOperand(0),
6949 Mask, LHSI->getName()+".mask");
6950 Value *And = InsertNewInstBefore(AndI, ICI);
6952 return new ICmpInst(*Context,
6953 TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6954 And, Context->getNullValue(And->getType()));
6959 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6960 case Instruction::AShr: {
6961 // Only handle equality comparisons of shift-by-constant.
6962 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6963 if (!ShAmt || !ICI.isEquality()) break;
6965 // Check that the shift amount is in range. If not, don't perform
6966 // undefined shifts. When the shift is visited it will be
6968 uint32_t TypeBits = RHSV.getBitWidth();
6969 if (ShAmt->uge(TypeBits))
6972 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6974 // If we are comparing against bits always shifted out, the
6975 // comparison cannot succeed.
6976 APInt Comp = RHSV << ShAmtVal;
6977 if (LHSI->getOpcode() == Instruction::LShr)
6978 Comp = Comp.lshr(ShAmtVal);
6980 Comp = Comp.ashr(ShAmtVal);
6982 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6983 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6984 Constant *Cst = Context->getConstantInt(Type::Int1Ty, IsICMP_NE);
6985 return ReplaceInstUsesWith(ICI, Cst);
6988 // Otherwise, check to see if the bits shifted out are known to be zero.
6989 // If so, we can compare against the unshifted value:
6990 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6991 if (LHSI->hasOneUse() &&
6992 MaskedValueIsZero(LHSI->getOperand(0),
6993 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6994 return new ICmpInst(*Context, ICI.getPredicate(), LHSI->getOperand(0),
6995 Context->getConstantExprShl(RHS, ShAmt));
6998 if (LHSI->hasOneUse()) {
6999 // Otherwise strength reduce the shift into an and.
7000 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
7001 Constant *Mask = Context->getConstantInt(Val);
7004 BinaryOperator::CreateAnd(LHSI->getOperand(0),
7005 Mask, LHSI->getName()+".mask");
7006 Value *And = InsertNewInstBefore(AndI, ICI);
7007 return new ICmpInst(*Context, ICI.getPredicate(), And,
7008 Context->getConstantExprShl(RHS, ShAmt));
7013 case Instruction::SDiv:
7014 case Instruction::UDiv:
7015 // Fold: icmp pred ([us]div X, C1), C2 -> range test
7016 // Fold this div into the comparison, producing a range check.
7017 // Determine, based on the divide type, what the range is being
7018 // checked. If there is an overflow on the low or high side, remember
7019 // it, otherwise compute the range [low, hi) bounding the new value.
7020 // See: InsertRangeTest above for the kinds of replacements possible.
7021 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
7022 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
7027 case Instruction::Add:
7028 // Fold: icmp pred (add, X, C1), C2
7030 if (!ICI.isEquality()) {
7031 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
7033 const APInt &LHSV = LHSC->getValue();
7035 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
7038 if (ICI.isSignedPredicate()) {
7039 if (CR.getLower().isSignBit()) {
7040 return new ICmpInst(*Context, ICmpInst::ICMP_SLT, LHSI->getOperand(0),
7041 Context->getConstantInt(CR.getUpper()));
7042 } else if (CR.getUpper().isSignBit()) {
7043 return new ICmpInst(*Context, ICmpInst::ICMP_SGE, LHSI->getOperand(0),
7044 Context->getConstantInt(CR.getLower()));
7047 if (CR.getLower().isMinValue()) {
7048 return new ICmpInst(*Context, ICmpInst::ICMP_ULT, LHSI->getOperand(0),
7049 Context->getConstantInt(CR.getUpper()));
7050 } else if (CR.getUpper().isMinValue()) {
7051 return new ICmpInst(*Context, ICmpInst::ICMP_UGE, LHSI->getOperand(0),
7052 Context->getConstantInt(CR.getLower()));
7059 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
7060 if (ICI.isEquality()) {
7061 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
7063 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
7064 // the second operand is a constant, simplify a bit.
7065 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
7066 switch (BO->getOpcode()) {
7067 case Instruction::SRem:
7068 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
7069 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
7070 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
7071 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
7072 Instruction *NewRem =
7073 BinaryOperator::CreateURem(BO->getOperand(0), BO->getOperand(1),
7075 InsertNewInstBefore(NewRem, ICI);
7076 return new ICmpInst(*Context, ICI.getPredicate(), NewRem,
7077 Context->getNullValue(BO->getType()));
7081 case Instruction::Add:
7082 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
7083 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7084 if (BO->hasOneUse())
7085 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7086 Context->getConstantExprSub(RHS, BOp1C));
7087 } else if (RHSV == 0) {
7088 // Replace ((add A, B) != 0) with (A != -B) if A or B is
7089 // efficiently invertible, or if the add has just this one use.
7090 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
7092 if (Value *NegVal = dyn_castNegVal(BOp1, Context))
7093 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, NegVal);
7094 else if (Value *NegVal = dyn_castNegVal(BOp0, Context))
7095 return new ICmpInst(*Context, ICI.getPredicate(), NegVal, BOp1);
7096 else if (BO->hasOneUse()) {
7097 Instruction *Neg = BinaryOperator::CreateNeg(BOp1);
7098 InsertNewInstBefore(Neg, ICI);
7100 return new ICmpInst(*Context, ICI.getPredicate(), BOp0, Neg);
7104 case Instruction::Xor:
7105 // For the xor case, we can xor two constants together, eliminating
7106 // the explicit xor.
7107 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
7108 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7109 Context->getConstantExprXor(RHS, BOC));
7112 case Instruction::Sub:
7113 // Replace (([sub|xor] A, B) != 0) with (A != B)
7115 return new ICmpInst(*Context, ICI.getPredicate(), BO->getOperand(0),
7119 case Instruction::Or:
7120 // If bits are being or'd in that are not present in the constant we
7121 // are comparing against, then the comparison could never succeed!
7122 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
7123 Constant *NotCI = Context->getConstantExprNot(RHS);
7124 if (!Context->getConstantExprAnd(BOC, NotCI)->isNullValue())
7125 return ReplaceInstUsesWith(ICI,
7126 Context->getConstantInt(Type::Int1Ty,
7131 case Instruction::And:
7132 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
7133 // If bits are being compared against that are and'd out, then the
7134 // comparison can never succeed!
7135 if ((RHSV & ~BOC->getValue()) != 0)
7136 return ReplaceInstUsesWith(ICI,
7137 Context->getConstantInt(Type::Int1Ty,
7140 // If we have ((X & C) == C), turn it into ((X & C) != 0).
7141 if (RHS == BOC && RHSV.isPowerOf2())
7142 return new ICmpInst(*Context, isICMP_NE ? ICmpInst::ICMP_EQ :
7143 ICmpInst::ICMP_NE, LHSI,
7144 Context->getNullValue(RHS->getType()));
7146 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
7147 if (BOC->getValue().isSignBit()) {
7148 Value *X = BO->getOperand(0);
7149 Constant *Zero = Context->getNullValue(X->getType());
7150 ICmpInst::Predicate pred = isICMP_NE ?
7151 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
7152 return new ICmpInst(*Context, pred, X, Zero);
7155 // ((X & ~7) == 0) --> X < 8
7156 if (RHSV == 0 && isHighOnes(BOC)) {
7157 Value *X = BO->getOperand(0);
7158 Constant *NegX = Context->getConstantExprNeg(BOC);
7159 ICmpInst::Predicate pred = isICMP_NE ?
7160 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
7161 return new ICmpInst(*Context, pred, X, NegX);
7166 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
7167 // Handle icmp {eq|ne} <intrinsic>, intcst.
7168 if (II->getIntrinsicID() == Intrinsic::bswap) {
7170 ICI.setOperand(0, II->getOperand(1));
7171 ICI.setOperand(1, Context->getConstantInt(RHSV.byteSwap()));
7179 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
7180 /// We only handle extending casts so far.
7182 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
7183 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
7184 Value *LHSCIOp = LHSCI->getOperand(0);
7185 const Type *SrcTy = LHSCIOp->getType();
7186 const Type *DestTy = LHSCI->getType();
7189 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
7190 // integer type is the same size as the pointer type.
7191 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
7192 getTargetData().getPointerSizeInBits() ==
7193 cast<IntegerType>(DestTy)->getBitWidth()) {
7195 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
7196 RHSOp = Context->getConstantExprIntToPtr(RHSC, SrcTy);
7197 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
7198 RHSOp = RHSC->getOperand(0);
7199 // If the pointer types don't match, insert a bitcast.
7200 if (LHSCIOp->getType() != RHSOp->getType())
7201 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
7205 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSOp);
7208 // The code below only handles extension cast instructions, so far.
7210 if (LHSCI->getOpcode() != Instruction::ZExt &&
7211 LHSCI->getOpcode() != Instruction::SExt)
7214 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
7215 bool isSignedCmp = ICI.isSignedPredicate();
7217 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
7218 // Not an extension from the same type?
7219 RHSCIOp = CI->getOperand(0);
7220 if (RHSCIOp->getType() != LHSCIOp->getType())
7223 // If the signedness of the two casts doesn't agree (i.e. one is a sext
7224 // and the other is a zext), then we can't handle this.
7225 if (CI->getOpcode() != LHSCI->getOpcode())
7228 // Deal with equality cases early.
7229 if (ICI.isEquality())
7230 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7232 // A signed comparison of sign extended values simplifies into a
7233 // signed comparison.
7234 if (isSignedCmp && isSignedExt)
7235 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, RHSCIOp);
7237 // The other three cases all fold into an unsigned comparison.
7238 return new ICmpInst(*Context, ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
7241 // If we aren't dealing with a constant on the RHS, exit early
7242 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
7246 // Compute the constant that would happen if we truncated to SrcTy then
7247 // reextended to DestTy.
7248 Constant *Res1 = Context->getConstantExprTrunc(CI, SrcTy);
7249 Constant *Res2 = Context->getConstantExprCast(LHSCI->getOpcode(),
7252 // If the re-extended constant didn't change...
7254 // Make sure that sign of the Cmp and the sign of the Cast are the same.
7255 // For example, we might have:
7256 // %A = sext i16 %X to i32
7257 // %B = icmp ugt i32 %A, 1330
7258 // It is incorrect to transform this into
7259 // %B = icmp ugt i16 %X, 1330
7260 // because %A may have negative value.
7262 // However, we allow this when the compare is EQ/NE, because they are
7264 if (isSignedExt == isSignedCmp || ICI.isEquality())
7265 return new ICmpInst(*Context, ICI.getPredicate(), LHSCIOp, Res1);
7269 // The re-extended constant changed so the constant cannot be represented
7270 // in the shorter type. Consequently, we cannot emit a simple comparison.
7272 // First, handle some easy cases. We know the result cannot be equal at this
7273 // point so handle the ICI.isEquality() cases
7274 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
7275 return ReplaceInstUsesWith(ICI, Context->getConstantIntFalse());
7276 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
7277 return ReplaceInstUsesWith(ICI, Context->getConstantIntTrue());
7279 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
7280 // should have been folded away previously and not enter in here.
7283 // We're performing a signed comparison.
7284 if (cast<ConstantInt>(CI)->getValue().isNegative())
7285 Result = Context->getConstantIntFalse(); // X < (small) --> false
7287 Result = Context->getConstantIntTrue(); // X < (large) --> true
7289 // We're performing an unsigned comparison.
7291 // We're performing an unsigned comp with a sign extended value.
7292 // This is true if the input is >= 0. [aka >s -1]
7293 Constant *NegOne = Context->getConstantIntAllOnesValue(SrcTy);
7294 Result = InsertNewInstBefore(new ICmpInst(*Context, ICmpInst::ICMP_SGT,
7295 LHSCIOp, NegOne, ICI.getName()), ICI);
7297 // Unsigned extend & unsigned compare -> always true.
7298 Result = Context->getConstantIntTrue();
7302 // Finally, return the value computed.
7303 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
7304 ICI.getPredicate() == ICmpInst::ICMP_SLT)
7305 return ReplaceInstUsesWith(ICI, Result);
7307 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
7308 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
7309 "ICmp should be folded!");
7310 if (Constant *CI = dyn_cast<Constant>(Result))
7311 return ReplaceInstUsesWith(ICI, Context->getConstantExprNot(CI));
7312 return BinaryOperator::CreateNot(Result);
7315 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
7316 return commonShiftTransforms(I);
7319 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
7320 return commonShiftTransforms(I);
7323 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
7324 if (Instruction *R = commonShiftTransforms(I))
7327 Value *Op0 = I.getOperand(0);
7329 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
7330 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
7331 if (CSI->isAllOnesValue())
7332 return ReplaceInstUsesWith(I, CSI);
7334 // See if we can turn a signed shr into an unsigned shr.
7335 if (MaskedValueIsZero(Op0,
7336 APInt::getSignBit(I.getType()->getScalarSizeInBits())))
7337 return BinaryOperator::CreateLShr(Op0, I.getOperand(1));
7339 // Arithmetic shifting an all-sign-bit value is a no-op.
7340 unsigned NumSignBits = ComputeNumSignBits(Op0);
7341 if (NumSignBits == Op0->getType()->getScalarSizeInBits())
7342 return ReplaceInstUsesWith(I, Op0);
7347 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
7348 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
7349 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
7351 // shl X, 0 == X and shr X, 0 == X
7352 // shl 0, X == 0 and shr 0, X == 0
7353 if (Op1 == Context->getNullValue(Op1->getType()) ||
7354 Op0 == Context->getNullValue(Op0->getType()))
7355 return ReplaceInstUsesWith(I, Op0);
7357 if (isa<UndefValue>(Op0)) {
7358 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
7359 return ReplaceInstUsesWith(I, Op0);
7360 else // undef << X -> 0, undef >>u X -> 0
7361 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7363 if (isa<UndefValue>(Op1)) {
7364 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
7365 return ReplaceInstUsesWith(I, Op0);
7366 else // X << undef, X >>u undef -> 0
7367 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7370 // See if we can fold away this shift.
7371 if (SimplifyDemandedInstructionBits(I))
7374 // Try to fold constant and into select arguments.
7375 if (isa<Constant>(Op0))
7376 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
7377 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7380 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
7381 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
7386 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
7387 BinaryOperator &I) {
7388 bool isLeftShift = I.getOpcode() == Instruction::Shl;
7390 // See if we can simplify any instructions used by the instruction whose sole
7391 // purpose is to compute bits we don't care about.
7392 uint32_t TypeBits = Op0->getType()->getScalarSizeInBits();
7394 // shl i32 X, 32 = 0 and srl i8 Y, 9 = 0, ... just don't eliminate
7397 if (Op1->uge(TypeBits)) {
7398 if (I.getOpcode() != Instruction::AShr)
7399 return ReplaceInstUsesWith(I, Context->getNullValue(Op0->getType()));
7401 I.setOperand(1, Context->getConstantInt(I.getType(), TypeBits-1));
7406 // ((X*C1) << C2) == (X * (C1 << C2))
7407 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
7408 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
7409 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
7410 return BinaryOperator::CreateMul(BO->getOperand(0),
7411 Context->getConstantExprShl(BOOp, Op1));
7413 // Try to fold constant and into select arguments.
7414 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
7415 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
7417 if (isa<PHINode>(Op0))
7418 if (Instruction *NV = FoldOpIntoPhi(I))
7421 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
7422 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
7423 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
7424 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
7425 // place. Don't try to do this transformation in this case. Also, we
7426 // require that the input operand is a shift-by-constant so that we have
7427 // confidence that the shifts will get folded together. We could do this
7428 // xform in more cases, but it is unlikely to be profitable.
7429 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
7430 isa<ConstantInt>(TrOp->getOperand(1))) {
7431 // Okay, we'll do this xform. Make the shift of shift.
7432 Constant *ShAmt = Context->getConstantExprZExt(Op1, TrOp->getType());
7433 Instruction *NSh = BinaryOperator::Create(I.getOpcode(), TrOp, ShAmt,
7435 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
7437 // For logical shifts, the truncation has the effect of making the high
7438 // part of the register be zeros. Emulate this by inserting an AND to
7439 // clear the top bits as needed. This 'and' will usually be zapped by
7440 // other xforms later if dead.
7441 unsigned SrcSize = TrOp->getType()->getScalarSizeInBits();
7442 unsigned DstSize = TI->getType()->getScalarSizeInBits();
7443 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
7445 // The mask we constructed says what the trunc would do if occurring
7446 // between the shifts. We want to know the effect *after* the second
7447 // shift. We know that it is a logical shift by a constant, so adjust the
7448 // mask as appropriate.
7449 if (I.getOpcode() == Instruction::Shl)
7450 MaskV <<= Op1->getZExtValue();
7452 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
7453 MaskV = MaskV.lshr(Op1->getZExtValue());
7457 BinaryOperator::CreateAnd(NSh, Context->getConstantInt(MaskV),
7459 InsertNewInstBefore(And, I); // shift1 & 0x00FF
7461 // Return the value truncated to the interesting size.
7462 return new TruncInst(And, I.getType());
7466 if (Op0->hasOneUse()) {
7467 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
7468 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7471 switch (Op0BO->getOpcode()) {
7473 case Instruction::Add:
7474 case Instruction::And:
7475 case Instruction::Or:
7476 case Instruction::Xor: {
7477 // These operators commute.
7478 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
7479 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
7480 match(Op0BO->getOperand(1), m_Shr(m_Value(V1),
7481 m_Specific(Op1)), *Context)){
7482 Instruction *YS = BinaryOperator::CreateShl(
7483 Op0BO->getOperand(0), Op1,
7485 InsertNewInstBefore(YS, I); // (Y << C)
7487 BinaryOperator::Create(Op0BO->getOpcode(), YS, V1,
7488 Op0BO->getOperand(1)->getName());
7489 InsertNewInstBefore(X, I); // (X + (Y << C))
7490 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7491 return BinaryOperator::CreateAnd(X, Context->getConstantInt(
7492 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7495 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
7496 Value *Op0BOOp1 = Op0BO->getOperand(1);
7497 if (isLeftShift && Op0BOOp1->hasOneUse() &&
7499 m_And(m_Shr(m_Value(V1), m_Specific(Op1)),
7500 m_ConstantInt(CC)), *Context) &&
7501 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse()) {
7502 Instruction *YS = BinaryOperator::CreateShl(
7503 Op0BO->getOperand(0), Op1,
7505 InsertNewInstBefore(YS, I); // (Y << C)
7507 BinaryOperator::CreateAnd(V1,
7508 Context->getConstantExprShl(CC, Op1),
7509 V1->getName()+".mask");
7510 InsertNewInstBefore(XM, I); // X & (CC << C)
7512 return BinaryOperator::Create(Op0BO->getOpcode(), YS, XM);
7517 case Instruction::Sub: {
7518 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
7519 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7520 match(Op0BO->getOperand(0), m_Shr(m_Value(V1),
7521 m_Specific(Op1)), *Context)){
7522 Instruction *YS = BinaryOperator::CreateShl(
7523 Op0BO->getOperand(1), Op1,
7525 InsertNewInstBefore(YS, I); // (Y << C)
7527 BinaryOperator::Create(Op0BO->getOpcode(), V1, YS,
7528 Op0BO->getOperand(0)->getName());
7529 InsertNewInstBefore(X, I); // (X + (Y << C))
7530 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
7531 return BinaryOperator::CreateAnd(X, Context->getConstantInt(
7532 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
7535 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
7536 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
7537 match(Op0BO->getOperand(0),
7538 m_And(m_Shr(m_Value(V1), m_Value(V2)),
7539 m_ConstantInt(CC)), *Context) && V2 == Op1 &&
7540 cast<BinaryOperator>(Op0BO->getOperand(0))
7541 ->getOperand(0)->hasOneUse()) {
7542 Instruction *YS = BinaryOperator::CreateShl(
7543 Op0BO->getOperand(1), Op1,
7545 InsertNewInstBefore(YS, I); // (Y << C)
7547 BinaryOperator::CreateAnd(V1,
7548 Context->getConstantExprShl(CC, Op1),
7549 V1->getName()+".mask");
7550 InsertNewInstBefore(XM, I); // X & (CC << C)
7552 return BinaryOperator::Create(Op0BO->getOpcode(), XM, YS);
7560 // If the operand is an bitwise operator with a constant RHS, and the
7561 // shift is the only use, we can pull it out of the shift.
7562 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
7563 bool isValid = true; // Valid only for And, Or, Xor
7564 bool highBitSet = false; // Transform if high bit of constant set?
7566 switch (Op0BO->getOpcode()) {
7567 default: isValid = false; break; // Do not perform transform!
7568 case Instruction::Add:
7569 isValid = isLeftShift;
7571 case Instruction::Or:
7572 case Instruction::Xor:
7575 case Instruction::And:
7580 // If this is a signed shift right, and the high bit is modified
7581 // by the logical operation, do not perform the transformation.
7582 // The highBitSet boolean indicates the value of the high bit of
7583 // the constant which would cause it to be modified for this
7586 if (isValid && I.getOpcode() == Instruction::AShr)
7587 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
7590 Constant *NewRHS = Context->getConstantExpr(I.getOpcode(), Op0C, Op1);
7592 Instruction *NewShift =
7593 BinaryOperator::Create(I.getOpcode(), Op0BO->getOperand(0), Op1);
7594 InsertNewInstBefore(NewShift, I);
7595 NewShift->takeName(Op0BO);
7597 return BinaryOperator::Create(Op0BO->getOpcode(), NewShift,
7604 // Find out if this is a shift of a shift by a constant.
7605 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
7606 if (ShiftOp && !ShiftOp->isShift())
7609 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
7610 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
7611 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
7612 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
7613 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
7614 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
7615 Value *X = ShiftOp->getOperand(0);
7617 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
7619 const IntegerType *Ty = cast<IntegerType>(I.getType());
7621 // Check for (X << c1) << c2 and (X >> c1) >> c2
7622 if (I.getOpcode() == ShiftOp->getOpcode()) {
7623 // If this is oversized composite shift, then unsigned shifts get 0, ashr
7625 if (AmtSum >= TypeBits) {
7626 if (I.getOpcode() != Instruction::AShr)
7627 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7628 AmtSum = TypeBits-1; // Saturate to 31 for i32 ashr.
7631 return BinaryOperator::Create(I.getOpcode(), X,
7632 Context->getConstantInt(Ty, AmtSum));
7633 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
7634 I.getOpcode() == Instruction::AShr) {
7635 if (AmtSum >= TypeBits)
7636 return ReplaceInstUsesWith(I, Context->getNullValue(I.getType()));
7638 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
7639 return BinaryOperator::CreateLShr(X, Context->getConstantInt(Ty, AmtSum));
7640 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
7641 I.getOpcode() == Instruction::LShr) {
7642 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
7643 if (AmtSum >= TypeBits)
7644 AmtSum = TypeBits-1;
7646 Instruction *Shift =
7647 BinaryOperator::CreateAShr(X, Context->getConstantInt(Ty, AmtSum));
7648 InsertNewInstBefore(Shift, I);
7650 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7651 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7654 // Okay, if we get here, one shift must be left, and the other shift must be
7655 // right. See if the amounts are equal.
7656 if (ShiftAmt1 == ShiftAmt2) {
7657 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
7658 if (I.getOpcode() == Instruction::Shl) {
7659 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
7660 return BinaryOperator::CreateAnd(X, Context->getConstantInt(Mask));
7662 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
7663 if (I.getOpcode() == Instruction::LShr) {
7664 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
7665 return BinaryOperator::CreateAnd(X, Context->getConstantInt(Mask));
7667 // We can simplify ((X << C) >>s C) into a trunc + sext.
7668 // NOTE: we could do this for any C, but that would make 'unusual' integer
7669 // types. For now, just stick to ones well-supported by the code
7671 const Type *SExtType = 0;
7672 switch (Ty->getBitWidth() - ShiftAmt1) {
7679 SExtType = Context->getIntegerType(Ty->getBitWidth() - ShiftAmt1);
7684 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
7685 InsertNewInstBefore(NewTrunc, I);
7686 return new SExtInst(NewTrunc, Ty);
7688 // Otherwise, we can't handle it yet.
7689 } else if (ShiftAmt1 < ShiftAmt2) {
7690 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
7692 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
7693 if (I.getOpcode() == Instruction::Shl) {
7694 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7695 ShiftOp->getOpcode() == Instruction::AShr);
7696 Instruction *Shift =
7697 BinaryOperator::CreateShl(X, Context->getConstantInt(Ty, ShiftDiff));
7698 InsertNewInstBefore(Shift, I);
7700 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7701 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7704 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
7705 if (I.getOpcode() == Instruction::LShr) {
7706 assert(ShiftOp->getOpcode() == Instruction::Shl);
7707 Instruction *Shift =
7708 BinaryOperator::CreateLShr(X, Context->getConstantInt(Ty, ShiftDiff));
7709 InsertNewInstBefore(Shift, I);
7711 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7712 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7715 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
7717 assert(ShiftAmt2 < ShiftAmt1);
7718 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
7720 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
7721 if (I.getOpcode() == Instruction::Shl) {
7722 assert(ShiftOp->getOpcode() == Instruction::LShr ||
7723 ShiftOp->getOpcode() == Instruction::AShr);
7724 Instruction *Shift =
7725 BinaryOperator::Create(ShiftOp->getOpcode(), X,
7726 Context->getConstantInt(Ty, ShiftDiff));
7727 InsertNewInstBefore(Shift, I);
7729 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
7730 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7733 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
7734 if (I.getOpcode() == Instruction::LShr) {
7735 assert(ShiftOp->getOpcode() == Instruction::Shl);
7736 Instruction *Shift =
7737 BinaryOperator::CreateShl(X, Context->getConstantInt(Ty, ShiftDiff));
7738 InsertNewInstBefore(Shift, I);
7740 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
7741 return BinaryOperator::CreateAnd(Shift, Context->getConstantInt(Mask));
7744 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
7751 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
7752 /// expression. If so, decompose it, returning some value X, such that Val is
7755 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
7756 int &Offset, LLVMContext *Context) {
7757 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
7758 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
7759 Offset = CI->getZExtValue();
7761 return Context->getConstantInt(Type::Int32Ty, 0);
7762 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
7763 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
7764 if (I->getOpcode() == Instruction::Shl) {
7765 // This is a value scaled by '1 << the shift amt'.
7766 Scale = 1U << RHS->getZExtValue();
7768 return I->getOperand(0);
7769 } else if (I->getOpcode() == Instruction::Mul) {
7770 // This value is scaled by 'RHS'.
7771 Scale = RHS->getZExtValue();
7773 return I->getOperand(0);
7774 } else if (I->getOpcode() == Instruction::Add) {
7775 // We have X+C. Check to see if we really have (X*C2)+C1,
7776 // where C1 is divisible by C2.
7779 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale,
7781 Offset += RHS->getZExtValue();
7788 // Otherwise, we can't look past this.
7795 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
7796 /// try to eliminate the cast by moving the type information into the alloc.
7797 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
7798 AllocationInst &AI) {
7799 const PointerType *PTy = cast<PointerType>(CI.getType());
7801 // Remove any uses of AI that are dead.
7802 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
7804 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
7805 Instruction *User = cast<Instruction>(*UI++);
7806 if (isInstructionTriviallyDead(User)) {
7807 while (UI != E && *UI == User)
7808 ++UI; // If this instruction uses AI more than once, don't break UI.
7811 DOUT << "IC: DCE: " << *User;
7812 EraseInstFromFunction(*User);
7816 // Get the type really allocated and the type casted to.
7817 const Type *AllocElTy = AI.getAllocatedType();
7818 const Type *CastElTy = PTy->getElementType();
7819 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
7821 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
7822 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
7823 if (CastElTyAlign < AllocElTyAlign) return 0;
7825 // If the allocation has multiple uses, only promote it if we are strictly
7826 // increasing the alignment of the resultant allocation. If we keep it the
7827 // same, we open the door to infinite loops of various kinds. (A reference
7828 // from a dbg.declare doesn't count as a use for this purpose.)
7829 if (!AI.hasOneUse() && !hasOneUsePlusDeclare(&AI) &&
7830 CastElTyAlign == AllocElTyAlign) return 0;
7832 uint64_t AllocElTySize = TD->getTypeAllocSize(AllocElTy);
7833 uint64_t CastElTySize = TD->getTypeAllocSize(CastElTy);
7834 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
7836 // See if we can satisfy the modulus by pulling a scale out of the array
7838 unsigned ArraySizeScale;
7840 Value *NumElements = // See if the array size is a decomposable linear expr.
7841 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale,
7842 ArrayOffset, Context);
7844 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7846 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7847 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7849 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7854 // If the allocation size is constant, form a constant mul expression
7855 Amt = Context->getConstantInt(Type::Int32Ty, Scale);
7856 if (isa<ConstantInt>(NumElements))
7857 Amt = Context->getConstantExprMul(cast<ConstantInt>(NumElements),
7858 cast<ConstantInt>(Amt));
7859 // otherwise multiply the amount and the number of elements
7861 Instruction *Tmp = BinaryOperator::CreateMul(Amt, NumElements, "tmp");
7862 Amt = InsertNewInstBefore(Tmp, AI);
7866 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7867 Value *Off = Context->getConstantInt(Type::Int32Ty, Offset, true);
7868 Instruction *Tmp = BinaryOperator::CreateAdd(Amt, Off, "tmp");
7869 Amt = InsertNewInstBefore(Tmp, AI);
7872 AllocationInst *New;
7873 if (isa<MallocInst>(AI))
7874 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7876 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7877 InsertNewInstBefore(New, AI);
7880 // If the allocation has one real use plus a dbg.declare, just remove the
7882 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(&AI)) {
7883 EraseInstFromFunction(*DI);
7885 // If the allocation has multiple real uses, insert a cast and change all
7886 // things that used it to use the new cast. This will also hack on CI, but it
7888 else if (!AI.hasOneUse()) {
7889 AddUsesToWorkList(AI);
7890 // New is the allocation instruction, pointer typed. AI is the original
7891 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7892 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7893 InsertNewInstBefore(NewCast, AI);
7894 AI.replaceAllUsesWith(NewCast);
7896 return ReplaceInstUsesWith(CI, New);
7899 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7900 /// and return it as type Ty without inserting any new casts and without
7901 /// changing the computed value. This is used by code that tries to decide
7902 /// whether promoting or shrinking integer operations to wider or smaller types
7903 /// will allow us to eliminate a truncate or extend.
7905 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7906 /// extension operation if Ty is larger.
7908 /// If CastOpc is a truncation, then Ty will be a type smaller than V. We
7909 /// should return true if trunc(V) can be computed by computing V in the smaller
7910 /// type. If V is an instruction, then trunc(inst(x,y)) can be computed as
7911 /// inst(trunc(x),trunc(y)), which only makes sense if x and y can be
7912 /// efficiently truncated.
7914 /// If CastOpc is a sext or zext, we are asking if the low bits of the value can
7915 /// bit computed in a larger type, which is then and'd or sext_in_reg'd to get
7916 /// the final result.
7917 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const Type *Ty,
7919 int &NumCastsRemoved){
7920 // We can always evaluate constants in another type.
7921 if (isa<Constant>(V))
7924 Instruction *I = dyn_cast<Instruction>(V);
7925 if (!I) return false;
7927 const Type *OrigTy = V->getType();
7929 // If this is an extension or truncate, we can often eliminate it.
7930 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7931 // If this is a cast from the destination type, we can trivially eliminate
7932 // it, and this will remove a cast overall.
7933 if (I->getOperand(0)->getType() == Ty) {
7934 // If the first operand is itself a cast, and is eliminable, do not count
7935 // this as an eliminable cast. We would prefer to eliminate those two
7937 if (!isa<CastInst>(I->getOperand(0)) && I->hasOneUse())
7943 // We can't extend or shrink something that has multiple uses: doing so would
7944 // require duplicating the instruction in general, which isn't profitable.
7945 if (!I->hasOneUse()) return false;
7947 unsigned Opc = I->getOpcode();
7949 case Instruction::Add:
7950 case Instruction::Sub:
7951 case Instruction::Mul:
7952 case Instruction::And:
7953 case Instruction::Or:
7954 case Instruction::Xor:
7955 // These operators can all arbitrarily be extended or truncated.
7956 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7958 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7961 case Instruction::Shl:
7962 // If we are truncating the result of this SHL, and if it's a shift of a
7963 // constant amount, we can always perform a SHL in a smaller type.
7964 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7965 uint32_t BitWidth = Ty->getScalarSizeInBits();
7966 if (BitWidth < OrigTy->getScalarSizeInBits() &&
7967 CI->getLimitedValue(BitWidth) < BitWidth)
7968 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7972 case Instruction::LShr:
7973 // If this is a truncate of a logical shr, we can truncate it to a smaller
7974 // lshr iff we know that the bits we would otherwise be shifting in are
7976 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7977 uint32_t OrigBitWidth = OrigTy->getScalarSizeInBits();
7978 uint32_t BitWidth = Ty->getScalarSizeInBits();
7979 if (BitWidth < OrigBitWidth &&
7980 MaskedValueIsZero(I->getOperand(0),
7981 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7982 CI->getLimitedValue(BitWidth) < BitWidth) {
7983 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7988 case Instruction::ZExt:
7989 case Instruction::SExt:
7990 case Instruction::Trunc:
7991 // If this is the same kind of case as our original (e.g. zext+zext), we
7992 // can safely replace it. Note that replacing it does not reduce the number
7993 // of casts in the input.
7997 // sext (zext ty1), ty2 -> zext ty2
7998 if (CastOpc == Instruction::SExt && Opc == Instruction::ZExt)
8001 case Instruction::Select: {
8002 SelectInst *SI = cast<SelectInst>(I);
8003 return CanEvaluateInDifferentType(SI->getTrueValue(), Ty, CastOpc,
8005 CanEvaluateInDifferentType(SI->getFalseValue(), Ty, CastOpc,
8008 case Instruction::PHI: {
8009 // We can change a phi if we can change all operands.
8010 PHINode *PN = cast<PHINode>(I);
8011 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
8012 if (!CanEvaluateInDifferentType(PN->getIncomingValue(i), Ty, CastOpc,
8018 // TODO: Can handle more cases here.
8025 /// EvaluateInDifferentType - Given an expression that
8026 /// CanEvaluateInDifferentType returns true for, actually insert the code to
8027 /// evaluate the expression.
8028 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
8030 if (Constant *C = dyn_cast<Constant>(V))
8031 return Context->getConstantExprIntegerCast(C, Ty,
8032 isSigned /*Sext or ZExt*/);
8034 // Otherwise, it must be an instruction.
8035 Instruction *I = cast<Instruction>(V);
8036 Instruction *Res = 0;
8037 unsigned Opc = I->getOpcode();
8039 case Instruction::Add:
8040 case Instruction::Sub:
8041 case Instruction::Mul:
8042 case Instruction::And:
8043 case Instruction::Or:
8044 case Instruction::Xor:
8045 case Instruction::AShr:
8046 case Instruction::LShr:
8047 case Instruction::Shl: {
8048 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
8049 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8050 Res = BinaryOperator::Create((Instruction::BinaryOps)Opc, LHS, RHS);
8053 case Instruction::Trunc:
8054 case Instruction::ZExt:
8055 case Instruction::SExt:
8056 // If the source type of the cast is the type we're trying for then we can
8057 // just return the source. There's no need to insert it because it is not
8059 if (I->getOperand(0)->getType() == Ty)
8060 return I->getOperand(0);
8062 // Otherwise, must be the same type of cast, so just reinsert a new one.
8063 Res = CastInst::Create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
8066 case Instruction::Select: {
8067 Value *True = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
8068 Value *False = EvaluateInDifferentType(I->getOperand(2), Ty, isSigned);
8069 Res = SelectInst::Create(I->getOperand(0), True, False);
8072 case Instruction::PHI: {
8073 PHINode *OPN = cast<PHINode>(I);
8074 PHINode *NPN = PHINode::Create(Ty);
8075 for (unsigned i = 0, e = OPN->getNumIncomingValues(); i != e; ++i) {
8076 Value *V =EvaluateInDifferentType(OPN->getIncomingValue(i), Ty, isSigned);
8077 NPN->addIncoming(V, OPN->getIncomingBlock(i));
8083 // TODO: Can handle more cases here.
8084 LLVM_UNREACHABLE("Unreachable!");
8089 return InsertNewInstBefore(Res, *I);
8092 /// @brief Implement the transforms common to all CastInst visitors.
8093 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
8094 Value *Src = CI.getOperand(0);
8096 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
8097 // eliminate it now.
8098 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
8099 if (Instruction::CastOps opc =
8100 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
8101 // The first cast (CSrc) is eliminable so we need to fix up or replace
8102 // the second cast (CI). CSrc will then have a good chance of being dead.
8103 return CastInst::Create(opc, CSrc->getOperand(0), CI.getType());
8107 // If we are casting a select then fold the cast into the select
8108 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
8109 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
8112 // If we are casting a PHI then fold the cast into the PHI
8113 if (isa<PHINode>(Src))
8114 if (Instruction *NV = FoldOpIntoPhi(CI))
8120 /// FindElementAtOffset - Given a type and a constant offset, determine whether
8121 /// or not there is a sequence of GEP indices into the type that will land us at
8122 /// the specified offset. If so, fill them into NewIndices and return the
8123 /// resultant element type, otherwise return null.
8124 static const Type *FindElementAtOffset(const Type *Ty, int64_t Offset,
8125 SmallVectorImpl<Value*> &NewIndices,
8126 const TargetData *TD,
8127 LLVMContext *Context) {
8128 if (!Ty->isSized()) return 0;
8130 // Start with the index over the outer type. Note that the type size
8131 // might be zero (even if the offset isn't zero) if the indexed type
8132 // is something like [0 x {int, int}]
8133 const Type *IntPtrTy = TD->getIntPtrType();
8134 int64_t FirstIdx = 0;
8135 if (int64_t TySize = TD->getTypeAllocSize(Ty)) {
8136 FirstIdx = Offset/TySize;
8137 Offset -= FirstIdx*TySize;
8139 // Handle hosts where % returns negative instead of values [0..TySize).
8143 assert(Offset >= 0);
8145 assert((uint64_t)Offset < (uint64_t)TySize && "Out of range offset");
8148 NewIndices.push_back(Context->getConstantInt(IntPtrTy, FirstIdx));
8150 // Index into the types. If we fail, set OrigBase to null.
8152 // Indexing into tail padding between struct/array elements.
8153 if (uint64_t(Offset*8) >= TD->getTypeSizeInBits(Ty))
8156 if (const StructType *STy = dyn_cast<StructType>(Ty)) {
8157 const StructLayout *SL = TD->getStructLayout(STy);
8158 assert(Offset < (int64_t)SL->getSizeInBytes() &&
8159 "Offset must stay within the indexed type");
8161 unsigned Elt = SL->getElementContainingOffset(Offset);
8162 NewIndices.push_back(Context->getConstantInt(Type::Int32Ty, Elt));
8164 Offset -= SL->getElementOffset(Elt);
8165 Ty = STy->getElementType(Elt);
8166 } else if (const ArrayType *AT = dyn_cast<ArrayType>(Ty)) {
8167 uint64_t EltSize = TD->getTypeAllocSize(AT->getElementType());
8168 assert(EltSize && "Cannot index into a zero-sized array");
8169 NewIndices.push_back(Context->getConstantInt(IntPtrTy,Offset/EltSize));
8171 Ty = AT->getElementType();
8173 // Otherwise, we can't index into the middle of this atomic type, bail.
8181 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
8182 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
8183 Value *Src = CI.getOperand(0);
8185 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
8186 // If casting the result of a getelementptr instruction with no offset, turn
8187 // this into a cast of the original pointer!
8188 if (GEP->hasAllZeroIndices()) {
8189 // Changing the cast operand is usually not a good idea but it is safe
8190 // here because the pointer operand is being replaced with another
8191 // pointer operand so the opcode doesn't need to change.
8193 CI.setOperand(0, GEP->getOperand(0));
8197 // If the GEP has a single use, and the base pointer is a bitcast, and the
8198 // GEP computes a constant offset, see if we can convert these three
8199 // instructions into fewer. This typically happens with unions and other
8200 // non-type-safe code.
8201 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
8202 if (GEP->hasAllConstantIndices()) {
8203 // We are guaranteed to get a constant from EmitGEPOffset.
8204 ConstantInt *OffsetV =
8205 cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
8206 int64_t Offset = OffsetV->getSExtValue();
8208 // Get the base pointer input of the bitcast, and the type it points to.
8209 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
8210 const Type *GEPIdxTy =
8211 cast<PointerType>(OrigBase->getType())->getElementType();
8212 SmallVector<Value*, 8> NewIndices;
8213 if (FindElementAtOffset(GEPIdxTy, Offset, NewIndices, TD, Context)) {
8214 // If we were able to index down into an element, create the GEP
8215 // and bitcast the result. This eliminates one bitcast, potentially
8217 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
8219 NewIndices.end(), "");
8220 InsertNewInstBefore(NGEP, CI);
8221 NGEP->takeName(GEP);
8223 if (isa<BitCastInst>(CI))
8224 return new BitCastInst(NGEP, CI.getType());
8225 assert(isa<PtrToIntInst>(CI));
8226 return new PtrToIntInst(NGEP, CI.getType());
8232 return commonCastTransforms(CI);
8235 /// isSafeIntegerType - Return true if this is a basic integer type, not a crazy
8236 /// type like i42. We don't want to introduce operations on random non-legal
8237 /// integer types where they don't already exist in the code. In the future,
8238 /// we should consider making this based off target-data, so that 32-bit targets
8239 /// won't get i64 operations etc.
8240 static bool isSafeIntegerType(const Type *Ty) {
8241 switch (Ty->getPrimitiveSizeInBits()) {
8252 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
8253 /// integer types. This function implements the common transforms for all those
8255 /// @brief Implement the transforms common to CastInst with integer operands
8256 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
8257 if (Instruction *Result = commonCastTransforms(CI))
8260 Value *Src = CI.getOperand(0);
8261 const Type *SrcTy = Src->getType();
8262 const Type *DestTy = CI.getType();
8263 uint32_t SrcBitSize = SrcTy->getScalarSizeInBits();
8264 uint32_t DestBitSize = DestTy->getScalarSizeInBits();
8266 // See if we can simplify any instructions used by the LHS whose sole
8267 // purpose is to compute bits we don't care about.
8268 if (SimplifyDemandedInstructionBits(CI))
8271 // If the source isn't an instruction or has more than one use then we
8272 // can't do anything more.
8273 Instruction *SrcI = dyn_cast<Instruction>(Src);
8274 if (!SrcI || !Src->hasOneUse())
8277 // Attempt to propagate the cast into the instruction for int->int casts.
8278 int NumCastsRemoved = 0;
8279 if (!isa<BitCastInst>(CI) &&
8280 // Only do this if the dest type is a simple type, don't convert the
8281 // expression tree to something weird like i93 unless the source is also
8283 (isSafeIntegerType(DestTy->getScalarType()) ||
8284 !isSafeIntegerType(SrcI->getType()->getScalarType())) &&
8285 CanEvaluateInDifferentType(SrcI, DestTy,
8286 CI.getOpcode(), NumCastsRemoved)) {
8287 // If this cast is a truncate, evaluting in a different type always
8288 // eliminates the cast, so it is always a win. If this is a zero-extension,
8289 // we need to do an AND to maintain the clear top-part of the computation,
8290 // so we require that the input have eliminated at least one cast. If this
8291 // is a sign extension, we insert two new casts (to do the extension) so we
8292 // require that two casts have been eliminated.
8293 bool DoXForm = false;
8294 bool JustReplace = false;
8295 switch (CI.getOpcode()) {
8297 // All the others use floating point so we shouldn't actually
8298 // get here because of the check above.
8299 LLVM_UNREACHABLE("Unknown cast type");
8300 case Instruction::Trunc:
8303 case Instruction::ZExt: {
8304 DoXForm = NumCastsRemoved >= 1;
8305 if (!DoXForm && 0) {
8306 // If it's unnecessary to issue an AND to clear the high bits, it's
8307 // always profitable to do this xform.
8308 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, false);
8309 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8310 if (MaskedValueIsZero(TryRes, Mask))
8311 return ReplaceInstUsesWith(CI, TryRes);
8313 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8314 if (TryI->use_empty())
8315 EraseInstFromFunction(*TryI);
8319 case Instruction::SExt: {
8320 DoXForm = NumCastsRemoved >= 2;
8321 if (!DoXForm && !isa<TruncInst>(SrcI) && 0) {
8322 // If we do not have to emit the truncate + sext pair, then it's always
8323 // profitable to do this xform.
8325 // It's not safe to eliminate the trunc + sext pair if one of the
8326 // eliminated cast is a truncate. e.g.
8327 // t2 = trunc i32 t1 to i16
8328 // t3 = sext i16 t2 to i32
8331 Value *TryRes = EvaluateInDifferentType(SrcI, DestTy, true);
8332 unsigned NumSignBits = ComputeNumSignBits(TryRes);
8333 if (NumSignBits > (DestBitSize - SrcBitSize))
8334 return ReplaceInstUsesWith(CI, TryRes);
8336 if (Instruction *TryI = dyn_cast<Instruction>(TryRes))
8337 if (TryI->use_empty())
8338 EraseInstFromFunction(*TryI);
8345 DOUT << "ICE: EvaluateInDifferentType converting expression type to avoid"
8347 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
8348 CI.getOpcode() == Instruction::SExt);
8350 // Just replace this cast with the result.
8351 return ReplaceInstUsesWith(CI, Res);
8353 assert(Res->getType() == DestTy);
8354 switch (CI.getOpcode()) {
8355 default: LLVM_UNREACHABLE("Unknown cast type!");
8356 case Instruction::Trunc:
8357 case Instruction::BitCast:
8358 // Just replace this cast with the result.
8359 return ReplaceInstUsesWith(CI, Res);
8360 case Instruction::ZExt: {
8361 assert(SrcBitSize < DestBitSize && "Not a zext?");
8363 // If the high bits are already zero, just replace this cast with the
8365 APInt Mask(APInt::getBitsSet(DestBitSize, SrcBitSize, DestBitSize));
8366 if (MaskedValueIsZero(Res, Mask))
8367 return ReplaceInstUsesWith(CI, Res);
8369 // We need to emit an AND to clear the high bits.
8370 Constant *C = Context->getConstantInt(APInt::getLowBitsSet(DestBitSize,
8372 return BinaryOperator::CreateAnd(Res, C);
8374 case Instruction::SExt: {
8375 // If the high bits are already filled with sign bit, just replace this
8376 // cast with the result.
8377 unsigned NumSignBits = ComputeNumSignBits(Res);
8378 if (NumSignBits > (DestBitSize - SrcBitSize))
8379 return ReplaceInstUsesWith(CI, Res);
8381 // We need to emit a cast to truncate, then a cast to sext.
8382 return CastInst::Create(Instruction::SExt,
8383 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
8390 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
8391 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
8393 switch (SrcI->getOpcode()) {
8394 case Instruction::Add:
8395 case Instruction::Mul:
8396 case Instruction::And:
8397 case Instruction::Or:
8398 case Instruction::Xor:
8399 // If we are discarding information, rewrite.
8400 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
8401 // Don't insert two casts if they cannot be eliminated. We allow
8402 // two casts to be inserted if the sizes are the same. This could
8403 // only be converting signedness, which is a noop.
8404 if (DestBitSize == SrcBitSize ||
8405 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
8406 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8407 Instruction::CastOps opcode = CI.getOpcode();
8408 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8409 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8410 return BinaryOperator::Create(
8411 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8415 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
8416 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
8417 SrcI->getOpcode() == Instruction::Xor &&
8418 Op1 == Context->getConstantIntTrue() &&
8419 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
8420 Value *New = InsertCastBefore(Instruction::ZExt, Op0, DestTy, CI);
8421 return BinaryOperator::CreateXor(New,
8422 Context->getConstantInt(CI.getType(), 1));
8425 case Instruction::SDiv:
8426 case Instruction::UDiv:
8427 case Instruction::SRem:
8428 case Instruction::URem:
8429 // If we are just changing the sign, rewrite.
8430 if (DestBitSize == SrcBitSize) {
8431 // Don't insert two casts if they cannot be eliminated. We allow
8432 // two casts to be inserted if the sizes are the same. This could
8433 // only be converting signedness, which is a noop.
8434 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
8435 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
8436 Value *Op0c = InsertCastBefore(Instruction::BitCast,
8437 Op0, DestTy, *SrcI);
8438 Value *Op1c = InsertCastBefore(Instruction::BitCast,
8439 Op1, DestTy, *SrcI);
8440 return BinaryOperator::Create(
8441 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
8446 case Instruction::Shl:
8447 // Allow changing the sign of the source operand. Do not allow
8448 // changing the size of the shift, UNLESS the shift amount is a
8449 // constant. We must not change variable sized shifts to a smaller
8450 // size, because it is undefined to shift more bits out than exist
8452 if (DestBitSize == SrcBitSize ||
8453 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
8454 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
8455 Instruction::BitCast : Instruction::Trunc);
8456 Value *Op0c = InsertCastBefore(opcode, Op0, DestTy, *SrcI);
8457 Value *Op1c = InsertCastBefore(opcode, Op1, DestTy, *SrcI);
8458 return BinaryOperator::CreateShl(Op0c, Op1c);
8461 case Instruction::AShr:
8462 // If this is a signed shr, and if all bits shifted in are about to be
8463 // truncated off, turn it into an unsigned shr to allow greater
8465 if (DestBitSize < SrcBitSize &&
8466 isa<ConstantInt>(Op1)) {
8467 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
8468 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
8469 // Insert the new logical shift right.
8470 return BinaryOperator::CreateLShr(Op0, Op1);
8478 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
8479 if (Instruction *Result = commonIntCastTransforms(CI))
8482 Value *Src = CI.getOperand(0);
8483 const Type *Ty = CI.getType();
8484 uint32_t DestBitWidth = Ty->getScalarSizeInBits();
8485 uint32_t SrcBitWidth = Src->getType()->getScalarSizeInBits();
8487 // Canonicalize trunc x to i1 -> (icmp ne (and x, 1), 0)
8488 if (DestBitWidth == 1 &&
8489 isa<VectorType>(Ty) == isa<VectorType>(Src->getType())) {
8490 Constant *One = Context->getConstantInt(Src->getType(), 1);
8491 Src = InsertNewInstBefore(BinaryOperator::CreateAnd(Src, One, "tmp"), CI);
8492 Value *Zero = Context->getNullValue(Src->getType());
8493 return new ICmpInst(*Context, ICmpInst::ICMP_NE, Src, Zero);
8496 // Optimize trunc(lshr(), c) to pull the shift through the truncate.
8497 ConstantInt *ShAmtV = 0;
8499 if (Src->hasOneUse() &&
8500 match(Src, m_LShr(m_Value(ShiftOp), m_ConstantInt(ShAmtV)), *Context)) {
8501 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
8503 // Get a mask for the bits shifting in.
8504 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
8505 if (MaskedValueIsZero(ShiftOp, Mask)) {
8506 if (ShAmt >= DestBitWidth) // All zeros.
8507 return ReplaceInstUsesWith(CI, Context->getNullValue(Ty));
8509 // Okay, we can shrink this. Truncate the input, then return a new
8511 Value *V1 = InsertCastBefore(Instruction::Trunc, ShiftOp, Ty, CI);
8512 Value *V2 = Context->getConstantExprTrunc(ShAmtV, Ty);
8513 return BinaryOperator::CreateLShr(V1, V2);
8520 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
8521 /// in order to eliminate the icmp.
8522 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
8524 // If we are just checking for a icmp eq of a single bit and zext'ing it
8525 // to an integer, then shift the bit to the appropriate place and then
8526 // cast to integer to avoid the comparison.
8527 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
8528 const APInt &Op1CV = Op1C->getValue();
8530 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
8531 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
8532 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
8533 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
8534 if (!DoXform) return ICI;
8536 Value *In = ICI->getOperand(0);
8537 Value *Sh = Context->getConstantInt(In->getType(),
8538 In->getType()->getScalarSizeInBits()-1);
8539 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In, Sh,
8540 In->getName()+".lobit"),
8542 if (In->getType() != CI.getType())
8543 In = CastInst::CreateIntegerCast(In, CI.getType(),
8544 false/*ZExt*/, "tmp", &CI);
8546 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
8547 Constant *One = Context->getConstantInt(In->getType(), 1);
8548 In = InsertNewInstBefore(BinaryOperator::CreateXor(In, One,
8549 In->getName()+".not"),
8553 return ReplaceInstUsesWith(CI, In);
8558 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
8559 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8560 // zext (X == 1) to i32 --> X iff X has only the low bit set.
8561 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
8562 // zext (X != 0) to i32 --> X iff X has only the low bit set.
8563 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
8564 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
8565 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
8566 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
8567 // This only works for EQ and NE
8568 ICI->isEquality()) {
8569 // If Op1C some other power of two, convert:
8570 uint32_t BitWidth = Op1C->getType()->getBitWidth();
8571 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8572 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
8573 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
8575 APInt KnownZeroMask(~KnownZero);
8576 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
8577 if (!DoXform) return ICI;
8579 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
8580 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
8581 // (X&4) == 2 --> false
8582 // (X&4) != 2 --> true
8583 Constant *Res = Context->getConstantInt(Type::Int1Ty, isNE);
8584 Res = Context->getConstantExprZExt(Res, CI.getType());
8585 return ReplaceInstUsesWith(CI, Res);
8588 uint32_t ShiftAmt = KnownZeroMask.logBase2();
8589 Value *In = ICI->getOperand(0);
8591 // Perform a logical shr by shiftamt.
8592 // Insert the shift to put the result in the low bit.
8593 In = InsertNewInstBefore(BinaryOperator::CreateLShr(In,
8594 Context->getConstantInt(In->getType(), ShiftAmt),
8595 In->getName()+".lobit"), CI);
8598 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
8599 Constant *One = Context->getConstantInt(In->getType(), 1);
8600 In = BinaryOperator::CreateXor(In, One, "tmp");
8601 InsertNewInstBefore(cast<Instruction>(In), CI);
8604 if (CI.getType() == In->getType())
8605 return ReplaceInstUsesWith(CI, In);
8607 return CastInst::CreateIntegerCast(In, CI.getType(), false/*ZExt*/);
8615 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
8616 // If one of the common conversion will work ..
8617 if (Instruction *Result = commonIntCastTransforms(CI))
8620 Value *Src = CI.getOperand(0);
8622 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
8623 // types and if the sizes are just right we can convert this into a logical
8624 // 'and' which will be much cheaper than the pair of casts.
8625 if (TruncInst *CSrc = dyn_cast<TruncInst>(Src)) { // A->B->C cast
8626 // Get the sizes of the types involved. We know that the intermediate type
8627 // will be smaller than A or C, but don't know the relation between A and C.
8628 Value *A = CSrc->getOperand(0);
8629 unsigned SrcSize = A->getType()->getScalarSizeInBits();
8630 unsigned MidSize = CSrc->getType()->getScalarSizeInBits();
8631 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8632 // If we're actually extending zero bits, then if
8633 // SrcSize < DstSize: zext(a & mask)
8634 // SrcSize == DstSize: a & mask
8635 // SrcSize > DstSize: trunc(a) & mask
8636 if (SrcSize < DstSize) {
8637 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8638 Constant *AndConst = Context->getConstantInt(A->getType(), AndValue);
8640 BinaryOperator::CreateAnd(A, AndConst, CSrc->getName()+".mask");
8641 InsertNewInstBefore(And, CI);
8642 return new ZExtInst(And, CI.getType());
8643 } else if (SrcSize == DstSize) {
8644 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
8645 return BinaryOperator::CreateAnd(A, Context->getConstantInt(A->getType(),
8647 } else if (SrcSize > DstSize) {
8648 Instruction *Trunc = new TruncInst(A, CI.getType(), "tmp");
8649 InsertNewInstBefore(Trunc, CI);
8650 APInt AndValue(APInt::getLowBitsSet(DstSize, MidSize));
8651 return BinaryOperator::CreateAnd(Trunc,
8652 Context->getConstantInt(Trunc->getType(),
8657 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
8658 return transformZExtICmp(ICI, CI);
8660 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
8661 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
8662 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
8663 // of the (zext icmp) will be transformed.
8664 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
8665 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
8666 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
8667 (transformZExtICmp(LHS, CI, false) ||
8668 transformZExtICmp(RHS, CI, false))) {
8669 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
8670 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
8671 return BinaryOperator::Create(Instruction::Or, LCast, RCast);
8675 // zext(trunc(t) & C) -> (t & zext(C)).
8676 if (SrcI && SrcI->getOpcode() == Instruction::And && SrcI->hasOneUse())
8677 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8678 if (TruncInst *TI = dyn_cast<TruncInst>(SrcI->getOperand(0))) {
8679 Value *TI0 = TI->getOperand(0);
8680 if (TI0->getType() == CI.getType())
8682 BinaryOperator::CreateAnd(TI0,
8683 Context->getConstantExprZExt(C, CI.getType()));
8686 // zext((trunc(t) & C) ^ C) -> ((t & zext(C)) ^ zext(C)).
8687 if (SrcI && SrcI->getOpcode() == Instruction::Xor && SrcI->hasOneUse())
8688 if (ConstantInt *C = dyn_cast<ConstantInt>(SrcI->getOperand(1)))
8689 if (BinaryOperator *And = dyn_cast<BinaryOperator>(SrcI->getOperand(0)))
8690 if (And->getOpcode() == Instruction::And && And->hasOneUse() &&
8691 And->getOperand(1) == C)
8692 if (TruncInst *TI = dyn_cast<TruncInst>(And->getOperand(0))) {
8693 Value *TI0 = TI->getOperand(0);
8694 if (TI0->getType() == CI.getType()) {
8695 Constant *ZC = Context->getConstantExprZExt(C, CI.getType());
8696 Instruction *NewAnd = BinaryOperator::CreateAnd(TI0, ZC, "tmp");
8697 InsertNewInstBefore(NewAnd, *And);
8698 return BinaryOperator::CreateXor(NewAnd, ZC);
8705 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
8706 if (Instruction *I = commonIntCastTransforms(CI))
8709 Value *Src = CI.getOperand(0);
8711 // Canonicalize sign-extend from i1 to a select.
8712 if (Src->getType() == Type::Int1Ty)
8713 return SelectInst::Create(Src,
8714 Context->getConstantIntAllOnesValue(CI.getType()),
8715 Context->getNullValue(CI.getType()));
8717 // See if the value being truncated is already sign extended. If so, just
8718 // eliminate the trunc/sext pair.
8719 if (getOpcode(Src) == Instruction::Trunc) {
8720 Value *Op = cast<User>(Src)->getOperand(0);
8721 unsigned OpBits = Op->getType()->getScalarSizeInBits();
8722 unsigned MidBits = Src->getType()->getScalarSizeInBits();
8723 unsigned DestBits = CI.getType()->getScalarSizeInBits();
8724 unsigned NumSignBits = ComputeNumSignBits(Op);
8726 if (OpBits == DestBits) {
8727 // Op is i32, Mid is i8, and Dest is i32. If Op has more than 24 sign
8728 // bits, it is already ready.
8729 if (NumSignBits > DestBits-MidBits)
8730 return ReplaceInstUsesWith(CI, Op);
8731 } else if (OpBits < DestBits) {
8732 // Op is i32, Mid is i8, and Dest is i64. If Op has more than 24 sign
8733 // bits, just sext from i32.
8734 if (NumSignBits > OpBits-MidBits)
8735 return new SExtInst(Op, CI.getType(), "tmp");
8737 // Op is i64, Mid is i8, and Dest is i32. If Op has more than 56 sign
8738 // bits, just truncate to i32.
8739 if (NumSignBits > OpBits-MidBits)
8740 return new TruncInst(Op, CI.getType(), "tmp");
8744 // If the input is a shl/ashr pair of a same constant, then this is a sign
8745 // extension from a smaller value. If we could trust arbitrary bitwidth
8746 // integers, we could turn this into a truncate to the smaller bit and then
8747 // use a sext for the whole extension. Since we don't, look deeper and check
8748 // for a truncate. If the source and dest are the same type, eliminate the
8749 // trunc and extend and just do shifts. For example, turn:
8750 // %a = trunc i32 %i to i8
8751 // %b = shl i8 %a, 6
8752 // %c = ashr i8 %b, 6
8753 // %d = sext i8 %c to i32
8755 // %a = shl i32 %i, 30
8756 // %d = ashr i32 %a, 30
8758 ConstantInt *BA = 0, *CA = 0;
8759 if (match(Src, m_AShr(m_Shl(m_Value(A), m_ConstantInt(BA)),
8760 m_ConstantInt(CA)), *Context) &&
8761 BA == CA && isa<TruncInst>(A)) {
8762 Value *I = cast<TruncInst>(A)->getOperand(0);
8763 if (I->getType() == CI.getType()) {
8764 unsigned MidSize = Src->getType()->getScalarSizeInBits();
8765 unsigned SrcDstSize = CI.getType()->getScalarSizeInBits();
8766 unsigned ShAmt = CA->getZExtValue()+SrcDstSize-MidSize;
8767 Constant *ShAmtV = Context->getConstantInt(CI.getType(), ShAmt);
8768 I = InsertNewInstBefore(BinaryOperator::CreateShl(I, ShAmtV,
8770 return BinaryOperator::CreateAShr(I, ShAmtV);
8777 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
8778 /// in the specified FP type without changing its value.
8779 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem,
8780 LLVMContext *Context) {
8782 APFloat F = CFP->getValueAPF();
8783 (void)F.convert(Sem, APFloat::rmNearestTiesToEven, &losesInfo);
8785 return Context->getConstantFP(F);
8789 /// LookThroughFPExtensions - If this is an fp extension instruction, look
8790 /// through it until we get the source value.
8791 static Value *LookThroughFPExtensions(Value *V, LLVMContext *Context) {
8792 if (Instruction *I = dyn_cast<Instruction>(V))
8793 if (I->getOpcode() == Instruction::FPExt)
8794 return LookThroughFPExtensions(I->getOperand(0), Context);
8796 // If this value is a constant, return the constant in the smallest FP type
8797 // that can accurately represent it. This allows us to turn
8798 // (float)((double)X+2.0) into x+2.0f.
8799 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
8800 if (CFP->getType() == Type::PPC_FP128Ty)
8801 return V; // No constant folding of this.
8802 // See if the value can be truncated to float and then reextended.
8803 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle, Context))
8805 if (CFP->getType() == Type::DoubleTy)
8806 return V; // Won't shrink.
8807 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble, Context))
8809 // Don't try to shrink to various long double types.
8815 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
8816 if (Instruction *I = commonCastTransforms(CI))
8819 // If we have fptrunc(fadd (fpextend x), (fpextend y)), where x and y are
8820 // smaller than the destination type, we can eliminate the truncate by doing
8821 // the add as the smaller type. This applies to fadd/fsub/fmul/fdiv as well as
8822 // many builtins (sqrt, etc).
8823 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
8824 if (OpI && OpI->hasOneUse()) {
8825 switch (OpI->getOpcode()) {
8827 case Instruction::FAdd:
8828 case Instruction::FSub:
8829 case Instruction::FMul:
8830 case Instruction::FDiv:
8831 case Instruction::FRem:
8832 const Type *SrcTy = OpI->getType();
8833 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0), Context);
8834 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1), Context);
8835 if (LHSTrunc->getType() != SrcTy &&
8836 RHSTrunc->getType() != SrcTy) {
8837 unsigned DstSize = CI.getType()->getScalarSizeInBits();
8838 // If the source types were both smaller than the destination type of
8839 // the cast, do this xform.
8840 if (LHSTrunc->getType()->getScalarSizeInBits() <= DstSize &&
8841 RHSTrunc->getType()->getScalarSizeInBits() <= DstSize) {
8842 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
8844 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
8846 return BinaryOperator::Create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
8855 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
8856 return commonCastTransforms(CI);
8859 Instruction *InstCombiner::visitFPToUI(FPToUIInst &FI) {
8860 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8862 return commonCastTransforms(FI);
8864 // fptoui(uitofp(X)) --> X
8865 // fptoui(sitofp(X)) --> X
8866 // This is safe if the intermediate type has enough bits in its mantissa to
8867 // accurately represent all values of X. For example, do not do this with
8868 // i64->float->i64. This is also safe for sitofp case, because any negative
8869 // 'X' value would cause an undefined result for the fptoui.
8870 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8871 OpI->getOperand(0)->getType() == FI.getType() &&
8872 (int)FI.getType()->getScalarSizeInBits() < /*extra bit for sign */
8873 OpI->getType()->getFPMantissaWidth())
8874 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8876 return commonCastTransforms(FI);
8879 Instruction *InstCombiner::visitFPToSI(FPToSIInst &FI) {
8880 Instruction *OpI = dyn_cast<Instruction>(FI.getOperand(0));
8882 return commonCastTransforms(FI);
8884 // fptosi(sitofp(X)) --> X
8885 // fptosi(uitofp(X)) --> X
8886 // This is safe if the intermediate type has enough bits in its mantissa to
8887 // accurately represent all values of X. For example, do not do this with
8888 // i64->float->i64. This is also safe for sitofp case, because any negative
8889 // 'X' value would cause an undefined result for the fptoui.
8890 if ((isa<UIToFPInst>(OpI) || isa<SIToFPInst>(OpI)) &&
8891 OpI->getOperand(0)->getType() == FI.getType() &&
8892 (int)FI.getType()->getScalarSizeInBits() <=
8893 OpI->getType()->getFPMantissaWidth())
8894 return ReplaceInstUsesWith(FI, OpI->getOperand(0));
8896 return commonCastTransforms(FI);
8899 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
8900 return commonCastTransforms(CI);
8903 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
8904 return commonCastTransforms(CI);
8907 Instruction *InstCombiner::visitPtrToInt(PtrToIntInst &CI) {
8908 // If the destination integer type is smaller than the intptr_t type for
8909 // this target, do a ptrtoint to intptr_t then do a trunc. This allows the
8910 // trunc to be exposed to other transforms. Don't do this for extending
8911 // ptrtoint's, because we don't know if the target sign or zero extends its
8913 if (CI.getType()->getScalarSizeInBits() < TD->getPointerSizeInBits()) {
8914 Value *P = InsertNewInstBefore(new PtrToIntInst(CI.getOperand(0),
8915 TD->getIntPtrType(),
8917 return new TruncInst(P, CI.getType());
8920 return commonPointerCastTransforms(CI);
8923 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
8924 // If the source integer type is larger than the intptr_t type for
8925 // this target, do a trunc to the intptr_t type, then inttoptr of it. This
8926 // allows the trunc to be exposed to other transforms. Don't do this for
8927 // extending inttoptr's, because we don't know if the target sign or zero
8928 // extends to pointers.
8929 if (CI.getOperand(0)->getType()->getScalarSizeInBits() >
8930 TD->getPointerSizeInBits()) {
8931 Value *P = InsertNewInstBefore(new TruncInst(CI.getOperand(0),
8932 TD->getIntPtrType(),
8934 return new IntToPtrInst(P, CI.getType());
8937 if (Instruction *I = commonCastTransforms(CI))
8940 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
8941 if (!DestPointee->isSized()) return 0;
8943 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
8946 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
8947 m_ConstantInt(Cst)), *Context)) {
8948 // If the source and destination operands have the same type, see if this
8949 // is a single-index GEP.
8950 if (X->getType() == CI.getType()) {
8951 // Get the size of the pointee type.
8952 uint64_t Size = TD->getTypeAllocSize(DestPointee);
8954 // Convert the constant to intptr type.
8955 APInt Offset = Cst->getValue();
8956 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8958 // If Offset is evenly divisible by Size, we can do this xform.
8959 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8960 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8961 return GetElementPtrInst::Create(X, Context->getConstantInt(Offset));
8964 // TODO: Could handle other cases, e.g. where add is indexing into field of
8966 } else if (CI.getOperand(0)->hasOneUse() &&
8967 match(CI.getOperand(0), m_Add(m_Value(X),
8968 m_ConstantInt(Cst)), *Context)) {
8969 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
8970 // "inttoptr+GEP" instead of "add+intptr".
8972 // Get the size of the pointee type.
8973 uint64_t Size = TD->getTypeAllocSize(DestPointee);
8975 // Convert the constant to intptr type.
8976 APInt Offset = Cst->getValue();
8977 Offset.sextOrTrunc(TD->getPointerSizeInBits());
8979 // If Offset is evenly divisible by Size, we can do this xform.
8980 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
8981 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
8983 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
8985 return GetElementPtrInst::Create(P,
8986 Context->getConstantInt(Offset), "tmp");
8992 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
8993 // If the operands are integer typed then apply the integer transforms,
8994 // otherwise just apply the common ones.
8995 Value *Src = CI.getOperand(0);
8996 const Type *SrcTy = Src->getType();
8997 const Type *DestTy = CI.getType();
8999 if (SrcTy->isInteger() && DestTy->isInteger()) {
9000 if (Instruction *Result = commonIntCastTransforms(CI))
9002 } else if (isa<PointerType>(SrcTy)) {
9003 if (Instruction *I = commonPointerCastTransforms(CI))
9006 if (Instruction *Result = commonCastTransforms(CI))
9011 // Get rid of casts from one type to the same type. These are useless and can
9012 // be replaced by the operand.
9013 if (DestTy == Src->getType())
9014 return ReplaceInstUsesWith(CI, Src);
9016 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
9017 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
9018 const Type *DstElTy = DstPTy->getElementType();
9019 const Type *SrcElTy = SrcPTy->getElementType();
9021 // If the address spaces don't match, don't eliminate the bitcast, which is
9022 // required for changing types.
9023 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
9026 // If we are casting a malloc or alloca to a pointer to a type of the same
9027 // size, rewrite the allocation instruction to allocate the "right" type.
9028 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
9029 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
9032 // If the source and destination are pointers, and this cast is equivalent
9033 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
9034 // This can enhance SROA and other transforms that want type-safe pointers.
9035 Constant *ZeroUInt = Context->getNullValue(Type::Int32Ty);
9036 unsigned NumZeros = 0;
9037 while (SrcElTy != DstElTy &&
9038 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
9039 SrcElTy->getNumContainedTypes() /* not "{}" */) {
9040 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
9044 // If we found a path from the src to dest, create the getelementptr now.
9045 if (SrcElTy == DstElTy) {
9046 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
9047 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
9048 ((Instruction*) NULL));
9052 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
9053 if (SVI->hasOneUse()) {
9054 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
9055 // a bitconvert to a vector with the same # elts.
9056 if (isa<VectorType>(DestTy) &&
9057 cast<VectorType>(DestTy)->getNumElements() ==
9058 SVI->getType()->getNumElements() &&
9059 SVI->getType()->getNumElements() ==
9060 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements()) {
9062 // If either of the operands is a cast from CI.getType(), then
9063 // evaluating the shuffle in the casted destination's type will allow
9064 // us to eliminate at least one cast.
9065 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
9066 Tmp->getOperand(0)->getType() == DestTy) ||
9067 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
9068 Tmp->getOperand(0)->getType() == DestTy)) {
9069 Value *LHS = InsertCastBefore(Instruction::BitCast,
9070 SVI->getOperand(0), DestTy, CI);
9071 Value *RHS = InsertCastBefore(Instruction::BitCast,
9072 SVI->getOperand(1), DestTy, CI);
9073 // Return a new shuffle vector. Use the same element ID's, as we
9074 // know the vector types match #elts.
9075 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
9083 /// GetSelectFoldableOperands - We want to turn code that looks like this:
9085 /// %D = select %cond, %C, %A
9087 /// %C = select %cond, %B, 0
9090 /// Assuming that the specified instruction is an operand to the select, return
9091 /// a bitmask indicating which operands of this instruction are foldable if they
9092 /// equal the other incoming value of the select.
9094 static unsigned GetSelectFoldableOperands(Instruction *I) {
9095 switch (I->getOpcode()) {
9096 case Instruction::Add:
9097 case Instruction::Mul:
9098 case Instruction::And:
9099 case Instruction::Or:
9100 case Instruction::Xor:
9101 return 3; // Can fold through either operand.
9102 case Instruction::Sub: // Can only fold on the amount subtracted.
9103 case Instruction::Shl: // Can only fold on the shift amount.
9104 case Instruction::LShr:
9105 case Instruction::AShr:
9108 return 0; // Cannot fold
9112 /// GetSelectFoldableConstant - For the same transformation as the previous
9113 /// function, return the identity constant that goes into the select.
9114 static Constant *GetSelectFoldableConstant(Instruction *I,
9115 LLVMContext *Context) {
9116 switch (I->getOpcode()) {
9117 default: LLVM_UNREACHABLE("This cannot happen!");
9118 case Instruction::Add:
9119 case Instruction::Sub:
9120 case Instruction::Or:
9121 case Instruction::Xor:
9122 case Instruction::Shl:
9123 case Instruction::LShr:
9124 case Instruction::AShr:
9125 return Context->getNullValue(I->getType());
9126 case Instruction::And:
9127 return Context->getAllOnesValue(I->getType());
9128 case Instruction::Mul:
9129 return Context->getConstantInt(I->getType(), 1);
9133 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
9134 /// have the same opcode and only one use each. Try to simplify this.
9135 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
9137 if (TI->getNumOperands() == 1) {
9138 // If this is a non-volatile load or a cast from the same type,
9141 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
9144 return 0; // unknown unary op.
9147 // Fold this by inserting a select from the input values.
9148 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
9149 FI->getOperand(0), SI.getName()+".v");
9150 InsertNewInstBefore(NewSI, SI);
9151 return CastInst::Create(Instruction::CastOps(TI->getOpcode()), NewSI,
9155 // Only handle binary operators here.
9156 if (!isa<BinaryOperator>(TI))
9159 // Figure out if the operations have any operands in common.
9160 Value *MatchOp, *OtherOpT, *OtherOpF;
9162 if (TI->getOperand(0) == FI->getOperand(0)) {
9163 MatchOp = TI->getOperand(0);
9164 OtherOpT = TI->getOperand(1);
9165 OtherOpF = FI->getOperand(1);
9166 MatchIsOpZero = true;
9167 } else if (TI->getOperand(1) == FI->getOperand(1)) {
9168 MatchOp = TI->getOperand(1);
9169 OtherOpT = TI->getOperand(0);
9170 OtherOpF = FI->getOperand(0);
9171 MatchIsOpZero = false;
9172 } else if (!TI->isCommutative()) {
9174 } else if (TI->getOperand(0) == FI->getOperand(1)) {
9175 MatchOp = TI->getOperand(0);
9176 OtherOpT = TI->getOperand(1);
9177 OtherOpF = FI->getOperand(0);
9178 MatchIsOpZero = true;
9179 } else if (TI->getOperand(1) == FI->getOperand(0)) {
9180 MatchOp = TI->getOperand(1);
9181 OtherOpT = TI->getOperand(0);
9182 OtherOpF = FI->getOperand(1);
9183 MatchIsOpZero = true;
9188 // If we reach here, they do have operations in common.
9189 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
9190 OtherOpF, SI.getName()+".v");
9191 InsertNewInstBefore(NewSI, SI);
9193 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
9195 return BinaryOperator::Create(BO->getOpcode(), MatchOp, NewSI);
9197 return BinaryOperator::Create(BO->getOpcode(), NewSI, MatchOp);
9199 LLVM_UNREACHABLE("Shouldn't get here");
9203 static bool isSelect01(Constant *C1, Constant *C2) {
9204 ConstantInt *C1I = dyn_cast<ConstantInt>(C1);
9207 ConstantInt *C2I = dyn_cast<ConstantInt>(C2);
9210 return (C1I->isZero() || C1I->isOne()) && (C2I->isZero() || C2I->isOne());
9213 /// FoldSelectIntoOp - Try fold the select into one of the operands to
9214 /// facilitate further optimization.
9215 Instruction *InstCombiner::FoldSelectIntoOp(SelectInst &SI, Value *TrueVal,
9217 // See the comment above GetSelectFoldableOperands for a description of the
9218 // transformation we are doing here.
9219 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal)) {
9220 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
9221 !isa<Constant>(FalseVal)) {
9222 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
9223 unsigned OpToFold = 0;
9224 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
9226 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
9231 Constant *C = GetSelectFoldableConstant(TVI, Context);
9232 Value *OOp = TVI->getOperand(2-OpToFold);
9233 // Avoid creating select between 2 constants unless it's selecting
9235 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9236 Instruction *NewSel = SelectInst::Create(SI.getCondition(), OOp, C);
9237 InsertNewInstBefore(NewSel, SI);
9238 NewSel->takeName(TVI);
9239 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
9240 return BinaryOperator::Create(BO->getOpcode(), FalseVal, NewSel);
9241 LLVM_UNREACHABLE("Unknown instruction!!");
9248 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal)) {
9249 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
9250 !isa<Constant>(TrueVal)) {
9251 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
9252 unsigned OpToFold = 0;
9253 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
9255 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
9260 Constant *C = GetSelectFoldableConstant(FVI, Context);
9261 Value *OOp = FVI->getOperand(2-OpToFold);
9262 // Avoid creating select between 2 constants unless it's selecting
9264 if (!isa<Constant>(OOp) || isSelect01(C, cast<Constant>(OOp))) {
9265 Instruction *NewSel = SelectInst::Create(SI.getCondition(), C, OOp);
9266 InsertNewInstBefore(NewSel, SI);
9267 NewSel->takeName(FVI);
9268 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
9269 return BinaryOperator::Create(BO->getOpcode(), TrueVal, NewSel);
9270 LLVM_UNREACHABLE("Unknown instruction!!");
9280 /// visitSelectInstWithICmp - Visit a SelectInst that has an
9281 /// ICmpInst as its first operand.
9283 Instruction *InstCombiner::visitSelectInstWithICmp(SelectInst &SI,
9285 bool Changed = false;
9286 ICmpInst::Predicate Pred = ICI->getPredicate();
9287 Value *CmpLHS = ICI->getOperand(0);
9288 Value *CmpRHS = ICI->getOperand(1);
9289 Value *TrueVal = SI.getTrueValue();
9290 Value *FalseVal = SI.getFalseValue();
9292 // Check cases where the comparison is with a constant that
9293 // can be adjusted to fit the min/max idiom. We may edit ICI in
9294 // place here, so make sure the select is the only user.
9295 if (ICI->hasOneUse())
9296 if (ConstantInt *CI = dyn_cast<ConstantInt>(CmpRHS)) {
9299 case ICmpInst::ICMP_ULT:
9300 case ICmpInst::ICMP_SLT: {
9301 // X < MIN ? T : F --> F
9302 if (CI->isMinValue(Pred == ICmpInst::ICMP_SLT))
9303 return ReplaceInstUsesWith(SI, FalseVal);
9304 // X < C ? X : C-1 --> X > C-1 ? C-1 : X
9305 Constant *AdjustedRHS = SubOne(CI, Context);
9306 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9307 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9308 Pred = ICmpInst::getSwappedPredicate(Pred);
9309 CmpRHS = AdjustedRHS;
9310 std::swap(FalseVal, TrueVal);
9311 ICI->setPredicate(Pred);
9312 ICI->setOperand(1, CmpRHS);
9313 SI.setOperand(1, TrueVal);
9314 SI.setOperand(2, FalseVal);
9319 case ICmpInst::ICMP_UGT:
9320 case ICmpInst::ICMP_SGT: {
9321 // X > MAX ? T : F --> F
9322 if (CI->isMaxValue(Pred == ICmpInst::ICMP_SGT))
9323 return ReplaceInstUsesWith(SI, FalseVal);
9324 // X > C ? X : C+1 --> X < C+1 ? C+1 : X
9325 Constant *AdjustedRHS = AddOne(CI, Context);
9326 if ((CmpLHS == TrueVal && AdjustedRHS == FalseVal) ||
9327 (CmpLHS == FalseVal && AdjustedRHS == TrueVal)) {
9328 Pred = ICmpInst::getSwappedPredicate(Pred);
9329 CmpRHS = AdjustedRHS;
9330 std::swap(FalseVal, TrueVal);
9331 ICI->setPredicate(Pred);
9332 ICI->setOperand(1, CmpRHS);
9333 SI.setOperand(1, TrueVal);
9334 SI.setOperand(2, FalseVal);
9341 // (x <s 0) ? -1 : 0 -> ashr x, 31 -> all ones if signed
9342 // (x >s -1) ? -1 : 0 -> ashr x, 31 -> all ones if not signed
9343 CmpInst::Predicate Pred = CmpInst::BAD_ICMP_PREDICATE;
9344 if (match(TrueVal, m_ConstantInt<-1>(), *Context) &&
9345 match(FalseVal, m_ConstantInt<0>(), *Context))
9346 Pred = ICI->getPredicate();
9347 else if (match(TrueVal, m_ConstantInt<0>(), *Context) &&
9348 match(FalseVal, m_ConstantInt<-1>(), *Context))
9349 Pred = CmpInst::getInversePredicate(ICI->getPredicate());
9351 if (Pred != CmpInst::BAD_ICMP_PREDICATE) {
9352 // If we are just checking for a icmp eq of a single bit and zext'ing it
9353 // to an integer, then shift the bit to the appropriate place and then
9354 // cast to integer to avoid the comparison.
9355 const APInt &Op1CV = CI->getValue();
9357 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
9358 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
9359 if ((Pred == ICmpInst::ICMP_SLT && Op1CV == 0) ||
9360 (Pred == ICmpInst::ICMP_SGT && Op1CV.isAllOnesValue())) {
9361 Value *In = ICI->getOperand(0);
9362 Value *Sh = Context->getConstantInt(In->getType(),
9363 In->getType()->getScalarSizeInBits()-1);
9364 In = InsertNewInstBefore(BinaryOperator::CreateAShr(In, Sh,
9365 In->getName()+".lobit"),
9367 if (In->getType() != SI.getType())
9368 In = CastInst::CreateIntegerCast(In, SI.getType(),
9369 true/*SExt*/, "tmp", ICI);
9371 if (Pred == ICmpInst::ICMP_SGT)
9372 In = InsertNewInstBefore(BinaryOperator::CreateNot(In,
9373 In->getName()+".not"), *ICI);
9375 return ReplaceInstUsesWith(SI, In);
9380 if (CmpLHS == TrueVal && CmpRHS == FalseVal) {
9381 // Transform (X == Y) ? X : Y -> Y
9382 if (Pred == ICmpInst::ICMP_EQ)
9383 return ReplaceInstUsesWith(SI, FalseVal);
9384 // Transform (X != Y) ? X : Y -> X
9385 if (Pred == ICmpInst::ICMP_NE)
9386 return ReplaceInstUsesWith(SI, TrueVal);
9387 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9389 } else if (CmpLHS == FalseVal && CmpRHS == TrueVal) {
9390 // Transform (X == Y) ? Y : X -> X
9391 if (Pred == ICmpInst::ICMP_EQ)
9392 return ReplaceInstUsesWith(SI, FalseVal);
9393 // Transform (X != Y) ? Y : X -> Y
9394 if (Pred == ICmpInst::ICMP_NE)
9395 return ReplaceInstUsesWith(SI, TrueVal);
9396 /// NOTE: if we wanted to, this is where to detect integer MIN/MAX
9399 /// NOTE: if we wanted to, this is where to detect integer ABS
9401 return Changed ? &SI : 0;
9404 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
9405 Value *CondVal = SI.getCondition();
9406 Value *TrueVal = SI.getTrueValue();
9407 Value *FalseVal = SI.getFalseValue();
9409 // select true, X, Y -> X
9410 // select false, X, Y -> Y
9411 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
9412 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
9414 // select C, X, X -> X
9415 if (TrueVal == FalseVal)
9416 return ReplaceInstUsesWith(SI, TrueVal);
9418 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
9419 return ReplaceInstUsesWith(SI, FalseVal);
9420 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
9421 return ReplaceInstUsesWith(SI, TrueVal);
9422 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
9423 if (isa<Constant>(TrueVal))
9424 return ReplaceInstUsesWith(SI, TrueVal);
9426 return ReplaceInstUsesWith(SI, FalseVal);
9429 if (SI.getType() == Type::Int1Ty) {
9430 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
9431 if (C->getZExtValue()) {
9432 // Change: A = select B, true, C --> A = or B, C
9433 return BinaryOperator::CreateOr(CondVal, FalseVal);
9435 // Change: A = select B, false, C --> A = and !B, C
9437 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9438 "not."+CondVal->getName()), SI);
9439 return BinaryOperator::CreateAnd(NotCond, FalseVal);
9441 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
9442 if (C->getZExtValue() == false) {
9443 // Change: A = select B, C, false --> A = and B, C
9444 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9446 // Change: A = select B, C, true --> A = or !B, C
9448 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9449 "not."+CondVal->getName()), SI);
9450 return BinaryOperator::CreateOr(NotCond, TrueVal);
9454 // select a, b, a -> a&b
9455 // select a, a, b -> a|b
9456 if (CondVal == TrueVal)
9457 return BinaryOperator::CreateOr(CondVal, FalseVal);
9458 else if (CondVal == FalseVal)
9459 return BinaryOperator::CreateAnd(CondVal, TrueVal);
9462 // Selecting between two integer constants?
9463 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
9464 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
9465 // select C, 1, 0 -> zext C to int
9466 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
9467 return CastInst::Create(Instruction::ZExt, CondVal, SI.getType());
9468 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
9469 // select C, 0, 1 -> zext !C to int
9471 InsertNewInstBefore(BinaryOperator::CreateNot(CondVal,
9472 "not."+CondVal->getName()), SI);
9473 return CastInst::Create(Instruction::ZExt, NotCond, SI.getType());
9476 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
9477 // If one of the constants is zero (we know they can't both be) and we
9478 // have an icmp instruction with zero, and we have an 'and' with the
9479 // non-constant value, eliminate this whole mess. This corresponds to
9480 // cases like this: ((X & 27) ? 27 : 0)
9481 if (TrueValC->isZero() || FalseValC->isZero())
9482 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
9483 cast<Constant>(IC->getOperand(1))->isNullValue())
9484 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
9485 if (ICA->getOpcode() == Instruction::And &&
9486 isa<ConstantInt>(ICA->getOperand(1)) &&
9487 (ICA->getOperand(1) == TrueValC ||
9488 ICA->getOperand(1) == FalseValC) &&
9489 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
9490 // Okay, now we know that everything is set up, we just don't
9491 // know whether we have a icmp_ne or icmp_eq and whether the
9492 // true or false val is the zero.
9493 bool ShouldNotVal = !TrueValC->isZero();
9494 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
9497 V = InsertNewInstBefore(BinaryOperator::Create(
9498 Instruction::Xor, V, ICA->getOperand(1)), SI);
9499 return ReplaceInstUsesWith(SI, V);
9504 // See if we are selecting two values based on a comparison of the two values.
9505 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
9506 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
9507 // Transform (X == Y) ? X : Y -> Y
9508 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9509 // This is not safe in general for floating point:
9510 // consider X== -0, Y== +0.
9511 // It becomes safe if either operand is a nonzero constant.
9512 ConstantFP *CFPt, *CFPf;
9513 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9514 !CFPt->getValueAPF().isZero()) ||
9515 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9516 !CFPf->getValueAPF().isZero()))
9517 return ReplaceInstUsesWith(SI, FalseVal);
9519 // Transform (X != Y) ? X : Y -> X
9520 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9521 return ReplaceInstUsesWith(SI, TrueVal);
9522 // NOTE: if we wanted to, this is where to detect MIN/MAX
9524 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
9525 // Transform (X == Y) ? Y : X -> X
9526 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
9527 // This is not safe in general for floating point:
9528 // consider X== -0, Y== +0.
9529 // It becomes safe if either operand is a nonzero constant.
9530 ConstantFP *CFPt, *CFPf;
9531 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
9532 !CFPt->getValueAPF().isZero()) ||
9533 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
9534 !CFPf->getValueAPF().isZero()))
9535 return ReplaceInstUsesWith(SI, FalseVal);
9537 // Transform (X != Y) ? Y : X -> Y
9538 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
9539 return ReplaceInstUsesWith(SI, TrueVal);
9540 // NOTE: if we wanted to, this is where to detect MIN/MAX
9542 // NOTE: if we wanted to, this is where to detect ABS
9545 // See if we are selecting two values based on a comparison of the two values.
9546 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal))
9547 if (Instruction *Result = visitSelectInstWithICmp(SI, ICI))
9550 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
9551 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
9552 if (TI->hasOneUse() && FI->hasOneUse()) {
9553 Instruction *AddOp = 0, *SubOp = 0;
9555 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
9556 if (TI->getOpcode() == FI->getOpcode())
9557 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
9560 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
9561 // even legal for FP.
9562 if ((TI->getOpcode() == Instruction::Sub &&
9563 FI->getOpcode() == Instruction::Add) ||
9564 (TI->getOpcode() == Instruction::FSub &&
9565 FI->getOpcode() == Instruction::FAdd)) {
9566 AddOp = FI; SubOp = TI;
9567 } else if ((FI->getOpcode() == Instruction::Sub &&
9568 TI->getOpcode() == Instruction::Add) ||
9569 (FI->getOpcode() == Instruction::FSub &&
9570 TI->getOpcode() == Instruction::FAdd)) {
9571 AddOp = TI; SubOp = FI;
9575 Value *OtherAddOp = 0;
9576 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
9577 OtherAddOp = AddOp->getOperand(1);
9578 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
9579 OtherAddOp = AddOp->getOperand(0);
9583 // So at this point we know we have (Y -> OtherAddOp):
9584 // select C, (add X, Y), (sub X, Z)
9585 Value *NegVal; // Compute -Z
9586 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
9587 NegVal = Context->getConstantExprNeg(C);
9589 NegVal = InsertNewInstBefore(
9590 BinaryOperator::CreateNeg(SubOp->getOperand(1), "tmp"), SI);
9593 Value *NewTrueOp = OtherAddOp;
9594 Value *NewFalseOp = NegVal;
9596 std::swap(NewTrueOp, NewFalseOp);
9597 Instruction *NewSel =
9598 SelectInst::Create(CondVal, NewTrueOp,
9599 NewFalseOp, SI.getName() + ".p");
9601 NewSel = InsertNewInstBefore(NewSel, SI);
9602 return BinaryOperator::CreateAdd(SubOp->getOperand(0), NewSel);
9607 // See if we can fold the select into one of our operands.
9608 if (SI.getType()->isInteger()) {
9609 Instruction *FoldI = FoldSelectIntoOp(SI, TrueVal, FalseVal);
9614 if (BinaryOperator::isNot(CondVal)) {
9615 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
9616 SI.setOperand(1, FalseVal);
9617 SI.setOperand(2, TrueVal);
9624 /// EnforceKnownAlignment - If the specified pointer points to an object that
9625 /// we control, modify the object's alignment to PrefAlign. This isn't
9626 /// often possible though. If alignment is important, a more reliable approach
9627 /// is to simply align all global variables and allocation instructions to
9628 /// their preferred alignment from the beginning.
9630 static unsigned EnforceKnownAlignment(Value *V,
9631 unsigned Align, unsigned PrefAlign) {
9633 User *U = dyn_cast<User>(V);
9634 if (!U) return Align;
9636 switch (getOpcode(U)) {
9638 case Instruction::BitCast:
9639 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9640 case Instruction::GetElementPtr: {
9641 // If all indexes are zero, it is just the alignment of the base pointer.
9642 bool AllZeroOperands = true;
9643 for (User::op_iterator i = U->op_begin() + 1, e = U->op_end(); i != e; ++i)
9644 if (!isa<Constant>(*i) ||
9645 !cast<Constant>(*i)->isNullValue()) {
9646 AllZeroOperands = false;
9650 if (AllZeroOperands) {
9651 // Treat this like a bitcast.
9652 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
9658 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
9659 // If there is a large requested alignment and we can, bump up the alignment
9661 if (!GV->isDeclaration()) {
9662 if (GV->getAlignment() >= PrefAlign)
9663 Align = GV->getAlignment();
9665 GV->setAlignment(PrefAlign);
9669 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
9670 // If there is a requested alignment and if this is an alloca, round up. We
9671 // don't do this for malloc, because some systems can't respect the request.
9672 if (isa<AllocaInst>(AI)) {
9673 if (AI->getAlignment() >= PrefAlign)
9674 Align = AI->getAlignment();
9676 AI->setAlignment(PrefAlign);
9685 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
9686 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
9687 /// and it is more than the alignment of the ultimate object, see if we can
9688 /// increase the alignment of the ultimate object, making this check succeed.
9689 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
9690 unsigned PrefAlign) {
9691 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
9692 sizeof(PrefAlign) * CHAR_BIT;
9693 APInt Mask = APInt::getAllOnesValue(BitWidth);
9694 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
9695 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
9696 unsigned TrailZ = KnownZero.countTrailingOnes();
9697 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
9699 if (PrefAlign > Align)
9700 Align = EnforceKnownAlignment(V, Align, PrefAlign);
9702 // We don't need to make any adjustment.
9706 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
9707 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
9708 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
9709 unsigned MinAlign = std::min(DstAlign, SrcAlign);
9710 unsigned CopyAlign = MI->getAlignment();
9712 if (CopyAlign < MinAlign) {
9713 MI->setAlignment(Context->getConstantInt(MI->getAlignmentType(),
9718 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
9720 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
9721 if (MemOpLength == 0) return 0;
9723 // Source and destination pointer types are always "i8*" for intrinsic. See
9724 // if the size is something we can handle with a single primitive load/store.
9725 // A single load+store correctly handles overlapping memory in the memmove
9727 unsigned Size = MemOpLength->getZExtValue();
9728 if (Size == 0) return MI; // Delete this mem transfer.
9730 if (Size > 8 || (Size&(Size-1)))
9731 return 0; // If not 1/2/4/8 bytes, exit.
9733 // Use an integer load+store unless we can find something better.
9735 Context->getPointerTypeUnqual(Context->getIntegerType(Size<<3));
9737 // Memcpy forces the use of i8* for the source and destination. That means
9738 // that if you're using memcpy to move one double around, you'll get a cast
9739 // from double* to i8*. We'd much rather use a double load+store rather than
9740 // an i64 load+store, here because this improves the odds that the source or
9741 // dest address will be promotable. See if we can find a better type than the
9742 // integer datatype.
9743 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
9744 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
9745 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
9746 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
9747 // down through these levels if so.
9748 while (!SrcETy->isSingleValueType()) {
9749 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
9750 if (STy->getNumElements() == 1)
9751 SrcETy = STy->getElementType(0);
9754 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
9755 if (ATy->getNumElements() == 1)
9756 SrcETy = ATy->getElementType();
9763 if (SrcETy->isSingleValueType())
9764 NewPtrTy = Context->getPointerTypeUnqual(SrcETy);
9769 // If the memcpy/memmove provides better alignment info than we can
9771 SrcAlign = std::max(SrcAlign, CopyAlign);
9772 DstAlign = std::max(DstAlign, CopyAlign);
9774 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
9775 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
9776 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
9777 InsertNewInstBefore(L, *MI);
9778 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
9780 // Set the size of the copy to 0, it will be deleted on the next iteration.
9781 MI->setOperand(3, Context->getNullValue(MemOpLength->getType()));
9785 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
9786 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
9787 if (MI->getAlignment() < Alignment) {
9788 MI->setAlignment(Context->getConstantInt(MI->getAlignmentType(),
9793 // Extract the length and alignment and fill if they are constant.
9794 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
9795 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
9796 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
9798 uint64_t Len = LenC->getZExtValue();
9799 Alignment = MI->getAlignment();
9801 // If the length is zero, this is a no-op
9802 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
9804 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
9805 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
9806 const Type *ITy = Context->getIntegerType(Len*8); // n=1 -> i8.
9808 Value *Dest = MI->getDest();
9809 Dest = InsertBitCastBefore(Dest, Context->getPointerTypeUnqual(ITy), *MI);
9811 // Alignment 0 is identity for alignment 1 for memset, but not store.
9812 if (Alignment == 0) Alignment = 1;
9814 // Extract the fill value and store.
9815 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
9816 InsertNewInstBefore(new StoreInst(Context->getConstantInt(ITy, Fill),
9817 Dest, false, Alignment), *MI);
9819 // Set the size of the copy to 0, it will be deleted on the next iteration.
9820 MI->setLength(Context->getNullValue(LenC->getType()));
9828 /// visitCallInst - CallInst simplification. This mostly only handles folding
9829 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
9830 /// the heavy lifting.
9832 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
9833 // If the caller function is nounwind, mark the call as nounwind, even if the
9835 if (CI.getParent()->getParent()->doesNotThrow() &&
9836 !CI.doesNotThrow()) {
9837 CI.setDoesNotThrow();
9843 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
9844 if (!II) return visitCallSite(&CI);
9846 // Intrinsics cannot occur in an invoke, so handle them here instead of in
9848 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
9849 bool Changed = false;
9851 // memmove/cpy/set of zero bytes is a noop.
9852 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
9853 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
9855 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
9856 if (CI->getZExtValue() == 1) {
9857 // Replace the instruction with just byte operations. We would
9858 // transform other cases to loads/stores, but we don't know if
9859 // alignment is sufficient.
9863 // If we have a memmove and the source operation is a constant global,
9864 // then the source and dest pointers can't alias, so we can change this
9865 // into a call to memcpy.
9866 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
9867 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
9868 if (GVSrc->isConstant()) {
9869 Module *M = CI.getParent()->getParent()->getParent();
9870 Intrinsic::ID MemCpyID = Intrinsic::memcpy;
9872 Tys[0] = CI.getOperand(3)->getType();
9874 Intrinsic::getDeclaration(M, MemCpyID, Tys, 1));
9878 // memmove(x,x,size) -> noop.
9879 if (MMI->getSource() == MMI->getDest())
9880 return EraseInstFromFunction(CI);
9883 // If we can determine a pointer alignment that is bigger than currently
9884 // set, update the alignment.
9885 if (isa<MemTransferInst>(MI)) {
9886 if (Instruction *I = SimplifyMemTransfer(MI))
9888 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
9889 if (Instruction *I = SimplifyMemSet(MSI))
9893 if (Changed) return II;
9896 switch (II->getIntrinsicID()) {
9898 case Intrinsic::bswap:
9899 // bswap(bswap(x)) -> x
9900 if (IntrinsicInst *Operand = dyn_cast<IntrinsicInst>(II->getOperand(1)))
9901 if (Operand->getIntrinsicID() == Intrinsic::bswap)
9902 return ReplaceInstUsesWith(CI, Operand->getOperand(1));
9904 case Intrinsic::ppc_altivec_lvx:
9905 case Intrinsic::ppc_altivec_lvxl:
9906 case Intrinsic::x86_sse_loadu_ps:
9907 case Intrinsic::x86_sse2_loadu_pd:
9908 case Intrinsic::x86_sse2_loadu_dq:
9909 // Turn PPC lvx -> load if the pointer is known aligned.
9910 // Turn X86 loadups -> load if the pointer is known aligned.
9911 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9912 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
9913 Context->getPointerTypeUnqual(II->getType()),
9915 return new LoadInst(Ptr);
9918 case Intrinsic::ppc_altivec_stvx:
9919 case Intrinsic::ppc_altivec_stvxl:
9920 // Turn stvx -> store if the pointer is known aligned.
9921 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
9922 const Type *OpPtrTy =
9923 Context->getPointerTypeUnqual(II->getOperand(1)->getType());
9924 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
9925 return new StoreInst(II->getOperand(1), Ptr);
9928 case Intrinsic::x86_sse_storeu_ps:
9929 case Intrinsic::x86_sse2_storeu_pd:
9930 case Intrinsic::x86_sse2_storeu_dq:
9931 // Turn X86 storeu -> store if the pointer is known aligned.
9932 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
9933 const Type *OpPtrTy =
9934 Context->getPointerTypeUnqual(II->getOperand(2)->getType());
9935 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
9936 return new StoreInst(II->getOperand(2), Ptr);
9940 case Intrinsic::x86_sse_cvttss2si: {
9941 // These intrinsics only demands the 0th element of its input vector. If
9942 // we can simplify the input based on that, do so now.
9944 cast<VectorType>(II->getOperand(1)->getType())->getNumElements();
9945 APInt DemandedElts(VWidth, 1);
9946 APInt UndefElts(VWidth, 0);
9947 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
9949 II->setOperand(1, V);
9955 case Intrinsic::ppc_altivec_vperm:
9956 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
9957 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
9958 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
9960 // Check that all of the elements are integer constants or undefs.
9961 bool AllEltsOk = true;
9962 for (unsigned i = 0; i != 16; ++i) {
9963 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
9964 !isa<UndefValue>(Mask->getOperand(i))) {
9971 // Cast the input vectors to byte vectors.
9972 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
9973 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
9974 Value *Result = Context->getUndef(Op0->getType());
9976 // Only extract each element once.
9977 Value *ExtractedElts[32];
9978 memset(ExtractedElts, 0, sizeof(ExtractedElts));
9980 for (unsigned i = 0; i != 16; ++i) {
9981 if (isa<UndefValue>(Mask->getOperand(i)))
9983 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
9984 Idx &= 31; // Match the hardware behavior.
9986 if (ExtractedElts[Idx] == 0) {
9988 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
9989 InsertNewInstBefore(Elt, CI);
9990 ExtractedElts[Idx] = Elt;
9993 // Insert this value into the result vector.
9994 Result = InsertElementInst::Create(Result, ExtractedElts[Idx],
9996 InsertNewInstBefore(cast<Instruction>(Result), CI);
9998 return CastInst::Create(Instruction::BitCast, Result, CI.getType());
10003 case Intrinsic::stackrestore: {
10004 // If the save is right next to the restore, remove the restore. This can
10005 // happen when variable allocas are DCE'd.
10006 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
10007 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
10008 BasicBlock::iterator BI = SS;
10010 return EraseInstFromFunction(CI);
10014 // Scan down this block to see if there is another stack restore in the
10015 // same block without an intervening call/alloca.
10016 BasicBlock::iterator BI = II;
10017 TerminatorInst *TI = II->getParent()->getTerminator();
10018 bool CannotRemove = false;
10019 for (++BI; &*BI != TI; ++BI) {
10020 if (isa<AllocaInst>(BI)) {
10021 CannotRemove = true;
10024 if (CallInst *BCI = dyn_cast<CallInst>(BI)) {
10025 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(BCI)) {
10026 // If there is a stackrestore below this one, remove this one.
10027 if (II->getIntrinsicID() == Intrinsic::stackrestore)
10028 return EraseInstFromFunction(CI);
10029 // Otherwise, ignore the intrinsic.
10031 // If we found a non-intrinsic call, we can't remove the stack
10033 CannotRemove = true;
10039 // If the stack restore is in a return/unwind block and if there are no
10040 // allocas or calls between the restore and the return, nuke the restore.
10041 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
10042 return EraseInstFromFunction(CI);
10047 return visitCallSite(II);
10050 // InvokeInst simplification
10052 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
10053 return visitCallSite(&II);
10056 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
10057 /// passed through the varargs area, we can eliminate the use of the cast.
10058 static bool isSafeToEliminateVarargsCast(const CallSite CS,
10059 const CastInst * const CI,
10060 const TargetData * const TD,
10062 if (!CI->isLosslessCast())
10065 // The size of ByVal arguments is derived from the type, so we
10066 // can't change to a type with a different size. If the size were
10067 // passed explicitly we could avoid this check.
10068 if (!CS.paramHasAttr(ix, Attribute::ByVal))
10071 const Type* SrcTy =
10072 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
10073 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
10074 if (!SrcTy->isSized() || !DstTy->isSized())
10076 if (TD->getTypeAllocSize(SrcTy) != TD->getTypeAllocSize(DstTy))
10081 // visitCallSite - Improvements for call and invoke instructions.
10083 Instruction *InstCombiner::visitCallSite(CallSite CS) {
10084 bool Changed = false;
10086 // If the callee is a constexpr cast of a function, attempt to move the cast
10087 // to the arguments of the call/invoke.
10088 if (transformConstExprCastCall(CS)) return 0;
10090 Value *Callee = CS.getCalledValue();
10092 if (Function *CalleeF = dyn_cast<Function>(Callee))
10093 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
10094 Instruction *OldCall = CS.getInstruction();
10095 // If the call and callee calling conventions don't match, this call must
10096 // be unreachable, as the call is undefined.
10097 new StoreInst(Context->getConstantIntTrue(),
10098 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)),
10100 if (!OldCall->use_empty())
10101 OldCall->replaceAllUsesWith(Context->getUndef(OldCall->getType()));
10102 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
10103 return EraseInstFromFunction(*OldCall);
10107 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
10108 // This instruction is not reachable, just remove it. We insert a store to
10109 // undef so that we know that this code is not reachable, despite the fact
10110 // that we can't modify the CFG here.
10111 new StoreInst(Context->getConstantIntTrue(),
10112 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)),
10113 CS.getInstruction());
10115 if (!CS.getInstruction()->use_empty())
10116 CS.getInstruction()->
10117 replaceAllUsesWith(Context->getUndef(CS.getInstruction()->getType()));
10119 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
10120 // Don't break the CFG, insert a dummy cond branch.
10121 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
10122 Context->getConstantIntTrue(), II);
10124 return EraseInstFromFunction(*CS.getInstruction());
10127 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
10128 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
10129 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
10130 return transformCallThroughTrampoline(CS);
10132 const PointerType *PTy = cast<PointerType>(Callee->getType());
10133 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10134 if (FTy->isVarArg()) {
10135 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
10136 // See if we can optimize any arguments passed through the varargs area of
10138 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
10139 E = CS.arg_end(); I != E; ++I, ++ix) {
10140 CastInst *CI = dyn_cast<CastInst>(*I);
10141 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
10142 *I = CI->getOperand(0);
10148 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
10149 // Inline asm calls cannot throw - mark them 'nounwind'.
10150 CS.setDoesNotThrow();
10154 return Changed ? CS.getInstruction() : 0;
10157 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
10158 // attempt to move the cast to the arguments of the call/invoke.
10160 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
10161 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
10162 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
10163 if (CE->getOpcode() != Instruction::BitCast ||
10164 !isa<Function>(CE->getOperand(0)))
10166 Function *Callee = cast<Function>(CE->getOperand(0));
10167 Instruction *Caller = CS.getInstruction();
10168 const AttrListPtr &CallerPAL = CS.getAttributes();
10170 // Okay, this is a cast from a function to a different type. Unless doing so
10171 // would cause a type conversion of one of our arguments, change this call to
10172 // be a direct call with arguments casted to the appropriate types.
10174 const FunctionType *FT = Callee->getFunctionType();
10175 const Type *OldRetTy = Caller->getType();
10176 const Type *NewRetTy = FT->getReturnType();
10178 if (isa<StructType>(NewRetTy))
10179 return false; // TODO: Handle multiple return values.
10181 // Check to see if we are changing the return type...
10182 if (OldRetTy != NewRetTy) {
10183 if (Callee->isDeclaration() &&
10184 // Conversion is ok if changing from one pointer type to another or from
10185 // a pointer to an integer of the same size.
10186 !((isa<PointerType>(OldRetTy) || OldRetTy == TD->getIntPtrType()) &&
10187 (isa<PointerType>(NewRetTy) || NewRetTy == TD->getIntPtrType())))
10188 return false; // Cannot transform this return value.
10190 if (!Caller->use_empty() &&
10191 // void -> non-void is handled specially
10192 NewRetTy != Type::VoidTy && !CastInst::isCastable(NewRetTy, OldRetTy))
10193 return false; // Cannot transform this return value.
10195 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
10196 Attributes RAttrs = CallerPAL.getRetAttributes();
10197 if (RAttrs & Attribute::typeIncompatible(NewRetTy))
10198 return false; // Attribute not compatible with transformed value.
10201 // If the callsite is an invoke instruction, and the return value is used by
10202 // a PHI node in a successor, we cannot change the return type of the call
10203 // because there is no place to put the cast instruction (without breaking
10204 // the critical edge). Bail out in this case.
10205 if (!Caller->use_empty())
10206 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
10207 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
10209 if (PHINode *PN = dyn_cast<PHINode>(*UI))
10210 if (PN->getParent() == II->getNormalDest() ||
10211 PN->getParent() == II->getUnwindDest())
10215 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
10216 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
10218 CallSite::arg_iterator AI = CS.arg_begin();
10219 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
10220 const Type *ParamTy = FT->getParamType(i);
10221 const Type *ActTy = (*AI)->getType();
10223 if (!CastInst::isCastable(ActTy, ParamTy))
10224 return false; // Cannot transform this parameter value.
10226 if (CallerPAL.getParamAttributes(i + 1)
10227 & Attribute::typeIncompatible(ParamTy))
10228 return false; // Attribute not compatible with transformed value.
10230 // Converting from one pointer type to another or between a pointer and an
10231 // integer of the same size is safe even if we do not have a body.
10232 bool isConvertible = ActTy == ParamTy ||
10233 ((isa<PointerType>(ParamTy) || ParamTy == TD->getIntPtrType()) &&
10234 (isa<PointerType>(ActTy) || ActTy == TD->getIntPtrType()));
10235 if (Callee->isDeclaration() && !isConvertible) return false;
10238 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
10239 Callee->isDeclaration())
10240 return false; // Do not delete arguments unless we have a function body.
10242 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
10243 !CallerPAL.isEmpty())
10244 // In this case we have more arguments than the new function type, but we
10245 // won't be dropping them. Check that these extra arguments have attributes
10246 // that are compatible with being a vararg call argument.
10247 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
10248 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
10250 Attributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
10251 if (PAttrs & Attribute::VarArgsIncompatible)
10255 // Okay, we decided that this is a safe thing to do: go ahead and start
10256 // inserting cast instructions as necessary...
10257 std::vector<Value*> Args;
10258 Args.reserve(NumActualArgs);
10259 SmallVector<AttributeWithIndex, 8> attrVec;
10260 attrVec.reserve(NumCommonArgs);
10262 // Get any return attributes.
10263 Attributes RAttrs = CallerPAL.getRetAttributes();
10265 // If the return value is not being used, the type may not be compatible
10266 // with the existing attributes. Wipe out any problematic attributes.
10267 RAttrs &= ~Attribute::typeIncompatible(NewRetTy);
10269 // Add the new return attributes.
10271 attrVec.push_back(AttributeWithIndex::get(0, RAttrs));
10273 AI = CS.arg_begin();
10274 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
10275 const Type *ParamTy = FT->getParamType(i);
10276 if ((*AI)->getType() == ParamTy) {
10277 Args.push_back(*AI);
10279 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
10280 false, ParamTy, false);
10281 CastInst *NewCast = CastInst::Create(opcode, *AI, ParamTy, "tmp");
10282 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
10285 // Add any parameter attributes.
10286 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10287 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10290 // If the function takes more arguments than the call was taking, add them
10292 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
10293 Args.push_back(Context->getNullValue(FT->getParamType(i)));
10295 // If we are removing arguments to the function, emit an obnoxious warning...
10296 if (FT->getNumParams() < NumActualArgs) {
10297 if (!FT->isVarArg()) {
10298 cerr << "WARNING: While resolving call to function '"
10299 << Callee->getName() << "' arguments were dropped!\n";
10301 // Add all of the arguments in their promoted form to the arg list...
10302 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
10303 const Type *PTy = getPromotedType((*AI)->getType());
10304 if (PTy != (*AI)->getType()) {
10305 // Must promote to pass through va_arg area!
10306 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
10308 Instruction *Cast = CastInst::Create(opcode, *AI, PTy, "tmp");
10309 InsertNewInstBefore(Cast, *Caller);
10310 Args.push_back(Cast);
10312 Args.push_back(*AI);
10315 // Add any parameter attributes.
10316 if (Attributes PAttrs = CallerPAL.getParamAttributes(i + 1))
10317 attrVec.push_back(AttributeWithIndex::get(i + 1, PAttrs));
10322 if (Attributes FnAttrs = CallerPAL.getFnAttributes())
10323 attrVec.push_back(AttributeWithIndex::get(~0, FnAttrs));
10325 if (NewRetTy == Type::VoidTy)
10326 Caller->setName(""); // Void type should not have a name.
10328 const AttrListPtr &NewCallerPAL = AttrListPtr::get(attrVec.begin(),attrVec.end());
10331 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10332 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
10333 Args.begin(), Args.end(),
10334 Caller->getName(), Caller);
10335 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
10336 cast<InvokeInst>(NC)->setAttributes(NewCallerPAL);
10338 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
10339 Caller->getName(), Caller);
10340 CallInst *CI = cast<CallInst>(Caller);
10341 if (CI->isTailCall())
10342 cast<CallInst>(NC)->setTailCall();
10343 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
10344 cast<CallInst>(NC)->setAttributes(NewCallerPAL);
10347 // Insert a cast of the return type as necessary.
10349 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
10350 if (NV->getType() != Type::VoidTy) {
10351 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
10353 NV = NC = CastInst::Create(opcode, NC, OldRetTy, "tmp");
10355 // If this is an invoke instruction, we should insert it after the first
10356 // non-phi, instruction in the normal successor block.
10357 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10358 BasicBlock::iterator I = II->getNormalDest()->getFirstNonPHI();
10359 InsertNewInstBefore(NC, *I);
10361 // Otherwise, it's a call, just insert cast right after the call instr
10362 InsertNewInstBefore(NC, *Caller);
10364 AddUsersToWorkList(*Caller);
10366 NV = Context->getUndef(Caller->getType());
10370 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10371 Caller->replaceAllUsesWith(NV);
10372 Caller->eraseFromParent();
10373 RemoveFromWorkList(Caller);
10377 // transformCallThroughTrampoline - Turn a call to a function created by the
10378 // init_trampoline intrinsic into a direct call to the underlying function.
10380 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
10381 Value *Callee = CS.getCalledValue();
10382 const PointerType *PTy = cast<PointerType>(Callee->getType());
10383 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
10384 const AttrListPtr &Attrs = CS.getAttributes();
10386 // If the call already has the 'nest' attribute somewhere then give up -
10387 // otherwise 'nest' would occur twice after splicing in the chain.
10388 if (Attrs.hasAttrSomewhere(Attribute::Nest))
10391 IntrinsicInst *Tramp =
10392 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
10394 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
10395 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
10396 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
10398 const AttrListPtr &NestAttrs = NestF->getAttributes();
10399 if (!NestAttrs.isEmpty()) {
10400 unsigned NestIdx = 1;
10401 const Type *NestTy = 0;
10402 Attributes NestAttr = Attribute::None;
10404 // Look for a parameter marked with the 'nest' attribute.
10405 for (FunctionType::param_iterator I = NestFTy->param_begin(),
10406 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
10407 if (NestAttrs.paramHasAttr(NestIdx, Attribute::Nest)) {
10408 // Record the parameter type and any other attributes.
10410 NestAttr = NestAttrs.getParamAttributes(NestIdx);
10415 Instruction *Caller = CS.getInstruction();
10416 std::vector<Value*> NewArgs;
10417 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
10419 SmallVector<AttributeWithIndex, 8> NewAttrs;
10420 NewAttrs.reserve(Attrs.getNumSlots() + 1);
10422 // Insert the nest argument into the call argument list, which may
10423 // mean appending it. Likewise for attributes.
10425 // Add any result attributes.
10426 if (Attributes Attr = Attrs.getRetAttributes())
10427 NewAttrs.push_back(AttributeWithIndex::get(0, Attr));
10431 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
10433 if (Idx == NestIdx) {
10434 // Add the chain argument and attributes.
10435 Value *NestVal = Tramp->getOperand(3);
10436 if (NestVal->getType() != NestTy)
10437 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
10438 NewArgs.push_back(NestVal);
10439 NewAttrs.push_back(AttributeWithIndex::get(NestIdx, NestAttr));
10445 // Add the original argument and attributes.
10446 NewArgs.push_back(*I);
10447 if (Attributes Attr = Attrs.getParamAttributes(Idx))
10449 (AttributeWithIndex::get(Idx + (Idx >= NestIdx), Attr));
10455 // Add any function attributes.
10456 if (Attributes Attr = Attrs.getFnAttributes())
10457 NewAttrs.push_back(AttributeWithIndex::get(~0, Attr));
10459 // The trampoline may have been bitcast to a bogus type (FTy).
10460 // Handle this by synthesizing a new function type, equal to FTy
10461 // with the chain parameter inserted.
10463 std::vector<const Type*> NewTypes;
10464 NewTypes.reserve(FTy->getNumParams()+1);
10466 // Insert the chain's type into the list of parameter types, which may
10467 // mean appending it.
10470 FunctionType::param_iterator I = FTy->param_begin(),
10471 E = FTy->param_end();
10474 if (Idx == NestIdx)
10475 // Add the chain's type.
10476 NewTypes.push_back(NestTy);
10481 // Add the original type.
10482 NewTypes.push_back(*I);
10488 // Replace the trampoline call with a direct call. Let the generic
10489 // code sort out any function type mismatches.
10490 FunctionType *NewFTy =
10491 Context->getFunctionType(FTy->getReturnType(), NewTypes,
10493 Constant *NewCallee =
10494 NestF->getType() == Context->getPointerTypeUnqual(NewFTy) ?
10495 NestF : Context->getConstantExprBitCast(NestF,
10496 Context->getPointerTypeUnqual(NewFTy));
10497 const AttrListPtr &NewPAL = AttrListPtr::get(NewAttrs.begin(),NewAttrs.end());
10499 Instruction *NewCaller;
10500 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
10501 NewCaller = InvokeInst::Create(NewCallee,
10502 II->getNormalDest(), II->getUnwindDest(),
10503 NewArgs.begin(), NewArgs.end(),
10504 Caller->getName(), Caller);
10505 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
10506 cast<InvokeInst>(NewCaller)->setAttributes(NewPAL);
10508 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
10509 Caller->getName(), Caller);
10510 if (cast<CallInst>(Caller)->isTailCall())
10511 cast<CallInst>(NewCaller)->setTailCall();
10512 cast<CallInst>(NewCaller)->
10513 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
10514 cast<CallInst>(NewCaller)->setAttributes(NewPAL);
10516 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
10517 Caller->replaceAllUsesWith(NewCaller);
10518 Caller->eraseFromParent();
10519 RemoveFromWorkList(Caller);
10524 // Replace the trampoline call with a direct call. Since there is no 'nest'
10525 // parameter, there is no need to adjust the argument list. Let the generic
10526 // code sort out any function type mismatches.
10527 Constant *NewCallee =
10528 NestF->getType() == PTy ? NestF :
10529 Context->getConstantExprBitCast(NestF, PTy);
10530 CS.setCalledFunction(NewCallee);
10531 return CS.getInstruction();
10534 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
10535 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
10536 /// and a single binop.
10537 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
10538 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10539 assert(isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst));
10540 unsigned Opc = FirstInst->getOpcode();
10541 Value *LHSVal = FirstInst->getOperand(0);
10542 Value *RHSVal = FirstInst->getOperand(1);
10544 const Type *LHSType = LHSVal->getType();
10545 const Type *RHSType = RHSVal->getType();
10547 // Scan to see if all operands are the same opcode, all have one use, and all
10548 // kill their operands (i.e. the operands have one use).
10549 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10550 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
10551 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
10552 // Verify type of the LHS matches so we don't fold cmp's of different
10553 // types or GEP's with different index types.
10554 I->getOperand(0)->getType() != LHSType ||
10555 I->getOperand(1)->getType() != RHSType)
10558 // If they are CmpInst instructions, check their predicates
10559 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
10560 if (cast<CmpInst>(I)->getPredicate() !=
10561 cast<CmpInst>(FirstInst)->getPredicate())
10564 // Keep track of which operand needs a phi node.
10565 if (I->getOperand(0) != LHSVal) LHSVal = 0;
10566 if (I->getOperand(1) != RHSVal) RHSVal = 0;
10569 // Otherwise, this is safe to transform!
10571 Value *InLHS = FirstInst->getOperand(0);
10572 Value *InRHS = FirstInst->getOperand(1);
10573 PHINode *NewLHS = 0, *NewRHS = 0;
10575 NewLHS = PHINode::Create(LHSType,
10576 FirstInst->getOperand(0)->getName() + ".pn");
10577 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
10578 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
10579 InsertNewInstBefore(NewLHS, PN);
10584 NewRHS = PHINode::Create(RHSType,
10585 FirstInst->getOperand(1)->getName() + ".pn");
10586 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
10587 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
10588 InsertNewInstBefore(NewRHS, PN);
10592 // Add all operands to the new PHIs.
10593 if (NewLHS || NewRHS) {
10594 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10595 Instruction *InInst = cast<Instruction>(PN.getIncomingValue(i));
10597 Value *NewInLHS = InInst->getOperand(0);
10598 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
10601 Value *NewInRHS = InInst->getOperand(1);
10602 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
10607 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10608 return BinaryOperator::Create(BinOp->getOpcode(), LHSVal, RHSVal);
10609 CmpInst *CIOp = cast<CmpInst>(FirstInst);
10610 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10614 Instruction *InstCombiner::FoldPHIArgGEPIntoPHI(PHINode &PN) {
10615 GetElementPtrInst *FirstInst =cast<GetElementPtrInst>(PN.getIncomingValue(0));
10617 SmallVector<Value*, 16> FixedOperands(FirstInst->op_begin(),
10618 FirstInst->op_end());
10619 // This is true if all GEP bases are allocas and if all indices into them are
10621 bool AllBasePointersAreAllocas = true;
10623 // Scan to see if all operands are the same opcode, all have one use, and all
10624 // kill their operands (i.e. the operands have one use).
10625 for (unsigned i = 1; i != PN.getNumIncomingValues(); ++i) {
10626 GetElementPtrInst *GEP= dyn_cast<GetElementPtrInst>(PN.getIncomingValue(i));
10627 if (!GEP || !GEP->hasOneUse() || GEP->getType() != FirstInst->getType() ||
10628 GEP->getNumOperands() != FirstInst->getNumOperands())
10631 // Keep track of whether or not all GEPs are of alloca pointers.
10632 if (AllBasePointersAreAllocas &&
10633 (!isa<AllocaInst>(GEP->getOperand(0)) ||
10634 !GEP->hasAllConstantIndices()))
10635 AllBasePointersAreAllocas = false;
10637 // Compare the operand lists.
10638 for (unsigned op = 0, e = FirstInst->getNumOperands(); op != e; ++op) {
10639 if (FirstInst->getOperand(op) == GEP->getOperand(op))
10642 // Don't merge two GEPs when two operands differ (introducing phi nodes)
10643 // if one of the PHIs has a constant for the index. The index may be
10644 // substantially cheaper to compute for the constants, so making it a
10645 // variable index could pessimize the path. This also handles the case
10646 // for struct indices, which must always be constant.
10647 if (isa<ConstantInt>(FirstInst->getOperand(op)) ||
10648 isa<ConstantInt>(GEP->getOperand(op)))
10651 if (FirstInst->getOperand(op)->getType() !=GEP->getOperand(op)->getType())
10653 FixedOperands[op] = 0; // Needs a PHI.
10657 // If all of the base pointers of the PHI'd GEPs are from allocas, don't
10658 // bother doing this transformation. At best, this will just save a bit of
10659 // offset calculation, but all the predecessors will have to materialize the
10660 // stack address into a register anyway. We'd actually rather *clone* the
10661 // load up into the predecessors so that we have a load of a gep of an alloca,
10662 // which can usually all be folded into the load.
10663 if (AllBasePointersAreAllocas)
10666 // Otherwise, this is safe to transform. Insert PHI nodes for each operand
10667 // that is variable.
10668 SmallVector<PHINode*, 16> OperandPhis(FixedOperands.size());
10670 bool HasAnyPHIs = false;
10671 for (unsigned i = 0, e = FixedOperands.size(); i != e; ++i) {
10672 if (FixedOperands[i]) continue; // operand doesn't need a phi.
10673 Value *FirstOp = FirstInst->getOperand(i);
10674 PHINode *NewPN = PHINode::Create(FirstOp->getType(),
10675 FirstOp->getName()+".pn");
10676 InsertNewInstBefore(NewPN, PN);
10678 NewPN->reserveOperandSpace(e);
10679 NewPN->addIncoming(FirstOp, PN.getIncomingBlock(0));
10680 OperandPhis[i] = NewPN;
10681 FixedOperands[i] = NewPN;
10686 // Add all operands to the new PHIs.
10688 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10689 GetElementPtrInst *InGEP =cast<GetElementPtrInst>(PN.getIncomingValue(i));
10690 BasicBlock *InBB = PN.getIncomingBlock(i);
10692 for (unsigned op = 0, e = OperandPhis.size(); op != e; ++op)
10693 if (PHINode *OpPhi = OperandPhis[op])
10694 OpPhi->addIncoming(InGEP->getOperand(op), InBB);
10698 Value *Base = FixedOperands[0];
10699 return GetElementPtrInst::Create(Base, FixedOperands.begin()+1,
10700 FixedOperands.end());
10704 /// isSafeAndProfitableToSinkLoad - Return true if we know that it is safe to
10705 /// sink the load out of the block that defines it. This means that it must be
10706 /// obvious the value of the load is not changed from the point of the load to
10707 /// the end of the block it is in.
10709 /// Finally, it is safe, but not profitable, to sink a load targetting a
10710 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
10712 static bool isSafeAndProfitableToSinkLoad(LoadInst *L) {
10713 BasicBlock::iterator BBI = L, E = L->getParent()->end();
10715 for (++BBI; BBI != E; ++BBI)
10716 if (BBI->mayWriteToMemory())
10719 // Check for non-address taken alloca. If not address-taken already, it isn't
10720 // profitable to do this xform.
10721 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
10722 bool isAddressTaken = false;
10723 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
10725 if (isa<LoadInst>(UI)) continue;
10726 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
10727 // If storing TO the alloca, then the address isn't taken.
10728 if (SI->getOperand(1) == AI) continue;
10730 isAddressTaken = true;
10734 if (!isAddressTaken && AI->isStaticAlloca())
10738 // If this load is a load from a GEP with a constant offset from an alloca,
10739 // then we don't want to sink it. In its present form, it will be
10740 // load [constant stack offset]. Sinking it will cause us to have to
10741 // materialize the stack addresses in each predecessor in a register only to
10742 // do a shared load from register in the successor.
10743 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(L->getOperand(0)))
10744 if (AllocaInst *AI = dyn_cast<AllocaInst>(GEP->getOperand(0)))
10745 if (AI->isStaticAlloca() && GEP->hasAllConstantIndices())
10752 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
10753 // operator and they all are only used by the PHI, PHI together their
10754 // inputs, and do the operation once, to the result of the PHI.
10755 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
10756 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
10758 // Scan the instruction, looking for input operations that can be folded away.
10759 // If all input operands to the phi are the same instruction (e.g. a cast from
10760 // the same type or "+42") we can pull the operation through the PHI, reducing
10761 // code size and simplifying code.
10762 Constant *ConstantOp = 0;
10763 const Type *CastSrcTy = 0;
10764 bool isVolatile = false;
10765 if (isa<CastInst>(FirstInst)) {
10766 CastSrcTy = FirstInst->getOperand(0)->getType();
10767 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
10768 // Can fold binop, compare or shift here if the RHS is a constant,
10769 // otherwise call FoldPHIArgBinOpIntoPHI.
10770 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
10771 if (ConstantOp == 0)
10772 return FoldPHIArgBinOpIntoPHI(PN);
10773 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
10774 isVolatile = LI->isVolatile();
10775 // We can't sink the load if the loaded value could be modified between the
10776 // load and the PHI.
10777 if (LI->getParent() != PN.getIncomingBlock(0) ||
10778 !isSafeAndProfitableToSinkLoad(LI))
10781 // If the PHI is of volatile loads and the load block has multiple
10782 // successors, sinking it would remove a load of the volatile value from
10783 // the path through the other successor.
10785 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10788 } else if (isa<GetElementPtrInst>(FirstInst)) {
10789 return FoldPHIArgGEPIntoPHI(PN);
10791 return 0; // Cannot fold this operation.
10794 // Check to see if all arguments are the same operation.
10795 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10796 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
10797 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
10798 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
10801 if (I->getOperand(0)->getType() != CastSrcTy)
10802 return 0; // Cast operation must match.
10803 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10804 // We can't sink the load if the loaded value could be modified between
10805 // the load and the PHI.
10806 if (LI->isVolatile() != isVolatile ||
10807 LI->getParent() != PN.getIncomingBlock(i) ||
10808 !isSafeAndProfitableToSinkLoad(LI))
10811 // If the PHI is of volatile loads and the load block has multiple
10812 // successors, sinking it would remove a load of the volatile value from
10813 // the path through the other successor.
10815 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
10818 } else if (I->getOperand(1) != ConstantOp) {
10823 // Okay, they are all the same operation. Create a new PHI node of the
10824 // correct type, and PHI together all of the LHS's of the instructions.
10825 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
10826 PN.getName()+".in");
10827 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
10829 Value *InVal = FirstInst->getOperand(0);
10830 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
10832 // Add all operands to the new PHI.
10833 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
10834 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
10835 if (NewInVal != InVal)
10837 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
10842 // The new PHI unions all of the same values together. This is really
10843 // common, so we handle it intelligently here for compile-time speed.
10847 InsertNewInstBefore(NewPN, PN);
10851 // Insert and return the new operation.
10852 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
10853 return CastInst::Create(FirstCI->getOpcode(), PhiVal, PN.getType());
10854 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
10855 return BinaryOperator::Create(BinOp->getOpcode(), PhiVal, ConstantOp);
10856 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
10857 return CmpInst::Create(*Context, CIOp->getOpcode(), CIOp->getPredicate(),
10858 PhiVal, ConstantOp);
10859 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
10861 // If this was a volatile load that we are merging, make sure to loop through
10862 // and mark all the input loads as non-volatile. If we don't do this, we will
10863 // insert a new volatile load and the old ones will not be deletable.
10865 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
10866 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
10868 return new LoadInst(PhiVal, "", isVolatile);
10871 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
10873 static bool DeadPHICycle(PHINode *PN,
10874 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
10875 if (PN->use_empty()) return true;
10876 if (!PN->hasOneUse()) return false;
10878 // Remember this node, and if we find the cycle, return.
10879 if (!PotentiallyDeadPHIs.insert(PN))
10882 // Don't scan crazily complex things.
10883 if (PotentiallyDeadPHIs.size() == 16)
10886 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
10887 return DeadPHICycle(PU, PotentiallyDeadPHIs);
10892 /// PHIsEqualValue - Return true if this phi node is always equal to
10893 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
10894 /// z = some value; x = phi (y, z); y = phi (x, z)
10895 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
10896 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
10897 // See if we already saw this PHI node.
10898 if (!ValueEqualPHIs.insert(PN))
10901 // Don't scan crazily complex things.
10902 if (ValueEqualPHIs.size() == 16)
10905 // Scan the operands to see if they are either phi nodes or are equal to
10907 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
10908 Value *Op = PN->getIncomingValue(i);
10909 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
10910 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
10912 } else if (Op != NonPhiInVal)
10920 // PHINode simplification
10922 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
10923 // If LCSSA is around, don't mess with Phi nodes
10924 if (MustPreserveLCSSA) return 0;
10926 if (Value *V = PN.hasConstantValue())
10927 return ReplaceInstUsesWith(PN, V);
10929 // If all PHI operands are the same operation, pull them through the PHI,
10930 // reducing code size.
10931 if (isa<Instruction>(PN.getIncomingValue(0)) &&
10932 isa<Instruction>(PN.getIncomingValue(1)) &&
10933 cast<Instruction>(PN.getIncomingValue(0))->getOpcode() ==
10934 cast<Instruction>(PN.getIncomingValue(1))->getOpcode() &&
10935 // FIXME: The hasOneUse check will fail for PHIs that use the value more
10936 // than themselves more than once.
10937 PN.getIncomingValue(0)->hasOneUse())
10938 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
10941 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
10942 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
10943 // PHI)... break the cycle.
10944 if (PN.hasOneUse()) {
10945 Instruction *PHIUser = cast<Instruction>(PN.use_back());
10946 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
10947 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
10948 PotentiallyDeadPHIs.insert(&PN);
10949 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
10950 return ReplaceInstUsesWith(PN, Context->getUndef(PN.getType()));
10953 // If this phi has a single use, and if that use just computes a value for
10954 // the next iteration of a loop, delete the phi. This occurs with unused
10955 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
10956 // common case here is good because the only other things that catch this
10957 // are induction variable analysis (sometimes) and ADCE, which is only run
10959 if (PHIUser->hasOneUse() &&
10960 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
10961 PHIUser->use_back() == &PN) {
10962 return ReplaceInstUsesWith(PN, Context->getUndef(PN.getType()));
10966 // We sometimes end up with phi cycles that non-obviously end up being the
10967 // same value, for example:
10968 // z = some value; x = phi (y, z); y = phi (x, z)
10969 // where the phi nodes don't necessarily need to be in the same block. Do a
10970 // quick check to see if the PHI node only contains a single non-phi value, if
10971 // so, scan to see if the phi cycle is actually equal to that value.
10973 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
10974 // Scan for the first non-phi operand.
10975 while (InValNo != NumOperandVals &&
10976 isa<PHINode>(PN.getIncomingValue(InValNo)))
10979 if (InValNo != NumOperandVals) {
10980 Value *NonPhiInVal = PN.getOperand(InValNo);
10982 // Scan the rest of the operands to see if there are any conflicts, if so
10983 // there is no need to recursively scan other phis.
10984 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
10985 Value *OpVal = PN.getIncomingValue(InValNo);
10986 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
10990 // If we scanned over all operands, then we have one unique value plus
10991 // phi values. Scan PHI nodes to see if they all merge in each other or
10993 if (InValNo == NumOperandVals) {
10994 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
10995 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
10996 return ReplaceInstUsesWith(PN, NonPhiInVal);
11003 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
11004 Instruction *InsertPoint,
11005 InstCombiner *IC) {
11006 unsigned PtrSize = DTy->getScalarSizeInBits();
11007 unsigned VTySize = V->getType()->getScalarSizeInBits();
11008 // We must cast correctly to the pointer type. Ensure that we
11009 // sign extend the integer value if it is smaller as this is
11010 // used for address computation.
11011 Instruction::CastOps opcode =
11012 (VTySize < PtrSize ? Instruction::SExt :
11013 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
11014 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
11018 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
11019 Value *PtrOp = GEP.getOperand(0);
11020 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
11021 // If so, eliminate the noop.
11022 if (GEP.getNumOperands() == 1)
11023 return ReplaceInstUsesWith(GEP, PtrOp);
11025 if (isa<UndefValue>(GEP.getOperand(0)))
11026 return ReplaceInstUsesWith(GEP, Context->getUndef(GEP.getType()));
11028 bool HasZeroPointerIndex = false;
11029 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
11030 HasZeroPointerIndex = C->isNullValue();
11032 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
11033 return ReplaceInstUsesWith(GEP, PtrOp);
11035 // Eliminate unneeded casts for indices.
11036 bool MadeChange = false;
11038 gep_type_iterator GTI = gep_type_begin(GEP);
11039 for (User::op_iterator i = GEP.op_begin() + 1, e = GEP.op_end();
11040 i != e; ++i, ++GTI) {
11041 if (isa<SequentialType>(*GTI)) {
11042 if (CastInst *CI = dyn_cast<CastInst>(*i)) {
11043 if (CI->getOpcode() == Instruction::ZExt ||
11044 CI->getOpcode() == Instruction::SExt) {
11045 const Type *SrcTy = CI->getOperand(0)->getType();
11046 // We can eliminate a cast from i32 to i64 iff the target
11047 // is a 32-bit pointer target.
11048 if (SrcTy->getScalarSizeInBits() >= TD->getPointerSizeInBits()) {
11050 *i = CI->getOperand(0);
11054 // If we are using a wider index than needed for this platform, shrink it
11055 // to what we need. If narrower, sign-extend it to what we need.
11056 // If the incoming value needs a cast instruction,
11057 // insert it. This explicit cast can make subsequent optimizations more
11060 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
11061 if (Constant *C = dyn_cast<Constant>(Op)) {
11062 *i = Context->getConstantExprTrunc(C, TD->getIntPtrType());
11065 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
11070 } else if (TD->getTypeSizeInBits(Op->getType()) < TD->getPointerSizeInBits()) {
11071 if (Constant *C = dyn_cast<Constant>(Op)) {
11072 *i = Context->getConstantExprSExt(C, TD->getIntPtrType());
11075 Op = InsertCastBefore(Instruction::SExt, Op, TD->getIntPtrType(),
11083 if (MadeChange) return &GEP;
11085 // Combine Indices - If the source pointer to this getelementptr instruction
11086 // is a getelementptr instruction, combine the indices of the two
11087 // getelementptr instructions into a single instruction.
11089 SmallVector<Value*, 8> SrcGEPOperands;
11090 if (User *Src = dyn_castGetElementPtr(PtrOp))
11091 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
11093 if (!SrcGEPOperands.empty()) {
11094 // Note that if our source is a gep chain itself that we wait for that
11095 // chain to be resolved before we perform this transformation. This
11096 // avoids us creating a TON of code in some cases.
11098 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
11099 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
11100 return 0; // Wait until our source is folded to completion.
11102 SmallVector<Value*, 8> Indices;
11104 // Find out whether the last index in the source GEP is a sequential idx.
11105 bool EndsWithSequential = false;
11106 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
11107 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
11108 EndsWithSequential = !isa<StructType>(*I);
11110 // Can we combine the two pointer arithmetics offsets?
11111 if (EndsWithSequential) {
11112 // Replace: gep (gep %P, long B), long A, ...
11113 // With: T = long A+B; gep %P, T, ...
11115 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
11116 if (SO1 == Context->getNullValue(SO1->getType())) {
11118 } else if (GO1 == Context->getNullValue(GO1->getType())) {
11121 // If they aren't the same type, convert both to an integer of the
11122 // target's pointer size.
11123 if (SO1->getType() != GO1->getType()) {
11124 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
11126 Context->getConstantExprIntegerCast(SO1C, GO1->getType(), true);
11127 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
11129 Context->getConstantExprIntegerCast(GO1C, SO1->getType(), true);
11131 unsigned PS = TD->getPointerSizeInBits();
11132 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
11133 // Convert GO1 to SO1's type.
11134 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
11136 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
11137 // Convert SO1 to GO1's type.
11138 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
11140 const Type *PT = TD->getIntPtrType();
11141 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
11142 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
11146 if (isa<Constant>(SO1) && isa<Constant>(GO1))
11147 Sum = Context->getConstantExprAdd(cast<Constant>(SO1),
11148 cast<Constant>(GO1));
11150 Sum = BinaryOperator::CreateAdd(SO1, GO1, PtrOp->getName()+".sum");
11151 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
11155 // Recycle the GEP we already have if possible.
11156 if (SrcGEPOperands.size() == 2) {
11157 GEP.setOperand(0, SrcGEPOperands[0]);
11158 GEP.setOperand(1, Sum);
11161 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11162 SrcGEPOperands.end()-1);
11163 Indices.push_back(Sum);
11164 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
11166 } else if (isa<Constant>(*GEP.idx_begin()) &&
11167 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
11168 SrcGEPOperands.size() != 1) {
11169 // Otherwise we can do the fold if the first index of the GEP is a zero
11170 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
11171 SrcGEPOperands.end());
11172 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
11175 if (!Indices.empty())
11176 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
11177 Indices.end(), GEP.getName());
11179 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
11180 // GEP of global variable. If all of the indices for this GEP are
11181 // constants, we can promote this to a constexpr instead of an instruction.
11183 // Scan for nonconstants...
11184 SmallVector<Constant*, 8> Indices;
11185 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
11186 for (; I != E && isa<Constant>(*I); ++I)
11187 Indices.push_back(cast<Constant>(*I));
11189 if (I == E) { // If they are all constants...
11190 Constant *CE = Context->getConstantExprGetElementPtr(GV,
11191 &Indices[0],Indices.size());
11193 // Replace all uses of the GEP with the new constexpr...
11194 return ReplaceInstUsesWith(GEP, CE);
11196 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
11197 if (!isa<PointerType>(X->getType())) {
11198 // Not interesting. Source pointer must be a cast from pointer.
11199 } else if (HasZeroPointerIndex) {
11200 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
11201 // into : GEP [10 x i8]* X, i32 0, ...
11203 // Likewise, transform: GEP (bitcast i8* X to [0 x i8]*), i32 0, ...
11204 // into : GEP i8* X, ...
11206 // This occurs when the program declares an array extern like "int X[];"
11207 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
11208 const PointerType *XTy = cast<PointerType>(X->getType());
11209 if (const ArrayType *CATy =
11210 dyn_cast<ArrayType>(CPTy->getElementType())) {
11211 // GEP (bitcast i8* X to [0 x i8]*), i32 0, ... ?
11212 if (CATy->getElementType() == XTy->getElementType()) {
11213 // -> GEP i8* X, ...
11214 SmallVector<Value*, 8> Indices(GEP.idx_begin()+1, GEP.idx_end());
11215 return GetElementPtrInst::Create(X, Indices.begin(), Indices.end(),
11217 } else if (const ArrayType *XATy =
11218 dyn_cast<ArrayType>(XTy->getElementType())) {
11219 // GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ... ?
11220 if (CATy->getElementType() == XATy->getElementType()) {
11221 // -> GEP [10 x i8]* X, i32 0, ...
11222 // At this point, we know that the cast source type is a pointer
11223 // to an array of the same type as the destination pointer
11224 // array. Because the array type is never stepped over (there
11225 // is a leading zero) we can fold the cast into this GEP.
11226 GEP.setOperand(0, X);
11231 } else if (GEP.getNumOperands() == 2) {
11232 // Transform things like:
11233 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
11234 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
11235 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
11236 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
11237 if (isa<ArrayType>(SrcElTy) &&
11238 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
11239 TD->getTypeAllocSize(ResElTy)) {
11241 Idx[0] = Context->getNullValue(Type::Int32Ty);
11242 Idx[1] = GEP.getOperand(1);
11243 Value *V = InsertNewInstBefore(
11244 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
11245 // V and GEP are both pointer types --> BitCast
11246 return new BitCastInst(V, GEP.getType());
11249 // Transform things like:
11250 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
11251 // (where tmp = 8*tmp2) into:
11252 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
11254 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
11255 uint64_t ArrayEltSize =
11256 TD->getTypeAllocSize(cast<ArrayType>(SrcElTy)->getElementType());
11258 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
11259 // allow either a mul, shift, or constant here.
11261 ConstantInt *Scale = 0;
11262 if (ArrayEltSize == 1) {
11263 NewIdx = GEP.getOperand(1);
11265 Context->getConstantInt(cast<IntegerType>(NewIdx->getType()), 1);
11266 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
11267 NewIdx = Context->getConstantInt(CI->getType(), 1);
11269 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
11270 if (Inst->getOpcode() == Instruction::Shl &&
11271 isa<ConstantInt>(Inst->getOperand(1))) {
11272 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
11273 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
11274 Scale = Context->getConstantInt(cast<IntegerType>(Inst->getType()),
11276 NewIdx = Inst->getOperand(0);
11277 } else if (Inst->getOpcode() == Instruction::Mul &&
11278 isa<ConstantInt>(Inst->getOperand(1))) {
11279 Scale = cast<ConstantInt>(Inst->getOperand(1));
11280 NewIdx = Inst->getOperand(0);
11284 // If the index will be to exactly the right offset with the scale taken
11285 // out, perform the transformation. Note, we don't know whether Scale is
11286 // signed or not. We'll use unsigned version of division/modulo
11287 // operation after making sure Scale doesn't have the sign bit set.
11288 if (ArrayEltSize && Scale && Scale->getSExtValue() >= 0LL &&
11289 Scale->getZExtValue() % ArrayEltSize == 0) {
11290 Scale = Context->getConstantInt(Scale->getType(),
11291 Scale->getZExtValue() / ArrayEltSize);
11292 if (Scale->getZExtValue() != 1) {
11294 Context->getConstantExprIntegerCast(Scale, NewIdx->getType(),
11296 Instruction *Sc = BinaryOperator::CreateMul(NewIdx, C, "idxscale");
11297 NewIdx = InsertNewInstBefore(Sc, GEP);
11300 // Insert the new GEP instruction.
11302 Idx[0] = Context->getNullValue(Type::Int32Ty);
11304 Instruction *NewGEP =
11305 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
11306 NewGEP = InsertNewInstBefore(NewGEP, GEP);
11307 // The NewGEP must be pointer typed, so must the old one -> BitCast
11308 return new BitCastInst(NewGEP, GEP.getType());
11314 /// See if we can simplify:
11315 /// X = bitcast A to B*
11316 /// Y = gep X, <...constant indices...>
11317 /// into a gep of the original struct. This is important for SROA and alias
11318 /// analysis of unions. If "A" is also a bitcast, wait for A/X to be merged.
11319 if (BitCastInst *BCI = dyn_cast<BitCastInst>(PtrOp)) {
11320 if (!isa<BitCastInst>(BCI->getOperand(0)) && GEP.hasAllConstantIndices()) {
11321 // Determine how much the GEP moves the pointer. We are guaranteed to get
11322 // a constant back from EmitGEPOffset.
11323 ConstantInt *OffsetV =
11324 cast<ConstantInt>(EmitGEPOffset(&GEP, GEP, *this));
11325 int64_t Offset = OffsetV->getSExtValue();
11327 // If this GEP instruction doesn't move the pointer, just replace the GEP
11328 // with a bitcast of the real input to the dest type.
11330 // If the bitcast is of an allocation, and the allocation will be
11331 // converted to match the type of the cast, don't touch this.
11332 if (isa<AllocationInst>(BCI->getOperand(0))) {
11333 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
11334 if (Instruction *I = visitBitCast(*BCI)) {
11337 BCI->getParent()->getInstList().insert(BCI, I);
11338 ReplaceInstUsesWith(*BCI, I);
11343 return new BitCastInst(BCI->getOperand(0), GEP.getType());
11346 // Otherwise, if the offset is non-zero, we need to find out if there is a
11347 // field at Offset in 'A's type. If so, we can pull the cast through the
11349 SmallVector<Value*, 8> NewIndices;
11351 cast<PointerType>(BCI->getOperand(0)->getType())->getElementType();
11352 if (FindElementAtOffset(InTy, Offset, NewIndices, TD, Context)) {
11353 Instruction *NGEP =
11354 GetElementPtrInst::Create(BCI->getOperand(0), NewIndices.begin(),
11356 if (NGEP->getType() == GEP.getType()) return NGEP;
11357 InsertNewInstBefore(NGEP, GEP);
11358 NGEP->takeName(&GEP);
11359 return new BitCastInst(NGEP, GEP.getType());
11367 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
11368 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
11369 if (AI.isArrayAllocation()) { // Check C != 1
11370 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
11371 const Type *NewTy =
11372 Context->getArrayType(AI.getAllocatedType(), C->getZExtValue());
11373 AllocationInst *New = 0;
11375 // Create and insert the replacement instruction...
11376 if (isa<MallocInst>(AI))
11377 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
11379 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
11380 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
11383 InsertNewInstBefore(New, AI);
11385 // Scan to the end of the allocation instructions, to skip over a block of
11386 // allocas if possible...also skip interleaved debug info
11388 BasicBlock::iterator It = New;
11389 while (isa<AllocationInst>(*It) || isa<DbgInfoIntrinsic>(*It)) ++It;
11391 // Now that I is pointing to the first non-allocation-inst in the block,
11392 // insert our getelementptr instruction...
11394 Value *NullIdx = Context->getNullValue(Type::Int32Ty);
11398 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
11399 New->getName()+".sub", It);
11401 // Now make everything use the getelementptr instead of the original
11403 return ReplaceInstUsesWith(AI, V);
11404 } else if (isa<UndefValue>(AI.getArraySize())) {
11405 return ReplaceInstUsesWith(AI, Context->getNullValue(AI.getType()));
11409 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized()) {
11410 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
11411 // Note that we only do this for alloca's, because malloc should allocate
11412 // and return a unique pointer, even for a zero byte allocation.
11413 if (TD->getTypeAllocSize(AI.getAllocatedType()) == 0)
11414 return ReplaceInstUsesWith(AI, Context->getNullValue(AI.getType()));
11416 // If the alignment is 0 (unspecified), assign it the preferred alignment.
11417 if (AI.getAlignment() == 0)
11418 AI.setAlignment(TD->getPrefTypeAlignment(AI.getAllocatedType()));
11424 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
11425 Value *Op = FI.getOperand(0);
11427 // free undef -> unreachable.
11428 if (isa<UndefValue>(Op)) {
11429 // Insert a new store to null because we cannot modify the CFG here.
11430 new StoreInst(Context->getConstantIntTrue(),
11431 Context->getUndef(Context->getPointerTypeUnqual(Type::Int1Ty)), &FI);
11432 return EraseInstFromFunction(FI);
11435 // If we have 'free null' delete the instruction. This can happen in stl code
11436 // when lots of inlining happens.
11437 if (isa<ConstantPointerNull>(Op))
11438 return EraseInstFromFunction(FI);
11440 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
11441 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
11442 FI.setOperand(0, CI->getOperand(0));
11446 // Change free (gep X, 0,0,0,0) into free(X)
11447 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11448 if (GEPI->hasAllZeroIndices()) {
11449 AddToWorkList(GEPI);
11450 FI.setOperand(0, GEPI->getOperand(0));
11455 // Change free(malloc) into nothing, if the malloc has a single use.
11456 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
11457 if (MI->hasOneUse()) {
11458 EraseInstFromFunction(FI);
11459 return EraseInstFromFunction(*MI);
11466 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
11467 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
11468 const TargetData *TD) {
11469 User *CI = cast<User>(LI.getOperand(0));
11470 Value *CastOp = CI->getOperand(0);
11471 LLVMContext *Context = IC.getContext();
11474 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
11475 // Instead of loading constant c string, use corresponding integer value
11476 // directly if string length is small enough.
11478 if (GetConstantStringInfo(CE->getOperand(0), Str) && !Str.empty()) {
11479 unsigned len = Str.length();
11480 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
11481 unsigned numBits = Ty->getPrimitiveSizeInBits();
11482 // Replace LI with immediate integer store.
11483 if ((numBits >> 3) == len + 1) {
11484 APInt StrVal(numBits, 0);
11485 APInt SingleChar(numBits, 0);
11486 if (TD->isLittleEndian()) {
11487 for (signed i = len-1; i >= 0; i--) {
11488 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11489 StrVal = (StrVal << 8) | SingleChar;
11492 for (unsigned i = 0; i < len; i++) {
11493 SingleChar = (uint64_t) Str[i] & UCHAR_MAX;
11494 StrVal = (StrVal << 8) | SingleChar;
11496 // Append NULL at the end.
11498 StrVal = (StrVal << 8) | SingleChar;
11500 Value *NL = Context->getConstantInt(StrVal);
11501 return IC.ReplaceInstUsesWith(LI, NL);
11507 const PointerType *DestTy = cast<PointerType>(CI->getType());
11508 const Type *DestPTy = DestTy->getElementType();
11509 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
11511 // If the address spaces don't match, don't eliminate the cast.
11512 if (DestTy->getAddressSpace() != SrcTy->getAddressSpace())
11515 const Type *SrcPTy = SrcTy->getElementType();
11517 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
11518 isa<VectorType>(DestPTy)) {
11519 // If the source is an array, the code below will not succeed. Check to
11520 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11522 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
11523 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
11524 if (ASrcTy->getNumElements() != 0) {
11526 Idxs[0] = Idxs[1] = Context->getNullValue(Type::Int32Ty);
11527 CastOp = Context->getConstantExprGetElementPtr(CSrc, Idxs, 2);
11528 SrcTy = cast<PointerType>(CastOp->getType());
11529 SrcPTy = SrcTy->getElementType();
11532 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
11533 isa<VectorType>(SrcPTy)) &&
11534 // Do not allow turning this into a load of an integer, which is then
11535 // casted to a pointer, this pessimizes pointer analysis a lot.
11536 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
11537 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
11538 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
11540 // Okay, we are casting from one integer or pointer type to another of
11541 // the same size. Instead of casting the pointer before the load, cast
11542 // the result of the loaded value.
11543 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
11545 LI.isVolatile()),LI);
11546 // Now cast the result of the load.
11547 return new BitCastInst(NewLoad, LI.getType());
11554 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
11555 Value *Op = LI.getOperand(0);
11557 // Attempt to improve the alignment.
11558 unsigned KnownAlign =
11559 GetOrEnforceKnownAlignment(Op, TD->getPrefTypeAlignment(LI.getType()));
11561 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
11562 LI.getAlignment()))
11563 LI.setAlignment(KnownAlign);
11565 // load (cast X) --> cast (load X) iff safe
11566 if (isa<CastInst>(Op))
11567 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11570 // None of the following transforms are legal for volatile loads.
11571 if (LI.isVolatile()) return 0;
11573 // Do really simple store-to-load forwarding and load CSE, to catch cases
11574 // where there are several consequtive memory accesses to the same location,
11575 // separated by a few arithmetic operations.
11576 BasicBlock::iterator BBI = &LI;
11577 if (Value *AvailableVal = FindAvailableLoadedValue(Op, LI.getParent(), BBI,6))
11578 return ReplaceInstUsesWith(LI, AvailableVal);
11580 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
11581 const Value *GEPI0 = GEPI->getOperand(0);
11582 // TODO: Consider a target hook for valid address spaces for this xform.
11583 if (isa<ConstantPointerNull>(GEPI0) &&
11584 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
11585 // Insert a new store to null instruction before the load to indicate
11586 // that this code is not reachable. We do this instead of inserting
11587 // an unreachable instruction directly because we cannot modify the
11589 new StoreInst(Context->getUndef(LI.getType()),
11590 Context->getNullValue(Op->getType()), &LI);
11591 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11595 if (Constant *C = dyn_cast<Constant>(Op)) {
11596 // load null/undef -> undef
11597 // TODO: Consider a target hook for valid address spaces for this xform.
11598 if (isa<UndefValue>(C) || (C->isNullValue() &&
11599 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
11600 // Insert a new store to null instruction before the load to indicate that
11601 // this code is not reachable. We do this instead of inserting an
11602 // unreachable instruction directly because we cannot modify the CFG.
11603 new StoreInst(Context->getUndef(LI.getType()),
11604 Context->getNullValue(Op->getType()), &LI);
11605 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11608 // Instcombine load (constant global) into the value loaded.
11609 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
11610 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11611 return ReplaceInstUsesWith(LI, GV->getInitializer());
11613 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
11614 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
11615 if (CE->getOpcode() == Instruction::GetElementPtr) {
11616 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
11617 if (GV->isConstant() && GV->hasDefinitiveInitializer())
11619 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE,
11621 return ReplaceInstUsesWith(LI, V);
11622 if (CE->getOperand(0)->isNullValue()) {
11623 // Insert a new store to null instruction before the load to indicate
11624 // that this code is not reachable. We do this instead of inserting
11625 // an unreachable instruction directly because we cannot modify the
11627 new StoreInst(Context->getUndef(LI.getType()),
11628 Context->getNullValue(Op->getType()), &LI);
11629 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11632 } else if (CE->isCast()) {
11633 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
11639 // If this load comes from anywhere in a constant global, and if the global
11640 // is all undef or zero, we know what it loads.
11641 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op->getUnderlyingObject())){
11642 if (GV->isConstant() && GV->hasDefinitiveInitializer()) {
11643 if (GV->getInitializer()->isNullValue())
11644 return ReplaceInstUsesWith(LI, Context->getNullValue(LI.getType()));
11645 else if (isa<UndefValue>(GV->getInitializer()))
11646 return ReplaceInstUsesWith(LI, Context->getUndef(LI.getType()));
11650 if (Op->hasOneUse()) {
11651 // Change select and PHI nodes to select values instead of addresses: this
11652 // helps alias analysis out a lot, allows many others simplifications, and
11653 // exposes redundancy in the code.
11655 // Note that we cannot do the transformation unless we know that the
11656 // introduced loads cannot trap! Something like this is valid as long as
11657 // the condition is always false: load (select bool %C, int* null, int* %G),
11658 // but it would not be valid if we transformed it to load from null
11659 // unconditionally.
11661 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
11662 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
11663 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
11664 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
11665 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
11666 SI->getOperand(1)->getName()+".val"), LI);
11667 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
11668 SI->getOperand(2)->getName()+".val"), LI);
11669 return SelectInst::Create(SI->getCondition(), V1, V2);
11672 // load (select (cond, null, P)) -> load P
11673 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
11674 if (C->isNullValue()) {
11675 LI.setOperand(0, SI->getOperand(2));
11679 // load (select (cond, P, null)) -> load P
11680 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
11681 if (C->isNullValue()) {
11682 LI.setOperand(0, SI->getOperand(1));
11690 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
11691 /// when possible. This makes it generally easy to do alias analysis and/or
11692 /// SROA/mem2reg of the memory object.
11693 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
11694 User *CI = cast<User>(SI.getOperand(1));
11695 Value *CastOp = CI->getOperand(0);
11696 LLVMContext *Context = IC.getContext();
11698 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
11699 const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType());
11700 if (SrcTy == 0) return 0;
11702 const Type *SrcPTy = SrcTy->getElementType();
11704 if (!DestPTy->isInteger() && !isa<PointerType>(DestPTy))
11707 /// NewGEPIndices - If SrcPTy is an aggregate type, we can emit a "noop gep"
11708 /// to its first element. This allows us to handle things like:
11709 /// store i32 xxx, (bitcast {foo*, float}* %P to i32*)
11710 /// on 32-bit hosts.
11711 SmallVector<Value*, 4> NewGEPIndices;
11713 // If the source is an array, the code below will not succeed. Check to
11714 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
11716 if (isa<ArrayType>(SrcPTy) || isa<StructType>(SrcPTy)) {
11717 // Index through pointer.
11718 Constant *Zero = Context->getNullValue(Type::Int32Ty);
11719 NewGEPIndices.push_back(Zero);
11722 if (const StructType *STy = dyn_cast<StructType>(SrcPTy)) {
11723 if (!STy->getNumElements()) /* Struct can be empty {} */
11725 NewGEPIndices.push_back(Zero);
11726 SrcPTy = STy->getElementType(0);
11727 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcPTy)) {
11728 NewGEPIndices.push_back(Zero);
11729 SrcPTy = ATy->getElementType();
11735 SrcTy = Context->getPointerType(SrcPTy, SrcTy->getAddressSpace());
11738 if (!SrcPTy->isInteger() && !isa<PointerType>(SrcPTy))
11741 // If the pointers point into different address spaces or if they point to
11742 // values with different sizes, we can't do the transformation.
11743 if (SrcTy->getAddressSpace() !=
11744 cast<PointerType>(CI->getType())->getAddressSpace() ||
11745 IC.getTargetData().getTypeSizeInBits(SrcPTy) !=
11746 IC.getTargetData().getTypeSizeInBits(DestPTy))
11749 // Okay, we are casting from one integer or pointer type to another of
11750 // the same size. Instead of casting the pointer before
11751 // the store, cast the value to be stored.
11753 Value *SIOp0 = SI.getOperand(0);
11754 Instruction::CastOps opcode = Instruction::BitCast;
11755 const Type* CastSrcTy = SIOp0->getType();
11756 const Type* CastDstTy = SrcPTy;
11757 if (isa<PointerType>(CastDstTy)) {
11758 if (CastSrcTy->isInteger())
11759 opcode = Instruction::IntToPtr;
11760 } else if (isa<IntegerType>(CastDstTy)) {
11761 if (isa<PointerType>(SIOp0->getType()))
11762 opcode = Instruction::PtrToInt;
11765 // SIOp0 is a pointer to aggregate and this is a store to the first field,
11766 // emit a GEP to index into its first field.
11767 if (!NewGEPIndices.empty()) {
11768 if (Constant *C = dyn_cast<Constant>(CastOp))
11769 CastOp = Context->getConstantExprGetElementPtr(C, &NewGEPIndices[0],
11770 NewGEPIndices.size());
11772 CastOp = IC.InsertNewInstBefore(
11773 GetElementPtrInst::Create(CastOp, NewGEPIndices.begin(),
11774 NewGEPIndices.end()), SI);
11777 if (Constant *C = dyn_cast<Constant>(SIOp0))
11778 NewCast = Context->getConstantExprCast(opcode, C, CastDstTy);
11780 NewCast = IC.InsertNewInstBefore(
11781 CastInst::Create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
11783 return new StoreInst(NewCast, CastOp);
11786 /// equivalentAddressValues - Test if A and B will obviously have the same
11787 /// value. This includes recognizing that %t0 and %t1 will have the same
11788 /// value in code like this:
11789 /// %t0 = getelementptr \@a, 0, 3
11790 /// store i32 0, i32* %t0
11791 /// %t1 = getelementptr \@a, 0, 3
11792 /// %t2 = load i32* %t1
11794 static bool equivalentAddressValues(Value *A, Value *B) {
11795 // Test if the values are trivially equivalent.
11796 if (A == B) return true;
11798 // Test if the values come form identical arithmetic instructions.
11799 if (isa<BinaryOperator>(A) ||
11800 isa<CastInst>(A) ||
11802 isa<GetElementPtrInst>(A))
11803 if (Instruction *BI = dyn_cast<Instruction>(B))
11804 if (cast<Instruction>(A)->isIdenticalTo(BI))
11807 // Otherwise they may not be equivalent.
11811 // If this instruction has two uses, one of which is a llvm.dbg.declare,
11812 // return the llvm.dbg.declare.
11813 DbgDeclareInst *InstCombiner::hasOneUsePlusDeclare(Value *V) {
11814 if (!V->hasNUses(2))
11816 for (Value::use_iterator UI = V->use_begin(), E = V->use_end();
11818 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI))
11820 if (isa<BitCastInst>(UI) && UI->hasOneUse()) {
11821 if (DbgDeclareInst *DI = dyn_cast<DbgDeclareInst>(UI->use_begin()))
11828 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
11829 Value *Val = SI.getOperand(0);
11830 Value *Ptr = SI.getOperand(1);
11832 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
11833 EraseInstFromFunction(SI);
11838 // If the RHS is an alloca with a single use, zapify the store, making the
11840 // If the RHS is an alloca with a two uses, the other one being a
11841 // llvm.dbg.declare, zapify the store and the declare, making the
11842 // alloca dead. We must do this to prevent declare's from affecting
11844 if (!SI.isVolatile()) {
11845 if (Ptr->hasOneUse()) {
11846 if (isa<AllocaInst>(Ptr)) {
11847 EraseInstFromFunction(SI);
11851 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
11852 if (isa<AllocaInst>(GEP->getOperand(0))) {
11853 if (GEP->getOperand(0)->hasOneUse()) {
11854 EraseInstFromFunction(SI);
11858 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(GEP->getOperand(0))) {
11859 EraseInstFromFunction(*DI);
11860 EraseInstFromFunction(SI);
11867 if (DbgDeclareInst *DI = hasOneUsePlusDeclare(Ptr)) {
11868 EraseInstFromFunction(*DI);
11869 EraseInstFromFunction(SI);
11875 // Attempt to improve the alignment.
11876 unsigned KnownAlign =
11877 GetOrEnforceKnownAlignment(Ptr, TD->getPrefTypeAlignment(Val->getType()));
11879 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
11880 SI.getAlignment()))
11881 SI.setAlignment(KnownAlign);
11883 // Do really simple DSE, to catch cases where there are several consecutive
11884 // stores to the same location, separated by a few arithmetic operations. This
11885 // situation often occurs with bitfield accesses.
11886 BasicBlock::iterator BBI = &SI;
11887 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
11890 // Don't count debug info directives, lest they affect codegen,
11891 // and we skip pointer-to-pointer bitcasts, which are NOPs.
11892 // It is necessary for correctness to skip those that feed into a
11893 // llvm.dbg.declare, as these are not present when debugging is off.
11894 if (isa<DbgInfoIntrinsic>(BBI) ||
11895 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
11900 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
11901 // Prev store isn't volatile, and stores to the same location?
11902 if (!PrevSI->isVolatile() &&equivalentAddressValues(PrevSI->getOperand(1),
11903 SI.getOperand(1))) {
11906 EraseInstFromFunction(*PrevSI);
11912 // If this is a load, we have to stop. However, if the loaded value is from
11913 // the pointer we're loading and is producing the pointer we're storing,
11914 // then *this* store is dead (X = load P; store X -> P).
11915 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
11916 if (LI == Val && equivalentAddressValues(LI->getOperand(0), Ptr) &&
11917 !SI.isVolatile()) {
11918 EraseInstFromFunction(SI);
11922 // Otherwise, this is a load from some other location. Stores before it
11923 // may not be dead.
11927 // Don't skip over loads or things that can modify memory.
11928 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
11933 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
11935 // store X, null -> turns into 'unreachable' in SimplifyCFG
11936 if (isa<ConstantPointerNull>(Ptr) &&
11937 cast<PointerType>(Ptr->getType())->getAddressSpace() == 0) {
11938 if (!isa<UndefValue>(Val)) {
11939 SI.setOperand(0, Context->getUndef(Val->getType()));
11940 if (Instruction *U = dyn_cast<Instruction>(Val))
11941 AddToWorkList(U); // Dropped a use.
11944 return 0; // Do not modify these!
11947 // store undef, Ptr -> noop
11948 if (isa<UndefValue>(Val)) {
11949 EraseInstFromFunction(SI);
11954 // If the pointer destination is a cast, see if we can fold the cast into the
11956 if (isa<CastInst>(Ptr))
11957 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11959 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
11961 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
11965 // If this store is the last instruction in the basic block (possibly
11966 // excepting debug info instructions and the pointer bitcasts that feed
11967 // into them), and if the block ends with an unconditional branch, try
11968 // to move it to the successor block.
11972 } while (isa<DbgInfoIntrinsic>(BBI) ||
11973 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType())));
11974 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
11975 if (BI->isUnconditional())
11976 if (SimplifyStoreAtEndOfBlock(SI))
11977 return 0; // xform done!
11982 /// SimplifyStoreAtEndOfBlock - Turn things like:
11983 /// if () { *P = v1; } else { *P = v2 }
11984 /// into a phi node with a store in the successor.
11986 /// Simplify things like:
11987 /// *P = v1; if () { *P = v2; }
11988 /// into a phi node with a store in the successor.
11990 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
11991 BasicBlock *StoreBB = SI.getParent();
11993 // Check to see if the successor block has exactly two incoming edges. If
11994 // so, see if the other predecessor contains a store to the same location.
11995 // if so, insert a PHI node (if needed) and move the stores down.
11996 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
11998 // Determine whether Dest has exactly two predecessors and, if so, compute
11999 // the other predecessor.
12000 pred_iterator PI = pred_begin(DestBB);
12001 BasicBlock *OtherBB = 0;
12002 if (*PI != StoreBB)
12005 if (PI == pred_end(DestBB))
12008 if (*PI != StoreBB) {
12013 if (++PI != pred_end(DestBB))
12016 // Bail out if all the relevant blocks aren't distinct (this can happen,
12017 // for example, if SI is in an infinite loop)
12018 if (StoreBB == DestBB || OtherBB == DestBB)
12021 // Verify that the other block ends in a branch and is not otherwise empty.
12022 BasicBlock::iterator BBI = OtherBB->getTerminator();
12023 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
12024 if (!OtherBr || BBI == OtherBB->begin())
12027 // If the other block ends in an unconditional branch, check for the 'if then
12028 // else' case. there is an instruction before the branch.
12029 StoreInst *OtherStore = 0;
12030 if (OtherBr->isUnconditional()) {
12032 // Skip over debugging info.
12033 while (isa<DbgInfoIntrinsic>(BBI) ||
12034 (isa<BitCastInst>(BBI) && isa<PointerType>(BBI->getType()))) {
12035 if (BBI==OtherBB->begin())
12039 // If this isn't a store, or isn't a store to the same location, bail out.
12040 OtherStore = dyn_cast<StoreInst>(BBI);
12041 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
12044 // Otherwise, the other block ended with a conditional branch. If one of the
12045 // destinations is StoreBB, then we have the if/then case.
12046 if (OtherBr->getSuccessor(0) != StoreBB &&
12047 OtherBr->getSuccessor(1) != StoreBB)
12050 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
12051 // if/then triangle. See if there is a store to the same ptr as SI that
12052 // lives in OtherBB.
12054 // Check to see if we find the matching store.
12055 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
12056 if (OtherStore->getOperand(1) != SI.getOperand(1))
12060 // If we find something that may be using or overwriting the stored
12061 // value, or if we run out of instructions, we can't do the xform.
12062 if (BBI->mayReadFromMemory() || BBI->mayWriteToMemory() ||
12063 BBI == OtherBB->begin())
12067 // In order to eliminate the store in OtherBr, we have to
12068 // make sure nothing reads or overwrites the stored value in
12070 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
12071 // FIXME: This should really be AA driven.
12072 if (I->mayReadFromMemory() || I->mayWriteToMemory())
12077 // Insert a PHI node now if we need it.
12078 Value *MergedVal = OtherStore->getOperand(0);
12079 if (MergedVal != SI.getOperand(0)) {
12080 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
12081 PN->reserveOperandSpace(2);
12082 PN->addIncoming(SI.getOperand(0), SI.getParent());
12083 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
12084 MergedVal = InsertNewInstBefore(PN, DestBB->front());
12087 // Advance to a place where it is safe to insert the new store and
12089 BBI = DestBB->getFirstNonPHI();
12090 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
12091 OtherStore->isVolatile()), *BBI);
12093 // Nuke the old stores.
12094 EraseInstFromFunction(SI);
12095 EraseInstFromFunction(*OtherStore);
12101 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
12102 // Change br (not X), label True, label False to: br X, label False, True
12104 BasicBlock *TrueDest;
12105 BasicBlock *FalseDest;
12106 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest), *Context) &&
12107 !isa<Constant>(X)) {
12108 // Swap Destinations and condition...
12109 BI.setCondition(X);
12110 BI.setSuccessor(0, FalseDest);
12111 BI.setSuccessor(1, TrueDest);
12115 // Cannonicalize fcmp_one -> fcmp_oeq
12116 FCmpInst::Predicate FPred; Value *Y;
12117 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
12118 TrueDest, FalseDest), *Context))
12119 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
12120 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
12121 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
12122 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
12123 Instruction *NewSCC = new FCmpInst(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);
12135 // Cannonicalize icmp_ne -> icmp_eq
12136 ICmpInst::Predicate IPred;
12137 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
12138 TrueDest, FalseDest), *Context))
12139 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
12140 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
12141 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
12142 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
12143 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
12144 Instruction *NewSCC = new ICmpInst(I, NewPred, X, Y, "");
12145 NewSCC->takeName(I);
12146 // Swap Destinations and condition...
12147 BI.setCondition(NewSCC);
12148 BI.setSuccessor(0, FalseDest);
12149 BI.setSuccessor(1, TrueDest);
12150 RemoveFromWorkList(I);
12151 I->eraseFromParent();;
12152 AddToWorkList(NewSCC);
12159 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
12160 Value *Cond = SI.getCondition();
12161 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
12162 if (I->getOpcode() == Instruction::Add)
12163 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
12164 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
12165 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
12167 Context->getConstantExprSub(cast<Constant>(SI.getOperand(i)),
12169 SI.setOperand(0, I->getOperand(0));
12177 Instruction *InstCombiner::visitExtractValueInst(ExtractValueInst &EV) {
12178 Value *Agg = EV.getAggregateOperand();
12180 if (!EV.hasIndices())
12181 return ReplaceInstUsesWith(EV, Agg);
12183 if (Constant *C = dyn_cast<Constant>(Agg)) {
12184 if (isa<UndefValue>(C))
12185 return ReplaceInstUsesWith(EV, Context->getUndef(EV.getType()));
12187 if (isa<ConstantAggregateZero>(C))
12188 return ReplaceInstUsesWith(EV, Context->getNullValue(EV.getType()));
12190 if (isa<ConstantArray>(C) || isa<ConstantStruct>(C)) {
12191 // Extract the element indexed by the first index out of the constant
12192 Value *V = C->getOperand(*EV.idx_begin());
12193 if (EV.getNumIndices() > 1)
12194 // Extract the remaining indices out of the constant indexed by the
12196 return ExtractValueInst::Create(V, EV.idx_begin() + 1, EV.idx_end());
12198 return ReplaceInstUsesWith(EV, V);
12200 return 0; // Can't handle other constants
12202 if (InsertValueInst *IV = dyn_cast<InsertValueInst>(Agg)) {
12203 // We're extracting from an insertvalue instruction, compare the indices
12204 const unsigned *exti, *exte, *insi, *inse;
12205 for (exti = EV.idx_begin(), insi = IV->idx_begin(),
12206 exte = EV.idx_end(), inse = IV->idx_end();
12207 exti != exte && insi != inse;
12209 if (*insi != *exti)
12210 // The insert and extract both reference distinctly different elements.
12211 // This means the extract is not influenced by the insert, and we can
12212 // replace the aggregate operand of the extract with the aggregate
12213 // operand of the insert. i.e., replace
12214 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12215 // %E = extractvalue { i32, { i32 } } %I, 0
12217 // %E = extractvalue { i32, { i32 } } %A, 0
12218 return ExtractValueInst::Create(IV->getAggregateOperand(),
12219 EV.idx_begin(), EV.idx_end());
12221 if (exti == exte && insi == inse)
12222 // Both iterators are at the end: Index lists are identical. Replace
12223 // %B = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12224 // %C = extractvalue { i32, { i32 } } %B, 1, 0
12226 return ReplaceInstUsesWith(EV, IV->getInsertedValueOperand());
12227 if (exti == exte) {
12228 // The extract list is a prefix of the insert list. i.e. replace
12229 // %I = insertvalue { i32, { i32 } } %A, i32 42, 1, 0
12230 // %E = extractvalue { i32, { i32 } } %I, 1
12232 // %X = extractvalue { i32, { i32 } } %A, 1
12233 // %E = insertvalue { i32 } %X, i32 42, 0
12234 // by switching the order of the insert and extract (though the
12235 // insertvalue should be left in, since it may have other uses).
12236 Value *NewEV = InsertNewInstBefore(
12237 ExtractValueInst::Create(IV->getAggregateOperand(),
12238 EV.idx_begin(), EV.idx_end()),
12240 return InsertValueInst::Create(NewEV, IV->getInsertedValueOperand(),
12244 // The insert list is a prefix of the extract list
12245 // We can simply remove the common indices from the extract and make it
12246 // operate on the inserted value instead of the insertvalue result.
12248 // %I = insertvalue { i32, { i32 } } %A, { i32 } { i32 42 }, 1
12249 // %E = extractvalue { i32, { i32 } } %I, 1, 0
12251 // %E extractvalue { i32 } { i32 42 }, 0
12252 return ExtractValueInst::Create(IV->getInsertedValueOperand(),
12255 // Can't simplify extracts from other values. Note that nested extracts are
12256 // already simplified implicitely by the above (extract ( extract (insert) )
12257 // will be translated into extract ( insert ( extract ) ) first and then just
12258 // the value inserted, if appropriate).
12262 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
12263 /// is to leave as a vector operation.
12264 static bool CheapToScalarize(Value *V, bool isConstant) {
12265 if (isa<ConstantAggregateZero>(V))
12267 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
12268 if (isConstant) return true;
12269 // If all elts are the same, we can extract.
12270 Constant *Op0 = C->getOperand(0);
12271 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12272 if (C->getOperand(i) != Op0)
12276 Instruction *I = dyn_cast<Instruction>(V);
12277 if (!I) return false;
12279 // Insert element gets simplified to the inserted element or is deleted if
12280 // this is constant idx extract element and its a constant idx insertelt.
12281 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
12282 isa<ConstantInt>(I->getOperand(2)))
12284 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
12286 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
12287 if (BO->hasOneUse() &&
12288 (CheapToScalarize(BO->getOperand(0), isConstant) ||
12289 CheapToScalarize(BO->getOperand(1), isConstant)))
12291 if (CmpInst *CI = dyn_cast<CmpInst>(I))
12292 if (CI->hasOneUse() &&
12293 (CheapToScalarize(CI->getOperand(0), isConstant) ||
12294 CheapToScalarize(CI->getOperand(1), isConstant)))
12300 /// Read and decode a shufflevector mask.
12302 /// It turns undef elements into values that are larger than the number of
12303 /// elements in the input.
12304 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
12305 unsigned NElts = SVI->getType()->getNumElements();
12306 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
12307 return std::vector<unsigned>(NElts, 0);
12308 if (isa<UndefValue>(SVI->getOperand(2)))
12309 return std::vector<unsigned>(NElts, 2*NElts);
12311 std::vector<unsigned> Result;
12312 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
12313 for (User::const_op_iterator i = CP->op_begin(), e = CP->op_end(); i!=e; ++i)
12314 if (isa<UndefValue>(*i))
12315 Result.push_back(NElts*2); // undef -> 8
12317 Result.push_back(cast<ConstantInt>(*i)->getZExtValue());
12321 /// FindScalarElement - Given a vector and an element number, see if the scalar
12322 /// value is already around as a register, for example if it were inserted then
12323 /// extracted from the vector.
12324 static Value *FindScalarElement(Value *V, unsigned EltNo,
12325 LLVMContext *Context) {
12326 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
12327 const VectorType *PTy = cast<VectorType>(V->getType());
12328 unsigned Width = PTy->getNumElements();
12329 if (EltNo >= Width) // Out of range access.
12330 return Context->getUndef(PTy->getElementType());
12332 if (isa<UndefValue>(V))
12333 return Context->getUndef(PTy->getElementType());
12334 else if (isa<ConstantAggregateZero>(V))
12335 return Context->getNullValue(PTy->getElementType());
12336 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
12337 return CP->getOperand(EltNo);
12338 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
12339 // If this is an insert to a variable element, we don't know what it is.
12340 if (!isa<ConstantInt>(III->getOperand(2)))
12342 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
12344 // If this is an insert to the element we are looking for, return the
12346 if (EltNo == IIElt)
12347 return III->getOperand(1);
12349 // Otherwise, the insertelement doesn't modify the value, recurse on its
12351 return FindScalarElement(III->getOperand(0), EltNo, Context);
12352 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
12353 unsigned LHSWidth =
12354 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12355 unsigned InEl = getShuffleMask(SVI)[EltNo];
12356 if (InEl < LHSWidth)
12357 return FindScalarElement(SVI->getOperand(0), InEl, Context);
12358 else if (InEl < LHSWidth*2)
12359 return FindScalarElement(SVI->getOperand(1), InEl - LHSWidth, Context);
12361 return Context->getUndef(PTy->getElementType());
12364 // Otherwise, we don't know.
12368 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
12369 // If vector val is undef, replace extract with scalar undef.
12370 if (isa<UndefValue>(EI.getOperand(0)))
12371 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12373 // If vector val is constant 0, replace extract with scalar 0.
12374 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
12375 return ReplaceInstUsesWith(EI, Context->getNullValue(EI.getType()));
12377 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
12378 // If vector val is constant with all elements the same, replace EI with
12379 // that element. When the elements are not identical, we cannot replace yet
12380 // (we do that below, but only when the index is constant).
12381 Constant *op0 = C->getOperand(0);
12382 for (unsigned i = 1; i < C->getNumOperands(); ++i)
12383 if (C->getOperand(i) != op0) {
12388 return ReplaceInstUsesWith(EI, op0);
12391 // If extracting a specified index from the vector, see if we can recursively
12392 // find a previously computed scalar that was inserted into the vector.
12393 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12394 unsigned IndexVal = IdxC->getZExtValue();
12395 unsigned VectorWidth =
12396 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
12398 // If this is extracting an invalid index, turn this into undef, to avoid
12399 // crashing the code below.
12400 if (IndexVal >= VectorWidth)
12401 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12403 // This instruction only demands the single element from the input vector.
12404 // If the input vector has a single use, simplify it based on this use
12406 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
12407 APInt UndefElts(VectorWidth, 0);
12408 APInt DemandedMask(VectorWidth, 1 << IndexVal);
12409 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
12410 DemandedMask, UndefElts)) {
12411 EI.setOperand(0, V);
12416 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal, Context))
12417 return ReplaceInstUsesWith(EI, Elt);
12419 // If the this extractelement is directly using a bitcast from a vector of
12420 // the same number of elements, see if we can find the source element from
12421 // it. In this case, we will end up needing to bitcast the scalars.
12422 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
12423 if (const VectorType *VT =
12424 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
12425 if (VT->getNumElements() == VectorWidth)
12426 if (Value *Elt = FindScalarElement(BCI->getOperand(0),
12427 IndexVal, Context))
12428 return new BitCastInst(Elt, EI.getType());
12432 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
12433 if (I->hasOneUse()) {
12434 // Push extractelement into predecessor operation if legal and
12435 // profitable to do so
12436 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
12437 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
12438 if (CheapToScalarize(BO, isConstantElt)) {
12439 ExtractElementInst *newEI0 =
12440 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
12441 EI.getName()+".lhs");
12442 ExtractElementInst *newEI1 =
12443 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
12444 EI.getName()+".rhs");
12445 InsertNewInstBefore(newEI0, EI);
12446 InsertNewInstBefore(newEI1, EI);
12447 return BinaryOperator::Create(BO->getOpcode(), newEI0, newEI1);
12449 } else if (isa<LoadInst>(I)) {
12451 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
12452 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
12453 Context->getPointerType(EI.getType(), AS),EI);
12454 GetElementPtrInst *GEP =
12455 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName()+".gep");
12456 InsertNewInstBefore(GEP, EI);
12457 return new LoadInst(GEP);
12460 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
12461 // Extracting the inserted element?
12462 if (IE->getOperand(2) == EI.getOperand(1))
12463 return ReplaceInstUsesWith(EI, IE->getOperand(1));
12464 // If the inserted and extracted elements are constants, they must not
12465 // be the same value, extract from the pre-inserted value instead.
12466 if (isa<Constant>(IE->getOperand(2)) &&
12467 isa<Constant>(EI.getOperand(1))) {
12468 AddUsesToWorkList(EI);
12469 EI.setOperand(0, IE->getOperand(0));
12472 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
12473 // If this is extracting an element from a shufflevector, figure out where
12474 // it came from and extract from the appropriate input element instead.
12475 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
12476 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
12478 unsigned LHSWidth =
12479 cast<VectorType>(SVI->getOperand(0)->getType())->getNumElements();
12481 if (SrcIdx < LHSWidth)
12482 Src = SVI->getOperand(0);
12483 else if (SrcIdx < LHSWidth*2) {
12484 SrcIdx -= LHSWidth;
12485 Src = SVI->getOperand(1);
12487 return ReplaceInstUsesWith(EI, Context->getUndef(EI.getType()));
12489 return new ExtractElementInst(Src, SrcIdx);
12496 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
12497 /// elements from either LHS or RHS, return the shuffle mask and true.
12498 /// Otherwise, return false.
12499 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
12500 std::vector<Constant*> &Mask,
12501 LLVMContext *Context) {
12502 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
12503 "Invalid CollectSingleShuffleElements");
12504 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12506 if (isa<UndefValue>(V)) {
12507 Mask.assign(NumElts, Context->getUndef(Type::Int32Ty));
12509 } else if (V == LHS) {
12510 for (unsigned i = 0; i != NumElts; ++i)
12511 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i));
12513 } else if (V == RHS) {
12514 for (unsigned i = 0; i != NumElts; ++i)
12515 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i+NumElts));
12517 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12518 // If this is an insert of an extract from some other vector, include it.
12519 Value *VecOp = IEI->getOperand(0);
12520 Value *ScalarOp = IEI->getOperand(1);
12521 Value *IdxOp = IEI->getOperand(2);
12523 if (!isa<ConstantInt>(IdxOp))
12525 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12527 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
12528 // Okay, we can handle this if the vector we are insertinting into is
12529 // transitively ok.
12530 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12531 // If so, update the mask to reflect the inserted undef.
12532 Mask[InsertedIdx] = Context->getUndef(Type::Int32Ty);
12535 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
12536 if (isa<ConstantInt>(EI->getOperand(1)) &&
12537 EI->getOperand(0)->getType() == V->getType()) {
12538 unsigned ExtractedIdx =
12539 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12541 // This must be extracting from either LHS or RHS.
12542 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
12543 // Okay, we can handle this if the vector we are insertinting into is
12544 // transitively ok.
12545 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask, Context)) {
12546 // If so, update the mask to reflect the inserted value.
12547 if (EI->getOperand(0) == LHS) {
12548 Mask[InsertedIdx % NumElts] =
12549 Context->getConstantInt(Type::Int32Ty, ExtractedIdx);
12551 assert(EI->getOperand(0) == RHS);
12552 Mask[InsertedIdx % NumElts] =
12553 Context->getConstantInt(Type::Int32Ty, ExtractedIdx+NumElts);
12562 // TODO: Handle shufflevector here!
12567 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
12568 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
12569 /// that computes V and the LHS value of the shuffle.
12570 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
12571 Value *&RHS, LLVMContext *Context) {
12572 assert(isa<VectorType>(V->getType()) &&
12573 (RHS == 0 || V->getType() == RHS->getType()) &&
12574 "Invalid shuffle!");
12575 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
12577 if (isa<UndefValue>(V)) {
12578 Mask.assign(NumElts, Context->getUndef(Type::Int32Ty));
12580 } else if (isa<ConstantAggregateZero>(V)) {
12581 Mask.assign(NumElts, Context->getConstantInt(Type::Int32Ty, 0));
12583 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
12584 // If this is an insert of an extract from some other vector, include it.
12585 Value *VecOp = IEI->getOperand(0);
12586 Value *ScalarOp = IEI->getOperand(1);
12587 Value *IdxOp = IEI->getOperand(2);
12589 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12590 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12591 EI->getOperand(0)->getType() == V->getType()) {
12592 unsigned ExtractedIdx =
12593 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12594 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12596 // Either the extracted from or inserted into vector must be RHSVec,
12597 // otherwise we'd end up with a shuffle of three inputs.
12598 if (EI->getOperand(0) == RHS || RHS == 0) {
12599 RHS = EI->getOperand(0);
12600 Value *V = CollectShuffleElements(VecOp, Mask, RHS, Context);
12601 Mask[InsertedIdx % NumElts] =
12602 Context->getConstantInt(Type::Int32Ty, NumElts+ExtractedIdx);
12606 if (VecOp == RHS) {
12607 Value *V = CollectShuffleElements(EI->getOperand(0), Mask,
12609 // Everything but the extracted element is replaced with the RHS.
12610 for (unsigned i = 0; i != NumElts; ++i) {
12611 if (i != InsertedIdx)
12612 Mask[i] = Context->getConstantInt(Type::Int32Ty, NumElts+i);
12617 // If this insertelement is a chain that comes from exactly these two
12618 // vectors, return the vector and the effective shuffle.
12619 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask,
12621 return EI->getOperand(0);
12626 // TODO: Handle shufflevector here!
12628 // Otherwise, can't do anything fancy. Return an identity vector.
12629 for (unsigned i = 0; i != NumElts; ++i)
12630 Mask.push_back(Context->getConstantInt(Type::Int32Ty, i));
12634 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
12635 Value *VecOp = IE.getOperand(0);
12636 Value *ScalarOp = IE.getOperand(1);
12637 Value *IdxOp = IE.getOperand(2);
12639 // Inserting an undef or into an undefined place, remove this.
12640 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
12641 ReplaceInstUsesWith(IE, VecOp);
12643 // If the inserted element was extracted from some other vector, and if the
12644 // indexes are constant, try to turn this into a shufflevector operation.
12645 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
12646 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
12647 EI->getOperand(0)->getType() == IE.getType()) {
12648 unsigned NumVectorElts = IE.getType()->getNumElements();
12649 unsigned ExtractedIdx =
12650 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
12651 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
12653 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
12654 return ReplaceInstUsesWith(IE, VecOp);
12656 if (InsertedIdx >= NumVectorElts) // Out of range insert.
12657 return ReplaceInstUsesWith(IE, Context->getUndef(IE.getType()));
12659 // If we are extracting a value from a vector, then inserting it right
12660 // back into the same place, just use the input vector.
12661 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
12662 return ReplaceInstUsesWith(IE, VecOp);
12664 // We could theoretically do this for ANY input. However, doing so could
12665 // turn chains of insertelement instructions into a chain of shufflevector
12666 // instructions, and right now we do not merge shufflevectors. As such,
12667 // only do this in a situation where it is clear that there is benefit.
12668 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
12669 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
12670 // the values of VecOp, except then one read from EIOp0.
12671 // Build a new shuffle mask.
12672 std::vector<Constant*> Mask;
12673 if (isa<UndefValue>(VecOp))
12674 Mask.assign(NumVectorElts, Context->getUndef(Type::Int32Ty));
12676 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
12677 Mask.assign(NumVectorElts, Context->getConstantInt(Type::Int32Ty,
12680 Mask[InsertedIdx] =
12681 Context->getConstantInt(Type::Int32Ty, ExtractedIdx);
12682 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
12683 Context->getConstantVector(Mask));
12686 // If this insertelement isn't used by some other insertelement, turn it
12687 // (and any insertelements it points to), into one big shuffle.
12688 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
12689 std::vector<Constant*> Mask;
12691 Value *LHS = CollectShuffleElements(&IE, Mask, RHS, Context);
12692 if (RHS == 0) RHS = Context->getUndef(LHS->getType());
12693 // We now have a shuffle of LHS, RHS, Mask.
12694 return new ShuffleVectorInst(LHS, RHS,
12695 Context->getConstantVector(Mask));
12700 unsigned VWidth = cast<VectorType>(VecOp->getType())->getNumElements();
12701 APInt UndefElts(VWidth, 0);
12702 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12703 if (SimplifyDemandedVectorElts(&IE, AllOnesEltMask, UndefElts))
12710 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
12711 Value *LHS = SVI.getOperand(0);
12712 Value *RHS = SVI.getOperand(1);
12713 std::vector<unsigned> Mask = getShuffleMask(&SVI);
12715 bool MadeChange = false;
12717 // Undefined shuffle mask -> undefined value.
12718 if (isa<UndefValue>(SVI.getOperand(2)))
12719 return ReplaceInstUsesWith(SVI, Context->getUndef(SVI.getType()));
12721 unsigned VWidth = cast<VectorType>(SVI.getType())->getNumElements();
12723 if (VWidth != cast<VectorType>(LHS->getType())->getNumElements())
12726 APInt UndefElts(VWidth, 0);
12727 APInt AllOnesEltMask(APInt::getAllOnesValue(VWidth));
12728 if (SimplifyDemandedVectorElts(&SVI, AllOnesEltMask, UndefElts)) {
12729 LHS = SVI.getOperand(0);
12730 RHS = SVI.getOperand(1);
12734 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
12735 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
12736 if (LHS == RHS || isa<UndefValue>(LHS)) {
12737 if (isa<UndefValue>(LHS) && LHS == RHS) {
12738 // shuffle(undef,undef,mask) -> undef.
12739 return ReplaceInstUsesWith(SVI, LHS);
12742 // Remap any references to RHS to use LHS.
12743 std::vector<Constant*> Elts;
12744 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12745 if (Mask[i] >= 2*e)
12746 Elts.push_back(Context->getUndef(Type::Int32Ty));
12748 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
12749 (Mask[i] < e && isa<UndefValue>(LHS))) {
12750 Mask[i] = 2*e; // Turn into undef.
12751 Elts.push_back(Context->getUndef(Type::Int32Ty));
12753 Mask[i] = Mask[i] % e; // Force to LHS.
12754 Elts.push_back(Context->getConstantInt(Type::Int32Ty, Mask[i]));
12758 SVI.setOperand(0, SVI.getOperand(1));
12759 SVI.setOperand(1, Context->getUndef(RHS->getType()));
12760 SVI.setOperand(2, Context->getConstantVector(Elts));
12761 LHS = SVI.getOperand(0);
12762 RHS = SVI.getOperand(1);
12766 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
12767 bool isLHSID = true, isRHSID = true;
12769 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
12770 if (Mask[i] >= e*2) continue; // Ignore undef values.
12771 // Is this an identity shuffle of the LHS value?
12772 isLHSID &= (Mask[i] == i);
12774 // Is this an identity shuffle of the RHS value?
12775 isRHSID &= (Mask[i]-e == i);
12778 // Eliminate identity shuffles.
12779 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
12780 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
12782 // If the LHS is a shufflevector itself, see if we can combine it with this
12783 // one without producing an unusual shuffle. Here we are really conservative:
12784 // we are absolutely afraid of producing a shuffle mask not in the input
12785 // program, because the code gen may not be smart enough to turn a merged
12786 // shuffle into two specific shuffles: it may produce worse code. As such,
12787 // we only merge two shuffles if the result is one of the two input shuffle
12788 // masks. In this case, merging the shuffles just removes one instruction,
12789 // which we know is safe. This is good for things like turning:
12790 // (splat(splat)) -> splat.
12791 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
12792 if (isa<UndefValue>(RHS)) {
12793 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
12795 std::vector<unsigned> NewMask;
12796 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
12797 if (Mask[i] >= 2*e)
12798 NewMask.push_back(2*e);
12800 NewMask.push_back(LHSMask[Mask[i]]);
12802 // If the result mask is equal to the src shuffle or this shuffle mask, do
12803 // the replacement.
12804 if (NewMask == LHSMask || NewMask == Mask) {
12805 unsigned LHSInNElts =
12806 cast<VectorType>(LHSSVI->getOperand(0)->getType())->getNumElements();
12807 std::vector<Constant*> Elts;
12808 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
12809 if (NewMask[i] >= LHSInNElts*2) {
12810 Elts.push_back(Context->getUndef(Type::Int32Ty));
12812 Elts.push_back(Context->getConstantInt(Type::Int32Ty, NewMask[i]));
12815 return new ShuffleVectorInst(LHSSVI->getOperand(0),
12816 LHSSVI->getOperand(1),
12817 Context->getConstantVector(Elts));
12822 return MadeChange ? &SVI : 0;
12828 /// TryToSinkInstruction - Try to move the specified instruction from its
12829 /// current block into the beginning of DestBlock, which can only happen if it's
12830 /// safe to move the instruction past all of the instructions between it and the
12831 /// end of its block.
12832 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
12833 assert(I->hasOneUse() && "Invariants didn't hold!");
12835 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
12836 if (isa<PHINode>(I) || I->mayHaveSideEffects() || isa<TerminatorInst>(I))
12839 // Do not sink alloca instructions out of the entry block.
12840 if (isa<AllocaInst>(I) && I->getParent() ==
12841 &DestBlock->getParent()->getEntryBlock())
12844 // We can only sink load instructions if there is nothing between the load and
12845 // the end of block that could change the value.
12846 if (I->mayReadFromMemory()) {
12847 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
12849 if (Scan->mayWriteToMemory())
12853 BasicBlock::iterator InsertPos = DestBlock->getFirstNonPHI();
12855 CopyPrecedingStopPoint(I, InsertPos);
12856 I->moveBefore(InsertPos);
12862 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
12863 /// all reachable code to the worklist.
12865 /// This has a couple of tricks to make the code faster and more powerful. In
12866 /// particular, we constant fold and DCE instructions as we go, to avoid adding
12867 /// them to the worklist (this significantly speeds up instcombine on code where
12868 /// many instructions are dead or constant). Additionally, if we find a branch
12869 /// whose condition is a known constant, we only visit the reachable successors.
12871 static void AddReachableCodeToWorklist(BasicBlock *BB,
12872 SmallPtrSet<BasicBlock*, 64> &Visited,
12874 const TargetData *TD) {
12875 SmallVector<BasicBlock*, 256> Worklist;
12876 Worklist.push_back(BB);
12878 while (!Worklist.empty()) {
12879 BB = Worklist.back();
12880 Worklist.pop_back();
12882 // We have now visited this block! If we've already been here, ignore it.
12883 if (!Visited.insert(BB)) continue;
12885 DbgInfoIntrinsic *DBI_Prev = NULL;
12886 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
12887 Instruction *Inst = BBI++;
12889 // DCE instruction if trivially dead.
12890 if (isInstructionTriviallyDead(Inst)) {
12892 DOUT << "IC: DCE: " << *Inst;
12893 Inst->eraseFromParent();
12897 // ConstantProp instruction if trivially constant.
12898 if (Constant *C = ConstantFoldInstruction(Inst, BB->getContext(), TD)) {
12899 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
12900 Inst->replaceAllUsesWith(C);
12902 Inst->eraseFromParent();
12906 // If there are two consecutive llvm.dbg.stoppoint calls then
12907 // it is likely that the optimizer deleted code in between these
12909 DbgInfoIntrinsic *DBI_Next = dyn_cast<DbgInfoIntrinsic>(Inst);
12912 && DBI_Prev->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint
12913 && DBI_Next->getIntrinsicID() == llvm::Intrinsic::dbg_stoppoint) {
12914 IC.RemoveFromWorkList(DBI_Prev);
12915 DBI_Prev->eraseFromParent();
12917 DBI_Prev = DBI_Next;
12922 IC.AddToWorkList(Inst);
12925 // Recursively visit successors. If this is a branch or switch on a
12926 // constant, only visit the reachable successor.
12927 TerminatorInst *TI = BB->getTerminator();
12928 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
12929 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
12930 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
12931 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
12932 Worklist.push_back(ReachableBB);
12935 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
12936 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
12937 // See if this is an explicit destination.
12938 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
12939 if (SI->getCaseValue(i) == Cond) {
12940 BasicBlock *ReachableBB = SI->getSuccessor(i);
12941 Worklist.push_back(ReachableBB);
12945 // Otherwise it is the default destination.
12946 Worklist.push_back(SI->getSuccessor(0));
12951 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
12952 Worklist.push_back(TI->getSuccessor(i));
12956 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
12957 bool Changed = false;
12958 TD = &getAnalysis<TargetData>();
12960 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
12961 << F.getNameStr() << "\n");
12964 // Do a depth-first traversal of the function, populate the worklist with
12965 // the reachable instructions. Ignore blocks that are not reachable. Keep
12966 // track of which blocks we visit.
12967 SmallPtrSet<BasicBlock*, 64> Visited;
12968 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
12970 // Do a quick scan over the function. If we find any blocks that are
12971 // unreachable, remove any instructions inside of them. This prevents
12972 // the instcombine code from having to deal with some bad special cases.
12973 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
12974 if (!Visited.count(BB)) {
12975 Instruction *Term = BB->getTerminator();
12976 while (Term != BB->begin()) { // Remove instrs bottom-up
12977 BasicBlock::iterator I = Term; --I;
12979 DOUT << "IC: DCE: " << *I;
12980 // A debug intrinsic shouldn't force another iteration if we weren't
12981 // going to do one without it.
12982 if (!isa<DbgInfoIntrinsic>(I)) {
12986 if (!I->use_empty())
12987 I->replaceAllUsesWith(Context->getUndef(I->getType()));
12988 I->eraseFromParent();
12993 while (!Worklist.empty()) {
12994 Instruction *I = RemoveOneFromWorkList();
12995 if (I == 0) continue; // skip null values.
12997 // Check to see if we can DCE the instruction.
12998 if (isInstructionTriviallyDead(I)) {
12999 // Add operands to the worklist.
13000 if (I->getNumOperands() < 4)
13001 AddUsesToWorkList(*I);
13004 DOUT << "IC: DCE: " << *I;
13006 I->eraseFromParent();
13007 RemoveFromWorkList(I);
13012 // Instruction isn't dead, see if we can constant propagate it.
13013 if (Constant *C = ConstantFoldInstruction(I, F.getContext(), TD)) {
13014 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
13016 // Add operands to the worklist.
13017 AddUsesToWorkList(*I);
13018 ReplaceInstUsesWith(*I, C);
13021 I->eraseFromParent();
13022 RemoveFromWorkList(I);
13028 (I->getType()->getTypeID() == Type::VoidTyID ||
13029 I->isTrapping())) {
13030 // See if we can constant fold its operands.
13031 for (User::op_iterator i = I->op_begin(), e = I->op_end(); i != e; ++i)
13032 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(i))
13033 if (Constant *NewC = ConstantFoldConstantExpression(CE,
13034 F.getContext(), TD))
13041 // See if we can trivially sink this instruction to a successor basic block.
13042 if (I->hasOneUse()) {
13043 BasicBlock *BB = I->getParent();
13044 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
13045 if (UserParent != BB) {
13046 bool UserIsSuccessor = false;
13047 // See if the user is one of our successors.
13048 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
13049 if (*SI == UserParent) {
13050 UserIsSuccessor = true;
13054 // If the user is one of our immediate successors, and if that successor
13055 // only has us as a predecessors (we'd have to split the critical edge
13056 // otherwise), we can keep going.
13057 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
13058 next(pred_begin(UserParent)) == pred_end(UserParent))
13059 // Okay, the CFG is simple enough, try to sink this instruction.
13060 Changed |= TryToSinkInstruction(I, UserParent);
13064 // Now that we have an instruction, try combining it to simplify it...
13068 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
13069 if (Instruction *Result = visit(*I)) {
13071 // Should we replace the old instruction with a new one?
13073 DOUT << "IC: Old = " << *I
13074 << " New = " << *Result;
13076 // Everything uses the new instruction now.
13077 I->replaceAllUsesWith(Result);
13079 // Push the new instruction and any users onto the worklist.
13080 AddToWorkList(Result);
13081 AddUsersToWorkList(*Result);
13083 // Move the name to the new instruction first.
13084 Result->takeName(I);
13086 // Insert the new instruction into the basic block...
13087 BasicBlock *InstParent = I->getParent();
13088 BasicBlock::iterator InsertPos = I;
13090 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
13091 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
13094 InstParent->getInstList().insert(InsertPos, Result);
13096 // Make sure that we reprocess all operands now that we reduced their
13098 AddUsesToWorkList(*I);
13100 // Instructions can end up on the worklist more than once. Make sure
13101 // we do not process an instruction that has been deleted.
13102 RemoveFromWorkList(I);
13104 // Erase the old instruction.
13105 InstParent->getInstList().erase(I);
13108 DOUT << "IC: Mod = " << OrigI
13109 << " New = " << *I;
13112 // If the instruction was modified, it's possible that it is now dead.
13113 // if so, remove it.
13114 if (isInstructionTriviallyDead(I)) {
13115 // Make sure we process all operands now that we are reducing their
13117 AddUsesToWorkList(*I);
13119 // Instructions may end up in the worklist more than once. Erase all
13120 // occurrences of this instruction.
13121 RemoveFromWorkList(I);
13122 I->eraseFromParent();
13125 AddUsersToWorkList(*I);
13132 assert(WorklistMap.empty() && "Worklist empty, but map not?");
13134 // Do an explicit clear, this shrinks the map if needed.
13135 WorklistMap.clear();
13140 bool InstCombiner::runOnFunction(Function &F) {
13141 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
13143 bool EverMadeChange = false;
13145 // Iterate while there is work to do.
13146 unsigned Iteration = 0;
13147 while (DoOneIteration(F, Iteration++))
13148 EverMadeChange = true;
13149 return EverMadeChange;
13152 FunctionPass *llvm::createInstructionCombiningPass() {
13153 return new InstCombiner();