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/Pass.h"
40 #include "llvm/DerivedTypes.h"
41 #include "llvm/GlobalVariable.h"
42 #include "llvm/Analysis/ConstantFolding.h"
43 #include "llvm/Target/TargetData.h"
44 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
45 #include "llvm/Transforms/Utils/Local.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/ConstantRange.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/GetElementPtrTypeIterator.h"
50 #include "llvm/Support/InstVisitor.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/PatternMatch.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/ADT/DenseMap.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/STLExtras.h"
63 using namespace llvm::PatternMatch;
65 STATISTIC(NumCombined , "Number of insts combined");
66 STATISTIC(NumConstProp, "Number of constant folds");
67 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
68 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
69 STATISTIC(NumSunkInst , "Number of instructions sunk");
72 class VISIBILITY_HIDDEN InstCombiner
73 : public FunctionPass,
74 public InstVisitor<InstCombiner, Instruction*> {
75 // Worklist of all of the instructions that need to be simplified.
76 std::vector<Instruction*> Worklist;
77 DenseMap<Instruction*, unsigned> WorklistMap;
79 bool MustPreserveLCSSA;
81 static char ID; // Pass identification, replacement for typeid
82 InstCombiner() : FunctionPass((intptr_t)&ID) {}
84 /// AddToWorkList - Add the specified instruction to the worklist if it
85 /// isn't already in it.
86 void AddToWorkList(Instruction *I) {
87 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
88 Worklist.push_back(I);
91 // RemoveFromWorkList - remove I from the worklist if it exists.
92 void RemoveFromWorkList(Instruction *I) {
93 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
94 if (It == WorklistMap.end()) return; // Not in worklist.
96 // Don't bother moving everything down, just null out the slot.
97 Worklist[It->second] = 0;
99 WorklistMap.erase(It);
102 Instruction *RemoveOneFromWorkList() {
103 Instruction *I = Worklist.back();
105 WorklistMap.erase(I);
110 /// AddUsersToWorkList - When an instruction is simplified, add all users of
111 /// the instruction to the work lists because they might get more simplified
114 void AddUsersToWorkList(Value &I) {
115 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
117 AddToWorkList(cast<Instruction>(*UI));
120 /// AddUsesToWorkList - When an instruction is simplified, add operands to
121 /// the work lists because they might get more simplified now.
123 void AddUsesToWorkList(Instruction &I) {
124 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
125 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
129 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
130 /// dead. Add all of its operands to the worklist, turning them into
131 /// undef's to reduce the number of uses of those instructions.
133 /// Return the specified operand before it is turned into an undef.
135 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
136 Value *R = I.getOperand(op);
138 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
139 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
141 // Set the operand to undef to drop the use.
142 I.setOperand(i, UndefValue::get(Op->getType()));
149 virtual bool runOnFunction(Function &F);
151 bool DoOneIteration(Function &F, unsigned ItNum);
153 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
154 AU.addRequired<TargetData>();
155 AU.addPreservedID(LCSSAID);
156 AU.setPreservesCFG();
159 TargetData &getTargetData() const { return *TD; }
161 // Visitation implementation - Implement instruction combining for different
162 // instruction types. The semantics are as follows:
164 // null - No change was made
165 // I - Change was made, I is still valid, I may be dead though
166 // otherwise - Change was made, replace I with returned instruction
168 Instruction *visitAdd(BinaryOperator &I);
169 Instruction *visitSub(BinaryOperator &I);
170 Instruction *visitMul(BinaryOperator &I);
171 Instruction *visitURem(BinaryOperator &I);
172 Instruction *visitSRem(BinaryOperator &I);
173 Instruction *visitFRem(BinaryOperator &I);
174 Instruction *commonRemTransforms(BinaryOperator &I);
175 Instruction *commonIRemTransforms(BinaryOperator &I);
176 Instruction *commonDivTransforms(BinaryOperator &I);
177 Instruction *commonIDivTransforms(BinaryOperator &I);
178 Instruction *visitUDiv(BinaryOperator &I);
179 Instruction *visitSDiv(BinaryOperator &I);
180 Instruction *visitFDiv(BinaryOperator &I);
181 Instruction *visitAnd(BinaryOperator &I);
182 Instruction *visitOr (BinaryOperator &I);
183 Instruction *visitXor(BinaryOperator &I);
184 Instruction *visitShl(BinaryOperator &I);
185 Instruction *visitAShr(BinaryOperator &I);
186 Instruction *visitLShr(BinaryOperator &I);
187 Instruction *commonShiftTransforms(BinaryOperator &I);
188 Instruction *visitFCmpInst(FCmpInst &I);
189 Instruction *visitICmpInst(ICmpInst &I);
190 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
191 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
194 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
195 ConstantInt *DivRHS);
197 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
198 ICmpInst::Predicate Cond, Instruction &I);
199 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
201 Instruction *commonCastTransforms(CastInst &CI);
202 Instruction *commonIntCastTransforms(CastInst &CI);
203 Instruction *commonPointerCastTransforms(CastInst &CI);
204 Instruction *visitTrunc(TruncInst &CI);
205 Instruction *visitZExt(ZExtInst &CI);
206 Instruction *visitSExt(SExtInst &CI);
207 Instruction *visitFPTrunc(FPTruncInst &CI);
208 Instruction *visitFPExt(CastInst &CI);
209 Instruction *visitFPToUI(CastInst &CI);
210 Instruction *visitFPToSI(CastInst &CI);
211 Instruction *visitUIToFP(CastInst &CI);
212 Instruction *visitSIToFP(CastInst &CI);
213 Instruction *visitPtrToInt(CastInst &CI);
214 Instruction *visitIntToPtr(IntToPtrInst &CI);
215 Instruction *visitBitCast(BitCastInst &CI);
216 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
218 Instruction *visitSelectInst(SelectInst &CI);
219 Instruction *visitCallInst(CallInst &CI);
220 Instruction *visitInvokeInst(InvokeInst &II);
221 Instruction *visitPHINode(PHINode &PN);
222 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
223 Instruction *visitAllocationInst(AllocationInst &AI);
224 Instruction *visitFreeInst(FreeInst &FI);
225 Instruction *visitLoadInst(LoadInst &LI);
226 Instruction *visitStoreInst(StoreInst &SI);
227 Instruction *visitBranchInst(BranchInst &BI);
228 Instruction *visitSwitchInst(SwitchInst &SI);
229 Instruction *visitInsertElementInst(InsertElementInst &IE);
230 Instruction *visitExtractElementInst(ExtractElementInst &EI);
231 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
233 // visitInstruction - Specify what to return for unhandled instructions...
234 Instruction *visitInstruction(Instruction &I) { return 0; }
237 Instruction *visitCallSite(CallSite CS);
238 bool transformConstExprCastCall(CallSite CS);
239 Instruction *transformCallThroughTrampoline(CallSite CS);
240 Instruction *transformZExtICmp(ICmpInst *ICI, Instruction &CI,
241 bool DoXform = true);
244 // InsertNewInstBefore - insert an instruction New before instruction Old
245 // in the program. Add the new instruction to the worklist.
247 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
248 assert(New && New->getParent() == 0 &&
249 "New instruction already inserted into a basic block!");
250 BasicBlock *BB = Old.getParent();
251 BB->getInstList().insert(&Old, New); // Insert inst
256 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
257 /// This also adds the cast to the worklist. Finally, this returns the
259 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
261 if (V->getType() == Ty) return V;
263 if (Constant *CV = dyn_cast<Constant>(V))
264 return ConstantExpr::getCast(opc, CV, Ty);
266 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
271 Value *InsertBitCastBefore(Value *V, const Type *Ty, Instruction &Pos) {
272 return InsertCastBefore(Instruction::BitCast, V, Ty, Pos);
276 // ReplaceInstUsesWith - This method is to be used when an instruction is
277 // found to be dead, replacable with another preexisting expression. Here
278 // we add all uses of I to the worklist, replace all uses of I with the new
279 // value, then return I, so that the inst combiner will know that I was
282 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
283 AddUsersToWorkList(I); // Add all modified instrs to worklist
285 I.replaceAllUsesWith(V);
288 // If we are replacing the instruction with itself, this must be in a
289 // segment of unreachable code, so just clobber the instruction.
290 I.replaceAllUsesWith(UndefValue::get(I.getType()));
295 // UpdateValueUsesWith - This method is to be used when an value is
296 // found to be replacable with another preexisting expression or was
297 // updated. Here we add all uses of I to the worklist, replace all uses of
298 // I with the new value (unless the instruction was just updated), then
299 // return true, so that the inst combiner will know that I was modified.
301 bool UpdateValueUsesWith(Value *Old, Value *New) {
302 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
304 Old->replaceAllUsesWith(New);
305 if (Instruction *I = dyn_cast<Instruction>(Old))
307 if (Instruction *I = dyn_cast<Instruction>(New))
312 // EraseInstFromFunction - When dealing with an instruction that has side
313 // effects or produces a void value, we can't rely on DCE to delete the
314 // instruction. Instead, visit methods should return the value returned by
316 Instruction *EraseInstFromFunction(Instruction &I) {
317 assert(I.use_empty() && "Cannot erase instruction that is used!");
318 AddUsesToWorkList(I);
319 RemoveFromWorkList(&I);
321 return 0; // Don't do anything with FI
325 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
326 /// InsertBefore instruction. This is specialized a bit to avoid inserting
327 /// casts that are known to not do anything...
329 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
330 Value *V, const Type *DestTy,
331 Instruction *InsertBefore);
333 /// SimplifyCommutative - This performs a few simplifications for
334 /// commutative operators.
335 bool SimplifyCommutative(BinaryOperator &I);
337 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
338 /// most-complex to least-complex order.
339 bool SimplifyCompare(CmpInst &I);
341 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
342 /// on the demanded bits.
343 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
344 APInt& KnownZero, APInt& KnownOne,
347 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
348 uint64_t &UndefElts, unsigned Depth = 0);
350 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
351 // PHI node as operand #0, see if we can fold the instruction into the PHI
352 // (which is only possible if all operands to the PHI are constants).
353 Instruction *FoldOpIntoPhi(Instruction &I);
355 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
356 // operator and they all are only used by the PHI, PHI together their
357 // inputs, and do the operation once, to the result of the PHI.
358 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
359 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
362 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
363 ConstantInt *AndRHS, BinaryOperator &TheAnd);
365 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
366 bool isSub, Instruction &I);
367 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
368 bool isSigned, bool Inside, Instruction &IB);
369 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
370 Instruction *MatchBSwap(BinaryOperator &I);
371 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
372 Instruction *SimplifyMemTransfer(MemIntrinsic *MI);
373 Instruction *SimplifyMemSet(MemSetInst *MI);
376 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
378 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
379 APInt& KnownOne, unsigned Depth = 0);
380 bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0);
381 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
383 int &NumCastsRemoved);
384 unsigned GetOrEnforceKnownAlignment(Value *V,
385 unsigned PrefAlign = 0);
389 char InstCombiner::ID = 0;
390 static RegisterPass<InstCombiner>
391 X("instcombine", "Combine redundant instructions");
393 // getComplexity: Assign a complexity or rank value to LLVM Values...
394 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
395 static unsigned getComplexity(Value *V) {
396 if (isa<Instruction>(V)) {
397 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
401 if (isa<Argument>(V)) return 3;
402 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
405 // isOnlyUse - Return true if this instruction will be deleted if we stop using
407 static bool isOnlyUse(Value *V) {
408 return V->hasOneUse() || isa<Constant>(V);
411 // getPromotedType - Return the specified type promoted as it would be to pass
412 // though a va_arg area...
413 static const Type *getPromotedType(const Type *Ty) {
414 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
415 if (ITy->getBitWidth() < 32)
416 return Type::Int32Ty;
421 /// getBitCastOperand - If the specified operand is a CastInst or a constant
422 /// expression bitcast, return the operand value, otherwise return null.
423 static Value *getBitCastOperand(Value *V) {
424 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
425 return I->getOperand(0);
426 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
427 if (CE->getOpcode() == Instruction::BitCast)
428 return CE->getOperand(0);
432 /// This function is a wrapper around CastInst::isEliminableCastPair. It
433 /// simply extracts arguments and returns what that function returns.
434 static Instruction::CastOps
435 isEliminableCastPair(
436 const CastInst *CI, ///< The first cast instruction
437 unsigned opcode, ///< The opcode of the second cast instruction
438 const Type *DstTy, ///< The target type for the second cast instruction
439 TargetData *TD ///< The target data for pointer size
442 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
443 const Type *MidTy = CI->getType(); // B from above
445 // Get the opcodes of the two Cast instructions
446 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
447 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
449 return Instruction::CastOps(
450 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
451 DstTy, TD->getIntPtrType()));
454 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
455 /// in any code being generated. It does not require codegen if V is simple
456 /// enough or if the cast can be folded into other casts.
457 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
458 const Type *Ty, TargetData *TD) {
459 if (V->getType() == Ty || isa<Constant>(V)) return false;
461 // If this is another cast that can be eliminated, it isn't codegen either.
462 if (const CastInst *CI = dyn_cast<CastInst>(V))
463 if (isEliminableCastPair(CI, opcode, Ty, TD))
468 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
469 /// InsertBefore instruction. This is specialized a bit to avoid inserting
470 /// casts that are known to not do anything...
472 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
473 Value *V, const Type *DestTy,
474 Instruction *InsertBefore) {
475 if (V->getType() == DestTy) return V;
476 if (Constant *C = dyn_cast<Constant>(V))
477 return ConstantExpr::getCast(opcode, C, DestTy);
479 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
482 // SimplifyCommutative - This performs a few simplifications for commutative
485 // 1. Order operands such that they are listed from right (least complex) to
486 // left (most complex). This puts constants before unary operators before
489 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
490 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
492 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
493 bool Changed = false;
494 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
495 Changed = !I.swapOperands();
497 if (!I.isAssociative()) return Changed;
498 Instruction::BinaryOps Opcode = I.getOpcode();
499 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
500 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
501 if (isa<Constant>(I.getOperand(1))) {
502 Constant *Folded = ConstantExpr::get(I.getOpcode(),
503 cast<Constant>(I.getOperand(1)),
504 cast<Constant>(Op->getOperand(1)));
505 I.setOperand(0, Op->getOperand(0));
506 I.setOperand(1, Folded);
508 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
509 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
510 isOnlyUse(Op) && isOnlyUse(Op1)) {
511 Constant *C1 = cast<Constant>(Op->getOperand(1));
512 Constant *C2 = cast<Constant>(Op1->getOperand(1));
514 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
515 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
516 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
520 I.setOperand(0, New);
521 I.setOperand(1, Folded);
528 /// SimplifyCompare - For a CmpInst this function just orders the operands
529 /// so that theyare listed from right (least complex) to left (most complex).
530 /// This puts constants before unary operators before binary operators.
531 bool InstCombiner::SimplifyCompare(CmpInst &I) {
532 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
535 // Compare instructions are not associative so there's nothing else we can do.
539 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
540 // if the LHS is a constant zero (which is the 'negate' form).
542 static inline Value *dyn_castNegVal(Value *V) {
543 if (BinaryOperator::isNeg(V))
544 return BinaryOperator::getNegArgument(V);
546 // Constants can be considered to be negated values if they can be folded.
547 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
548 return ConstantExpr::getNeg(C);
552 static inline Value *dyn_castNotVal(Value *V) {
553 if (BinaryOperator::isNot(V))
554 return BinaryOperator::getNotArgument(V);
556 // Constants can be considered to be not'ed values...
557 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
558 return ConstantInt::get(~C->getValue());
562 // dyn_castFoldableMul - If this value is a multiply that can be folded into
563 // other computations (because it has a constant operand), return the
564 // non-constant operand of the multiply, and set CST to point to the multiplier.
565 // Otherwise, return null.
567 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
568 if (V->hasOneUse() && V->getType()->isInteger())
569 if (Instruction *I = dyn_cast<Instruction>(V)) {
570 if (I->getOpcode() == Instruction::Mul)
571 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
572 return I->getOperand(0);
573 if (I->getOpcode() == Instruction::Shl)
574 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
575 // The multiplier is really 1 << CST.
576 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
577 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
578 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
579 return I->getOperand(0);
585 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
586 /// expression, return it.
587 static User *dyn_castGetElementPtr(Value *V) {
588 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
589 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
590 if (CE->getOpcode() == Instruction::GetElementPtr)
591 return cast<User>(V);
595 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
596 /// opcode value. Otherwise return UserOp1.
597 static unsigned getOpcode(User *U) {
598 if (Instruction *I = dyn_cast<Instruction>(U))
599 return I->getOpcode();
600 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U))
601 return CE->getOpcode();
602 // Use UserOp1 to mean there's no opcode.
603 return Instruction::UserOp1;
606 /// AddOne - Add one to a ConstantInt
607 static ConstantInt *AddOne(ConstantInt *C) {
608 APInt Val(C->getValue());
609 return ConstantInt::get(++Val);
611 /// SubOne - Subtract one from a ConstantInt
612 static ConstantInt *SubOne(ConstantInt *C) {
613 APInt Val(C->getValue());
614 return ConstantInt::get(--Val);
616 /// Add - Add two ConstantInts together
617 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
618 return ConstantInt::get(C1->getValue() + C2->getValue());
620 /// And - Bitwise AND two ConstantInts together
621 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
622 return ConstantInt::get(C1->getValue() & C2->getValue());
624 /// Subtract - Subtract one ConstantInt from another
625 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
626 return ConstantInt::get(C1->getValue() - C2->getValue());
628 /// Multiply - Multiply two ConstantInts together
629 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
630 return ConstantInt::get(C1->getValue() * C2->getValue());
632 /// MultiplyOverflows - True if the multiply can not be expressed in an int
634 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
635 uint32_t W = C1->getBitWidth();
636 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
645 APInt MulExt = LHSExt * RHSExt;
648 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
649 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
650 return MulExt.slt(Min) || MulExt.sgt(Max);
652 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
655 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
656 /// known to be either zero or one and return them in the KnownZero/KnownOne
657 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
659 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
660 /// we cannot optimize based on the assumption that it is zero without changing
661 /// it to be an explicit zero. If we don't change it to zero, other code could
662 /// optimized based on the contradictory assumption that it is non-zero.
663 /// Because instcombine aggressively folds operations with undef args anyway,
664 /// this won't lose us code quality.
665 void InstCombiner::ComputeMaskedBits(Value *V, const APInt &Mask,
666 APInt& KnownZero, APInt& KnownOne,
668 assert(V && "No Value?");
669 assert(Depth <= 6 && "Limit Search Depth");
670 uint32_t BitWidth = Mask.getBitWidth();
671 assert((V->getType()->isInteger() || isa<PointerType>(V->getType())) &&
672 "Not integer or pointer type!");
673 assert((!TD || TD->getTypeSizeInBits(V->getType()) == BitWidth) &&
674 (!isa<IntegerType>(V->getType()) ||
675 V->getType()->getPrimitiveSizeInBits() == BitWidth) &&
676 KnownZero.getBitWidth() == BitWidth &&
677 KnownOne.getBitWidth() == BitWidth &&
678 "V, Mask, KnownOne and KnownZero should have same BitWidth");
679 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
680 // We know all of the bits for a constant!
681 KnownOne = CI->getValue() & Mask;
682 KnownZero = ~KnownOne & Mask;
685 // Null is all-zeros.
686 if (isa<ConstantPointerNull>(V)) {
691 // The address of an aligned GlobalValue has trailing zeros.
692 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
693 unsigned Align = GV->getAlignment();
694 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
695 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
697 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
698 CountTrailingZeros_32(Align));
705 KnownZero.clear(); KnownOne.clear(); // Start out not knowing anything.
707 if (Depth == 6 || Mask == 0)
708 return; // Limit search depth.
710 User *I = dyn_cast<User>(V);
713 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
714 switch (getOpcode(I)) {
716 case Instruction::And: {
717 // If either the LHS or the RHS are Zero, the result is zero.
718 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
719 APInt Mask2(Mask & ~KnownZero);
720 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
721 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
722 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
724 // Output known-1 bits are only known if set in both the LHS & RHS.
725 KnownOne &= KnownOne2;
726 // Output known-0 are known to be clear if zero in either the LHS | RHS.
727 KnownZero |= KnownZero2;
730 case Instruction::Or: {
731 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
732 APInt Mask2(Mask & ~KnownOne);
733 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
734 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
735 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
737 // Output known-0 bits are only known if clear in both the LHS & RHS.
738 KnownZero &= KnownZero2;
739 // Output known-1 are known to be set if set in either the LHS | RHS.
740 KnownOne |= KnownOne2;
743 case Instruction::Xor: {
744 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
745 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
746 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
747 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
749 // Output known-0 bits are known if clear or set in both the LHS & RHS.
750 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
751 // Output known-1 are known to be set if set in only one of the LHS, RHS.
752 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
753 KnownZero = KnownZeroOut;
756 case Instruction::Mul: {
757 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
758 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
759 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
760 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
761 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
763 // If low bits are zero in either operand, output low known-0 bits.
764 // Also compute a conserative estimate for high known-0 bits.
765 // More trickiness is possible, but this is sufficient for the
766 // interesting case of alignment computation.
768 unsigned TrailZ = KnownZero.countTrailingOnes() +
769 KnownZero2.countTrailingOnes();
770 unsigned LeadZ = std::max(KnownZero.countLeadingOnes() +
771 KnownZero2.countLeadingOnes(),
772 BitWidth) - BitWidth;
774 TrailZ = std::min(TrailZ, BitWidth);
775 LeadZ = std::min(LeadZ, BitWidth);
776 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) |
777 APInt::getHighBitsSet(BitWidth, LeadZ);
781 case Instruction::UDiv: {
782 // For the purposes of computing leading zeros we can conservatively
783 // treat a udiv as a logical right shift by the power of 2 known to
784 // be less than the denominator.
785 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
786 ComputeMaskedBits(I->getOperand(0),
787 AllOnes, KnownZero2, KnownOne2, Depth+1);
788 unsigned LeadZ = KnownZero2.countLeadingOnes();
792 ComputeMaskedBits(I->getOperand(1),
793 AllOnes, KnownZero2, KnownOne2, Depth+1);
794 unsigned RHSUnknownLeadingOnes = KnownOne2.countLeadingZeros();
795 if (RHSUnknownLeadingOnes != BitWidth)
796 LeadZ = std::min(BitWidth,
797 LeadZ + BitWidth - RHSUnknownLeadingOnes - 1);
799 KnownZero = APInt::getHighBitsSet(BitWidth, LeadZ) & Mask;
802 case Instruction::Select:
803 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
804 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
805 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
806 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
808 // Only known if known in both the LHS and RHS.
809 KnownOne &= KnownOne2;
810 KnownZero &= KnownZero2;
812 case Instruction::FPTrunc:
813 case Instruction::FPExt:
814 case Instruction::FPToUI:
815 case Instruction::FPToSI:
816 case Instruction::SIToFP:
817 case Instruction::UIToFP:
818 return; // Can't work with floating point.
819 case Instruction::PtrToInt:
820 case Instruction::IntToPtr:
821 // We can't handle these if we don't know the pointer size.
823 // Fall through and handle them the same as zext/trunc.
824 case Instruction::ZExt:
825 case Instruction::Trunc: {
826 // All these have integer operands
827 const Type *SrcTy = I->getOperand(0)->getType();
828 uint32_t SrcBitWidth = TD ?
829 TD->getTypeSizeInBits(SrcTy) :
830 SrcTy->getPrimitiveSizeInBits();
832 MaskIn.zextOrTrunc(SrcBitWidth);
833 KnownZero.zextOrTrunc(SrcBitWidth);
834 KnownOne.zextOrTrunc(SrcBitWidth);
835 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
836 KnownZero.zextOrTrunc(BitWidth);
837 KnownOne.zextOrTrunc(BitWidth);
838 // Any top bits are known to be zero.
839 if (BitWidth > SrcBitWidth)
840 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
843 case Instruction::BitCast: {
844 const Type *SrcTy = I->getOperand(0)->getType();
845 if (SrcTy->isInteger() || isa<PointerType>(SrcTy)) {
846 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
851 case Instruction::SExt: {
852 // Compute the bits in the result that are not present in the input.
853 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
854 uint32_t SrcBitWidth = SrcTy->getBitWidth();
857 MaskIn.trunc(SrcBitWidth);
858 KnownZero.trunc(SrcBitWidth);
859 KnownOne.trunc(SrcBitWidth);
860 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
861 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
862 KnownZero.zext(BitWidth);
863 KnownOne.zext(BitWidth);
865 // If the sign bit of the input is known set or clear, then we know the
866 // top bits of the result.
867 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
868 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
869 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
870 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
873 case Instruction::Shl:
874 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
875 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
876 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
877 APInt Mask2(Mask.lshr(ShiftAmt));
878 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
879 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
880 KnownZero <<= ShiftAmt;
881 KnownOne <<= ShiftAmt;
882 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
886 case Instruction::LShr:
887 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
888 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
889 // Compute the new bits that are at the top now.
890 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
892 // Unsigned shift right.
893 APInt Mask2(Mask.shl(ShiftAmt));
894 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
895 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
896 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
897 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
898 // high bits known zero.
899 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
903 case Instruction::AShr:
904 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
905 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
906 // Compute the new bits that are at the top now.
907 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
909 // Signed shift right.
910 APInt Mask2(Mask.shl(ShiftAmt));
911 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
912 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
913 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
914 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
916 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
917 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
918 KnownZero |= HighBits;
919 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
920 KnownOne |= HighBits;
924 case Instruction::Sub: {
925 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
926 // We know that the top bits of C-X are clear if X contains less bits
927 // than C (i.e. no wrap-around can happen). For example, 20-X is
928 // positive if we can prove that X is >= 0 and < 16.
929 if (!CLHS->getValue().isNegative()) {
930 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
931 // NLZ can't be BitWidth with no sign bit
932 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
933 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero2, KnownOne2,
936 // If all of the MaskV bits are known to be zero, then we know the
937 // output top bits are zero, because we now know that the output is
939 if ((KnownZero2 & MaskV) == MaskV) {
940 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
941 // Top bits known zero.
942 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
948 case Instruction::Add: {
949 // Output known-0 bits are known if clear or set in both the low clear bits
950 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
952 APInt Mask2 = APInt::getLowBitsSet(BitWidth, Mask.countTrailingOnes());
953 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
954 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
955 unsigned KnownZeroOut = KnownZero2.countTrailingOnes();
957 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero2, KnownOne2, Depth+1);
958 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
959 KnownZeroOut = std::min(KnownZeroOut,
960 KnownZero2.countTrailingOnes());
962 KnownZero |= APInt::getLowBitsSet(BitWidth, KnownZeroOut);
965 case Instruction::SRem:
966 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
967 APInt RA = Rem->getValue();
968 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
969 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
970 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
971 ComputeMaskedBits(I->getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
973 // The sign of a remainder is equal to the sign of the first
974 // operand (zero being positive).
975 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
976 KnownZero2 |= ~LowBits;
977 else if (KnownOne2[BitWidth-1])
978 KnownOne2 |= ~LowBits;
980 KnownZero |= KnownZero2 & Mask;
981 KnownOne |= KnownOne2 & Mask;
983 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
987 case Instruction::URem: {
988 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
989 APInt RA = Rem->getValue();
990 if (RA.isPowerOf2()) {
991 APInt LowBits = (RA - 1);
992 APInt Mask2 = LowBits & Mask;
993 KnownZero |= ~LowBits & Mask;
994 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
995 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1000 // Since the result is less than or equal to either operand, any leading
1001 // zero bits in either operand must also exist in the result.
1002 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1003 ComputeMaskedBits(I->getOperand(0), AllOnes, KnownZero, KnownOne,
1005 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
1008 uint32_t Leaders = std::max(KnownZero.countLeadingOnes(),
1009 KnownZero2.countLeadingOnes());
1011 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
1015 case Instruction::Alloca:
1016 case Instruction::Malloc: {
1017 AllocationInst *AI = cast<AllocationInst>(V);
1018 unsigned Align = AI->getAlignment();
1019 if (Align == 0 && TD) {
1020 if (isa<AllocaInst>(AI))
1021 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
1022 else if (isa<MallocInst>(AI)) {
1023 // Malloc returns maximally aligned memory.
1024 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
1027 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
1030 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
1035 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
1036 CountTrailingZeros_32(Align));
1039 case Instruction::GetElementPtr: {
1040 // Analyze all of the subscripts of this getelementptr instruction
1041 // to determine if we can prove known low zero bits.
1042 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
1043 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1044 ComputeMaskedBits(I->getOperand(0), LocalMask,
1045 LocalKnownZero, LocalKnownOne, Depth+1);
1046 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1048 gep_type_iterator GTI = gep_type_begin(I);
1049 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1050 Value *Index = I->getOperand(i);
1051 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
1052 // Handle struct member offset arithmetic.
1054 const StructLayout *SL = TD->getStructLayout(STy);
1055 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1056 uint64_t Offset = SL->getElementOffset(Idx);
1057 TrailZ = std::min(TrailZ,
1058 CountTrailingZeros_64(Offset));
1060 // Handle array index arithmetic.
1061 const Type *IndexedTy = GTI.getIndexedType();
1062 if (!IndexedTy->isSized()) return;
1063 unsigned GEPOpiBits = Index->getType()->getPrimitiveSizeInBits();
1064 uint64_t TypeSize = TD ? TD->getABITypeSize(IndexedTy) : 1;
1065 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
1066 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1067 ComputeMaskedBits(Index, LocalMask,
1068 LocalKnownZero, LocalKnownOne, Depth+1);
1069 TrailZ = std::min(TrailZ,
1070 CountTrailingZeros_64(TypeSize) +
1071 LocalKnownZero.countTrailingOnes());
1075 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
1078 case Instruction::PHI: {
1079 PHINode *P = cast<PHINode>(I);
1080 // Handle the case of a simple two-predecessor recurrence PHI.
1081 // There's a lot more that could theoretically be done here, but
1082 // this is sufficient to catch some interesting cases.
1083 if (P->getNumIncomingValues() == 2) {
1084 for (unsigned i = 0; i != 2; ++i) {
1085 Value *L = P->getIncomingValue(i);
1086 Value *R = P->getIncomingValue(!i);
1087 User *LU = dyn_cast<User>(L);
1088 unsigned Opcode = LU ? getOpcode(LU) : (unsigned)Instruction::UserOp1;
1089 // Check for operations that have the property that if
1090 // both their operands have low zero bits, the result
1091 // will have low zero bits.
1092 if (Opcode == Instruction::Add ||
1093 Opcode == Instruction::Sub ||
1094 Opcode == Instruction::And ||
1095 Opcode == Instruction::Or ||
1096 Opcode == Instruction::Mul) {
1097 Value *LL = LU->getOperand(0);
1098 Value *LR = LU->getOperand(1);
1099 // Find a recurrence.
1106 // Ok, we have a PHI of the form L op= R. Check for low
1108 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1109 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, Depth+1);
1110 Mask2 = APInt::getLowBitsSet(BitWidth,
1111 KnownZero2.countTrailingOnes());
1114 ComputeMaskedBits(L, Mask2, KnownZero2, KnownOne2, Depth+1);
1116 APInt::getLowBitsSet(BitWidth,
1117 KnownZero2.countTrailingOnes());
1124 case Instruction::Call:
1125 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(I)) {
1126 switch (II->getIntrinsicID()) {
1128 case Intrinsic::ctpop:
1129 case Intrinsic::ctlz:
1130 case Intrinsic::cttz: {
1131 unsigned LowBits = Log2_32(BitWidth)+1;
1132 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - LowBits);
1141 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1142 /// this predicate to simplify operations downstream. Mask is known to be zero
1143 /// for bits that V cannot have.
1144 bool InstCombiner::MaskedValueIsZero(Value *V, const APInt& Mask,
1146 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1147 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
1148 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1149 return (KnownZero & Mask) == Mask;
1152 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1153 /// specified instruction is a constant integer. If so, check to see if there
1154 /// are any bits set in the constant that are not demanded. If so, shrink the
1155 /// constant and return true.
1156 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
1158 assert(I && "No instruction?");
1159 assert(OpNo < I->getNumOperands() && "Operand index too large");
1161 // If the operand is not a constant integer, nothing to do.
1162 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
1163 if (!OpC) return false;
1165 // If there are no bits set that aren't demanded, nothing to do.
1166 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
1167 if ((~Demanded & OpC->getValue()) == 0)
1170 // This instruction is producing bits that are not demanded. Shrink the RHS.
1171 Demanded &= OpC->getValue();
1172 I->setOperand(OpNo, ConstantInt::get(Demanded));
1176 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
1177 // set of known zero and one bits, compute the maximum and minimum values that
1178 // could have the specified known zero and known one bits, returning them in
1180 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
1181 const APInt& KnownZero,
1182 const APInt& KnownOne,
1183 APInt& Min, APInt& Max) {
1184 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
1185 assert(KnownZero.getBitWidth() == BitWidth &&
1186 KnownOne.getBitWidth() == BitWidth &&
1187 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
1188 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1189 APInt UnknownBits = ~(KnownZero|KnownOne);
1191 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
1192 // bit if it is unknown.
1194 Max = KnownOne|UnknownBits;
1196 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
1197 Min.set(BitWidth-1);
1198 Max.clear(BitWidth-1);
1202 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
1203 // a set of known zero and one bits, compute the maximum and minimum values that
1204 // could have the specified known zero and known one bits, returning them in
1206 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
1207 const APInt &KnownZero,
1208 const APInt &KnownOne,
1209 APInt &Min, APInt &Max) {
1210 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
1211 assert(KnownZero.getBitWidth() == BitWidth &&
1212 KnownOne.getBitWidth() == BitWidth &&
1213 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
1214 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1215 APInt UnknownBits = ~(KnownZero|KnownOne);
1217 // The minimum value is when the unknown bits are all zeros.
1219 // The maximum value is when the unknown bits are all ones.
1220 Max = KnownOne|UnknownBits;
1223 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
1224 /// value based on the demanded bits. When this function is called, it is known
1225 /// that only the bits set in DemandedMask of the result of V are ever used
1226 /// downstream. Consequently, depending on the mask and V, it may be possible
1227 /// to replace V with a constant or one of its operands. In such cases, this
1228 /// function does the replacement and returns true. In all other cases, it
1229 /// returns false after analyzing the expression and setting KnownOne and known
1230 /// to be one in the expression. KnownZero contains all the bits that are known
1231 /// to be zero in the expression. These are provided to potentially allow the
1232 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
1233 /// the expression. KnownOne and KnownZero always follow the invariant that
1234 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
1235 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
1236 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
1237 /// and KnownOne must all be the same.
1238 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
1239 APInt& KnownZero, APInt& KnownOne,
1241 assert(V != 0 && "Null pointer of Value???");
1242 assert(Depth <= 6 && "Limit Search Depth");
1243 uint32_t BitWidth = DemandedMask.getBitWidth();
1244 const IntegerType *VTy = cast<IntegerType>(V->getType());
1245 assert(VTy->getBitWidth() == BitWidth &&
1246 KnownZero.getBitWidth() == BitWidth &&
1247 KnownOne.getBitWidth() == BitWidth &&
1248 "Value *V, DemandedMask, KnownZero and KnownOne \
1249 must have same BitWidth");
1250 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1251 // We know all of the bits for a constant!
1252 KnownOne = CI->getValue() & DemandedMask;
1253 KnownZero = ~KnownOne & DemandedMask;
1259 if (!V->hasOneUse()) { // Other users may use these bits.
1260 if (Depth != 0) { // Not at the root.
1261 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1262 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1265 // If this is the root being simplified, allow it to have multiple uses,
1266 // just set the DemandedMask to all bits.
1267 DemandedMask = APInt::getAllOnesValue(BitWidth);
1268 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1269 if (V != UndefValue::get(VTy))
1270 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1272 } else if (Depth == 6) { // Limit search depth.
1276 Instruction *I = dyn_cast<Instruction>(V);
1277 if (!I) return false; // Only analyze instructions.
1279 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1280 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
1281 switch (I->getOpcode()) {
1283 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1285 case Instruction::And:
1286 // If either the LHS or the RHS are Zero, the result is zero.
1287 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1288 RHSKnownZero, RHSKnownOne, Depth+1))
1290 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1291 "Bits known to be one AND zero?");
1293 // If something is known zero on the RHS, the bits aren't demanded on the
1295 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
1296 LHSKnownZero, LHSKnownOne, Depth+1))
1298 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1299 "Bits known to be one AND zero?");
1301 // If all of the demanded bits are known 1 on one side, return the other.
1302 // These bits cannot contribute to the result of the 'and'.
1303 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1304 (DemandedMask & ~LHSKnownZero))
1305 return UpdateValueUsesWith(I, I->getOperand(0));
1306 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1307 (DemandedMask & ~RHSKnownZero))
1308 return UpdateValueUsesWith(I, I->getOperand(1));
1310 // If all of the demanded bits in the inputs are known zeros, return zero.
1311 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1312 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1314 // If the RHS is a constant, see if we can simplify it.
1315 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1316 return UpdateValueUsesWith(I, I);
1318 // Output known-1 bits are only known if set in both the LHS & RHS.
1319 RHSKnownOne &= LHSKnownOne;
1320 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1321 RHSKnownZero |= LHSKnownZero;
1323 case Instruction::Or:
1324 // If either the LHS or the RHS are One, the result is One.
1325 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1326 RHSKnownZero, RHSKnownOne, Depth+1))
1328 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1329 "Bits known to be one AND zero?");
1330 // If something is known one on the RHS, the bits aren't demanded on the
1332 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1333 LHSKnownZero, LHSKnownOne, Depth+1))
1335 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1336 "Bits known to be one AND zero?");
1338 // If all of the demanded bits are known zero on one side, return the other.
1339 // These bits cannot contribute to the result of the 'or'.
1340 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1341 (DemandedMask & ~LHSKnownOne))
1342 return UpdateValueUsesWith(I, I->getOperand(0));
1343 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1344 (DemandedMask & ~RHSKnownOne))
1345 return UpdateValueUsesWith(I, I->getOperand(1));
1347 // If all of the potentially set bits on one side are known to be set on
1348 // the other side, just use the 'other' side.
1349 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1350 (DemandedMask & (~RHSKnownZero)))
1351 return UpdateValueUsesWith(I, I->getOperand(0));
1352 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1353 (DemandedMask & (~LHSKnownZero)))
1354 return UpdateValueUsesWith(I, I->getOperand(1));
1356 // If the RHS is a constant, see if we can simplify it.
1357 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1358 return UpdateValueUsesWith(I, I);
1360 // Output known-0 bits are only known if clear in both the LHS & RHS.
1361 RHSKnownZero &= LHSKnownZero;
1362 // Output known-1 are known to be set if set in either the LHS | RHS.
1363 RHSKnownOne |= LHSKnownOne;
1365 case Instruction::Xor: {
1366 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1367 RHSKnownZero, RHSKnownOne, Depth+1))
1369 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1370 "Bits known to be one AND zero?");
1371 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1372 LHSKnownZero, LHSKnownOne, Depth+1))
1374 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1375 "Bits known to be one AND zero?");
1377 // If all of the demanded bits are known zero on one side, return the other.
1378 // These bits cannot contribute to the result of the 'xor'.
1379 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1380 return UpdateValueUsesWith(I, I->getOperand(0));
1381 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1382 return UpdateValueUsesWith(I, I->getOperand(1));
1384 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1385 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1386 (RHSKnownOne & LHSKnownOne);
1387 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1388 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1389 (RHSKnownOne & LHSKnownZero);
1391 // If all of the demanded bits are known to be zero on one side or the
1392 // other, turn this into an *inclusive* or.
1393 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1394 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1396 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1398 InsertNewInstBefore(Or, *I);
1399 return UpdateValueUsesWith(I, Or);
1402 // If all of the demanded bits on one side are known, and all of the set
1403 // bits on that side are also known to be set on the other side, turn this
1404 // into an AND, as we know the bits will be cleared.
1405 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1406 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1408 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1409 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1411 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1412 InsertNewInstBefore(And, *I);
1413 return UpdateValueUsesWith(I, And);
1417 // If the RHS is a constant, see if we can simplify it.
1418 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1419 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1420 return UpdateValueUsesWith(I, I);
1422 RHSKnownZero = KnownZeroOut;
1423 RHSKnownOne = KnownOneOut;
1426 case Instruction::Select:
1427 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1428 RHSKnownZero, RHSKnownOne, Depth+1))
1430 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1431 LHSKnownZero, LHSKnownOne, Depth+1))
1433 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1434 "Bits known to be one AND zero?");
1435 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1436 "Bits known to be one AND zero?");
1438 // If the operands are constants, see if we can simplify them.
1439 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1440 return UpdateValueUsesWith(I, I);
1441 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1442 return UpdateValueUsesWith(I, I);
1444 // Only known if known in both the LHS and RHS.
1445 RHSKnownOne &= LHSKnownOne;
1446 RHSKnownZero &= LHSKnownZero;
1448 case Instruction::Trunc: {
1450 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1451 DemandedMask.zext(truncBf);
1452 RHSKnownZero.zext(truncBf);
1453 RHSKnownOne.zext(truncBf);
1454 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1455 RHSKnownZero, RHSKnownOne, Depth+1))
1457 DemandedMask.trunc(BitWidth);
1458 RHSKnownZero.trunc(BitWidth);
1459 RHSKnownOne.trunc(BitWidth);
1460 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1461 "Bits known to be one AND zero?");
1464 case Instruction::BitCast:
1465 if (!I->getOperand(0)->getType()->isInteger())
1468 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1469 RHSKnownZero, RHSKnownOne, Depth+1))
1471 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1472 "Bits known to be one AND zero?");
1474 case Instruction::ZExt: {
1475 // Compute the bits in the result that are not present in the input.
1476 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1477 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1479 DemandedMask.trunc(SrcBitWidth);
1480 RHSKnownZero.trunc(SrcBitWidth);
1481 RHSKnownOne.trunc(SrcBitWidth);
1482 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1483 RHSKnownZero, RHSKnownOne, Depth+1))
1485 DemandedMask.zext(BitWidth);
1486 RHSKnownZero.zext(BitWidth);
1487 RHSKnownOne.zext(BitWidth);
1488 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1489 "Bits known to be one AND zero?");
1490 // The top bits are known to be zero.
1491 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1494 case Instruction::SExt: {
1495 // Compute the bits in the result that are not present in the input.
1496 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1497 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1499 APInt InputDemandedBits = DemandedMask &
1500 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1502 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1503 // If any of the sign extended bits are demanded, we know that the sign
1505 if ((NewBits & DemandedMask) != 0)
1506 InputDemandedBits.set(SrcBitWidth-1);
1508 InputDemandedBits.trunc(SrcBitWidth);
1509 RHSKnownZero.trunc(SrcBitWidth);
1510 RHSKnownOne.trunc(SrcBitWidth);
1511 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1512 RHSKnownZero, RHSKnownOne, Depth+1))
1514 InputDemandedBits.zext(BitWidth);
1515 RHSKnownZero.zext(BitWidth);
1516 RHSKnownOne.zext(BitWidth);
1517 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1518 "Bits known to be one AND zero?");
1520 // If the sign bit of the input is known set or clear, then we know the
1521 // top bits of the result.
1523 // If the input sign bit is known zero, or if the NewBits are not demanded
1524 // convert this into a zero extension.
1525 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1527 // Convert to ZExt cast
1528 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1529 return UpdateValueUsesWith(I, NewCast);
1530 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1531 RHSKnownOne |= NewBits;
1535 case Instruction::Add: {
1536 // Figure out what the input bits are. If the top bits of the and result
1537 // are not demanded, then the add doesn't demand them from its input
1539 uint32_t NLZ = DemandedMask.countLeadingZeros();
1541 // If there is a constant on the RHS, there are a variety of xformations
1543 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1544 // If null, this should be simplified elsewhere. Some of the xforms here
1545 // won't work if the RHS is zero.
1549 // If the top bit of the output is demanded, demand everything from the
1550 // input. Otherwise, we demand all the input bits except NLZ top bits.
1551 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1553 // Find information about known zero/one bits in the input.
1554 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1555 LHSKnownZero, LHSKnownOne, Depth+1))
1558 // If the RHS of the add has bits set that can't affect the input, reduce
1560 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1561 return UpdateValueUsesWith(I, I);
1563 // Avoid excess work.
1564 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1567 // Turn it into OR if input bits are zero.
1568 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1570 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1572 InsertNewInstBefore(Or, *I);
1573 return UpdateValueUsesWith(I, Or);
1576 // We can say something about the output known-zero and known-one bits,
1577 // depending on potential carries from the input constant and the
1578 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1579 // bits set and the RHS constant is 0x01001, then we know we have a known
1580 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1582 // To compute this, we first compute the potential carry bits. These are
1583 // the bits which may be modified. I'm not aware of a better way to do
1585 const APInt& RHSVal = RHS->getValue();
1586 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1588 // Now that we know which bits have carries, compute the known-1/0 sets.
1590 // Bits are known one if they are known zero in one operand and one in the
1591 // other, and there is no input carry.
1592 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1593 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1595 // Bits are known zero if they are known zero in both operands and there
1596 // is no input carry.
1597 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1599 // If the high-bits of this ADD are not demanded, then it does not demand
1600 // the high bits of its LHS or RHS.
1601 if (DemandedMask[BitWidth-1] == 0) {
1602 // Right fill the mask of bits for this ADD to demand the most
1603 // significant bit and all those below it.
1604 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1605 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1606 LHSKnownZero, LHSKnownOne, Depth+1))
1608 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1609 LHSKnownZero, LHSKnownOne, Depth+1))
1615 case Instruction::Sub:
1616 // If the high-bits of this SUB are not demanded, then it does not demand
1617 // the high bits of its LHS or RHS.
1618 if (DemandedMask[BitWidth-1] == 0) {
1619 // Right fill the mask of bits for this SUB to demand the most
1620 // significant bit and all those below it.
1621 uint32_t NLZ = DemandedMask.countLeadingZeros();
1622 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1623 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1624 LHSKnownZero, LHSKnownOne, Depth+1))
1626 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1627 LHSKnownZero, LHSKnownOne, Depth+1))
1630 // Otherwise just hand the sub off to ComputeMaskedBits to fill in
1631 // the known zeros and ones.
1632 ComputeMaskedBits(V, DemandedMask, RHSKnownZero, RHSKnownOne, Depth);
1634 case Instruction::Shl:
1635 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1636 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1637 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1638 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1639 RHSKnownZero, RHSKnownOne, Depth+1))
1641 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1642 "Bits known to be one AND zero?");
1643 RHSKnownZero <<= ShiftAmt;
1644 RHSKnownOne <<= ShiftAmt;
1645 // low bits known zero.
1647 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1650 case Instruction::LShr:
1651 // For a logical shift right
1652 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1653 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1655 // Unsigned shift right.
1656 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1657 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1658 RHSKnownZero, RHSKnownOne, Depth+1))
1660 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1661 "Bits known to be one AND zero?");
1662 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1663 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1665 // Compute the new bits that are at the top now.
1666 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1667 RHSKnownZero |= HighBits; // high bits known zero.
1671 case Instruction::AShr:
1672 // If this is an arithmetic shift right and only the low-bit is set, we can
1673 // always convert this into a logical shr, even if the shift amount is
1674 // variable. The low bit of the shift cannot be an input sign bit unless
1675 // the shift amount is >= the size of the datatype, which is undefined.
1676 if (DemandedMask == 1) {
1677 // Perform the logical shift right.
1678 Value *NewVal = BinaryOperator::createLShr(
1679 I->getOperand(0), I->getOperand(1), I->getName());
1680 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1681 return UpdateValueUsesWith(I, NewVal);
1684 // If the sign bit is the only bit demanded by this ashr, then there is no
1685 // need to do it, the shift doesn't change the high bit.
1686 if (DemandedMask.isSignBit())
1687 return UpdateValueUsesWith(I, I->getOperand(0));
1689 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1690 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1692 // Signed shift right.
1693 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1694 // If any of the "high bits" are demanded, we should set the sign bit as
1696 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1697 DemandedMaskIn.set(BitWidth-1);
1698 if (SimplifyDemandedBits(I->getOperand(0),
1700 RHSKnownZero, RHSKnownOne, Depth+1))
1702 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1703 "Bits known to be one AND zero?");
1704 // Compute the new bits that are at the top now.
1705 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1706 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1707 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1709 // Handle the sign bits.
1710 APInt SignBit(APInt::getSignBit(BitWidth));
1711 // Adjust to where it is now in the mask.
1712 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1714 // If the input sign bit is known to be zero, or if none of the top bits
1715 // are demanded, turn this into an unsigned shift right.
1716 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1717 (HighBits & ~DemandedMask) == HighBits) {
1718 // Perform the logical shift right.
1719 Value *NewVal = BinaryOperator::createLShr(
1720 I->getOperand(0), SA, I->getName());
1721 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1722 return UpdateValueUsesWith(I, NewVal);
1723 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1724 RHSKnownOne |= HighBits;
1728 case Instruction::SRem:
1729 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1730 APInt RA = Rem->getValue();
1731 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1732 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) : ~RA;
1733 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1734 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1735 LHSKnownZero, LHSKnownOne, Depth+1))
1738 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1739 LHSKnownZero |= ~LowBits;
1740 else if (LHSKnownOne[BitWidth-1])
1741 LHSKnownOne |= ~LowBits;
1743 KnownZero |= LHSKnownZero & DemandedMask;
1744 KnownOne |= LHSKnownOne & DemandedMask;
1746 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1750 case Instruction::URem: {
1751 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1752 APInt RA = Rem->getValue();
1753 if (RA.isPowerOf2()) {
1754 APInt LowBits = (RA - 1);
1755 APInt Mask2 = LowBits & DemandedMask;
1756 KnownZero |= ~LowBits & DemandedMask;
1757 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1758 KnownZero, KnownOne, Depth+1))
1761 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1766 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1767 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1768 if (SimplifyDemandedBits(I->getOperand(0), AllOnes,
1769 KnownZero2, KnownOne2, Depth+1))
1772 uint32_t Leaders = KnownZero2.countLeadingOnes();
1773 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1774 KnownZero2, KnownOne2, Depth+1))
1777 Leaders = std::max(Leaders,
1778 KnownZero2.countLeadingOnes());
1779 KnownZero = APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1784 // If the client is only demanding bits that we know, return the known
1786 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1787 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1792 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1793 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1794 /// actually used by the caller. This method analyzes which elements of the
1795 /// operand are undef and returns that information in UndefElts.
1797 /// If the information about demanded elements can be used to simplify the
1798 /// operation, the operation is simplified, then the resultant value is
1799 /// returned. This returns null if no change was made.
1800 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1801 uint64_t &UndefElts,
1803 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1804 assert(VWidth <= 64 && "Vector too wide to analyze!");
1805 uint64_t EltMask = ~0ULL >> (64-VWidth);
1806 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1807 "Invalid DemandedElts!");
1809 if (isa<UndefValue>(V)) {
1810 // If the entire vector is undefined, just return this info.
1811 UndefElts = EltMask;
1813 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1814 UndefElts = EltMask;
1815 return UndefValue::get(V->getType());
1819 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1820 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1821 Constant *Undef = UndefValue::get(EltTy);
1823 std::vector<Constant*> Elts;
1824 for (unsigned i = 0; i != VWidth; ++i)
1825 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1826 Elts.push_back(Undef);
1827 UndefElts |= (1ULL << i);
1828 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1829 Elts.push_back(Undef);
1830 UndefElts |= (1ULL << i);
1831 } else { // Otherwise, defined.
1832 Elts.push_back(CP->getOperand(i));
1835 // If we changed the constant, return it.
1836 Constant *NewCP = ConstantVector::get(Elts);
1837 return NewCP != CP ? NewCP : 0;
1838 } else if (isa<ConstantAggregateZero>(V)) {
1839 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1841 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1842 Constant *Zero = Constant::getNullValue(EltTy);
1843 Constant *Undef = UndefValue::get(EltTy);
1844 std::vector<Constant*> Elts;
1845 for (unsigned i = 0; i != VWidth; ++i)
1846 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1847 UndefElts = DemandedElts ^ EltMask;
1848 return ConstantVector::get(Elts);
1851 if (!V->hasOneUse()) { // Other users may use these bits.
1852 if (Depth != 0) { // Not at the root.
1853 // TODO: Just compute the UndefElts information recursively.
1857 } else if (Depth == 10) { // Limit search depth.
1861 Instruction *I = dyn_cast<Instruction>(V);
1862 if (!I) return false; // Only analyze instructions.
1864 bool MadeChange = false;
1865 uint64_t UndefElts2;
1867 switch (I->getOpcode()) {
1870 case Instruction::InsertElement: {
1871 // If this is a variable index, we don't know which element it overwrites.
1872 // demand exactly the same input as we produce.
1873 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1875 // Note that we can't propagate undef elt info, because we don't know
1876 // which elt is getting updated.
1877 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1878 UndefElts2, Depth+1);
1879 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1883 // If this is inserting an element that isn't demanded, remove this
1885 unsigned IdxNo = Idx->getZExtValue();
1886 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1887 return AddSoonDeadInstToWorklist(*I, 0);
1889 // Otherwise, the element inserted overwrites whatever was there, so the
1890 // input demanded set is simpler than the output set.
1891 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1892 DemandedElts & ~(1ULL << IdxNo),
1893 UndefElts, Depth+1);
1894 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1896 // The inserted element is defined.
1897 UndefElts |= 1ULL << IdxNo;
1900 case Instruction::BitCast: {
1901 // Vector->vector casts only.
1902 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1904 unsigned InVWidth = VTy->getNumElements();
1905 uint64_t InputDemandedElts = 0;
1908 if (VWidth == InVWidth) {
1909 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1910 // elements as are demanded of us.
1912 InputDemandedElts = DemandedElts;
1913 } else if (VWidth > InVWidth) {
1917 // If there are more elements in the result than there are in the source,
1918 // then an input element is live if any of the corresponding output
1919 // elements are live.
1920 Ratio = VWidth/InVWidth;
1921 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1922 if (DemandedElts & (1ULL << OutIdx))
1923 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1929 // If there are more elements in the source than there are in the result,
1930 // then an input element is live if the corresponding output element is
1932 Ratio = InVWidth/VWidth;
1933 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1934 if (DemandedElts & (1ULL << InIdx/Ratio))
1935 InputDemandedElts |= 1ULL << InIdx;
1938 // div/rem demand all inputs, because they don't want divide by zero.
1939 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1940 UndefElts2, Depth+1);
1942 I->setOperand(0, TmpV);
1946 UndefElts = UndefElts2;
1947 if (VWidth > InVWidth) {
1948 assert(0 && "Unimp");
1949 // If there are more elements in the result than there are in the source,
1950 // then an output element is undef if the corresponding input element is
1952 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1953 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1954 UndefElts |= 1ULL << OutIdx;
1955 } else if (VWidth < InVWidth) {
1956 assert(0 && "Unimp");
1957 // If there are more elements in the source than there are in the result,
1958 // then a result element is undef if all of the corresponding input
1959 // elements are undef.
1960 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1961 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1962 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1963 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1967 case Instruction::And:
1968 case Instruction::Or:
1969 case Instruction::Xor:
1970 case Instruction::Add:
1971 case Instruction::Sub:
1972 case Instruction::Mul:
1973 // div/rem demand all inputs, because they don't want divide by zero.
1974 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1975 UndefElts, Depth+1);
1976 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1977 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1978 UndefElts2, Depth+1);
1979 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1981 // Output elements are undefined if both are undefined. Consider things
1982 // like undef&0. The result is known zero, not undef.
1983 UndefElts &= UndefElts2;
1986 case Instruction::Call: {
1987 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1989 switch (II->getIntrinsicID()) {
1992 // Binary vector operations that work column-wise. A dest element is a
1993 // function of the corresponding input elements from the two inputs.
1994 case Intrinsic::x86_sse_sub_ss:
1995 case Intrinsic::x86_sse_mul_ss:
1996 case Intrinsic::x86_sse_min_ss:
1997 case Intrinsic::x86_sse_max_ss:
1998 case Intrinsic::x86_sse2_sub_sd:
1999 case Intrinsic::x86_sse2_mul_sd:
2000 case Intrinsic::x86_sse2_min_sd:
2001 case Intrinsic::x86_sse2_max_sd:
2002 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
2003 UndefElts, Depth+1);
2004 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
2005 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
2006 UndefElts2, Depth+1);
2007 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
2009 // If only the low elt is demanded and this is a scalarizable intrinsic,
2010 // scalarize it now.
2011 if (DemandedElts == 1) {
2012 switch (II->getIntrinsicID()) {
2014 case Intrinsic::x86_sse_sub_ss:
2015 case Intrinsic::x86_sse_mul_ss:
2016 case Intrinsic::x86_sse2_sub_sd:
2017 case Intrinsic::x86_sse2_mul_sd:
2018 // TODO: Lower MIN/MAX/ABS/etc
2019 Value *LHS = II->getOperand(1);
2020 Value *RHS = II->getOperand(2);
2021 // Extract the element as scalars.
2022 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
2023 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
2025 switch (II->getIntrinsicID()) {
2026 default: assert(0 && "Case stmts out of sync!");
2027 case Intrinsic::x86_sse_sub_ss:
2028 case Intrinsic::x86_sse2_sub_sd:
2029 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
2030 II->getName()), *II);
2032 case Intrinsic::x86_sse_mul_ss:
2033 case Intrinsic::x86_sse2_mul_sd:
2034 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
2035 II->getName()), *II);
2040 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
2042 InsertNewInstBefore(New, *II);
2043 AddSoonDeadInstToWorklist(*II, 0);
2048 // Output elements are undefined if both are undefined. Consider things
2049 // like undef&0. The result is known zero, not undef.
2050 UndefElts &= UndefElts2;
2056 return MadeChange ? I : 0;
2059 /// @returns true if the specified compare predicate is
2060 /// true when both operands are equal...
2061 /// @brief Determine if the icmp Predicate is true when both operands are equal
2062 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
2063 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
2064 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
2065 pred == ICmpInst::ICMP_SLE;
2068 /// @returns true if the specified compare instruction is
2069 /// true when both operands are equal...
2070 /// @brief Determine if the ICmpInst returns true when both operands are equal
2071 static bool isTrueWhenEqual(ICmpInst &ICI) {
2072 return isTrueWhenEqual(ICI.getPredicate());
2075 /// AssociativeOpt - Perform an optimization on an associative operator. This
2076 /// function is designed to check a chain of associative operators for a
2077 /// potential to apply a certain optimization. Since the optimization may be
2078 /// applicable if the expression was reassociated, this checks the chain, then
2079 /// reassociates the expression as necessary to expose the optimization
2080 /// opportunity. This makes use of a special Functor, which must define
2081 /// 'shouldApply' and 'apply' methods.
2083 template<typename Functor>
2084 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
2085 unsigned Opcode = Root.getOpcode();
2086 Value *LHS = Root.getOperand(0);
2088 // Quick check, see if the immediate LHS matches...
2089 if (F.shouldApply(LHS))
2090 return F.apply(Root);
2092 // Otherwise, if the LHS is not of the same opcode as the root, return.
2093 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2094 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
2095 // Should we apply this transform to the RHS?
2096 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
2098 // If not to the RHS, check to see if we should apply to the LHS...
2099 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
2100 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
2104 // If the functor wants to apply the optimization to the RHS of LHSI,
2105 // reassociate the expression from ((? op A) op B) to (? op (A op B))
2107 BasicBlock *BB = Root.getParent();
2109 // Now all of the instructions are in the current basic block, go ahead
2110 // and perform the reassociation.
2111 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
2113 // First move the selected RHS to the LHS of the root...
2114 Root.setOperand(0, LHSI->getOperand(1));
2116 // Make what used to be the LHS of the root be the user of the root...
2117 Value *ExtraOperand = TmpLHSI->getOperand(1);
2118 if (&Root == TmpLHSI) {
2119 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
2122 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
2123 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
2124 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
2125 BasicBlock::iterator ARI = &Root; ++ARI;
2126 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
2129 // Now propagate the ExtraOperand down the chain of instructions until we
2131 while (TmpLHSI != LHSI) {
2132 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
2133 // Move the instruction to immediately before the chain we are
2134 // constructing to avoid breaking dominance properties.
2135 NextLHSI->getParent()->getInstList().remove(NextLHSI);
2136 BB->getInstList().insert(ARI, NextLHSI);
2139 Value *NextOp = NextLHSI->getOperand(1);
2140 NextLHSI->setOperand(1, ExtraOperand);
2142 ExtraOperand = NextOp;
2145 // Now that the instructions are reassociated, have the functor perform
2146 // the transformation...
2147 return F.apply(Root);
2150 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
2157 // AddRHS - Implements: X + X --> X << 1
2160 AddRHS(Value *rhs) : RHS(rhs) {}
2161 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2162 Instruction *apply(BinaryOperator &Add) const {
2163 return BinaryOperator::createShl(Add.getOperand(0),
2164 ConstantInt::get(Add.getType(), 1));
2168 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
2170 struct AddMaskingAnd {
2172 AddMaskingAnd(Constant *c) : C2(c) {}
2173 bool shouldApply(Value *LHS) const {
2175 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
2176 ConstantExpr::getAnd(C1, C2)->isNullValue();
2178 Instruction *apply(BinaryOperator &Add) const {
2179 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
2185 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
2187 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
2188 if (Constant *SOC = dyn_cast<Constant>(SO))
2189 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
2191 return IC->InsertNewInstBefore(CastInst::create(
2192 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
2195 // Figure out if the constant is the left or the right argument.
2196 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
2197 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
2199 if (Constant *SOC = dyn_cast<Constant>(SO)) {
2201 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
2202 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
2205 Value *Op0 = SO, *Op1 = ConstOperand;
2207 std::swap(Op0, Op1);
2209 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2210 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
2211 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2212 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
2213 SO->getName()+".cmp");
2215 assert(0 && "Unknown binary instruction type!");
2218 return IC->InsertNewInstBefore(New, I);
2221 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
2222 // constant as the other operand, try to fold the binary operator into the
2223 // select arguments. This also works for Cast instructions, which obviously do
2224 // not have a second operand.
2225 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
2227 // Don't modify shared select instructions
2228 if (!SI->hasOneUse()) return 0;
2229 Value *TV = SI->getOperand(1);
2230 Value *FV = SI->getOperand(2);
2232 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2233 // Bool selects with constant operands can be folded to logical ops.
2234 if (SI->getType() == Type::Int1Ty) return 0;
2236 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2237 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2239 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
2246 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
2247 /// node as operand #0, see if we can fold the instruction into the PHI (which
2248 /// is only possible if all operands to the PHI are constants).
2249 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
2250 PHINode *PN = cast<PHINode>(I.getOperand(0));
2251 unsigned NumPHIValues = PN->getNumIncomingValues();
2252 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
2254 // Check to see if all of the operands of the PHI are constants. If there is
2255 // one non-constant value, remember the BB it is. If there is more than one
2256 // or if *it* is a PHI, bail out.
2257 BasicBlock *NonConstBB = 0;
2258 for (unsigned i = 0; i != NumPHIValues; ++i)
2259 if (!isa<Constant>(PN->getIncomingValue(i))) {
2260 if (NonConstBB) return 0; // More than one non-const value.
2261 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2262 NonConstBB = PN->getIncomingBlock(i);
2264 // If the incoming non-constant value is in I's block, we have an infinite
2266 if (NonConstBB == I.getParent())
2270 // If there is exactly one non-constant value, we can insert a copy of the
2271 // operation in that block. However, if this is a critical edge, we would be
2272 // inserting the computation one some other paths (e.g. inside a loop). Only
2273 // do this if the pred block is unconditionally branching into the phi block.
2275 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2276 if (!BI || !BI->isUnconditional()) return 0;
2279 // Okay, we can do the transformation: create the new PHI node.
2280 PHINode *NewPN = PHINode::Create(I.getType(), "");
2281 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2282 InsertNewInstBefore(NewPN, *PN);
2283 NewPN->takeName(PN);
2285 // Next, add all of the operands to the PHI.
2286 if (I.getNumOperands() == 2) {
2287 Constant *C = cast<Constant>(I.getOperand(1));
2288 for (unsigned i = 0; i != NumPHIValues; ++i) {
2290 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2291 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2292 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2294 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2296 assert(PN->getIncomingBlock(i) == NonConstBB);
2297 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2298 InV = BinaryOperator::create(BO->getOpcode(),
2299 PN->getIncomingValue(i), C, "phitmp",
2300 NonConstBB->getTerminator());
2301 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2302 InV = CmpInst::create(CI->getOpcode(),
2304 PN->getIncomingValue(i), C, "phitmp",
2305 NonConstBB->getTerminator());
2307 assert(0 && "Unknown binop!");
2309 AddToWorkList(cast<Instruction>(InV));
2311 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2314 CastInst *CI = cast<CastInst>(&I);
2315 const Type *RetTy = CI->getType();
2316 for (unsigned i = 0; i != NumPHIValues; ++i) {
2318 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2319 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2321 assert(PN->getIncomingBlock(i) == NonConstBB);
2322 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
2323 I.getType(), "phitmp",
2324 NonConstBB->getTerminator());
2325 AddToWorkList(cast<Instruction>(InV));
2327 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2330 return ReplaceInstUsesWith(I, NewPN);
2334 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
2335 /// value is never equal to -0.0.
2337 /// Note that this function will need to be revisited when we support nondefault
2340 static bool CannotBeNegativeZero(const Value *V) {
2341 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2342 return !CFP->getValueAPF().isNegZero();
2344 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2345 if (const Instruction *I = dyn_cast<Instruction>(V)) {
2346 if (I->getOpcode() == Instruction::Add &&
2347 isa<ConstantFP>(I->getOperand(1)) &&
2348 cast<ConstantFP>(I->getOperand(1))->isNullValue())
2351 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2352 if (II->getIntrinsicID() == Intrinsic::sqrt)
2353 return CannotBeNegativeZero(II->getOperand(1));
2355 if (const CallInst *CI = dyn_cast<CallInst>(I))
2356 if (const Function *F = CI->getCalledFunction()) {
2357 if (F->isDeclaration()) {
2358 switch (F->getNameLen()) {
2359 case 3: // abs(x) != -0.0
2360 if (!strcmp(F->getNameStart(), "abs")) return true;
2362 case 4: // abs[lf](x) != -0.0
2363 if (!strcmp(F->getNameStart(), "absf")) return true;
2364 if (!strcmp(F->getNameStart(), "absl")) return true;
2375 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2376 bool Changed = SimplifyCommutative(I);
2377 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2379 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2380 // X + undef -> undef
2381 if (isa<UndefValue>(RHS))
2382 return ReplaceInstUsesWith(I, RHS);
2385 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2386 if (RHSC->isNullValue())
2387 return ReplaceInstUsesWith(I, LHS);
2388 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2389 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2390 (I.getType())->getValueAPF()))
2391 return ReplaceInstUsesWith(I, LHS);
2394 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2395 // X + (signbit) --> X ^ signbit
2396 const APInt& Val = CI->getValue();
2397 uint32_t BitWidth = Val.getBitWidth();
2398 if (Val == APInt::getSignBit(BitWidth))
2399 return BinaryOperator::createXor(LHS, RHS);
2401 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2402 // (X & 254)+1 -> (X&254)|1
2403 if (!isa<VectorType>(I.getType())) {
2404 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2405 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2406 KnownZero, KnownOne))
2411 if (isa<PHINode>(LHS))
2412 if (Instruction *NV = FoldOpIntoPhi(I))
2415 ConstantInt *XorRHS = 0;
2417 if (isa<ConstantInt>(RHSC) &&
2418 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2419 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2420 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2422 uint32_t Size = TySizeBits / 2;
2423 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2424 APInt CFF80Val(-C0080Val);
2426 if (TySizeBits > Size) {
2427 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2428 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2429 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2430 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2431 // This is a sign extend if the top bits are known zero.
2432 if (!MaskedValueIsZero(XorLHS,
2433 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2434 Size = 0; // Not a sign ext, but can't be any others either.
2439 C0080Val = APIntOps::lshr(C0080Val, Size);
2440 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2441 } while (Size >= 1);
2443 // FIXME: This shouldn't be necessary. When the backends can handle types
2444 // with funny bit widths then this whole cascade of if statements should
2445 // be removed. It is just here to get the size of the "middle" type back
2446 // up to something that the back ends can handle.
2447 const Type *MiddleType = 0;
2450 case 32: MiddleType = Type::Int32Ty; break;
2451 case 16: MiddleType = Type::Int16Ty; break;
2452 case 8: MiddleType = Type::Int8Ty; break;
2455 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2456 InsertNewInstBefore(NewTrunc, I);
2457 return new SExtInst(NewTrunc, I.getType(), I.getName());
2463 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2464 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2466 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2467 if (RHSI->getOpcode() == Instruction::Sub)
2468 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2469 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2471 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2472 if (LHSI->getOpcode() == Instruction::Sub)
2473 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2474 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2479 // -A + -B --> -(A + B)
2480 if (Value *LHSV = dyn_castNegVal(LHS)) {
2481 if (LHS->getType()->isIntOrIntVector()) {
2482 if (Value *RHSV = dyn_castNegVal(RHS)) {
2483 Instruction *NewAdd = BinaryOperator::createAdd(LHSV, RHSV, "sum");
2484 InsertNewInstBefore(NewAdd, I);
2485 return BinaryOperator::createNeg(NewAdd);
2489 return BinaryOperator::createSub(RHS, LHSV);
2493 if (!isa<Constant>(RHS))
2494 if (Value *V = dyn_castNegVal(RHS))
2495 return BinaryOperator::createSub(LHS, V);
2499 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2500 if (X == RHS) // X*C + X --> X * (C+1)
2501 return BinaryOperator::createMul(RHS, AddOne(C2));
2503 // X*C1 + X*C2 --> X * (C1+C2)
2505 if (X == dyn_castFoldableMul(RHS, C1))
2506 return BinaryOperator::createMul(X, Add(C1, C2));
2509 // X + X*C --> X * (C+1)
2510 if (dyn_castFoldableMul(RHS, C2) == LHS)
2511 return BinaryOperator::createMul(LHS, AddOne(C2));
2513 // X + ~X --> -1 since ~X = -X-1
2514 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2515 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2518 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2519 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2520 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2523 // W*X + Y*Z --> W * (X+Z) iff W == Y
2524 if (I.getType()->isIntOrIntVector()) {
2525 Value *W, *X, *Y, *Z;
2526 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2527 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2531 } else if (Y == X) {
2533 } else if (X == Z) {
2540 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2541 LHS->getName()), I);
2542 return BinaryOperator::createMul(W, NewAdd);
2547 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2549 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2550 return BinaryOperator::createSub(SubOne(CRHS), X);
2552 // (X & FF00) + xx00 -> (X+xx00) & FF00
2553 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2554 Constant *Anded = And(CRHS, C2);
2555 if (Anded == CRHS) {
2556 // See if all bits from the first bit set in the Add RHS up are included
2557 // in the mask. First, get the rightmost bit.
2558 const APInt& AddRHSV = CRHS->getValue();
2560 // Form a mask of all bits from the lowest bit added through the top.
2561 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2563 // See if the and mask includes all of these bits.
2564 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2566 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2567 // Okay, the xform is safe. Insert the new add pronto.
2568 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2569 LHS->getName()), I);
2570 return BinaryOperator::createAnd(NewAdd, C2);
2575 // Try to fold constant add into select arguments.
2576 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2577 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2581 // add (cast *A to intptrtype) B ->
2582 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2584 CastInst *CI = dyn_cast<CastInst>(LHS);
2587 CI = dyn_cast<CastInst>(RHS);
2590 if (CI && CI->getType()->isSized() &&
2591 (CI->getType()->getPrimitiveSizeInBits() ==
2592 TD->getIntPtrType()->getPrimitiveSizeInBits())
2593 && isa<PointerType>(CI->getOperand(0)->getType())) {
2595 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2596 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2597 PointerType::get(Type::Int8Ty, AS), I);
2598 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2599 return new PtrToIntInst(I2, CI->getType());
2603 // add (select X 0 (sub n A)) A --> select X A n
2605 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2608 SI = dyn_cast<SelectInst>(RHS);
2611 if (SI && SI->hasOneUse()) {
2612 Value *TV = SI->getTrueValue();
2613 Value *FV = SI->getFalseValue();
2616 // Can we fold the add into the argument of the select?
2617 // We check both true and false select arguments for a matching subtract.
2618 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2619 A == Other) // Fold the add into the true select value.
2620 return SelectInst::Create(SI->getCondition(), N, A);
2621 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2622 A == Other) // Fold the add into the false select value.
2623 return SelectInst::Create(SI->getCondition(), A, N);
2627 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2628 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2629 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2630 return ReplaceInstUsesWith(I, LHS);
2632 return Changed ? &I : 0;
2635 // isSignBit - Return true if the value represented by the constant only has the
2636 // highest order bit set.
2637 static bool isSignBit(ConstantInt *CI) {
2638 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2639 return CI->getValue() == APInt::getSignBit(NumBits);
2642 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2643 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2645 if (Op0 == Op1) // sub X, X -> 0
2646 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2648 // If this is a 'B = x-(-A)', change to B = x+A...
2649 if (Value *V = dyn_castNegVal(Op1))
2650 return BinaryOperator::createAdd(Op0, V);
2652 if (isa<UndefValue>(Op0))
2653 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2654 if (isa<UndefValue>(Op1))
2655 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2657 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2658 // Replace (-1 - A) with (~A)...
2659 if (C->isAllOnesValue())
2660 return BinaryOperator::createNot(Op1);
2662 // C - ~X == X + (1+C)
2664 if (match(Op1, m_Not(m_Value(X))))
2665 return BinaryOperator::createAdd(X, AddOne(C));
2667 // -(X >>u 31) -> (X >>s 31)
2668 // -(X >>s 31) -> (X >>u 31)
2670 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2671 if (SI->getOpcode() == Instruction::LShr) {
2672 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2673 // Check to see if we are shifting out everything but the sign bit.
2674 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2675 SI->getType()->getPrimitiveSizeInBits()-1) {
2676 // Ok, the transformation is safe. Insert AShr.
2677 return BinaryOperator::create(Instruction::AShr,
2678 SI->getOperand(0), CU, SI->getName());
2682 else if (SI->getOpcode() == Instruction::AShr) {
2683 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2684 // Check to see if we are shifting out everything but the sign bit.
2685 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2686 SI->getType()->getPrimitiveSizeInBits()-1) {
2687 // Ok, the transformation is safe. Insert LShr.
2688 return BinaryOperator::createLShr(
2689 SI->getOperand(0), CU, SI->getName());
2696 // Try to fold constant sub into select arguments.
2697 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2698 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2701 if (isa<PHINode>(Op0))
2702 if (Instruction *NV = FoldOpIntoPhi(I))
2706 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2707 if (Op1I->getOpcode() == Instruction::Add &&
2708 !Op0->getType()->isFPOrFPVector()) {
2709 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2710 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2711 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2712 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2713 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2714 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2715 // C1-(X+C2) --> (C1-C2)-X
2716 return BinaryOperator::createSub(Subtract(CI1, CI2),
2717 Op1I->getOperand(0));
2721 if (Op1I->hasOneUse()) {
2722 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2723 // is not used by anyone else...
2725 if (Op1I->getOpcode() == Instruction::Sub &&
2726 !Op1I->getType()->isFPOrFPVector()) {
2727 // Swap the two operands of the subexpr...
2728 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2729 Op1I->setOperand(0, IIOp1);
2730 Op1I->setOperand(1, IIOp0);
2732 // Create the new top level add instruction...
2733 return BinaryOperator::createAdd(Op0, Op1);
2736 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2738 if (Op1I->getOpcode() == Instruction::And &&
2739 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2740 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2743 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2744 return BinaryOperator::createAnd(Op0, NewNot);
2747 // 0 - (X sdiv C) -> (X sdiv -C)
2748 if (Op1I->getOpcode() == Instruction::SDiv)
2749 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2751 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2752 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2753 ConstantExpr::getNeg(DivRHS));
2755 // X - X*C --> X * (1-C)
2756 ConstantInt *C2 = 0;
2757 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2758 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2759 return BinaryOperator::createMul(Op0, CP1);
2762 // X - ((X / Y) * Y) --> X % Y
2763 if (Op1I->getOpcode() == Instruction::Mul)
2764 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2765 if (Op0 == I->getOperand(0) &&
2766 Op1I->getOperand(1) == I->getOperand(1)) {
2767 if (I->getOpcode() == Instruction::SDiv)
2768 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2769 if (I->getOpcode() == Instruction::UDiv)
2770 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2775 if (!Op0->getType()->isFPOrFPVector())
2776 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2777 if (Op0I->getOpcode() == Instruction::Add) {
2778 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2779 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2780 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2781 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2782 } else if (Op0I->getOpcode() == Instruction::Sub) {
2783 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2784 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2789 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2790 if (X == Op1) // X*C - X --> X * (C-1)
2791 return BinaryOperator::createMul(Op1, SubOne(C1));
2793 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2794 if (X == dyn_castFoldableMul(Op1, C2))
2795 return BinaryOperator::createMul(X, Subtract(C1, C2));
2800 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2801 /// comparison only checks the sign bit. If it only checks the sign bit, set
2802 /// TrueIfSigned if the result of the comparison is true when the input value is
2804 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2805 bool &TrueIfSigned) {
2807 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2808 TrueIfSigned = true;
2809 return RHS->isZero();
2810 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2811 TrueIfSigned = true;
2812 return RHS->isAllOnesValue();
2813 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2814 TrueIfSigned = false;
2815 return RHS->isAllOnesValue();
2816 case ICmpInst::ICMP_UGT:
2817 // True if LHS u> RHS and RHS == high-bit-mask - 1
2818 TrueIfSigned = true;
2819 return RHS->getValue() ==
2820 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2821 case ICmpInst::ICMP_UGE:
2822 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2823 TrueIfSigned = true;
2824 return RHS->getValue() ==
2825 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2831 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2832 bool Changed = SimplifyCommutative(I);
2833 Value *Op0 = I.getOperand(0);
2835 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2836 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2838 // Simplify mul instructions with a constant RHS...
2839 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2840 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2842 // ((X << C1)*C2) == (X * (C2 << C1))
2843 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2844 if (SI->getOpcode() == Instruction::Shl)
2845 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2846 return BinaryOperator::createMul(SI->getOperand(0),
2847 ConstantExpr::getShl(CI, ShOp));
2850 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2851 if (CI->equalsInt(1)) // X * 1 == X
2852 return ReplaceInstUsesWith(I, Op0);
2853 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2854 return BinaryOperator::createNeg(Op0, I.getName());
2856 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2857 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2858 return BinaryOperator::createShl(Op0,
2859 ConstantInt::get(Op0->getType(), Val.logBase2()));
2861 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2862 if (Op1F->isNullValue())
2863 return ReplaceInstUsesWith(I, Op1);
2865 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2866 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2867 // We need a better interface for long double here.
2868 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2869 if (Op1F->isExactlyValue(1.0))
2870 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2873 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2874 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2875 isa<ConstantInt>(Op0I->getOperand(1))) {
2876 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2877 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2879 InsertNewInstBefore(Add, I);
2880 Value *C1C2 = ConstantExpr::getMul(Op1,
2881 cast<Constant>(Op0I->getOperand(1)));
2882 return BinaryOperator::createAdd(Add, C1C2);
2886 // Try to fold constant mul into select arguments.
2887 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2888 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2891 if (isa<PHINode>(Op0))
2892 if (Instruction *NV = FoldOpIntoPhi(I))
2896 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2897 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2898 return BinaryOperator::createMul(Op0v, Op1v);
2900 // If one of the operands of the multiply is a cast from a boolean value, then
2901 // we know the bool is either zero or one, so this is a 'masking' multiply.
2902 // See if we can simplify things based on how the boolean was originally
2904 CastInst *BoolCast = 0;
2905 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2906 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2909 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2910 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2913 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2914 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2915 const Type *SCOpTy = SCIOp0->getType();
2918 // If the icmp is true iff the sign bit of X is set, then convert this
2919 // multiply into a shift/and combination.
2920 if (isa<ConstantInt>(SCIOp1) &&
2921 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2923 // Shift the X value right to turn it into "all signbits".
2924 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2925 SCOpTy->getPrimitiveSizeInBits()-1);
2927 InsertNewInstBefore(
2928 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2929 BoolCast->getOperand(0)->getName()+
2932 // If the multiply type is not the same as the source type, sign extend
2933 // or truncate to the multiply type.
2934 if (I.getType() != V->getType()) {
2935 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2936 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2937 Instruction::CastOps opcode =
2938 (SrcBits == DstBits ? Instruction::BitCast :
2939 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2940 V = InsertCastBefore(opcode, V, I.getType(), I);
2943 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2944 return BinaryOperator::createAnd(V, OtherOp);
2949 return Changed ? &I : 0;
2952 /// This function implements the transforms on div instructions that work
2953 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2954 /// used by the visitors to those instructions.
2955 /// @brief Transforms common to all three div instructions
2956 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2957 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2959 // undef / X -> 0 for integer.
2960 // undef / X -> undef for FP (the undef could be a snan).
2961 if (isa<UndefValue>(Op0)) {
2962 if (Op0->getType()->isFPOrFPVector())
2963 return ReplaceInstUsesWith(I, Op0);
2964 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2967 // X / undef -> undef
2968 if (isa<UndefValue>(Op1))
2969 return ReplaceInstUsesWith(I, Op1);
2971 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2972 // This does not apply for fdiv.
2973 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2974 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2975 // the same basic block, then we replace the select with Y, and the
2976 // condition of the select with false (if the cond value is in the same BB).
2977 // If the select has uses other than the div, this allows them to be
2978 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2979 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2980 if (ST->isNullValue()) {
2981 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2982 if (CondI && CondI->getParent() == I.getParent())
2983 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2984 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2985 I.setOperand(1, SI->getOperand(2));
2987 UpdateValueUsesWith(SI, SI->getOperand(2));
2991 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2992 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2993 if (ST->isNullValue()) {
2994 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2995 if (CondI && CondI->getParent() == I.getParent())
2996 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2997 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2998 I.setOperand(1, SI->getOperand(1));
3000 UpdateValueUsesWith(SI, SI->getOperand(1));
3008 /// This function implements the transforms common to both integer division
3009 /// instructions (udiv and sdiv). It is called by the visitors to those integer
3010 /// division instructions.
3011 /// @brief Common integer divide transforms
3012 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
3013 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3015 if (Instruction *Common = commonDivTransforms(I))
3018 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3020 if (RHS->equalsInt(1))
3021 return ReplaceInstUsesWith(I, Op0);
3023 // (X / C1) / C2 -> X / (C1*C2)
3024 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
3025 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
3026 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
3027 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
3028 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3030 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
3031 Multiply(RHS, LHSRHS));
3034 if (!RHS->isZero()) { // avoid X udiv 0
3035 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3036 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3038 if (isa<PHINode>(Op0))
3039 if (Instruction *NV = FoldOpIntoPhi(I))
3044 // 0 / X == 0, we don't need to preserve faults!
3045 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
3046 if (LHS->equalsInt(0))
3047 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3052 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
3053 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3055 // Handle the integer div common cases
3056 if (Instruction *Common = commonIDivTransforms(I))
3059 // X udiv C^2 -> X >> C
3060 // Check to see if this is an unsigned division with an exact power of 2,
3061 // if so, convert to a right shift.
3062 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3063 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3064 return BinaryOperator::createLShr(Op0,
3065 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3068 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3069 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3070 if (RHSI->getOpcode() == Instruction::Shl &&
3071 isa<ConstantInt>(RHSI->getOperand(0))) {
3072 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3073 if (C1.isPowerOf2()) {
3074 Value *N = RHSI->getOperand(1);
3075 const Type *NTy = N->getType();
3076 if (uint32_t C2 = C1.logBase2()) {
3077 Constant *C2V = ConstantInt::get(NTy, C2);
3078 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
3080 return BinaryOperator::createLShr(Op0, N);
3085 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3086 // where C1&C2 are powers of two.
3087 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3088 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3089 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3090 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3091 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3092 // Compute the shift amounts
3093 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3094 // Construct the "on true" case of the select
3095 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3096 Instruction *TSI = BinaryOperator::createLShr(
3097 Op0, TC, SI->getName()+".t");
3098 TSI = InsertNewInstBefore(TSI, I);
3100 // Construct the "on false" case of the select
3101 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3102 Instruction *FSI = BinaryOperator::createLShr(
3103 Op0, FC, SI->getName()+".f");
3104 FSI = InsertNewInstBefore(FSI, I);
3106 // construct the select instruction and return it.
3107 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3113 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3114 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3116 // Handle the integer div common cases
3117 if (Instruction *Common = commonIDivTransforms(I))
3120 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3122 if (RHS->isAllOnesValue())
3123 return BinaryOperator::createNeg(Op0);
3126 if (Value *LHSNeg = dyn_castNegVal(Op0))
3127 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
3130 // If the sign bits of both operands are zero (i.e. we can prove they are
3131 // unsigned inputs), turn this into a udiv.
3132 if (I.getType()->isInteger()) {
3133 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3134 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3135 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3136 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
3143 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3144 return commonDivTransforms(I);
3147 /// This function implements the transforms on rem instructions that work
3148 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3149 /// is used by the visitors to those instructions.
3150 /// @brief Transforms common to all three rem instructions
3151 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3152 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3154 // 0 % X == 0 for integer, we don't need to preserve faults!
3155 if (Constant *LHS = dyn_cast<Constant>(Op0))
3156 if (LHS->isNullValue())
3157 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3159 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3160 if (I.getType()->isFPOrFPVector())
3161 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3162 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3164 if (isa<UndefValue>(Op1))
3165 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3167 // Handle cases involving: rem X, (select Cond, Y, Z)
3168 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3169 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
3170 // the same basic block, then we replace the select with Y, and the
3171 // condition of the select with false (if the cond value is in the same
3172 // BB). If the select has uses other than the div, this allows them to be
3174 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3175 if (ST->isNullValue()) {
3176 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3177 if (CondI && CondI->getParent() == I.getParent())
3178 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3179 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3180 I.setOperand(1, SI->getOperand(2));
3182 UpdateValueUsesWith(SI, SI->getOperand(2));
3185 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
3186 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3187 if (ST->isNullValue()) {
3188 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3189 if (CondI && CondI->getParent() == I.getParent())
3190 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3191 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3192 I.setOperand(1, SI->getOperand(1));
3194 UpdateValueUsesWith(SI, SI->getOperand(1));
3202 /// This function implements the transforms common to both integer remainder
3203 /// instructions (urem and srem). It is called by the visitors to those integer
3204 /// remainder instructions.
3205 /// @brief Common integer remainder transforms
3206 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3207 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3209 if (Instruction *common = commonRemTransforms(I))
3212 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3213 // X % 0 == undef, we don't need to preserve faults!
3214 if (RHS->equalsInt(0))
3215 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3217 if (RHS->equalsInt(1)) // X % 1 == 0
3218 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3220 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3221 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3222 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3224 } else if (isa<PHINode>(Op0I)) {
3225 if (Instruction *NV = FoldOpIntoPhi(I))
3229 // See if we can fold away this rem instruction.
3230 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3231 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3232 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3233 KnownZero, KnownOne))
3241 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3242 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3244 if (Instruction *common = commonIRemTransforms(I))
3247 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3248 // X urem C^2 -> X and C
3249 // Check to see if this is an unsigned remainder with an exact power of 2,
3250 // if so, convert to a bitwise and.
3251 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3252 if (C->getValue().isPowerOf2())
3253 return BinaryOperator::createAnd(Op0, SubOne(C));
3256 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3257 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3258 if (RHSI->getOpcode() == Instruction::Shl &&
3259 isa<ConstantInt>(RHSI->getOperand(0))) {
3260 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3261 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3262 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
3264 return BinaryOperator::createAnd(Op0, Add);
3269 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3270 // where C1&C2 are powers of two.
3271 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3272 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3273 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3274 // STO == 0 and SFO == 0 handled above.
3275 if ((STO->getValue().isPowerOf2()) &&
3276 (SFO->getValue().isPowerOf2())) {
3277 Value *TrueAnd = InsertNewInstBefore(
3278 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3279 Value *FalseAnd = InsertNewInstBefore(
3280 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3281 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3289 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3290 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3292 // Handle the integer rem common cases
3293 if (Instruction *common = commonIRemTransforms(I))
3296 if (Value *RHSNeg = dyn_castNegVal(Op1))
3297 if (!isa<ConstantInt>(RHSNeg) ||
3298 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
3300 AddUsesToWorkList(I);
3301 I.setOperand(1, RHSNeg);
3305 // If the sign bits of both operands are zero (i.e. we can prove they are
3306 // unsigned inputs), turn this into a urem.
3307 if (I.getType()->isInteger()) {
3308 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3309 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3310 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3311 return BinaryOperator::createURem(Op0, Op1, I.getName());
3318 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3319 return commonRemTransforms(I);
3322 // isMaxValueMinusOne - return true if this is Max-1
3323 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
3324 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3326 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
3327 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
3330 // isMinValuePlusOne - return true if this is Min+1
3331 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3333 return C->getValue() == 1; // unsigned
3335 // Calculate 1111111111000000000000
3336 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3337 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
3340 // isOneBitSet - Return true if there is exactly one bit set in the specified
3342 static bool isOneBitSet(const ConstantInt *CI) {
3343 return CI->getValue().isPowerOf2();
3346 // isHighOnes - Return true if the constant is of the form 1+0+.
3347 // This is the same as lowones(~X).
3348 static bool isHighOnes(const ConstantInt *CI) {
3349 return (~CI->getValue() + 1).isPowerOf2();
3352 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3353 /// are carefully arranged to allow folding of expressions such as:
3355 /// (A < B) | (A > B) --> (A != B)
3357 /// Note that this is only valid if the first and second predicates have the
3358 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3360 /// Three bits are used to represent the condition, as follows:
3365 /// <=> Value Definition
3366 /// 000 0 Always false
3373 /// 111 7 Always true
3375 static unsigned getICmpCode(const ICmpInst *ICI) {
3376 switch (ICI->getPredicate()) {
3378 case ICmpInst::ICMP_UGT: return 1; // 001
3379 case ICmpInst::ICMP_SGT: return 1; // 001
3380 case ICmpInst::ICMP_EQ: return 2; // 010
3381 case ICmpInst::ICMP_UGE: return 3; // 011
3382 case ICmpInst::ICMP_SGE: return 3; // 011
3383 case ICmpInst::ICMP_ULT: return 4; // 100
3384 case ICmpInst::ICMP_SLT: return 4; // 100
3385 case ICmpInst::ICMP_NE: return 5; // 101
3386 case ICmpInst::ICMP_ULE: return 6; // 110
3387 case ICmpInst::ICMP_SLE: return 6; // 110
3390 assert(0 && "Invalid ICmp predicate!");
3395 /// getICmpValue - This is the complement of getICmpCode, which turns an
3396 /// opcode and two operands into either a constant true or false, or a brand
3397 /// new ICmp instruction. The sign is passed in to determine which kind
3398 /// of predicate to use in new icmp instructions.
3399 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3401 default: assert(0 && "Illegal ICmp code!");
3402 case 0: return ConstantInt::getFalse();
3405 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3407 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3408 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3411 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3413 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3416 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3418 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3419 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3422 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3424 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3425 case 7: return ConstantInt::getTrue();
3429 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3430 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3431 (ICmpInst::isSignedPredicate(p1) &&
3432 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3433 (ICmpInst::isSignedPredicate(p2) &&
3434 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3438 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3439 struct FoldICmpLogical {
3442 ICmpInst::Predicate pred;
3443 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3444 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3445 pred(ICI->getPredicate()) {}
3446 bool shouldApply(Value *V) const {
3447 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3448 if (PredicatesFoldable(pred, ICI->getPredicate()))
3449 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3450 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3453 Instruction *apply(Instruction &Log) const {
3454 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3455 if (ICI->getOperand(0) != LHS) {
3456 assert(ICI->getOperand(1) == LHS);
3457 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3460 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3461 unsigned LHSCode = getICmpCode(ICI);
3462 unsigned RHSCode = getICmpCode(RHSICI);
3464 switch (Log.getOpcode()) {
3465 case Instruction::And: Code = LHSCode & RHSCode; break;
3466 case Instruction::Or: Code = LHSCode | RHSCode; break;
3467 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3468 default: assert(0 && "Illegal logical opcode!"); return 0;
3471 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3472 ICmpInst::isSignedPredicate(ICI->getPredicate());
3474 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3475 if (Instruction *I = dyn_cast<Instruction>(RV))
3477 // Otherwise, it's a constant boolean value...
3478 return IC.ReplaceInstUsesWith(Log, RV);
3481 } // end anonymous namespace
3483 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3484 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3485 // guaranteed to be a binary operator.
3486 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3488 ConstantInt *AndRHS,
3489 BinaryOperator &TheAnd) {
3490 Value *X = Op->getOperand(0);
3491 Constant *Together = 0;
3493 Together = And(AndRHS, OpRHS);
3495 switch (Op->getOpcode()) {
3496 case Instruction::Xor:
3497 if (Op->hasOneUse()) {
3498 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3499 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3500 InsertNewInstBefore(And, TheAnd);
3502 return BinaryOperator::createXor(And, Together);
3505 case Instruction::Or:
3506 if (Together == AndRHS) // (X | C) & C --> C
3507 return ReplaceInstUsesWith(TheAnd, AndRHS);
3509 if (Op->hasOneUse() && Together != OpRHS) {
3510 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3511 Instruction *Or = BinaryOperator::createOr(X, Together);
3512 InsertNewInstBefore(Or, TheAnd);
3514 return BinaryOperator::createAnd(Or, AndRHS);
3517 case Instruction::Add:
3518 if (Op->hasOneUse()) {
3519 // Adding a one to a single bit bit-field should be turned into an XOR
3520 // of the bit. First thing to check is to see if this AND is with a
3521 // single bit constant.
3522 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3524 // If there is only one bit set...
3525 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3526 // Ok, at this point, we know that we are masking the result of the
3527 // ADD down to exactly one bit. If the constant we are adding has
3528 // no bits set below this bit, then we can eliminate the ADD.
3529 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3531 // Check to see if any bits below the one bit set in AndRHSV are set.
3532 if ((AddRHS & (AndRHSV-1)) == 0) {
3533 // If not, the only thing that can effect the output of the AND is
3534 // the bit specified by AndRHSV. If that bit is set, the effect of
3535 // the XOR is to toggle the bit. If it is clear, then the ADD has
3537 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3538 TheAnd.setOperand(0, X);
3541 // Pull the XOR out of the AND.
3542 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3543 InsertNewInstBefore(NewAnd, TheAnd);
3544 NewAnd->takeName(Op);
3545 return BinaryOperator::createXor(NewAnd, AndRHS);
3552 case Instruction::Shl: {
3553 // We know that the AND will not produce any of the bits shifted in, so if
3554 // the anded constant includes them, clear them now!
3556 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3557 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3558 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3559 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3561 if (CI->getValue() == ShlMask) {
3562 // Masking out bits that the shift already masks
3563 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3564 } else if (CI != AndRHS) { // Reducing bits set in and.
3565 TheAnd.setOperand(1, CI);
3570 case Instruction::LShr:
3572 // We know that the AND will not produce any of the bits shifted in, so if
3573 // the anded constant includes them, clear them now! This only applies to
3574 // unsigned shifts, because a signed shr may bring in set bits!
3576 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3577 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3578 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3579 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3581 if (CI->getValue() == ShrMask) {
3582 // Masking out bits that the shift already masks.
3583 return ReplaceInstUsesWith(TheAnd, Op);
3584 } else if (CI != AndRHS) {
3585 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3590 case Instruction::AShr:
3592 // See if this is shifting in some sign extension, then masking it out
3594 if (Op->hasOneUse()) {
3595 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3596 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3597 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3598 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3599 if (C == AndRHS) { // Masking out bits shifted in.
3600 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3601 // Make the argument unsigned.
3602 Value *ShVal = Op->getOperand(0);
3603 ShVal = InsertNewInstBefore(
3604 BinaryOperator::createLShr(ShVal, OpRHS,
3605 Op->getName()), TheAnd);
3606 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3615 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3616 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3617 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3618 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3619 /// insert new instructions.
3620 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3621 bool isSigned, bool Inside,
3623 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3624 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3625 "Lo is not <= Hi in range emission code!");
3628 if (Lo == Hi) // Trivially false.
3629 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3631 // V >= Min && V < Hi --> V < Hi
3632 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3633 ICmpInst::Predicate pred = (isSigned ?
3634 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3635 return new ICmpInst(pred, V, Hi);
3638 // Emit V-Lo <u Hi-Lo
3639 Constant *NegLo = ConstantExpr::getNeg(Lo);
3640 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3641 InsertNewInstBefore(Add, IB);
3642 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3643 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3646 if (Lo == Hi) // Trivially true.
3647 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3649 // V < Min || V >= Hi -> V > Hi-1
3650 Hi = SubOne(cast<ConstantInt>(Hi));
3651 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3652 ICmpInst::Predicate pred = (isSigned ?
3653 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3654 return new ICmpInst(pred, V, Hi);
3657 // Emit V-Lo >u Hi-1-Lo
3658 // Note that Hi has already had one subtracted from it, above.
3659 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3660 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3661 InsertNewInstBefore(Add, IB);
3662 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3663 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3666 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3667 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3668 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3669 // not, since all 1s are not contiguous.
3670 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3671 const APInt& V = Val->getValue();
3672 uint32_t BitWidth = Val->getType()->getBitWidth();
3673 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3675 // look for the first zero bit after the run of ones
3676 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3677 // look for the first non-zero bit
3678 ME = V.getActiveBits();
3682 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3683 /// where isSub determines whether the operator is a sub. If we can fold one of
3684 /// the following xforms:
3686 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3687 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3688 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3690 /// return (A +/- B).
3692 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3693 ConstantInt *Mask, bool isSub,
3695 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3696 if (!LHSI || LHSI->getNumOperands() != 2 ||
3697 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3699 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3701 switch (LHSI->getOpcode()) {
3703 case Instruction::And:
3704 if (And(N, Mask) == Mask) {
3705 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3706 if ((Mask->getValue().countLeadingZeros() +
3707 Mask->getValue().countPopulation()) ==
3708 Mask->getValue().getBitWidth())
3711 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3712 // part, we don't need any explicit masks to take them out of A. If that
3713 // is all N is, ignore it.
3714 uint32_t MB = 0, ME = 0;
3715 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3716 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3717 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3718 if (MaskedValueIsZero(RHS, Mask))
3723 case Instruction::Or:
3724 case Instruction::Xor:
3725 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3726 if ((Mask->getValue().countLeadingZeros() +
3727 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3728 && And(N, Mask)->isZero())
3735 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3737 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3738 return InsertNewInstBefore(New, I);
3741 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3742 bool Changed = SimplifyCommutative(I);
3743 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3745 if (isa<UndefValue>(Op1)) // X & undef -> 0
3746 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3750 return ReplaceInstUsesWith(I, Op1);
3752 // See if we can simplify any instructions used by the instruction whose sole
3753 // purpose is to compute bits we don't care about.
3754 if (!isa<VectorType>(I.getType())) {
3755 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3756 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3757 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3758 KnownZero, KnownOne))
3761 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3762 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3763 return ReplaceInstUsesWith(I, I.getOperand(0));
3764 } else if (isa<ConstantAggregateZero>(Op1)) {
3765 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3769 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3770 const APInt& AndRHSMask = AndRHS->getValue();
3771 APInt NotAndRHS(~AndRHSMask);
3773 // Optimize a variety of ((val OP C1) & C2) combinations...
3774 if (isa<BinaryOperator>(Op0)) {
3775 Instruction *Op0I = cast<Instruction>(Op0);
3776 Value *Op0LHS = Op0I->getOperand(0);
3777 Value *Op0RHS = Op0I->getOperand(1);
3778 switch (Op0I->getOpcode()) {
3779 case Instruction::Xor:
3780 case Instruction::Or:
3781 // If the mask is only needed on one incoming arm, push it up.
3782 if (Op0I->hasOneUse()) {
3783 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3784 // Not masking anything out for the LHS, move to RHS.
3785 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3786 Op0RHS->getName()+".masked");
3787 InsertNewInstBefore(NewRHS, I);
3788 return BinaryOperator::create(
3789 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3791 if (!isa<Constant>(Op0RHS) &&
3792 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3793 // Not masking anything out for the RHS, move to LHS.
3794 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3795 Op0LHS->getName()+".masked");
3796 InsertNewInstBefore(NewLHS, I);
3797 return BinaryOperator::create(
3798 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3803 case Instruction::Add:
3804 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3805 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3806 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3807 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3808 return BinaryOperator::createAnd(V, AndRHS);
3809 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3810 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3813 case Instruction::Sub:
3814 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3815 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3816 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3817 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3818 return BinaryOperator::createAnd(V, AndRHS);
3822 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3823 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3825 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3826 // If this is an integer truncation or change from signed-to-unsigned, and
3827 // if the source is an and/or with immediate, transform it. This
3828 // frequently occurs for bitfield accesses.
3829 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3830 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3831 CastOp->getNumOperands() == 2)
3832 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3833 if (CastOp->getOpcode() == Instruction::And) {
3834 // Change: and (cast (and X, C1) to T), C2
3835 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3836 // This will fold the two constants together, which may allow
3837 // other simplifications.
3838 Instruction *NewCast = CastInst::createTruncOrBitCast(
3839 CastOp->getOperand(0), I.getType(),
3840 CastOp->getName()+".shrunk");
3841 NewCast = InsertNewInstBefore(NewCast, I);
3842 // trunc_or_bitcast(C1)&C2
3843 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3844 C3 = ConstantExpr::getAnd(C3, AndRHS);
3845 return BinaryOperator::createAnd(NewCast, C3);
3846 } else if (CastOp->getOpcode() == Instruction::Or) {
3847 // Change: and (cast (or X, C1) to T), C2
3848 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3849 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3850 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3851 return ReplaceInstUsesWith(I, AndRHS);
3857 // Try to fold constant and into select arguments.
3858 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3859 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3861 if (isa<PHINode>(Op0))
3862 if (Instruction *NV = FoldOpIntoPhi(I))
3866 Value *Op0NotVal = dyn_castNotVal(Op0);
3867 Value *Op1NotVal = dyn_castNotVal(Op1);
3869 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3870 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3872 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3873 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3874 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3875 I.getName()+".demorgan");
3876 InsertNewInstBefore(Or, I);
3877 return BinaryOperator::createNot(Or);
3881 Value *A = 0, *B = 0, *C = 0, *D = 0;
3882 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3883 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3884 return ReplaceInstUsesWith(I, Op1);
3886 // (A|B) & ~(A&B) -> A^B
3887 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3888 if ((A == C && B == D) || (A == D && B == C))
3889 return BinaryOperator::createXor(A, B);
3893 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3894 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3895 return ReplaceInstUsesWith(I, Op0);
3897 // ~(A&B) & (A|B) -> A^B
3898 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3899 if ((A == C && B == D) || (A == D && B == C))
3900 return BinaryOperator::createXor(A, B);
3904 if (Op0->hasOneUse() &&
3905 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3906 if (A == Op1) { // (A^B)&A -> A&(A^B)
3907 I.swapOperands(); // Simplify below
3908 std::swap(Op0, Op1);
3909 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3910 cast<BinaryOperator>(Op0)->swapOperands();
3911 I.swapOperands(); // Simplify below
3912 std::swap(Op0, Op1);
3915 if (Op1->hasOneUse() &&
3916 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3917 if (B == Op0) { // B&(A^B) -> B&(B^A)
3918 cast<BinaryOperator>(Op1)->swapOperands();
3921 if (A == Op0) { // A&(A^B) -> A & ~B
3922 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3923 InsertNewInstBefore(NotB, I);
3924 return BinaryOperator::createAnd(A, NotB);
3929 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3930 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3931 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3934 Value *LHSVal, *RHSVal;
3935 ConstantInt *LHSCst, *RHSCst;
3936 ICmpInst::Predicate LHSCC, RHSCC;
3937 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3938 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3939 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3940 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3941 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3942 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3943 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3944 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3946 // Don't try to fold ICMP_SLT + ICMP_ULT.
3947 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3948 ICmpInst::isSignedPredicate(LHSCC) ==
3949 ICmpInst::isSignedPredicate(RHSCC))) {
3950 // Ensure that the larger constant is on the RHS.
3951 ICmpInst::Predicate GT;
3952 if (ICmpInst::isSignedPredicate(LHSCC) ||
3953 (ICmpInst::isEquality(LHSCC) &&
3954 ICmpInst::isSignedPredicate(RHSCC)))
3955 GT = ICmpInst::ICMP_SGT;
3957 GT = ICmpInst::ICMP_UGT;
3959 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3960 ICmpInst *LHS = cast<ICmpInst>(Op0);
3961 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3962 std::swap(LHS, RHS);
3963 std::swap(LHSCst, RHSCst);
3964 std::swap(LHSCC, RHSCC);
3967 // At this point, we know we have have two icmp instructions
3968 // comparing a value against two constants and and'ing the result
3969 // together. Because of the above check, we know that we only have
3970 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3971 // (from the FoldICmpLogical check above), that the two constants
3972 // are not equal and that the larger constant is on the RHS
3973 assert(LHSCst != RHSCst && "Compares not folded above?");
3976 default: assert(0 && "Unknown integer condition code!");
3977 case ICmpInst::ICMP_EQ:
3979 default: assert(0 && "Unknown integer condition code!");
3980 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3981 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3982 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3983 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3984 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3985 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3986 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3987 return ReplaceInstUsesWith(I, LHS);
3989 case ICmpInst::ICMP_NE:
3991 default: assert(0 && "Unknown integer condition code!");
3992 case ICmpInst::ICMP_ULT:
3993 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3994 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3995 break; // (X != 13 & X u< 15) -> no change
3996 case ICmpInst::ICMP_SLT:
3997 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3998 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3999 break; // (X != 13 & X s< 15) -> no change
4000 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
4001 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
4002 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
4003 return ReplaceInstUsesWith(I, RHS);
4004 case ICmpInst::ICMP_NE:
4005 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
4006 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4007 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4008 LHSVal->getName()+".off");
4009 InsertNewInstBefore(Add, I);
4010 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
4011 ConstantInt::get(Add->getType(), 1));
4013 break; // (X != 13 & X != 15) -> no change
4016 case ICmpInst::ICMP_ULT:
4018 default: assert(0 && "Unknown integer condition code!");
4019 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
4020 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
4021 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4022 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
4024 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
4025 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
4026 return ReplaceInstUsesWith(I, LHS);
4027 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
4031 case ICmpInst::ICMP_SLT:
4033 default: assert(0 && "Unknown integer condition code!");
4034 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
4035 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
4036 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4037 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
4039 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
4040 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
4041 return ReplaceInstUsesWith(I, LHS);
4042 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
4046 case ICmpInst::ICMP_UGT:
4048 default: assert(0 && "Unknown integer condition code!");
4049 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
4050 return ReplaceInstUsesWith(I, LHS);
4051 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
4052 return ReplaceInstUsesWith(I, RHS);
4053 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
4055 case ICmpInst::ICMP_NE:
4056 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
4057 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4058 break; // (X u> 13 & X != 15) -> no change
4059 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
4060 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
4062 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
4066 case ICmpInst::ICMP_SGT:
4068 default: assert(0 && "Unknown integer condition code!");
4069 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
4070 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
4071 return ReplaceInstUsesWith(I, RHS);
4072 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
4074 case ICmpInst::ICMP_NE:
4075 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
4076 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4077 break; // (X s> 13 & X != 15) -> no change
4078 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
4079 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
4081 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4089 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4090 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4091 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4092 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4093 const Type *SrcTy = Op0C->getOperand(0)->getType();
4094 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4095 // Only do this if the casts both really cause code to be generated.
4096 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4098 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4100 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
4101 Op1C->getOperand(0),
4103 InsertNewInstBefore(NewOp, I);
4104 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4108 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4109 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4110 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4111 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4112 SI0->getOperand(1) == SI1->getOperand(1) &&
4113 (SI0->hasOneUse() || SI1->hasOneUse())) {
4114 Instruction *NewOp =
4115 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
4117 SI0->getName()), I);
4118 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4119 SI1->getOperand(1));
4123 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4124 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4125 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4126 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4127 RHS->getPredicate() == FCmpInst::FCMP_ORD)
4128 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4129 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4130 // If either of the constants are nans, then the whole thing returns
4132 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4133 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4134 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4135 RHS->getOperand(0));
4140 return Changed ? &I : 0;
4143 /// CollectBSwapParts - Look to see if the specified value defines a single byte
4144 /// in the result. If it does, and if the specified byte hasn't been filled in
4145 /// yet, fill it in and return false.
4146 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
4147 Instruction *I = dyn_cast<Instruction>(V);
4148 if (I == 0) return true;
4150 // If this is an or instruction, it is an inner node of the bswap.
4151 if (I->getOpcode() == Instruction::Or)
4152 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
4153 CollectBSwapParts(I->getOperand(1), ByteValues);
4155 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
4156 // If this is a shift by a constant int, and it is "24", then its operand
4157 // defines a byte. We only handle unsigned types here.
4158 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
4159 // Not shifting the entire input by N-1 bytes?
4160 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
4161 8*(ByteValues.size()-1))
4165 if (I->getOpcode() == Instruction::Shl) {
4166 // X << 24 defines the top byte with the lowest of the input bytes.
4167 DestNo = ByteValues.size()-1;
4169 // X >>u 24 defines the low byte with the highest of the input bytes.
4173 // If the destination byte value is already defined, the values are or'd
4174 // together, which isn't a bswap (unless it's an or of the same bits).
4175 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
4177 ByteValues[DestNo] = I->getOperand(0);
4181 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
4183 Value *Shift = 0, *ShiftLHS = 0;
4184 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
4185 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
4186 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
4188 Instruction *SI = cast<Instruction>(Shift);
4190 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
4191 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
4192 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
4195 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
4197 if (AndAmt->getValue().getActiveBits() > 64)
4199 uint64_t AndAmtVal = AndAmt->getZExtValue();
4200 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
4201 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
4203 // Unknown mask for bswap.
4204 if (DestByte == ByteValues.size()) return true;
4206 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
4208 if (SI->getOpcode() == Instruction::Shl)
4209 SrcByte = DestByte - ShiftBytes;
4211 SrcByte = DestByte + ShiftBytes;
4213 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
4214 if (SrcByte != ByteValues.size()-DestByte-1)
4217 // If the destination byte value is already defined, the values are or'd
4218 // together, which isn't a bswap (unless it's an or of the same bits).
4219 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
4221 ByteValues[DestByte] = SI->getOperand(0);
4225 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4226 /// If so, insert the new bswap intrinsic and return it.
4227 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4228 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4229 if (!ITy || ITy->getBitWidth() % 16)
4230 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4232 /// ByteValues - For each byte of the result, we keep track of which value
4233 /// defines each byte.
4234 SmallVector<Value*, 8> ByteValues;
4235 ByteValues.resize(ITy->getBitWidth()/8);
4237 // Try to find all the pieces corresponding to the bswap.
4238 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
4239 CollectBSwapParts(I.getOperand(1), ByteValues))
4242 // Check to see if all of the bytes come from the same value.
4243 Value *V = ByteValues[0];
4244 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4246 // Check to make sure that all of the bytes come from the same value.
4247 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4248 if (ByteValues[i] != V)
4250 const Type *Tys[] = { ITy };
4251 Module *M = I.getParent()->getParent()->getParent();
4252 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4253 return CallInst::Create(F, V);
4257 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4258 bool Changed = SimplifyCommutative(I);
4259 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4261 if (isa<UndefValue>(Op1)) // X | undef -> -1
4262 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4266 return ReplaceInstUsesWith(I, Op0);
4268 // See if we can simplify any instructions used by the instruction whose sole
4269 // purpose is to compute bits we don't care about.
4270 if (!isa<VectorType>(I.getType())) {
4271 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4272 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4273 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4274 KnownZero, KnownOne))
4276 } else if (isa<ConstantAggregateZero>(Op1)) {
4277 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4278 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4279 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4280 return ReplaceInstUsesWith(I, I.getOperand(1));
4286 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4287 ConstantInt *C1 = 0; Value *X = 0;
4288 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4289 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4290 Instruction *Or = BinaryOperator::createOr(X, RHS);
4291 InsertNewInstBefore(Or, I);
4293 return BinaryOperator::createAnd(Or,
4294 ConstantInt::get(RHS->getValue() | C1->getValue()));
4297 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4298 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4299 Instruction *Or = BinaryOperator::createOr(X, RHS);
4300 InsertNewInstBefore(Or, I);
4302 return BinaryOperator::createXor(Or,
4303 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4306 // Try to fold constant and into select arguments.
4307 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4308 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4310 if (isa<PHINode>(Op0))
4311 if (Instruction *NV = FoldOpIntoPhi(I))
4315 Value *A = 0, *B = 0;
4316 ConstantInt *C1 = 0, *C2 = 0;
4318 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4319 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4320 return ReplaceInstUsesWith(I, Op1);
4321 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4322 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4323 return ReplaceInstUsesWith(I, Op0);
4325 // (A | B) | C and A | (B | C) -> bswap if possible.
4326 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4327 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4328 match(Op1, m_Or(m_Value(), m_Value())) ||
4329 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4330 match(Op1, m_Shift(m_Value(), m_Value())))) {
4331 if (Instruction *BSwap = MatchBSwap(I))
4335 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4336 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4337 MaskedValueIsZero(Op1, C1->getValue())) {
4338 Instruction *NOr = BinaryOperator::createOr(A, Op1);
4339 InsertNewInstBefore(NOr, I);
4341 return BinaryOperator::createXor(NOr, C1);
4344 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4345 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4346 MaskedValueIsZero(Op0, C1->getValue())) {
4347 Instruction *NOr = BinaryOperator::createOr(A, Op0);
4348 InsertNewInstBefore(NOr, I);
4350 return BinaryOperator::createXor(NOr, C1);
4354 Value *C = 0, *D = 0;
4355 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4356 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4357 Value *V1 = 0, *V2 = 0, *V3 = 0;
4358 C1 = dyn_cast<ConstantInt>(C);
4359 C2 = dyn_cast<ConstantInt>(D);
4360 if (C1 && C2) { // (A & C1)|(B & C2)
4361 // If we have: ((V + N) & C1) | (V & C2)
4362 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4363 // replace with V+N.
4364 if (C1->getValue() == ~C2->getValue()) {
4365 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4366 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4367 // Add commutes, try both ways.
4368 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4369 return ReplaceInstUsesWith(I, A);
4370 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4371 return ReplaceInstUsesWith(I, A);
4373 // Or commutes, try both ways.
4374 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4375 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4376 // Add commutes, try both ways.
4377 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4378 return ReplaceInstUsesWith(I, B);
4379 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4380 return ReplaceInstUsesWith(I, B);
4383 V1 = 0; V2 = 0; V3 = 0;
4386 // Check to see if we have any common things being and'ed. If so, find the
4387 // terms for V1 & (V2|V3).
4388 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4389 if (A == B) // (A & C)|(A & D) == A & (C|D)
4390 V1 = A, V2 = C, V3 = D;
4391 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4392 V1 = A, V2 = B, V3 = C;
4393 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4394 V1 = C, V2 = A, V3 = D;
4395 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4396 V1 = C, V2 = A, V3 = B;
4400 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4401 return BinaryOperator::createAnd(V1, Or);
4406 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4407 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4408 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4409 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4410 SI0->getOperand(1) == SI1->getOperand(1) &&
4411 (SI0->hasOneUse() || SI1->hasOneUse())) {
4412 Instruction *NewOp =
4413 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4415 SI0->getName()), I);
4416 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4417 SI1->getOperand(1));
4421 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4422 if (A == Op1) // ~A | A == -1
4423 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4427 // Note, A is still live here!
4428 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4430 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4432 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4433 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4434 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4435 I.getName()+".demorgan"), I);
4436 return BinaryOperator::createNot(And);
4440 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4441 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4442 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4445 Value *LHSVal, *RHSVal;
4446 ConstantInt *LHSCst, *RHSCst;
4447 ICmpInst::Predicate LHSCC, RHSCC;
4448 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4449 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4450 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4451 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4452 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4453 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4454 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4455 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4456 // We can't fold (ugt x, C) | (sgt x, C2).
4457 PredicatesFoldable(LHSCC, RHSCC)) {
4458 // Ensure that the larger constant is on the RHS.
4459 ICmpInst *LHS = cast<ICmpInst>(Op0);
4461 if (ICmpInst::isSignedPredicate(LHSCC))
4462 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4464 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4467 std::swap(LHS, RHS);
4468 std::swap(LHSCst, RHSCst);
4469 std::swap(LHSCC, RHSCC);
4472 // At this point, we know we have have two icmp instructions
4473 // comparing a value against two constants and or'ing the result
4474 // together. Because of the above check, we know that we only have
4475 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4476 // FoldICmpLogical check above), that the two constants are not
4478 assert(LHSCst != RHSCst && "Compares not folded above?");
4481 default: assert(0 && "Unknown integer condition code!");
4482 case ICmpInst::ICMP_EQ:
4484 default: assert(0 && "Unknown integer condition code!");
4485 case ICmpInst::ICMP_EQ:
4486 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4487 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4488 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4489 LHSVal->getName()+".off");
4490 InsertNewInstBefore(Add, I);
4491 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4492 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4494 break; // (X == 13 | X == 15) -> no change
4495 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4496 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4498 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4499 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4500 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4501 return ReplaceInstUsesWith(I, RHS);
4504 case ICmpInst::ICMP_NE:
4506 default: assert(0 && "Unknown integer condition code!");
4507 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4508 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4509 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4510 return ReplaceInstUsesWith(I, LHS);
4511 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4512 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4513 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4514 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4517 case ICmpInst::ICMP_ULT:
4519 default: assert(0 && "Unknown integer condition code!");
4520 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4522 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4523 // If RHSCst is [us]MAXINT, it is always false. Not handling
4524 // this can cause overflow.
4525 if (RHSCst->isMaxValue(false))
4526 return ReplaceInstUsesWith(I, LHS);
4527 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4529 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4531 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4532 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4533 return ReplaceInstUsesWith(I, RHS);
4534 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4538 case ICmpInst::ICMP_SLT:
4540 default: assert(0 && "Unknown integer condition code!");
4541 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4543 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4544 // If RHSCst is [us]MAXINT, it is always false. Not handling
4545 // this can cause overflow.
4546 if (RHSCst->isMaxValue(true))
4547 return ReplaceInstUsesWith(I, LHS);
4548 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4550 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4552 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4553 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4554 return ReplaceInstUsesWith(I, RHS);
4555 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4559 case ICmpInst::ICMP_UGT:
4561 default: assert(0 && "Unknown integer condition code!");
4562 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4563 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4564 return ReplaceInstUsesWith(I, LHS);
4565 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4567 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4568 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4569 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4570 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4574 case ICmpInst::ICMP_SGT:
4576 default: assert(0 && "Unknown integer condition code!");
4577 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4578 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4579 return ReplaceInstUsesWith(I, LHS);
4580 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4582 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4583 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4584 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4585 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4593 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4594 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4595 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4596 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4597 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4598 !isa<ICmpInst>(Op1C->getOperand(0))) {
4599 const Type *SrcTy = Op0C->getOperand(0)->getType();
4600 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4601 // Only do this if the casts both really cause code to be
4603 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4605 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4607 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4608 Op1C->getOperand(0),
4610 InsertNewInstBefore(NewOp, I);
4611 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4618 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4619 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4620 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4621 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4622 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4623 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4624 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4625 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4626 // If either of the constants are nans, then the whole thing returns
4628 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4629 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4631 // Otherwise, no need to compare the two constants, compare the
4633 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4634 RHS->getOperand(0));
4639 return Changed ? &I : 0;
4644 // XorSelf - Implements: X ^ X --> 0
4647 XorSelf(Value *rhs) : RHS(rhs) {}
4648 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4649 Instruction *apply(BinaryOperator &Xor) const {
4656 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4657 bool Changed = SimplifyCommutative(I);
4658 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4660 if (isa<UndefValue>(Op1)) {
4661 if (isa<UndefValue>(Op0))
4662 // Handle undef ^ undef -> 0 special case. This is a common
4664 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4665 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4668 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4669 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4670 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4671 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4674 // See if we can simplify any instructions used by the instruction whose sole
4675 // purpose is to compute bits we don't care about.
4676 if (!isa<VectorType>(I.getType())) {
4677 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4678 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4679 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4680 KnownZero, KnownOne))
4682 } else if (isa<ConstantAggregateZero>(Op1)) {
4683 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4686 // Is this a ~ operation?
4687 if (Value *NotOp = dyn_castNotVal(&I)) {
4688 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4689 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4690 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4691 if (Op0I->getOpcode() == Instruction::And ||
4692 Op0I->getOpcode() == Instruction::Or) {
4693 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4694 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4696 BinaryOperator::createNot(Op0I->getOperand(1),
4697 Op0I->getOperand(1)->getName()+".not");
4698 InsertNewInstBefore(NotY, I);
4699 if (Op0I->getOpcode() == Instruction::And)
4700 return BinaryOperator::createOr(Op0NotVal, NotY);
4702 return BinaryOperator::createAnd(Op0NotVal, NotY);
4709 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4710 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4711 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4712 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4713 return new ICmpInst(ICI->getInversePredicate(),
4714 ICI->getOperand(0), ICI->getOperand(1));
4716 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4717 return new FCmpInst(FCI->getInversePredicate(),
4718 FCI->getOperand(0), FCI->getOperand(1));
4721 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4722 // ~(c-X) == X-c-1 == X+(-c-1)
4723 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4724 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4725 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4726 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4727 ConstantInt::get(I.getType(), 1));
4728 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4731 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4732 if (Op0I->getOpcode() == Instruction::Add) {
4733 // ~(X-c) --> (-c-1)-X
4734 if (RHS->isAllOnesValue()) {
4735 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4736 return BinaryOperator::createSub(
4737 ConstantExpr::getSub(NegOp0CI,
4738 ConstantInt::get(I.getType(), 1)),
4739 Op0I->getOperand(0));
4740 } else if (RHS->getValue().isSignBit()) {
4741 // (X + C) ^ signbit -> (X + C + signbit)
4742 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4743 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4746 } else if (Op0I->getOpcode() == Instruction::Or) {
4747 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4748 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4749 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4750 // Anything in both C1 and C2 is known to be zero, remove it from
4752 Constant *CommonBits = And(Op0CI, RHS);
4753 NewRHS = ConstantExpr::getAnd(NewRHS,
4754 ConstantExpr::getNot(CommonBits));
4755 AddToWorkList(Op0I);
4756 I.setOperand(0, Op0I->getOperand(0));
4757 I.setOperand(1, NewRHS);
4764 // Try to fold constant and into select arguments.
4765 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4766 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4768 if (isa<PHINode>(Op0))
4769 if (Instruction *NV = FoldOpIntoPhi(I))
4773 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4775 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4777 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4779 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4782 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4785 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4786 if (A == Op0) { // B^(B|A) == (A|B)^B
4787 Op1I->swapOperands();
4789 std::swap(Op0, Op1);
4790 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4791 I.swapOperands(); // Simplified below.
4792 std::swap(Op0, Op1);
4794 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4795 if (Op0 == A) // A^(A^B) == B
4796 return ReplaceInstUsesWith(I, B);
4797 else if (Op0 == B) // A^(B^A) == B
4798 return ReplaceInstUsesWith(I, A);
4799 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4800 if (A == Op0) { // A^(A&B) -> A^(B&A)
4801 Op1I->swapOperands();
4804 if (B == Op0) { // A^(B&A) -> (B&A)^A
4805 I.swapOperands(); // Simplified below.
4806 std::swap(Op0, Op1);
4811 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4814 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4815 if (A == Op1) // (B|A)^B == (A|B)^B
4817 if (B == Op1) { // (A|B)^B == A & ~B
4819 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4820 return BinaryOperator::createAnd(A, NotB);
4822 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4823 if (Op1 == A) // (A^B)^A == B
4824 return ReplaceInstUsesWith(I, B);
4825 else if (Op1 == B) // (B^A)^A == B
4826 return ReplaceInstUsesWith(I, A);
4827 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4828 if (A == Op1) // (A&B)^A -> (B&A)^A
4830 if (B == Op1 && // (B&A)^A == ~B & A
4831 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4833 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4834 return BinaryOperator::createAnd(N, Op1);
4839 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4840 if (Op0I && Op1I && Op0I->isShift() &&
4841 Op0I->getOpcode() == Op1I->getOpcode() &&
4842 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4843 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4844 Instruction *NewOp =
4845 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4846 Op1I->getOperand(0),
4847 Op0I->getName()), I);
4848 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4849 Op1I->getOperand(1));
4853 Value *A, *B, *C, *D;
4854 // (A & B)^(A | B) -> A ^ B
4855 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4856 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4857 if ((A == C && B == D) || (A == D && B == C))
4858 return BinaryOperator::createXor(A, B);
4860 // (A | B)^(A & B) -> A ^ B
4861 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4862 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4863 if ((A == C && B == D) || (A == D && B == C))
4864 return BinaryOperator::createXor(A, B);
4868 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4869 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4870 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4871 // (X & Y)^(X & Y) -> (Y^Z) & X
4872 Value *X = 0, *Y = 0, *Z = 0;
4874 X = A, Y = B, Z = D;
4876 X = A, Y = B, Z = C;
4878 X = B, Y = A, Z = D;
4880 X = B, Y = A, Z = C;
4883 Instruction *NewOp =
4884 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4885 return BinaryOperator::createAnd(NewOp, X);
4890 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4891 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4892 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4895 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4896 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4897 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4898 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4899 const Type *SrcTy = Op0C->getOperand(0)->getType();
4900 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4901 // Only do this if the casts both really cause code to be generated.
4902 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4904 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4906 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4907 Op1C->getOperand(0),
4909 InsertNewInstBefore(NewOp, I);
4910 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4914 return Changed ? &I : 0;
4917 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4918 /// overflowed for this type.
4919 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4920 ConstantInt *In2, bool IsSigned = false) {
4921 Result = cast<ConstantInt>(Add(In1, In2));
4924 if (In2->getValue().isNegative())
4925 return Result->getValue().sgt(In1->getValue());
4927 return Result->getValue().slt(In1->getValue());
4929 return Result->getValue().ult(In1->getValue());
4932 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4933 /// code necessary to compute the offset from the base pointer (without adding
4934 /// in the base pointer). Return the result as a signed integer of intptr size.
4935 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4936 TargetData &TD = IC.getTargetData();
4937 gep_type_iterator GTI = gep_type_begin(GEP);
4938 const Type *IntPtrTy = TD.getIntPtrType();
4939 Value *Result = Constant::getNullValue(IntPtrTy);
4941 // Build a mask for high order bits.
4942 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4943 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4945 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4946 Value *Op = GEP->getOperand(i);
4947 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4948 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4949 if (OpC->isZero()) continue;
4951 // Handle a struct index, which adds its field offset to the pointer.
4952 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4953 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4955 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4956 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4958 Result = IC.InsertNewInstBefore(
4959 BinaryOperator::createAdd(Result,
4960 ConstantInt::get(IntPtrTy, Size),
4961 GEP->getName()+".offs"), I);
4965 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4966 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4967 Scale = ConstantExpr::getMul(OC, Scale);
4968 if (Constant *RC = dyn_cast<Constant>(Result))
4969 Result = ConstantExpr::getAdd(RC, Scale);
4971 // Emit an add instruction.
4972 Result = IC.InsertNewInstBefore(
4973 BinaryOperator::createAdd(Result, Scale,
4974 GEP->getName()+".offs"), I);
4978 // Convert to correct type.
4979 if (Op->getType() != IntPtrTy) {
4980 if (Constant *OpC = dyn_cast<Constant>(Op))
4981 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4983 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4984 Op->getName()+".c"), I);
4987 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4988 if (Constant *OpC = dyn_cast<Constant>(Op))
4989 Op = ConstantExpr::getMul(OpC, Scale);
4990 else // We'll let instcombine(mul) convert this to a shl if possible.
4991 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4992 GEP->getName()+".idx"), I);
4995 // Emit an add instruction.
4996 if (isa<Constant>(Op) && isa<Constant>(Result))
4997 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4998 cast<Constant>(Result));
5000 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
5001 GEP->getName()+".offs"), I);
5007 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
5008 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
5009 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
5010 /// complex, and scales are involved. The above expression would also be legal
5011 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
5012 /// later form is less amenable to optimization though, and we are allowed to
5013 /// generate the first by knowing that pointer arithmetic doesn't overflow.
5015 /// If we can't emit an optimized form for this expression, this returns null.
5017 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
5019 TargetData &TD = IC.getTargetData();
5020 gep_type_iterator GTI = gep_type_begin(GEP);
5022 // Check to see if this gep only has a single variable index. If so, and if
5023 // any constant indices are a multiple of its scale, then we can compute this
5024 // in terms of the scale of the variable index. For example, if the GEP
5025 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
5026 // because the expression will cross zero at the same point.
5027 unsigned i, e = GEP->getNumOperands();
5029 for (i = 1; i != e; ++i, ++GTI) {
5030 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
5031 // Compute the aggregate offset of constant indices.
5032 if (CI->isZero()) continue;
5034 // Handle a struct index, which adds its field offset to the pointer.
5035 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5036 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5038 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5039 Offset += Size*CI->getSExtValue();
5042 // Found our variable index.
5047 // If there are no variable indices, we must have a constant offset, just
5048 // evaluate it the general way.
5049 if (i == e) return 0;
5051 Value *VariableIdx = GEP->getOperand(i);
5052 // Determine the scale factor of the variable element. For example, this is
5053 // 4 if the variable index is into an array of i32.
5054 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
5056 // Verify that there are no other variable indices. If so, emit the hard way.
5057 for (++i, ++GTI; i != e; ++i, ++GTI) {
5058 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
5061 // Compute the aggregate offset of constant indices.
5062 if (CI->isZero()) continue;
5064 // Handle a struct index, which adds its field offset to the pointer.
5065 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5066 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5068 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5069 Offset += Size*CI->getSExtValue();
5073 // Okay, we know we have a single variable index, which must be a
5074 // pointer/array/vector index. If there is no offset, life is simple, return
5076 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5078 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5079 // we don't need to bother extending: the extension won't affect where the
5080 // computation crosses zero.
5081 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5082 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5083 VariableIdx->getNameStart(), &I);
5087 // Otherwise, there is an index. The computation we will do will be modulo
5088 // the pointer size, so get it.
5089 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5091 Offset &= PtrSizeMask;
5092 VariableScale &= PtrSizeMask;
5094 // To do this transformation, any constant index must be a multiple of the
5095 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5096 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5097 // multiple of the variable scale.
5098 int64_t NewOffs = Offset / (int64_t)VariableScale;
5099 if (Offset != NewOffs*(int64_t)VariableScale)
5102 // Okay, we can do this evaluation. Start by converting the index to intptr.
5103 const Type *IntPtrTy = TD.getIntPtrType();
5104 if (VariableIdx->getType() != IntPtrTy)
5105 VariableIdx = CastInst::createIntegerCast(VariableIdx, IntPtrTy,
5107 VariableIdx->getNameStart(), &I);
5108 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5109 return BinaryOperator::createAdd(VariableIdx, OffsetVal, "offset", &I);
5113 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5114 /// else. At this point we know that the GEP is on the LHS of the comparison.
5115 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5116 ICmpInst::Predicate Cond,
5118 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5120 // Look through bitcasts.
5121 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5122 RHS = BCI->getOperand(0);
5124 Value *PtrBase = GEPLHS->getOperand(0);
5125 if (PtrBase == RHS) {
5126 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5127 // This transformation (ignoring the base and scales) is valid because we
5128 // know pointers can't overflow. See if we can output an optimized form.
5129 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5131 // If not, synthesize the offset the hard way.
5133 Offset = EmitGEPOffset(GEPLHS, I, *this);
5134 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5135 Constant::getNullValue(Offset->getType()));
5136 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5137 // If the base pointers are different, but the indices are the same, just
5138 // compare the base pointer.
5139 if (PtrBase != GEPRHS->getOperand(0)) {
5140 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5141 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5142 GEPRHS->getOperand(0)->getType();
5144 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5145 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5146 IndicesTheSame = false;
5150 // If all indices are the same, just compare the base pointers.
5152 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5153 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5155 // Otherwise, the base pointers are different and the indices are
5156 // different, bail out.
5160 // If one of the GEPs has all zero indices, recurse.
5161 bool AllZeros = true;
5162 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5163 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5164 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5169 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5170 ICmpInst::getSwappedPredicate(Cond), I);
5172 // If the other GEP has all zero indices, recurse.
5174 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5175 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5176 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5181 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5183 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5184 // If the GEPs only differ by one index, compare it.
5185 unsigned NumDifferences = 0; // Keep track of # differences.
5186 unsigned DiffOperand = 0; // The operand that differs.
5187 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5188 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5189 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5190 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5191 // Irreconcilable differences.
5195 if (NumDifferences++) break;
5200 if (NumDifferences == 0) // SAME GEP?
5201 return ReplaceInstUsesWith(I, // No comparison is needed here.
5202 ConstantInt::get(Type::Int1Ty,
5203 isTrueWhenEqual(Cond)));
5205 else if (NumDifferences == 1) {
5206 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5207 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5208 // Make sure we do a signed comparison here.
5209 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5213 // Only lower this if the icmp is the only user of the GEP or if we expect
5214 // the result to fold to a constant!
5215 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5216 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5217 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5218 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5219 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5220 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5226 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5227 bool Changed = SimplifyCompare(I);
5228 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5230 // Fold trivial predicates.
5231 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5232 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5233 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5234 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5236 // Simplify 'fcmp pred X, X'
5238 switch (I.getPredicate()) {
5239 default: assert(0 && "Unknown predicate!");
5240 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5241 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5242 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5243 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5244 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5245 case FCmpInst::FCMP_OLT: // True if ordered and less than
5246 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5247 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5249 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5250 case FCmpInst::FCMP_ULT: // True if unordered or less than
5251 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5252 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5253 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5254 I.setPredicate(FCmpInst::FCMP_UNO);
5255 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5258 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5259 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5260 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5261 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5262 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5263 I.setPredicate(FCmpInst::FCMP_ORD);
5264 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5269 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5270 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5272 // Handle fcmp with constant RHS
5273 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5274 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5275 switch (LHSI->getOpcode()) {
5276 case Instruction::PHI:
5277 if (Instruction *NV = FoldOpIntoPhi(I))
5280 case Instruction::Select:
5281 // If either operand of the select is a constant, we can fold the
5282 // comparison into the select arms, which will cause one to be
5283 // constant folded and the select turned into a bitwise or.
5284 Value *Op1 = 0, *Op2 = 0;
5285 if (LHSI->hasOneUse()) {
5286 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5287 // Fold the known value into the constant operand.
5288 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5289 // Insert a new FCmp of the other select operand.
5290 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5291 LHSI->getOperand(2), RHSC,
5293 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5294 // Fold the known value into the constant operand.
5295 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5296 // Insert a new FCmp of the other select operand.
5297 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5298 LHSI->getOperand(1), RHSC,
5304 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5309 return Changed ? &I : 0;
5312 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5313 bool Changed = SimplifyCompare(I);
5314 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5315 const Type *Ty = Op0->getType();
5319 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5320 isTrueWhenEqual(I)));
5322 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5323 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5325 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5326 // addresses never equal each other! We already know that Op0 != Op1.
5327 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5328 isa<ConstantPointerNull>(Op0)) &&
5329 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5330 isa<ConstantPointerNull>(Op1)))
5331 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5332 !isTrueWhenEqual(I)));
5334 // icmp's with boolean values can always be turned into bitwise operations
5335 if (Ty == Type::Int1Ty) {
5336 switch (I.getPredicate()) {
5337 default: assert(0 && "Invalid icmp instruction!");
5338 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5339 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
5340 InsertNewInstBefore(Xor, I);
5341 return BinaryOperator::createNot(Xor);
5343 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5344 return BinaryOperator::createXor(Op0, Op1);
5346 case ICmpInst::ICMP_UGT:
5347 case ICmpInst::ICMP_SGT:
5348 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5350 case ICmpInst::ICMP_ULT:
5351 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5352 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
5353 InsertNewInstBefore(Not, I);
5354 return BinaryOperator::createAnd(Not, Op1);
5356 case ICmpInst::ICMP_UGE:
5357 case ICmpInst::ICMP_SGE:
5358 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5360 case ICmpInst::ICMP_ULE:
5361 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5362 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
5363 InsertNewInstBefore(Not, I);
5364 return BinaryOperator::createOr(Not, Op1);
5369 // See if we are doing a comparison between a constant and an instruction that
5370 // can be folded into the comparison.
5371 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5374 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5375 if (I.isEquality() && CI->isNullValue() &&
5376 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5377 // (icmp cond A B) if cond is equality
5378 return new ICmpInst(I.getPredicate(), A, B);
5381 switch (I.getPredicate()) {
5383 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5384 if (CI->isMinValue(false))
5385 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5386 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5387 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5388 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5389 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5390 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5391 if (CI->isMinValue(true))
5392 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5393 ConstantInt::getAllOnesValue(Op0->getType()));
5397 case ICmpInst::ICMP_SLT:
5398 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5399 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5400 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5401 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5402 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5403 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5406 case ICmpInst::ICMP_UGT:
5407 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5408 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5409 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5410 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5411 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5412 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5414 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5415 if (CI->isMaxValue(true))
5416 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5417 ConstantInt::getNullValue(Op0->getType()));
5420 case ICmpInst::ICMP_SGT:
5421 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5422 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5423 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5424 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5425 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5426 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5429 case ICmpInst::ICMP_ULE:
5430 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5431 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5432 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5433 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5434 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5435 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5438 case ICmpInst::ICMP_SLE:
5439 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5440 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5441 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5442 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5443 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5444 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5447 case ICmpInst::ICMP_UGE:
5448 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5449 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5450 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5451 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5452 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5453 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5456 case ICmpInst::ICMP_SGE:
5457 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5458 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5459 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5460 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5461 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5462 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5466 // If we still have a icmp le or icmp ge instruction, turn it into the
5467 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5468 // already been handled above, this requires little checking.
5470 switch (I.getPredicate()) {
5472 case ICmpInst::ICMP_ULE:
5473 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5474 case ICmpInst::ICMP_SLE:
5475 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5476 case ICmpInst::ICMP_UGE:
5477 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5478 case ICmpInst::ICMP_SGE:
5479 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5482 // See if we can fold the comparison based on bits known to be zero or one
5483 // in the input. If this comparison is a normal comparison, it demands all
5484 // bits, if it is a sign bit comparison, it only demands the sign bit.
5487 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5489 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5490 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5491 if (SimplifyDemandedBits(Op0,
5492 isSignBit ? APInt::getSignBit(BitWidth)
5493 : APInt::getAllOnesValue(BitWidth),
5494 KnownZero, KnownOne, 0))
5497 // Given the known and unknown bits, compute a range that the LHS could be
5499 if ((KnownOne | KnownZero) != 0) {
5500 // Compute the Min, Max and RHS values based on the known bits. For the
5501 // EQ and NE we use unsigned values.
5502 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5503 const APInt& RHSVal = CI->getValue();
5504 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5505 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5508 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5511 switch (I.getPredicate()) { // LE/GE have been folded already.
5512 default: assert(0 && "Unknown icmp opcode!");
5513 case ICmpInst::ICMP_EQ:
5514 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5515 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5517 case ICmpInst::ICMP_NE:
5518 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5519 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5521 case ICmpInst::ICMP_ULT:
5522 if (Max.ult(RHSVal))
5523 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5524 if (Min.uge(RHSVal))
5525 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5527 case ICmpInst::ICMP_UGT:
5528 if (Min.ugt(RHSVal))
5529 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5530 if (Max.ule(RHSVal))
5531 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5533 case ICmpInst::ICMP_SLT:
5534 if (Max.slt(RHSVal))
5535 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5536 if (Min.sgt(RHSVal))
5537 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5539 case ICmpInst::ICMP_SGT:
5540 if (Min.sgt(RHSVal))
5541 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5542 if (Max.sle(RHSVal))
5543 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5548 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5549 // instruction, see if that instruction also has constants so that the
5550 // instruction can be folded into the icmp
5551 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5552 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5556 // Handle icmp with constant (but not simple integer constant) RHS
5557 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5558 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5559 switch (LHSI->getOpcode()) {
5560 case Instruction::GetElementPtr:
5561 if (RHSC->isNullValue()) {
5562 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5563 bool isAllZeros = true;
5564 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5565 if (!isa<Constant>(LHSI->getOperand(i)) ||
5566 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5571 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5572 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5576 case Instruction::PHI:
5577 if (Instruction *NV = FoldOpIntoPhi(I))
5580 case Instruction::Select: {
5581 // If either operand of the select is a constant, we can fold the
5582 // comparison into the select arms, which will cause one to be
5583 // constant folded and the select turned into a bitwise or.
5584 Value *Op1 = 0, *Op2 = 0;
5585 if (LHSI->hasOneUse()) {
5586 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5587 // Fold the known value into the constant operand.
5588 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5589 // Insert a new ICmp of the other select operand.
5590 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5591 LHSI->getOperand(2), RHSC,
5593 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5594 // Fold the known value into the constant operand.
5595 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5596 // Insert a new ICmp of the other select operand.
5597 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5598 LHSI->getOperand(1), RHSC,
5604 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5607 case Instruction::Malloc:
5608 // If we have (malloc != null), and if the malloc has a single use, we
5609 // can assume it is successful and remove the malloc.
5610 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5611 AddToWorkList(LHSI);
5612 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5613 !isTrueWhenEqual(I)));
5619 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5620 if (User *GEP = dyn_castGetElementPtr(Op0))
5621 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5623 if (User *GEP = dyn_castGetElementPtr(Op1))
5624 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5625 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5628 // Test to see if the operands of the icmp are casted versions of other
5629 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5631 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5632 if (isa<PointerType>(Op0->getType()) &&
5633 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5634 // We keep moving the cast from the left operand over to the right
5635 // operand, where it can often be eliminated completely.
5636 Op0 = CI->getOperand(0);
5638 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5639 // so eliminate it as well.
5640 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5641 Op1 = CI2->getOperand(0);
5643 // If Op1 is a constant, we can fold the cast into the constant.
5644 if (Op0->getType() != Op1->getType()) {
5645 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5646 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5648 // Otherwise, cast the RHS right before the icmp
5649 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5652 return new ICmpInst(I.getPredicate(), Op0, Op1);
5656 if (isa<CastInst>(Op0)) {
5657 // Handle the special case of: icmp (cast bool to X), <cst>
5658 // This comes up when you have code like
5661 // For generality, we handle any zero-extension of any operand comparison
5662 // with a constant or another cast from the same type.
5663 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5664 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5668 // ~x < ~y --> y < x
5670 if (match(Op0, m_Not(m_Value(A))) &&
5671 match(Op1, m_Not(m_Value(B))))
5672 return new ICmpInst(I.getPredicate(), B, A);
5675 if (I.isEquality()) {
5676 Value *A, *B, *C, *D;
5678 // -x == -y --> x == y
5679 if (match(Op0, m_Neg(m_Value(A))) &&
5680 match(Op1, m_Neg(m_Value(B))))
5681 return new ICmpInst(I.getPredicate(), A, B);
5683 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5684 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5685 Value *OtherVal = A == Op1 ? B : A;
5686 return new ICmpInst(I.getPredicate(), OtherVal,
5687 Constant::getNullValue(A->getType()));
5690 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5691 // A^c1 == C^c2 --> A == C^(c1^c2)
5692 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5693 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5694 if (Op1->hasOneUse()) {
5695 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5696 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5697 return new ICmpInst(I.getPredicate(), A,
5698 InsertNewInstBefore(Xor, I));
5701 // A^B == A^D -> B == D
5702 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5703 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5704 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5705 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5709 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5710 (A == Op0 || B == Op0)) {
5711 // A == (A^B) -> B == 0
5712 Value *OtherVal = A == Op0 ? B : A;
5713 return new ICmpInst(I.getPredicate(), OtherVal,
5714 Constant::getNullValue(A->getType()));
5716 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5717 // (A-B) == A -> B == 0
5718 return new ICmpInst(I.getPredicate(), B,
5719 Constant::getNullValue(B->getType()));
5721 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5722 // A == (A-B) -> B == 0
5723 return new ICmpInst(I.getPredicate(), B,
5724 Constant::getNullValue(B->getType()));
5727 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5728 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5729 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5730 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5731 Value *X = 0, *Y = 0, *Z = 0;
5734 X = B; Y = D; Z = A;
5735 } else if (A == D) {
5736 X = B; Y = C; Z = A;
5737 } else if (B == C) {
5738 X = A; Y = D; Z = B;
5739 } else if (B == D) {
5740 X = A; Y = C; Z = B;
5743 if (X) { // Build (X^Y) & Z
5744 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5745 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5746 I.setOperand(0, Op1);
5747 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5752 return Changed ? &I : 0;
5756 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5757 /// and CmpRHS are both known to be integer constants.
5758 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5759 ConstantInt *DivRHS) {
5760 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5761 const APInt &CmpRHSV = CmpRHS->getValue();
5763 // FIXME: If the operand types don't match the type of the divide
5764 // then don't attempt this transform. The code below doesn't have the
5765 // logic to deal with a signed divide and an unsigned compare (and
5766 // vice versa). This is because (x /s C1) <s C2 produces different
5767 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5768 // (x /u C1) <u C2. Simply casting the operands and result won't
5769 // work. :( The if statement below tests that condition and bails
5771 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5772 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5774 if (DivRHS->isZero())
5775 return 0; // The ProdOV computation fails on divide by zero.
5777 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5778 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5779 // C2 (CI). By solving for X we can turn this into a range check
5780 // instead of computing a divide.
5781 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5783 // Determine if the product overflows by seeing if the product is
5784 // not equal to the divide. Make sure we do the same kind of divide
5785 // as in the LHS instruction that we're folding.
5786 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5787 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5789 // Get the ICmp opcode
5790 ICmpInst::Predicate Pred = ICI.getPredicate();
5792 // Figure out the interval that is being checked. For example, a comparison
5793 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5794 // Compute this interval based on the constants involved and the signedness of
5795 // the compare/divide. This computes a half-open interval, keeping track of
5796 // whether either value in the interval overflows. After analysis each
5797 // overflow variable is set to 0 if it's corresponding bound variable is valid
5798 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5799 int LoOverflow = 0, HiOverflow = 0;
5800 ConstantInt *LoBound = 0, *HiBound = 0;
5803 if (!DivIsSigned) { // udiv
5804 // e.g. X/5 op 3 --> [15, 20)
5806 HiOverflow = LoOverflow = ProdOV;
5808 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5809 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5810 if (CmpRHSV == 0) { // (X / pos) op 0
5811 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5812 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5814 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5815 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5816 HiOverflow = LoOverflow = ProdOV;
5818 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5819 } else { // (X / pos) op neg
5820 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5821 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5822 LoOverflow = AddWithOverflow(LoBound, Prod,
5823 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5824 HiBound = AddOne(Prod);
5825 HiOverflow = ProdOV ? -1 : 0;
5827 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5828 if (CmpRHSV == 0) { // (X / neg) op 0
5829 // e.g. X/-5 op 0 --> [-4, 5)
5830 LoBound = AddOne(DivRHS);
5831 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5832 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5833 HiOverflow = 1; // [INTMIN+1, overflow)
5834 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5836 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5837 // e.g. X/-5 op 3 --> [-19, -14)
5838 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5840 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5841 HiBound = AddOne(Prod);
5842 } else { // (X / neg) op neg
5843 // e.g. X/-5 op -3 --> [15, 20)
5845 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5846 HiBound = Subtract(Prod, DivRHS);
5849 // Dividing by a negative swaps the condition. LT <-> GT
5850 Pred = ICmpInst::getSwappedPredicate(Pred);
5853 Value *X = DivI->getOperand(0);
5855 default: assert(0 && "Unhandled icmp opcode!");
5856 case ICmpInst::ICMP_EQ:
5857 if (LoOverflow && HiOverflow)
5858 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5859 else if (HiOverflow)
5860 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5861 ICmpInst::ICMP_UGE, X, LoBound);
5862 else if (LoOverflow)
5863 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5864 ICmpInst::ICMP_ULT, X, HiBound);
5866 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5867 case ICmpInst::ICMP_NE:
5868 if (LoOverflow && HiOverflow)
5869 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5870 else if (HiOverflow)
5871 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5872 ICmpInst::ICMP_ULT, X, LoBound);
5873 else if (LoOverflow)
5874 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5875 ICmpInst::ICMP_UGE, X, HiBound);
5877 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5878 case ICmpInst::ICMP_ULT:
5879 case ICmpInst::ICMP_SLT:
5880 if (LoOverflow == +1) // Low bound is greater than input range.
5881 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5882 if (LoOverflow == -1) // Low bound is less than input range.
5883 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5884 return new ICmpInst(Pred, X, LoBound);
5885 case ICmpInst::ICMP_UGT:
5886 case ICmpInst::ICMP_SGT:
5887 if (HiOverflow == +1) // High bound greater than input range.
5888 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5889 else if (HiOverflow == -1) // High bound less than input range.
5890 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5891 if (Pred == ICmpInst::ICMP_UGT)
5892 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5894 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5899 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5901 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5904 const APInt &RHSV = RHS->getValue();
5906 switch (LHSI->getOpcode()) {
5907 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5908 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5909 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5911 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5912 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5913 Value *CompareVal = LHSI->getOperand(0);
5915 // If the sign bit of the XorCST is not set, there is no change to
5916 // the operation, just stop using the Xor.
5917 if (!XorCST->getValue().isNegative()) {
5918 ICI.setOperand(0, CompareVal);
5919 AddToWorkList(LHSI);
5923 // Was the old condition true if the operand is positive?
5924 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5926 // If so, the new one isn't.
5927 isTrueIfPositive ^= true;
5929 if (isTrueIfPositive)
5930 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5932 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5936 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5937 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5938 LHSI->getOperand(0)->hasOneUse()) {
5939 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5941 // If the LHS is an AND of a truncating cast, we can widen the
5942 // and/compare to be the input width without changing the value
5943 // produced, eliminating a cast.
5944 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5945 // We can do this transformation if either the AND constant does not
5946 // have its sign bit set or if it is an equality comparison.
5947 // Extending a relational comparison when we're checking the sign
5948 // bit would not work.
5949 if (Cast->hasOneUse() &&
5950 (ICI.isEquality() ||
5951 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5953 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5954 APInt NewCST = AndCST->getValue();
5955 NewCST.zext(BitWidth);
5957 NewCI.zext(BitWidth);
5958 Instruction *NewAnd =
5959 BinaryOperator::createAnd(Cast->getOperand(0),
5960 ConstantInt::get(NewCST),LHSI->getName());
5961 InsertNewInstBefore(NewAnd, ICI);
5962 return new ICmpInst(ICI.getPredicate(), NewAnd,
5963 ConstantInt::get(NewCI));
5967 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5968 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5969 // happens a LOT in code produced by the C front-end, for bitfield
5971 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5972 if (Shift && !Shift->isShift())
5976 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5977 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5978 const Type *AndTy = AndCST->getType(); // Type of the and.
5980 // We can fold this as long as we can't shift unknown bits
5981 // into the mask. This can only happen with signed shift
5982 // rights, as they sign-extend.
5984 bool CanFold = Shift->isLogicalShift();
5986 // To test for the bad case of the signed shr, see if any
5987 // of the bits shifted in could be tested after the mask.
5988 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5989 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5991 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5992 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5993 AndCST->getValue()) == 0)
5999 if (Shift->getOpcode() == Instruction::Shl)
6000 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
6002 NewCst = ConstantExpr::getShl(RHS, ShAmt);
6004 // Check to see if we are shifting out any of the bits being
6006 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
6007 // If we shifted bits out, the fold is not going to work out.
6008 // As a special case, check to see if this means that the
6009 // result is always true or false now.
6010 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6011 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6012 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6013 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6015 ICI.setOperand(1, NewCst);
6016 Constant *NewAndCST;
6017 if (Shift->getOpcode() == Instruction::Shl)
6018 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
6020 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
6021 LHSI->setOperand(1, NewAndCST);
6022 LHSI->setOperand(0, Shift->getOperand(0));
6023 AddToWorkList(Shift); // Shift is dead.
6024 AddUsesToWorkList(ICI);
6030 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
6031 // preferable because it allows the C<<Y expression to be hoisted out
6032 // of a loop if Y is invariant and X is not.
6033 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
6034 ICI.isEquality() && !Shift->isArithmeticShift() &&
6035 isa<Instruction>(Shift->getOperand(0))) {
6038 if (Shift->getOpcode() == Instruction::LShr) {
6039 NS = BinaryOperator::createShl(AndCST,
6040 Shift->getOperand(1), "tmp");
6042 // Insert a logical shift.
6043 NS = BinaryOperator::createLShr(AndCST,
6044 Shift->getOperand(1), "tmp");
6046 InsertNewInstBefore(cast<Instruction>(NS), ICI);
6048 // Compute X & (C << Y).
6049 Instruction *NewAnd =
6050 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
6051 InsertNewInstBefore(NewAnd, ICI);
6053 ICI.setOperand(0, NewAnd);
6059 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
6060 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6063 uint32_t TypeBits = RHSV.getBitWidth();
6065 // Check that the shift amount is in range. If not, don't perform
6066 // undefined shifts. When the shift is visited it will be
6068 if (ShAmt->uge(TypeBits))
6071 if (ICI.isEquality()) {
6072 // If we are comparing against bits always shifted out, the
6073 // comparison cannot succeed.
6075 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
6076 if (Comp != RHS) {// Comparing against a bit that we know is zero.
6077 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6078 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6079 return ReplaceInstUsesWith(ICI, Cst);
6082 if (LHSI->hasOneUse()) {
6083 // Otherwise strength reduce the shift into an and.
6084 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6086 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6089 BinaryOperator::createAnd(LHSI->getOperand(0),
6090 Mask, LHSI->getName()+".mask");
6091 Value *And = InsertNewInstBefore(AndI, ICI);
6092 return new ICmpInst(ICI.getPredicate(), And,
6093 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6097 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6098 bool TrueIfSigned = false;
6099 if (LHSI->hasOneUse() &&
6100 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6101 // (X << 31) <s 0 --> (X&1) != 0
6102 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6103 (TypeBits-ShAmt->getZExtValue()-1));
6105 BinaryOperator::createAnd(LHSI->getOperand(0),
6106 Mask, LHSI->getName()+".mask");
6107 Value *And = InsertNewInstBefore(AndI, ICI);
6109 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6110 And, Constant::getNullValue(And->getType()));
6115 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6116 case Instruction::AShr: {
6117 // Only handle equality comparisons of shift-by-constant.
6118 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6119 if (!ShAmt || !ICI.isEquality()) break;
6121 // Check that the shift amount is in range. If not, don't perform
6122 // undefined shifts. When the shift is visited it will be
6124 uint32_t TypeBits = RHSV.getBitWidth();
6125 if (ShAmt->uge(TypeBits))
6128 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6130 // If we are comparing against bits always shifted out, the
6131 // comparison cannot succeed.
6132 APInt Comp = RHSV << ShAmtVal;
6133 if (LHSI->getOpcode() == Instruction::LShr)
6134 Comp = Comp.lshr(ShAmtVal);
6136 Comp = Comp.ashr(ShAmtVal);
6138 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6139 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6140 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6141 return ReplaceInstUsesWith(ICI, Cst);
6144 // Otherwise, check to see if the bits shifted out are known to be zero.
6145 // If so, we can compare against the unshifted value:
6146 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6147 if (LHSI->hasOneUse() &&
6148 MaskedValueIsZero(LHSI->getOperand(0),
6149 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6150 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6151 ConstantExpr::getShl(RHS, ShAmt));
6154 if (LHSI->hasOneUse()) {
6155 // Otherwise strength reduce the shift into an and.
6156 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6157 Constant *Mask = ConstantInt::get(Val);
6160 BinaryOperator::createAnd(LHSI->getOperand(0),
6161 Mask, LHSI->getName()+".mask");
6162 Value *And = InsertNewInstBefore(AndI, ICI);
6163 return new ICmpInst(ICI.getPredicate(), And,
6164 ConstantExpr::getShl(RHS, ShAmt));
6169 case Instruction::SDiv:
6170 case Instruction::UDiv:
6171 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6172 // Fold this div into the comparison, producing a range check.
6173 // Determine, based on the divide type, what the range is being
6174 // checked. If there is an overflow on the low or high side, remember
6175 // it, otherwise compute the range [low, hi) bounding the new value.
6176 // See: InsertRangeTest above for the kinds of replacements possible.
6177 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6178 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6183 case Instruction::Add:
6184 // Fold: icmp pred (add, X, C1), C2
6186 if (!ICI.isEquality()) {
6187 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6189 const APInt &LHSV = LHSC->getValue();
6191 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6194 if (ICI.isSignedPredicate()) {
6195 if (CR.getLower().isSignBit()) {
6196 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6197 ConstantInt::get(CR.getUpper()));
6198 } else if (CR.getUpper().isSignBit()) {
6199 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6200 ConstantInt::get(CR.getLower()));
6203 if (CR.getLower().isMinValue()) {
6204 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6205 ConstantInt::get(CR.getUpper()));
6206 } else if (CR.getUpper().isMinValue()) {
6207 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6208 ConstantInt::get(CR.getLower()));
6215 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6216 if (ICI.isEquality()) {
6217 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6219 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6220 // the second operand is a constant, simplify a bit.
6221 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6222 switch (BO->getOpcode()) {
6223 case Instruction::SRem:
6224 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6225 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6226 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6227 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6228 Instruction *NewRem =
6229 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
6231 InsertNewInstBefore(NewRem, ICI);
6232 return new ICmpInst(ICI.getPredicate(), NewRem,
6233 Constant::getNullValue(BO->getType()));
6237 case Instruction::Add:
6238 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6239 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6240 if (BO->hasOneUse())
6241 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6242 Subtract(RHS, BOp1C));
6243 } else if (RHSV == 0) {
6244 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6245 // efficiently invertible, or if the add has just this one use.
6246 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6248 if (Value *NegVal = dyn_castNegVal(BOp1))
6249 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6250 else if (Value *NegVal = dyn_castNegVal(BOp0))
6251 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6252 else if (BO->hasOneUse()) {
6253 Instruction *Neg = BinaryOperator::createNeg(BOp1);
6254 InsertNewInstBefore(Neg, ICI);
6256 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6260 case Instruction::Xor:
6261 // For the xor case, we can xor two constants together, eliminating
6262 // the explicit xor.
6263 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6264 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6265 ConstantExpr::getXor(RHS, BOC));
6268 case Instruction::Sub:
6269 // Replace (([sub|xor] A, B) != 0) with (A != B)
6271 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6275 case Instruction::Or:
6276 // If bits are being or'd in that are not present in the constant we
6277 // are comparing against, then the comparison could never succeed!
6278 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6279 Constant *NotCI = ConstantExpr::getNot(RHS);
6280 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6281 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6286 case Instruction::And:
6287 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6288 // If bits are being compared against that are and'd out, then the
6289 // comparison can never succeed!
6290 if ((RHSV & ~BOC->getValue()) != 0)
6291 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6294 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6295 if (RHS == BOC && RHSV.isPowerOf2())
6296 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6297 ICmpInst::ICMP_NE, LHSI,
6298 Constant::getNullValue(RHS->getType()));
6300 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6301 if (isSignBit(BOC)) {
6302 Value *X = BO->getOperand(0);
6303 Constant *Zero = Constant::getNullValue(X->getType());
6304 ICmpInst::Predicate pred = isICMP_NE ?
6305 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6306 return new ICmpInst(pred, X, Zero);
6309 // ((X & ~7) == 0) --> X < 8
6310 if (RHSV == 0 && isHighOnes(BOC)) {
6311 Value *X = BO->getOperand(0);
6312 Constant *NegX = ConstantExpr::getNeg(BOC);
6313 ICmpInst::Predicate pred = isICMP_NE ?
6314 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6315 return new ICmpInst(pred, X, NegX);
6320 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6321 // Handle icmp {eq|ne} <intrinsic>, intcst.
6322 if (II->getIntrinsicID() == Intrinsic::bswap) {
6324 ICI.setOperand(0, II->getOperand(1));
6325 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6329 } else { // Not a ICMP_EQ/ICMP_NE
6330 // If the LHS is a cast from an integral value of the same size,
6331 // then since we know the RHS is a constant, try to simlify.
6332 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6333 Value *CastOp = Cast->getOperand(0);
6334 const Type *SrcTy = CastOp->getType();
6335 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6336 if (SrcTy->isInteger() &&
6337 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6338 // If this is an unsigned comparison, try to make the comparison use
6339 // smaller constant values.
6340 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6341 // X u< 128 => X s> -1
6342 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6343 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6344 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6345 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6346 // X u> 127 => X s< 0
6347 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6348 Constant::getNullValue(SrcTy));
6356 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6357 /// We only handle extending casts so far.
6359 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6360 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6361 Value *LHSCIOp = LHSCI->getOperand(0);
6362 const Type *SrcTy = LHSCIOp->getType();
6363 const Type *DestTy = LHSCI->getType();
6366 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6367 // integer type is the same size as the pointer type.
6368 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6369 getTargetData().getPointerSizeInBits() ==
6370 cast<IntegerType>(DestTy)->getBitWidth()) {
6372 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6373 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6374 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6375 RHSOp = RHSC->getOperand(0);
6376 // If the pointer types don't match, insert a bitcast.
6377 if (LHSCIOp->getType() != RHSOp->getType())
6378 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6382 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6385 // The code below only handles extension cast instructions, so far.
6387 if (LHSCI->getOpcode() != Instruction::ZExt &&
6388 LHSCI->getOpcode() != Instruction::SExt)
6391 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6392 bool isSignedCmp = ICI.isSignedPredicate();
6394 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6395 // Not an extension from the same type?
6396 RHSCIOp = CI->getOperand(0);
6397 if (RHSCIOp->getType() != LHSCIOp->getType())
6400 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6401 // and the other is a zext), then we can't handle this.
6402 if (CI->getOpcode() != LHSCI->getOpcode())
6405 // Deal with equality cases early.
6406 if (ICI.isEquality())
6407 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6409 // A signed comparison of sign extended values simplifies into a
6410 // signed comparison.
6411 if (isSignedCmp && isSignedExt)
6412 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6414 // The other three cases all fold into an unsigned comparison.
6415 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6418 // If we aren't dealing with a constant on the RHS, exit early
6419 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6423 // Compute the constant that would happen if we truncated to SrcTy then
6424 // reextended to DestTy.
6425 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6426 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6428 // If the re-extended constant didn't change...
6430 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6431 // For example, we might have:
6432 // %A = sext short %X to uint
6433 // %B = icmp ugt uint %A, 1330
6434 // It is incorrect to transform this into
6435 // %B = icmp ugt short %X, 1330
6436 // because %A may have negative value.
6438 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6439 // OR operation is EQ/NE.
6440 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6441 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6446 // The re-extended constant changed so the constant cannot be represented
6447 // in the shorter type. Consequently, we cannot emit a simple comparison.
6449 // First, handle some easy cases. We know the result cannot be equal at this
6450 // point so handle the ICI.isEquality() cases
6451 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6452 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6453 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6454 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6456 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6457 // should have been folded away previously and not enter in here.
6460 // We're performing a signed comparison.
6461 if (cast<ConstantInt>(CI)->getValue().isNegative())
6462 Result = ConstantInt::getFalse(); // X < (small) --> false
6464 Result = ConstantInt::getTrue(); // X < (large) --> true
6466 // We're performing an unsigned comparison.
6468 // We're performing an unsigned comp with a sign extended value.
6469 // This is true if the input is >= 0. [aka >s -1]
6470 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6471 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6472 NegOne, ICI.getName()), ICI);
6474 // Unsigned extend & unsigned compare -> always true.
6475 Result = ConstantInt::getTrue();
6479 // Finally, return the value computed.
6480 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6481 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6482 return ReplaceInstUsesWith(ICI, Result);
6484 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6485 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6486 "ICmp should be folded!");
6487 if (Constant *CI = dyn_cast<Constant>(Result))
6488 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6490 return BinaryOperator::createNot(Result);
6494 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6495 return commonShiftTransforms(I);
6498 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6499 return commonShiftTransforms(I);
6502 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6503 if (Instruction *R = commonShiftTransforms(I))
6506 Value *Op0 = I.getOperand(0);
6508 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6509 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6510 if (CSI->isAllOnesValue())
6511 return ReplaceInstUsesWith(I, CSI);
6513 // See if we can turn a signed shr into an unsigned shr.
6514 if (MaskedValueIsZero(Op0,
6515 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6516 return BinaryOperator::createLShr(Op0, I.getOperand(1));
6521 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6522 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6523 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6525 // shl X, 0 == X and shr X, 0 == X
6526 // shl 0, X == 0 and shr 0, X == 0
6527 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6528 Op0 == Constant::getNullValue(Op0->getType()))
6529 return ReplaceInstUsesWith(I, Op0);
6531 if (isa<UndefValue>(Op0)) {
6532 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6533 return ReplaceInstUsesWith(I, Op0);
6534 else // undef << X -> 0, undef >>u X -> 0
6535 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6537 if (isa<UndefValue>(Op1)) {
6538 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6539 return ReplaceInstUsesWith(I, Op0);
6540 else // X << undef, X >>u undef -> 0
6541 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6544 // Try to fold constant and into select arguments.
6545 if (isa<Constant>(Op0))
6546 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6547 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6550 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6551 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6556 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6557 BinaryOperator &I) {
6558 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6560 // See if we can simplify any instructions used by the instruction whose sole
6561 // purpose is to compute bits we don't care about.
6562 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6563 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6564 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6565 KnownZero, KnownOne))
6568 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6569 // of a signed value.
6571 if (Op1->uge(TypeBits)) {
6572 if (I.getOpcode() != Instruction::AShr)
6573 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6575 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6580 // ((X*C1) << C2) == (X * (C1 << C2))
6581 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6582 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6583 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6584 return BinaryOperator::createMul(BO->getOperand(0),
6585 ConstantExpr::getShl(BOOp, Op1));
6587 // Try to fold constant and into select arguments.
6588 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6589 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6591 if (isa<PHINode>(Op0))
6592 if (Instruction *NV = FoldOpIntoPhi(I))
6595 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6596 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6597 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6598 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6599 // place. Don't try to do this transformation in this case. Also, we
6600 // require that the input operand is a shift-by-constant so that we have
6601 // confidence that the shifts will get folded together. We could do this
6602 // xform in more cases, but it is unlikely to be profitable.
6603 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6604 isa<ConstantInt>(TrOp->getOperand(1))) {
6605 // Okay, we'll do this xform. Make the shift of shift.
6606 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6607 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6609 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6611 // For logical shifts, the truncation has the effect of making the high
6612 // part of the register be zeros. Emulate this by inserting an AND to
6613 // clear the top bits as needed. This 'and' will usually be zapped by
6614 // other xforms later if dead.
6615 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6616 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6617 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6619 // The mask we constructed says what the trunc would do if occurring
6620 // between the shifts. We want to know the effect *after* the second
6621 // shift. We know that it is a logical shift by a constant, so adjust the
6622 // mask as appropriate.
6623 if (I.getOpcode() == Instruction::Shl)
6624 MaskV <<= Op1->getZExtValue();
6626 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6627 MaskV = MaskV.lshr(Op1->getZExtValue());
6630 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6632 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6634 // Return the value truncated to the interesting size.
6635 return new TruncInst(And, I.getType());
6639 if (Op0->hasOneUse()) {
6640 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6641 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6644 switch (Op0BO->getOpcode()) {
6646 case Instruction::Add:
6647 case Instruction::And:
6648 case Instruction::Or:
6649 case Instruction::Xor: {
6650 // These operators commute.
6651 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6652 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6653 match(Op0BO->getOperand(1),
6654 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6655 Instruction *YS = BinaryOperator::createShl(
6656 Op0BO->getOperand(0), Op1,
6658 InsertNewInstBefore(YS, I); // (Y << C)
6660 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6661 Op0BO->getOperand(1)->getName());
6662 InsertNewInstBefore(X, I); // (X + (Y << C))
6663 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6664 return BinaryOperator::createAnd(X, ConstantInt::get(
6665 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6668 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6669 Value *Op0BOOp1 = Op0BO->getOperand(1);
6670 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6672 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6673 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6675 Instruction *YS = BinaryOperator::createShl(
6676 Op0BO->getOperand(0), Op1,
6678 InsertNewInstBefore(YS, I); // (Y << C)
6680 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6681 V1->getName()+".mask");
6682 InsertNewInstBefore(XM, I); // X & (CC << C)
6684 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6689 case Instruction::Sub: {
6690 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6691 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6692 match(Op0BO->getOperand(0),
6693 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6694 Instruction *YS = BinaryOperator::createShl(
6695 Op0BO->getOperand(1), Op1,
6697 InsertNewInstBefore(YS, I); // (Y << C)
6699 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6700 Op0BO->getOperand(0)->getName());
6701 InsertNewInstBefore(X, I); // (X + (Y << C))
6702 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6703 return BinaryOperator::createAnd(X, ConstantInt::get(
6704 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6707 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6708 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6709 match(Op0BO->getOperand(0),
6710 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6711 m_ConstantInt(CC))) && V2 == Op1 &&
6712 cast<BinaryOperator>(Op0BO->getOperand(0))
6713 ->getOperand(0)->hasOneUse()) {
6714 Instruction *YS = BinaryOperator::createShl(
6715 Op0BO->getOperand(1), Op1,
6717 InsertNewInstBefore(YS, I); // (Y << C)
6719 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6720 V1->getName()+".mask");
6721 InsertNewInstBefore(XM, I); // X & (CC << C)
6723 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6731 // If the operand is an bitwise operator with a constant RHS, and the
6732 // shift is the only use, we can pull it out of the shift.
6733 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6734 bool isValid = true; // Valid only for And, Or, Xor
6735 bool highBitSet = false; // Transform if high bit of constant set?
6737 switch (Op0BO->getOpcode()) {
6738 default: isValid = false; break; // Do not perform transform!
6739 case Instruction::Add:
6740 isValid = isLeftShift;
6742 case Instruction::Or:
6743 case Instruction::Xor:
6746 case Instruction::And:
6751 // If this is a signed shift right, and the high bit is modified
6752 // by the logical operation, do not perform the transformation.
6753 // The highBitSet boolean indicates the value of the high bit of
6754 // the constant which would cause it to be modified for this
6757 if (isValid && I.getOpcode() == Instruction::AShr)
6758 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6761 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6763 Instruction *NewShift =
6764 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6765 InsertNewInstBefore(NewShift, I);
6766 NewShift->takeName(Op0BO);
6768 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6775 // Find out if this is a shift of a shift by a constant.
6776 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6777 if (ShiftOp && !ShiftOp->isShift())
6780 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6781 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6782 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6783 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6784 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6785 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6786 Value *X = ShiftOp->getOperand(0);
6788 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6789 if (AmtSum > TypeBits)
6792 const IntegerType *Ty = cast<IntegerType>(I.getType());
6794 // Check for (X << c1) << c2 and (X >> c1) >> c2
6795 if (I.getOpcode() == ShiftOp->getOpcode()) {
6796 return BinaryOperator::create(I.getOpcode(), X,
6797 ConstantInt::get(Ty, AmtSum));
6798 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6799 I.getOpcode() == Instruction::AShr) {
6800 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6801 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6802 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6803 I.getOpcode() == Instruction::LShr) {
6804 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6805 Instruction *Shift =
6806 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6807 InsertNewInstBefore(Shift, I);
6809 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6810 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6813 // Okay, if we get here, one shift must be left, and the other shift must be
6814 // right. See if the amounts are equal.
6815 if (ShiftAmt1 == ShiftAmt2) {
6816 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6817 if (I.getOpcode() == Instruction::Shl) {
6818 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6819 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6821 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6822 if (I.getOpcode() == Instruction::LShr) {
6823 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6824 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6826 // We can simplify ((X << C) >>s C) into a trunc + sext.
6827 // NOTE: we could do this for any C, but that would make 'unusual' integer
6828 // types. For now, just stick to ones well-supported by the code
6830 const Type *SExtType = 0;
6831 switch (Ty->getBitWidth() - ShiftAmt1) {
6838 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6843 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6844 InsertNewInstBefore(NewTrunc, I);
6845 return new SExtInst(NewTrunc, Ty);
6847 // Otherwise, we can't handle it yet.
6848 } else if (ShiftAmt1 < ShiftAmt2) {
6849 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6851 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6852 if (I.getOpcode() == Instruction::Shl) {
6853 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6854 ShiftOp->getOpcode() == Instruction::AShr);
6855 Instruction *Shift =
6856 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6857 InsertNewInstBefore(Shift, I);
6859 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6860 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6863 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6864 if (I.getOpcode() == Instruction::LShr) {
6865 assert(ShiftOp->getOpcode() == Instruction::Shl);
6866 Instruction *Shift =
6867 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6868 InsertNewInstBefore(Shift, I);
6870 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6871 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6874 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6876 assert(ShiftAmt2 < ShiftAmt1);
6877 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6879 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6880 if (I.getOpcode() == Instruction::Shl) {
6881 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6882 ShiftOp->getOpcode() == Instruction::AShr);
6883 Instruction *Shift =
6884 BinaryOperator::create(ShiftOp->getOpcode(), X,
6885 ConstantInt::get(Ty, ShiftDiff));
6886 InsertNewInstBefore(Shift, I);
6888 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6889 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6892 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6893 if (I.getOpcode() == Instruction::LShr) {
6894 assert(ShiftOp->getOpcode() == Instruction::Shl);
6895 Instruction *Shift =
6896 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6897 InsertNewInstBefore(Shift, I);
6899 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6900 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6903 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6910 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6911 /// expression. If so, decompose it, returning some value X, such that Val is
6914 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6916 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6917 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6918 Offset = CI->getZExtValue();
6920 return ConstantInt::get(Type::Int32Ty, 0);
6921 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6922 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6923 if (I->getOpcode() == Instruction::Shl) {
6924 // This is a value scaled by '1 << the shift amt'.
6925 Scale = 1U << RHS->getZExtValue();
6927 return I->getOperand(0);
6928 } else if (I->getOpcode() == Instruction::Mul) {
6929 // This value is scaled by 'RHS'.
6930 Scale = RHS->getZExtValue();
6932 return I->getOperand(0);
6933 } else if (I->getOpcode() == Instruction::Add) {
6934 // We have X+C. Check to see if we really have (X*C2)+C1,
6935 // where C1 is divisible by C2.
6938 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6939 Offset += RHS->getZExtValue();
6946 // Otherwise, we can't look past this.
6953 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6954 /// try to eliminate the cast by moving the type information into the alloc.
6955 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6956 AllocationInst &AI) {
6957 const PointerType *PTy = cast<PointerType>(CI.getType());
6959 // Remove any uses of AI that are dead.
6960 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6962 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6963 Instruction *User = cast<Instruction>(*UI++);
6964 if (isInstructionTriviallyDead(User)) {
6965 while (UI != E && *UI == User)
6966 ++UI; // If this instruction uses AI more than once, don't break UI.
6969 DOUT << "IC: DCE: " << *User;
6970 EraseInstFromFunction(*User);
6974 // Get the type really allocated and the type casted to.
6975 const Type *AllocElTy = AI.getAllocatedType();
6976 const Type *CastElTy = PTy->getElementType();
6977 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6979 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6980 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6981 if (CastElTyAlign < AllocElTyAlign) return 0;
6983 // If the allocation has multiple uses, only promote it if we are strictly
6984 // increasing the alignment of the resultant allocation. If we keep it the
6985 // same, we open the door to infinite loops of various kinds.
6986 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6988 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6989 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6990 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6992 // See if we can satisfy the modulus by pulling a scale out of the array
6994 unsigned ArraySizeScale;
6996 Value *NumElements = // See if the array size is a decomposable linear expr.
6997 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6999 // If we can now satisfy the modulus, by using a non-1 scale, we really can
7001 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
7002 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
7004 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
7009 // If the allocation size is constant, form a constant mul expression
7010 Amt = ConstantInt::get(Type::Int32Ty, Scale);
7011 if (isa<ConstantInt>(NumElements))
7012 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
7013 // otherwise multiply the amount and the number of elements
7014 else if (Scale != 1) {
7015 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
7016 Amt = InsertNewInstBefore(Tmp, AI);
7020 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
7021 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
7022 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
7023 Amt = InsertNewInstBefore(Tmp, AI);
7026 AllocationInst *New;
7027 if (isa<MallocInst>(AI))
7028 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
7030 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
7031 InsertNewInstBefore(New, AI);
7034 // If the allocation has multiple uses, insert a cast and change all things
7035 // that used it to use the new cast. This will also hack on CI, but it will
7037 if (!AI.hasOneUse()) {
7038 AddUsesToWorkList(AI);
7039 // New is the allocation instruction, pointer typed. AI is the original
7040 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
7041 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
7042 InsertNewInstBefore(NewCast, AI);
7043 AI.replaceAllUsesWith(NewCast);
7045 return ReplaceInstUsesWith(CI, New);
7048 /// CanEvaluateInDifferentType - Return true if we can take the specified value
7049 /// and return it as type Ty without inserting any new casts and without
7050 /// changing the computed value. This is used by code that tries to decide
7051 /// whether promoting or shrinking integer operations to wider or smaller types
7052 /// will allow us to eliminate a truncate or extend.
7054 /// This is a truncation operation if Ty is smaller than V->getType(), or an
7055 /// extension operation if Ty is larger.
7056 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
7058 int &NumCastsRemoved) {
7059 // We can always evaluate constants in another type.
7060 if (isa<ConstantInt>(V))
7063 Instruction *I = dyn_cast<Instruction>(V);
7064 if (!I) return false;
7066 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
7068 // If this is an extension or truncate, we can often eliminate it.
7069 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
7070 // If this is a cast from the destination type, we can trivially eliminate
7071 // it, and this will remove a cast overall.
7072 if (I->getOperand(0)->getType() == Ty) {
7073 // If the first operand is itself a cast, and is eliminable, do not count
7074 // this as an eliminable cast. We would prefer to eliminate those two
7076 if (!isa<CastInst>(I->getOperand(0)))
7082 // We can't extend or shrink something that has multiple uses: doing so would
7083 // require duplicating the instruction in general, which isn't profitable.
7084 if (!I->hasOneUse()) return false;
7086 switch (I->getOpcode()) {
7087 case Instruction::Add:
7088 case Instruction::Sub:
7089 case Instruction::And:
7090 case Instruction::Or:
7091 case Instruction::Xor:
7092 // These operators can all arbitrarily be extended or truncated.
7093 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7095 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7098 case Instruction::Mul:
7099 // A multiply can be truncated by truncating its operands.
7100 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
7101 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7103 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7106 case Instruction::Shl:
7107 // If we are truncating the result of this SHL, and if it's a shift of a
7108 // constant amount, we can always perform a SHL in a smaller type.
7109 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7110 uint32_t BitWidth = Ty->getBitWidth();
7111 if (BitWidth < OrigTy->getBitWidth() &&
7112 CI->getLimitedValue(BitWidth) < BitWidth)
7113 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7117 case Instruction::LShr:
7118 // If this is a truncate of a logical shr, we can truncate it to a smaller
7119 // lshr iff we know that the bits we would otherwise be shifting in are
7121 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7122 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7123 uint32_t BitWidth = Ty->getBitWidth();
7124 if (BitWidth < OrigBitWidth &&
7125 MaskedValueIsZero(I->getOperand(0),
7126 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7127 CI->getLimitedValue(BitWidth) < BitWidth) {
7128 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7133 case Instruction::ZExt:
7134 case Instruction::SExt:
7135 case Instruction::Trunc:
7136 // If this is the same kind of case as our original (e.g. zext+zext), we
7137 // can safely replace it. Note that replacing it does not reduce the number
7138 // of casts in the input.
7139 if (I->getOpcode() == CastOpc)
7144 // TODO: Can handle more cases here.
7151 /// EvaluateInDifferentType - Given an expression that
7152 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7153 /// evaluate the expression.
7154 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7156 if (Constant *C = dyn_cast<Constant>(V))
7157 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7159 // Otherwise, it must be an instruction.
7160 Instruction *I = cast<Instruction>(V);
7161 Instruction *Res = 0;
7162 switch (I->getOpcode()) {
7163 case Instruction::Add:
7164 case Instruction::Sub:
7165 case Instruction::Mul:
7166 case Instruction::And:
7167 case Instruction::Or:
7168 case Instruction::Xor:
7169 case Instruction::AShr:
7170 case Instruction::LShr:
7171 case Instruction::Shl: {
7172 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7173 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7174 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
7175 LHS, RHS, I->getName());
7178 case Instruction::Trunc:
7179 case Instruction::ZExt:
7180 case Instruction::SExt:
7181 // If the source type of the cast is the type we're trying for then we can
7182 // just return the source. There's no need to insert it because it is not
7184 if (I->getOperand(0)->getType() == Ty)
7185 return I->getOperand(0);
7187 // Otherwise, must be the same type of case, so just reinsert a new one.
7188 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7192 // TODO: Can handle more cases here.
7193 assert(0 && "Unreachable!");
7197 return InsertNewInstBefore(Res, *I);
7200 /// @brief Implement the transforms common to all CastInst visitors.
7201 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7202 Value *Src = CI.getOperand(0);
7204 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7205 // eliminate it now.
7206 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7207 if (Instruction::CastOps opc =
7208 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7209 // The first cast (CSrc) is eliminable so we need to fix up or replace
7210 // the second cast (CI). CSrc will then have a good chance of being dead.
7211 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
7215 // If we are casting a select then fold the cast into the select
7216 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7217 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7220 // If we are casting a PHI then fold the cast into the PHI
7221 if (isa<PHINode>(Src))
7222 if (Instruction *NV = FoldOpIntoPhi(CI))
7228 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7229 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7230 Value *Src = CI.getOperand(0);
7232 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7233 // If casting the result of a getelementptr instruction with no offset, turn
7234 // this into a cast of the original pointer!
7235 if (GEP->hasAllZeroIndices()) {
7236 // Changing the cast operand is usually not a good idea but it is safe
7237 // here because the pointer operand is being replaced with another
7238 // pointer operand so the opcode doesn't need to change.
7240 CI.setOperand(0, GEP->getOperand(0));
7244 // If the GEP has a single use, and the base pointer is a bitcast, and the
7245 // GEP computes a constant offset, see if we can convert these three
7246 // instructions into fewer. This typically happens with unions and other
7247 // non-type-safe code.
7248 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7249 if (GEP->hasAllConstantIndices()) {
7250 // We are guaranteed to get a constant from EmitGEPOffset.
7251 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7252 int64_t Offset = OffsetV->getSExtValue();
7254 // Get the base pointer input of the bitcast, and the type it points to.
7255 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7256 const Type *GEPIdxTy =
7257 cast<PointerType>(OrigBase->getType())->getElementType();
7258 if (GEPIdxTy->isSized()) {
7259 SmallVector<Value*, 8> NewIndices;
7261 // Start with the index over the outer type. Note that the type size
7262 // might be zero (even if the offset isn't zero) if the indexed type
7263 // is something like [0 x {int, int}]
7264 const Type *IntPtrTy = TD->getIntPtrType();
7265 int64_t FirstIdx = 0;
7266 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7267 FirstIdx = Offset/TySize;
7270 // Handle silly modulus not returning values values [0..TySize).
7274 assert(Offset >= 0);
7276 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7279 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7281 // Index into the types. If we fail, set OrigBase to null.
7283 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7284 const StructLayout *SL = TD->getStructLayout(STy);
7285 if (Offset < (int64_t)SL->getSizeInBytes()) {
7286 unsigned Elt = SL->getElementContainingOffset(Offset);
7287 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7289 Offset -= SL->getElementOffset(Elt);
7290 GEPIdxTy = STy->getElementType(Elt);
7292 // Otherwise, we can't index into this, bail out.
7296 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7297 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7298 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7299 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7302 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7304 GEPIdxTy = STy->getElementType();
7306 // Otherwise, we can't index into this, bail out.
7312 // If we were able to index down into an element, create the GEP
7313 // and bitcast the result. This eliminates one bitcast, potentially
7315 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7317 NewIndices.end(), "");
7318 InsertNewInstBefore(NGEP, CI);
7319 NGEP->takeName(GEP);
7321 if (isa<BitCastInst>(CI))
7322 return new BitCastInst(NGEP, CI.getType());
7323 assert(isa<PtrToIntInst>(CI));
7324 return new PtrToIntInst(NGEP, CI.getType());
7331 return commonCastTransforms(CI);
7336 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7337 /// integer types. This function implements the common transforms for all those
7339 /// @brief Implement the transforms common to CastInst with integer operands
7340 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7341 if (Instruction *Result = commonCastTransforms(CI))
7344 Value *Src = CI.getOperand(0);
7345 const Type *SrcTy = Src->getType();
7346 const Type *DestTy = CI.getType();
7347 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7348 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7350 // See if we can simplify any instructions used by the LHS whose sole
7351 // purpose is to compute bits we don't care about.
7352 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7353 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7354 KnownZero, KnownOne))
7357 // If the source isn't an instruction or has more than one use then we
7358 // can't do anything more.
7359 Instruction *SrcI = dyn_cast<Instruction>(Src);
7360 if (!SrcI || !Src->hasOneUse())
7363 // Attempt to propagate the cast into the instruction for int->int casts.
7364 int NumCastsRemoved = 0;
7365 if (!isa<BitCastInst>(CI) &&
7366 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7367 CI.getOpcode(), NumCastsRemoved)) {
7368 // If this cast is a truncate, evaluting in a different type always
7369 // eliminates the cast, so it is always a win. If this is a zero-extension,
7370 // we need to do an AND to maintain the clear top-part of the computation,
7371 // so we require that the input have eliminated at least one cast. If this
7372 // is a sign extension, we insert two new casts (to do the extension) so we
7373 // require that two casts have been eliminated.
7375 switch (CI.getOpcode()) {
7377 // All the others use floating point so we shouldn't actually
7378 // get here because of the check above.
7379 assert(0 && "Unknown cast type");
7380 case Instruction::Trunc:
7383 case Instruction::ZExt:
7384 DoXForm = NumCastsRemoved >= 1;
7386 case Instruction::SExt:
7387 DoXForm = NumCastsRemoved >= 2;
7392 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7393 CI.getOpcode() == Instruction::SExt);
7394 assert(Res->getType() == DestTy);
7395 switch (CI.getOpcode()) {
7396 default: assert(0 && "Unknown cast type!");
7397 case Instruction::Trunc:
7398 case Instruction::BitCast:
7399 // Just replace this cast with the result.
7400 return ReplaceInstUsesWith(CI, Res);
7401 case Instruction::ZExt: {
7402 // We need to emit an AND to clear the high bits.
7403 assert(SrcBitSize < DestBitSize && "Not a zext?");
7404 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7406 return BinaryOperator::createAnd(Res, C);
7408 case Instruction::SExt:
7409 // We need to emit a cast to truncate, then a cast to sext.
7410 return CastInst::create(Instruction::SExt,
7411 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7417 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7418 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7420 switch (SrcI->getOpcode()) {
7421 case Instruction::Add:
7422 case Instruction::Mul:
7423 case Instruction::And:
7424 case Instruction::Or:
7425 case Instruction::Xor:
7426 // If we are discarding information, rewrite.
7427 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7428 // Don't insert two casts if they cannot be eliminated. We allow
7429 // two casts to be inserted if the sizes are the same. This could
7430 // only be converting signedness, which is a noop.
7431 if (DestBitSize == SrcBitSize ||
7432 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7433 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7434 Instruction::CastOps opcode = CI.getOpcode();
7435 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7436 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7437 return BinaryOperator::create(
7438 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7442 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7443 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7444 SrcI->getOpcode() == Instruction::Xor &&
7445 Op1 == ConstantInt::getTrue() &&
7446 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7447 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7448 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
7451 case Instruction::SDiv:
7452 case Instruction::UDiv:
7453 case Instruction::SRem:
7454 case Instruction::URem:
7455 // If we are just changing the sign, rewrite.
7456 if (DestBitSize == SrcBitSize) {
7457 // Don't insert two casts if they cannot be eliminated. We allow
7458 // two casts to be inserted if the sizes are the same. This could
7459 // only be converting signedness, which is a noop.
7460 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7461 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7462 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7464 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7466 return BinaryOperator::create(
7467 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7472 case Instruction::Shl:
7473 // Allow changing the sign of the source operand. Do not allow
7474 // changing the size of the shift, UNLESS the shift amount is a
7475 // constant. We must not change variable sized shifts to a smaller
7476 // size, because it is undefined to shift more bits out than exist
7478 if (DestBitSize == SrcBitSize ||
7479 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7480 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7481 Instruction::BitCast : Instruction::Trunc);
7482 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7483 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7484 return BinaryOperator::createShl(Op0c, Op1c);
7487 case Instruction::AShr:
7488 // If this is a signed shr, and if all bits shifted in are about to be
7489 // truncated off, turn it into an unsigned shr to allow greater
7491 if (DestBitSize < SrcBitSize &&
7492 isa<ConstantInt>(Op1)) {
7493 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7494 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7495 // Insert the new logical shift right.
7496 return BinaryOperator::createLShr(Op0, Op1);
7504 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7505 if (Instruction *Result = commonIntCastTransforms(CI))
7508 Value *Src = CI.getOperand(0);
7509 const Type *Ty = CI.getType();
7510 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7511 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7513 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7514 switch (SrcI->getOpcode()) {
7516 case Instruction::LShr:
7517 // We can shrink lshr to something smaller if we know the bits shifted in
7518 // are already zeros.
7519 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7520 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7522 // Get a mask for the bits shifting in.
7523 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7524 Value* SrcIOp0 = SrcI->getOperand(0);
7525 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7526 if (ShAmt >= DestBitWidth) // All zeros.
7527 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7529 // Okay, we can shrink this. Truncate the input, then return a new
7531 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7532 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7534 return BinaryOperator::createLShr(V1, V2);
7536 } else { // This is a variable shr.
7538 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7539 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7540 // loop-invariant and CSE'd.
7541 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7542 Value *One = ConstantInt::get(SrcI->getType(), 1);
7544 Value *V = InsertNewInstBefore(
7545 BinaryOperator::createShl(One, SrcI->getOperand(1),
7547 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7548 SrcI->getOperand(0),
7550 Value *Zero = Constant::getNullValue(V->getType());
7551 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7561 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7562 /// in order to eliminate the icmp.
7563 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7565 // If we are just checking for a icmp eq of a single bit and zext'ing it
7566 // to an integer, then shift the bit to the appropriate place and then
7567 // cast to integer to avoid the comparison.
7568 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7569 const APInt &Op1CV = Op1C->getValue();
7571 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7572 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7573 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7574 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7575 if (!DoXform) return ICI;
7577 Value *In = ICI->getOperand(0);
7578 Value *Sh = ConstantInt::get(In->getType(),
7579 In->getType()->getPrimitiveSizeInBits()-1);
7580 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7581 In->getName()+".lobit"),
7583 if (In->getType() != CI.getType())
7584 In = CastInst::createIntegerCast(In, CI.getType(),
7585 false/*ZExt*/, "tmp", &CI);
7587 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7588 Constant *One = ConstantInt::get(In->getType(), 1);
7589 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7590 In->getName()+".not"),
7594 return ReplaceInstUsesWith(CI, In);
7599 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7600 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7601 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7602 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7603 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7604 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7605 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7606 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7607 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7608 // This only works for EQ and NE
7609 ICI->isEquality()) {
7610 // If Op1C some other power of two, convert:
7611 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7612 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7613 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7614 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7616 APInt KnownZeroMask(~KnownZero);
7617 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7618 if (!DoXform) return ICI;
7620 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7621 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7622 // (X&4) == 2 --> false
7623 // (X&4) != 2 --> true
7624 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7625 Res = ConstantExpr::getZExt(Res, CI.getType());
7626 return ReplaceInstUsesWith(CI, Res);
7629 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7630 Value *In = ICI->getOperand(0);
7632 // Perform a logical shr by shiftamt.
7633 // Insert the shift to put the result in the low bit.
7634 In = InsertNewInstBefore(BinaryOperator::createLShr(In,
7635 ConstantInt::get(In->getType(), ShiftAmt),
7636 In->getName()+".lobit"), CI);
7639 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7640 Constant *One = ConstantInt::get(In->getType(), 1);
7641 In = BinaryOperator::createXor(In, One, "tmp");
7642 InsertNewInstBefore(cast<Instruction>(In), CI);
7645 if (CI.getType() == In->getType())
7646 return ReplaceInstUsesWith(CI, In);
7648 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7656 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7657 // If one of the common conversion will work ..
7658 if (Instruction *Result = commonIntCastTransforms(CI))
7661 Value *Src = CI.getOperand(0);
7663 // If this is a cast of a cast
7664 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7665 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7666 // types and if the sizes are just right we can convert this into a logical
7667 // 'and' which will be much cheaper than the pair of casts.
7668 if (isa<TruncInst>(CSrc)) {
7669 // Get the sizes of the types involved
7670 Value *A = CSrc->getOperand(0);
7671 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7672 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7673 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7674 // If we're actually extending zero bits and the trunc is a no-op
7675 if (MidSize < DstSize && SrcSize == DstSize) {
7676 // Replace both of the casts with an And of the type mask.
7677 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7678 Constant *AndConst = ConstantInt::get(AndValue);
7680 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7681 // Unfortunately, if the type changed, we need to cast it back.
7682 if (And->getType() != CI.getType()) {
7683 And->setName(CSrc->getName()+".mask");
7684 InsertNewInstBefore(And, CI);
7685 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7692 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7693 return transformZExtICmp(ICI, CI);
7695 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7696 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7697 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7698 // of the (zext icmp) will be transformed.
7699 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7700 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7701 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7702 (transformZExtICmp(LHS, CI, false) ||
7703 transformZExtICmp(RHS, CI, false))) {
7704 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7705 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7706 return BinaryOperator::create(Instruction::Or, LCast, RCast);
7713 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7714 if (Instruction *I = commonIntCastTransforms(CI))
7717 Value *Src = CI.getOperand(0);
7719 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7720 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7721 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7722 // If we are just checking for a icmp eq of a single bit and zext'ing it
7723 // to an integer, then shift the bit to the appropriate place and then
7724 // cast to integer to avoid the comparison.
7725 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7726 const APInt &Op1CV = Op1C->getValue();
7728 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7729 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7730 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7731 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7732 Value *In = ICI->getOperand(0);
7733 Value *Sh = ConstantInt::get(In->getType(),
7734 In->getType()->getPrimitiveSizeInBits()-1);
7735 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7736 In->getName()+".lobit"),
7738 if (In->getType() != CI.getType())
7739 In = CastInst::createIntegerCast(In, CI.getType(),
7740 true/*SExt*/, "tmp", &CI);
7742 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7743 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7744 In->getName()+".not"), CI);
7746 return ReplaceInstUsesWith(CI, In);
7754 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7755 /// in the specified FP type without changing its value.
7756 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7757 APFloat F = CFP->getValueAPF();
7758 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7759 return ConstantFP::get(F);
7763 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7764 /// through it until we get the source value.
7765 static Value *LookThroughFPExtensions(Value *V) {
7766 if (Instruction *I = dyn_cast<Instruction>(V))
7767 if (I->getOpcode() == Instruction::FPExt)
7768 return LookThroughFPExtensions(I->getOperand(0));
7770 // If this value is a constant, return the constant in the smallest FP type
7771 // that can accurately represent it. This allows us to turn
7772 // (float)((double)X+2.0) into x+2.0f.
7773 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7774 if (CFP->getType() == Type::PPC_FP128Ty)
7775 return V; // No constant folding of this.
7776 // See if the value can be truncated to float and then reextended.
7777 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7779 if (CFP->getType() == Type::DoubleTy)
7780 return V; // Won't shrink.
7781 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7783 // Don't try to shrink to various long double types.
7789 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7790 if (Instruction *I = commonCastTransforms(CI))
7793 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7794 // smaller than the destination type, we can eliminate the truncate by doing
7795 // the add as the smaller type. This applies to add/sub/mul/div as well as
7796 // many builtins (sqrt, etc).
7797 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7798 if (OpI && OpI->hasOneUse()) {
7799 switch (OpI->getOpcode()) {
7801 case Instruction::Add:
7802 case Instruction::Sub:
7803 case Instruction::Mul:
7804 case Instruction::FDiv:
7805 case Instruction::FRem:
7806 const Type *SrcTy = OpI->getType();
7807 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7808 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7809 if (LHSTrunc->getType() != SrcTy &&
7810 RHSTrunc->getType() != SrcTy) {
7811 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7812 // If the source types were both smaller than the destination type of
7813 // the cast, do this xform.
7814 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7815 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7816 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7818 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7820 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7829 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7830 return commonCastTransforms(CI);
7833 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7834 return commonCastTransforms(CI);
7837 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7838 return commonCastTransforms(CI);
7841 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7842 return commonCastTransforms(CI);
7845 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7846 return commonCastTransforms(CI);
7849 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7850 return commonPointerCastTransforms(CI);
7853 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7854 if (Instruction *I = commonCastTransforms(CI))
7857 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7858 if (!DestPointee->isSized()) return 0;
7860 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7863 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7864 m_ConstantInt(Cst)))) {
7865 // If the source and destination operands have the same type, see if this
7866 // is a single-index GEP.
7867 if (X->getType() == CI.getType()) {
7868 // Get the size of the pointee type.
7869 uint64_t Size = TD->getABITypeSize(DestPointee);
7871 // Convert the constant to intptr type.
7872 APInt Offset = Cst->getValue();
7873 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7875 // If Offset is evenly divisible by Size, we can do this xform.
7876 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7877 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7878 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7881 // TODO: Could handle other cases, e.g. where add is indexing into field of
7883 } else if (CI.getOperand(0)->hasOneUse() &&
7884 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7885 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7886 // "inttoptr+GEP" instead of "add+intptr".
7888 // Get the size of the pointee type.
7889 uint64_t Size = TD->getABITypeSize(DestPointee);
7891 // Convert the constant to intptr type.
7892 APInt Offset = Cst->getValue();
7893 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7895 // If Offset is evenly divisible by Size, we can do this xform.
7896 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7897 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7899 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7901 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7907 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7908 // If the operands are integer typed then apply the integer transforms,
7909 // otherwise just apply the common ones.
7910 Value *Src = CI.getOperand(0);
7911 const Type *SrcTy = Src->getType();
7912 const Type *DestTy = CI.getType();
7914 if (SrcTy->isInteger() && DestTy->isInteger()) {
7915 if (Instruction *Result = commonIntCastTransforms(CI))
7917 } else if (isa<PointerType>(SrcTy)) {
7918 if (Instruction *I = commonPointerCastTransforms(CI))
7921 if (Instruction *Result = commonCastTransforms(CI))
7926 // Get rid of casts from one type to the same type. These are useless and can
7927 // be replaced by the operand.
7928 if (DestTy == Src->getType())
7929 return ReplaceInstUsesWith(CI, Src);
7931 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7932 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7933 const Type *DstElTy = DstPTy->getElementType();
7934 const Type *SrcElTy = SrcPTy->getElementType();
7936 // If the address spaces don't match, don't eliminate the bitcast, which is
7937 // required for changing types.
7938 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
7941 // If we are casting a malloc or alloca to a pointer to a type of the same
7942 // size, rewrite the allocation instruction to allocate the "right" type.
7943 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7944 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7947 // If the source and destination are pointers, and this cast is equivalent
7948 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7949 // This can enhance SROA and other transforms that want type-safe pointers.
7950 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7951 unsigned NumZeros = 0;
7952 while (SrcElTy != DstElTy &&
7953 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7954 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7955 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7959 // If we found a path from the src to dest, create the getelementptr now.
7960 if (SrcElTy == DstElTy) {
7961 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7962 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
7963 ((Instruction*) NULL));
7967 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7968 if (SVI->hasOneUse()) {
7969 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7970 // a bitconvert to a vector with the same # elts.
7971 if (isa<VectorType>(DestTy) &&
7972 cast<VectorType>(DestTy)->getNumElements() ==
7973 SVI->getType()->getNumElements()) {
7975 // If either of the operands is a cast from CI.getType(), then
7976 // evaluating the shuffle in the casted destination's type will allow
7977 // us to eliminate at least one cast.
7978 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7979 Tmp->getOperand(0)->getType() == DestTy) ||
7980 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7981 Tmp->getOperand(0)->getType() == DestTy)) {
7982 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7983 SVI->getOperand(0), DestTy, &CI);
7984 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7985 SVI->getOperand(1), DestTy, &CI);
7986 // Return a new shuffle vector. Use the same element ID's, as we
7987 // know the vector types match #elts.
7988 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7996 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7998 /// %D = select %cond, %C, %A
8000 /// %C = select %cond, %B, 0
8003 /// Assuming that the specified instruction is an operand to the select, return
8004 /// a bitmask indicating which operands of this instruction are foldable if they
8005 /// equal the other incoming value of the select.
8007 static unsigned GetSelectFoldableOperands(Instruction *I) {
8008 switch (I->getOpcode()) {
8009 case Instruction::Add:
8010 case Instruction::Mul:
8011 case Instruction::And:
8012 case Instruction::Or:
8013 case Instruction::Xor:
8014 return 3; // Can fold through either operand.
8015 case Instruction::Sub: // Can only fold on the amount subtracted.
8016 case Instruction::Shl: // Can only fold on the shift amount.
8017 case Instruction::LShr:
8018 case Instruction::AShr:
8021 return 0; // Cannot fold
8025 /// GetSelectFoldableConstant - For the same transformation as the previous
8026 /// function, return the identity constant that goes into the select.
8027 static Constant *GetSelectFoldableConstant(Instruction *I) {
8028 switch (I->getOpcode()) {
8029 default: assert(0 && "This cannot happen!"); abort();
8030 case Instruction::Add:
8031 case Instruction::Sub:
8032 case Instruction::Or:
8033 case Instruction::Xor:
8034 case Instruction::Shl:
8035 case Instruction::LShr:
8036 case Instruction::AShr:
8037 return Constant::getNullValue(I->getType());
8038 case Instruction::And:
8039 return Constant::getAllOnesValue(I->getType());
8040 case Instruction::Mul:
8041 return ConstantInt::get(I->getType(), 1);
8045 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
8046 /// have the same opcode and only one use each. Try to simplify this.
8047 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
8049 if (TI->getNumOperands() == 1) {
8050 // If this is a non-volatile load or a cast from the same type,
8053 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
8056 return 0; // unknown unary op.
8059 // Fold this by inserting a select from the input values.
8060 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
8061 FI->getOperand(0), SI.getName()+".v");
8062 InsertNewInstBefore(NewSI, SI);
8063 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
8067 // Only handle binary operators here.
8068 if (!isa<BinaryOperator>(TI))
8071 // Figure out if the operations have any operands in common.
8072 Value *MatchOp, *OtherOpT, *OtherOpF;
8074 if (TI->getOperand(0) == FI->getOperand(0)) {
8075 MatchOp = TI->getOperand(0);
8076 OtherOpT = TI->getOperand(1);
8077 OtherOpF = FI->getOperand(1);
8078 MatchIsOpZero = true;
8079 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8080 MatchOp = TI->getOperand(1);
8081 OtherOpT = TI->getOperand(0);
8082 OtherOpF = FI->getOperand(0);
8083 MatchIsOpZero = false;
8084 } else if (!TI->isCommutative()) {
8086 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8087 MatchOp = TI->getOperand(0);
8088 OtherOpT = TI->getOperand(1);
8089 OtherOpF = FI->getOperand(0);
8090 MatchIsOpZero = true;
8091 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8092 MatchOp = TI->getOperand(1);
8093 OtherOpT = TI->getOperand(0);
8094 OtherOpF = FI->getOperand(1);
8095 MatchIsOpZero = true;
8100 // If we reach here, they do have operations in common.
8101 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8102 OtherOpF, SI.getName()+".v");
8103 InsertNewInstBefore(NewSI, SI);
8105 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8107 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
8109 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
8111 assert(0 && "Shouldn't get here");
8115 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8116 Value *CondVal = SI.getCondition();
8117 Value *TrueVal = SI.getTrueValue();
8118 Value *FalseVal = SI.getFalseValue();
8120 // select true, X, Y -> X
8121 // select false, X, Y -> Y
8122 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8123 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8125 // select C, X, X -> X
8126 if (TrueVal == FalseVal)
8127 return ReplaceInstUsesWith(SI, TrueVal);
8129 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8130 return ReplaceInstUsesWith(SI, FalseVal);
8131 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8132 return ReplaceInstUsesWith(SI, TrueVal);
8133 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8134 if (isa<Constant>(TrueVal))
8135 return ReplaceInstUsesWith(SI, TrueVal);
8137 return ReplaceInstUsesWith(SI, FalseVal);
8140 if (SI.getType() == Type::Int1Ty) {
8141 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8142 if (C->getZExtValue()) {
8143 // Change: A = select B, true, C --> A = or B, C
8144 return BinaryOperator::createOr(CondVal, FalseVal);
8146 // Change: A = select B, false, C --> A = and !B, C
8148 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
8149 "not."+CondVal->getName()), SI);
8150 return BinaryOperator::createAnd(NotCond, FalseVal);
8152 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8153 if (C->getZExtValue() == false) {
8154 // Change: A = select B, C, false --> A = and B, C
8155 return BinaryOperator::createAnd(CondVal, TrueVal);
8157 // Change: A = select B, C, true --> A = or !B, C
8159 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
8160 "not."+CondVal->getName()), SI);
8161 return BinaryOperator::createOr(NotCond, TrueVal);
8165 // select a, b, a -> a&b
8166 // select a, a, b -> a|b
8167 if (CondVal == TrueVal)
8168 return BinaryOperator::createOr(CondVal, FalseVal);
8169 else if (CondVal == FalseVal)
8170 return BinaryOperator::createAnd(CondVal, TrueVal);
8173 // Selecting between two integer constants?
8174 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8175 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8176 // select C, 1, 0 -> zext C to int
8177 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8178 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
8179 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8180 // select C, 0, 1 -> zext !C to int
8182 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
8183 "not."+CondVal->getName()), SI);
8184 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
8187 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8189 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8191 // (x <s 0) ? -1 : 0 -> ashr x, 31
8192 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8193 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8194 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8195 // The comparison constant and the result are not neccessarily the
8196 // same width. Make an all-ones value by inserting a AShr.
8197 Value *X = IC->getOperand(0);
8198 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8199 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8200 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
8202 InsertNewInstBefore(SRA, SI);
8204 // Finally, convert to the type of the select RHS. We figure out
8205 // if this requires a SExt, Trunc or BitCast based on the sizes.
8206 Instruction::CastOps opc = Instruction::BitCast;
8207 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8208 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8209 if (SRASize < SISize)
8210 opc = Instruction::SExt;
8211 else if (SRASize > SISize)
8212 opc = Instruction::Trunc;
8213 return CastInst::create(opc, SRA, SI.getType());
8218 // If one of the constants is zero (we know they can't both be) and we
8219 // have an icmp instruction with zero, and we have an 'and' with the
8220 // non-constant value, eliminate this whole mess. This corresponds to
8221 // cases like this: ((X & 27) ? 27 : 0)
8222 if (TrueValC->isZero() || FalseValC->isZero())
8223 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8224 cast<Constant>(IC->getOperand(1))->isNullValue())
8225 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8226 if (ICA->getOpcode() == Instruction::And &&
8227 isa<ConstantInt>(ICA->getOperand(1)) &&
8228 (ICA->getOperand(1) == TrueValC ||
8229 ICA->getOperand(1) == FalseValC) &&
8230 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8231 // Okay, now we know that everything is set up, we just don't
8232 // know whether we have a icmp_ne or icmp_eq and whether the
8233 // true or false val is the zero.
8234 bool ShouldNotVal = !TrueValC->isZero();
8235 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8238 V = InsertNewInstBefore(BinaryOperator::create(
8239 Instruction::Xor, V, ICA->getOperand(1)), SI);
8240 return ReplaceInstUsesWith(SI, V);
8245 // See if we are selecting two values based on a comparison of the two values.
8246 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8247 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8248 // Transform (X == Y) ? X : Y -> Y
8249 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8250 // This is not safe in general for floating point:
8251 // consider X== -0, Y== +0.
8252 // It becomes safe if either operand is a nonzero constant.
8253 ConstantFP *CFPt, *CFPf;
8254 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8255 !CFPt->getValueAPF().isZero()) ||
8256 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8257 !CFPf->getValueAPF().isZero()))
8258 return ReplaceInstUsesWith(SI, FalseVal);
8260 // Transform (X != Y) ? X : Y -> X
8261 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8262 return ReplaceInstUsesWith(SI, TrueVal);
8263 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8265 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8266 // Transform (X == Y) ? Y : X -> X
8267 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8268 // This is not safe in general for floating point:
8269 // consider X== -0, Y== +0.
8270 // It becomes safe if either operand is a nonzero constant.
8271 ConstantFP *CFPt, *CFPf;
8272 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8273 !CFPt->getValueAPF().isZero()) ||
8274 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8275 !CFPf->getValueAPF().isZero()))
8276 return ReplaceInstUsesWith(SI, FalseVal);
8278 // Transform (X != Y) ? Y : X -> Y
8279 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8280 return ReplaceInstUsesWith(SI, TrueVal);
8281 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8285 // See if we are selecting two values based on a comparison of the two values.
8286 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8287 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8288 // Transform (X == Y) ? X : Y -> Y
8289 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8290 return ReplaceInstUsesWith(SI, FalseVal);
8291 // Transform (X != Y) ? X : Y -> X
8292 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8293 return ReplaceInstUsesWith(SI, TrueVal);
8294 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8296 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8297 // Transform (X == Y) ? Y : X -> X
8298 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8299 return ReplaceInstUsesWith(SI, FalseVal);
8300 // Transform (X != Y) ? Y : X -> Y
8301 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8302 return ReplaceInstUsesWith(SI, TrueVal);
8303 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8307 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8308 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8309 if (TI->hasOneUse() && FI->hasOneUse()) {
8310 Instruction *AddOp = 0, *SubOp = 0;
8312 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8313 if (TI->getOpcode() == FI->getOpcode())
8314 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8317 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8318 // even legal for FP.
8319 if (TI->getOpcode() == Instruction::Sub &&
8320 FI->getOpcode() == Instruction::Add) {
8321 AddOp = FI; SubOp = TI;
8322 } else if (FI->getOpcode() == Instruction::Sub &&
8323 TI->getOpcode() == Instruction::Add) {
8324 AddOp = TI; SubOp = FI;
8328 Value *OtherAddOp = 0;
8329 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8330 OtherAddOp = AddOp->getOperand(1);
8331 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8332 OtherAddOp = AddOp->getOperand(0);
8336 // So at this point we know we have (Y -> OtherAddOp):
8337 // select C, (add X, Y), (sub X, Z)
8338 Value *NegVal; // Compute -Z
8339 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8340 NegVal = ConstantExpr::getNeg(C);
8342 NegVal = InsertNewInstBefore(
8343 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
8346 Value *NewTrueOp = OtherAddOp;
8347 Value *NewFalseOp = NegVal;
8349 std::swap(NewTrueOp, NewFalseOp);
8350 Instruction *NewSel =
8351 SelectInst::Create(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
8353 NewSel = InsertNewInstBefore(NewSel, SI);
8354 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
8359 // See if we can fold the select into one of our operands.
8360 if (SI.getType()->isInteger()) {
8361 // See the comment above GetSelectFoldableOperands for a description of the
8362 // transformation we are doing here.
8363 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8364 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8365 !isa<Constant>(FalseVal))
8366 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8367 unsigned OpToFold = 0;
8368 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8370 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8375 Constant *C = GetSelectFoldableConstant(TVI);
8376 Instruction *NewSel =
8377 SelectInst::Create(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
8378 InsertNewInstBefore(NewSel, SI);
8379 NewSel->takeName(TVI);
8380 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8381 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
8383 assert(0 && "Unknown instruction!!");
8388 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8389 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8390 !isa<Constant>(TrueVal))
8391 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8392 unsigned OpToFold = 0;
8393 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8395 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8400 Constant *C = GetSelectFoldableConstant(FVI);
8401 Instruction *NewSel =
8402 SelectInst::Create(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
8403 InsertNewInstBefore(NewSel, SI);
8404 NewSel->takeName(FVI);
8405 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8406 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
8408 assert(0 && "Unknown instruction!!");
8413 if (BinaryOperator::isNot(CondVal)) {
8414 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8415 SI.setOperand(1, FalseVal);
8416 SI.setOperand(2, TrueVal);
8423 /// EnforceKnownAlignment - If the specified pointer points to an object that
8424 /// we control, modify the object's alignment to PrefAlign. This isn't
8425 /// often possible though. If alignment is important, a more reliable approach
8426 /// is to simply align all global variables and allocation instructions to
8427 /// their preferred alignment from the beginning.
8429 static unsigned EnforceKnownAlignment(Value *V,
8430 unsigned Align, unsigned PrefAlign) {
8432 User *U = dyn_cast<User>(V);
8433 if (!U) return Align;
8435 switch (getOpcode(U)) {
8437 case Instruction::BitCast:
8438 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8439 case Instruction::GetElementPtr: {
8440 // If all indexes are zero, it is just the alignment of the base pointer.
8441 bool AllZeroOperands = true;
8442 for (unsigned i = 1, e = U->getNumOperands(); i != e; ++i)
8443 if (!isa<Constant>(U->getOperand(i)) ||
8444 !cast<Constant>(U->getOperand(i))->isNullValue()) {
8445 AllZeroOperands = false;
8449 if (AllZeroOperands) {
8450 // Treat this like a bitcast.
8451 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8457 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8458 // If there is a large requested alignment and we can, bump up the alignment
8460 if (!GV->isDeclaration()) {
8461 GV->setAlignment(PrefAlign);
8464 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8465 // If there is a requested alignment and if this is an alloca, round up. We
8466 // don't do this for malloc, because some systems can't respect the request.
8467 if (isa<AllocaInst>(AI)) {
8468 AI->setAlignment(PrefAlign);
8476 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8477 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8478 /// and it is more than the alignment of the ultimate object, see if we can
8479 /// increase the alignment of the ultimate object, making this check succeed.
8480 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8481 unsigned PrefAlign) {
8482 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8483 sizeof(PrefAlign) * CHAR_BIT;
8484 APInt Mask = APInt::getAllOnesValue(BitWidth);
8485 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8486 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8487 unsigned TrailZ = KnownZero.countTrailingOnes();
8488 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8490 if (PrefAlign > Align)
8491 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8493 // We don't need to make any adjustment.
8497 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8498 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8499 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8500 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8501 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8503 if (CopyAlign < MinAlign) {
8504 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8508 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8510 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8511 if (MemOpLength == 0) return 0;
8513 // Source and destination pointer types are always "i8*" for intrinsic. See
8514 // if the size is something we can handle with a single primitive load/store.
8515 // A single load+store correctly handles overlapping memory in the memmove
8517 unsigned Size = MemOpLength->getZExtValue();
8518 if (Size == 0) return MI; // Delete this mem transfer.
8520 if (Size > 8 || (Size&(Size-1)))
8521 return 0; // If not 1/2/4/8 bytes, exit.
8523 // Use an integer load+store unless we can find something better.
8524 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8526 // Memcpy forces the use of i8* for the source and destination. That means
8527 // that if you're using memcpy to move one double around, you'll get a cast
8528 // from double* to i8*. We'd much rather use a double load+store rather than
8529 // an i64 load+store, here because this improves the odds that the source or
8530 // dest address will be promotable. See if we can find a better type than the
8531 // integer datatype.
8532 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8533 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8534 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8535 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8536 // down through these levels if so.
8537 while (!SrcETy->isFirstClassType()) {
8538 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8539 if (STy->getNumElements() == 1)
8540 SrcETy = STy->getElementType(0);
8543 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8544 if (ATy->getNumElements() == 1)
8545 SrcETy = ATy->getElementType();
8552 if (SrcETy->isFirstClassType())
8553 NewPtrTy = PointerType::getUnqual(SrcETy);
8558 // If the memcpy/memmove provides better alignment info than we can
8560 SrcAlign = std::max(SrcAlign, CopyAlign);
8561 DstAlign = std::max(DstAlign, CopyAlign);
8563 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8564 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8565 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8566 InsertNewInstBefore(L, *MI);
8567 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8569 // Set the size of the copy to 0, it will be deleted on the next iteration.
8570 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8574 Instruction *InstCombiner::SimplifyMemSet(MemSetInst *MI) {
8575 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8576 if (MI->getAlignment()->getZExtValue() < Alignment) {
8577 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8581 // Extract the length and alignment and fill if they are constant.
8582 ConstantInt *LenC = dyn_cast<ConstantInt>(MI->getLength());
8583 ConstantInt *FillC = dyn_cast<ConstantInt>(MI->getValue());
8584 if (!LenC || !FillC || FillC->getType() != Type::Int8Ty)
8586 uint64_t Len = LenC->getZExtValue();
8587 Alignment = MI->getAlignment()->getZExtValue();
8589 // If the length is zero, this is a no-op
8590 if (Len == 0) return MI; // memset(d,c,0,a) -> noop
8592 // memset(s,c,n) -> store s, c (for n=1,2,4,8)
8593 if (Len <= 8 && isPowerOf2_32((uint32_t)Len)) {
8594 const Type *ITy = IntegerType::get(Len*8); // n=1 -> i8.
8596 Value *Dest = MI->getDest();
8597 Dest = InsertBitCastBefore(Dest, PointerType::getUnqual(ITy), *MI);
8599 // Alignment 0 is identity for alignment 1 for memset, but not store.
8600 if (Alignment == 0) Alignment = 1;
8602 // Extract the fill value and store.
8603 uint64_t Fill = FillC->getZExtValue()*0x0101010101010101ULL;
8604 InsertNewInstBefore(new StoreInst(ConstantInt::get(ITy, Fill), Dest, false,
8607 // Set the size of the copy to 0, it will be deleted on the next iteration.
8608 MI->setLength(Constant::getNullValue(LenC->getType()));
8616 /// visitCallInst - CallInst simplification. This mostly only handles folding
8617 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8618 /// the heavy lifting.
8620 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8621 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8622 if (!II) return visitCallSite(&CI);
8624 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8626 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8627 bool Changed = false;
8629 // memmove/cpy/set of zero bytes is a noop.
8630 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8631 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8633 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8634 if (CI->getZExtValue() == 1) {
8635 // Replace the instruction with just byte operations. We would
8636 // transform other cases to loads/stores, but we don't know if
8637 // alignment is sufficient.
8641 // If we have a memmove and the source operation is a constant global,
8642 // then the source and dest pointers can't alias, so we can change this
8643 // into a call to memcpy.
8644 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8645 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8646 if (GVSrc->isConstant()) {
8647 Module *M = CI.getParent()->getParent()->getParent();
8648 Intrinsic::ID MemCpyID;
8649 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8650 MemCpyID = Intrinsic::memcpy_i32;
8652 MemCpyID = Intrinsic::memcpy_i64;
8653 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8658 // If we can determine a pointer alignment that is bigger than currently
8659 // set, update the alignment.
8660 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8661 if (Instruction *I = SimplifyMemTransfer(MI))
8663 } else if (MemSetInst *MSI = dyn_cast<MemSetInst>(MI)) {
8664 if (Instruction *I = SimplifyMemSet(MSI))
8668 if (Changed) return II;
8670 switch (II->getIntrinsicID()) {
8672 case Intrinsic::ppc_altivec_lvx:
8673 case Intrinsic::ppc_altivec_lvxl:
8674 case Intrinsic::x86_sse_loadu_ps:
8675 case Intrinsic::x86_sse2_loadu_pd:
8676 case Intrinsic::x86_sse2_loadu_dq:
8677 // Turn PPC lvx -> load if the pointer is known aligned.
8678 // Turn X86 loadups -> load if the pointer is known aligned.
8679 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8680 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8681 PointerType::getUnqual(II->getType()),
8683 return new LoadInst(Ptr);
8686 case Intrinsic::ppc_altivec_stvx:
8687 case Intrinsic::ppc_altivec_stvxl:
8688 // Turn stvx -> store if the pointer is known aligned.
8689 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8690 const Type *OpPtrTy =
8691 PointerType::getUnqual(II->getOperand(1)->getType());
8692 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8693 return new StoreInst(II->getOperand(1), Ptr);
8696 case Intrinsic::x86_sse_storeu_ps:
8697 case Intrinsic::x86_sse2_storeu_pd:
8698 case Intrinsic::x86_sse2_storeu_dq:
8699 case Intrinsic::x86_sse2_storel_dq:
8700 // Turn X86 storeu -> store if the pointer is known aligned.
8701 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8702 const Type *OpPtrTy =
8703 PointerType::getUnqual(II->getOperand(2)->getType());
8704 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8705 return new StoreInst(II->getOperand(2), Ptr);
8709 case Intrinsic::x86_sse_cvttss2si: {
8710 // These intrinsics only demands the 0th element of its input vector. If
8711 // we can simplify the input based on that, do so now.
8713 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8715 II->setOperand(1, V);
8721 case Intrinsic::ppc_altivec_vperm:
8722 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8723 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8724 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8726 // Check that all of the elements are integer constants or undefs.
8727 bool AllEltsOk = true;
8728 for (unsigned i = 0; i != 16; ++i) {
8729 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8730 !isa<UndefValue>(Mask->getOperand(i))) {
8737 // Cast the input vectors to byte vectors.
8738 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8739 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8740 Value *Result = UndefValue::get(Op0->getType());
8742 // Only extract each element once.
8743 Value *ExtractedElts[32];
8744 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8746 for (unsigned i = 0; i != 16; ++i) {
8747 if (isa<UndefValue>(Mask->getOperand(i)))
8749 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8750 Idx &= 31; // Match the hardware behavior.
8752 if (ExtractedElts[Idx] == 0) {
8754 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8755 InsertNewInstBefore(Elt, CI);
8756 ExtractedElts[Idx] = Elt;
8759 // Insert this value into the result vector.
8760 Result = InsertElementInst::Create(Result, ExtractedElts[Idx], i, "tmp");
8761 InsertNewInstBefore(cast<Instruction>(Result), CI);
8763 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8768 case Intrinsic::stackrestore: {
8769 // If the save is right next to the restore, remove the restore. This can
8770 // happen when variable allocas are DCE'd.
8771 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8772 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8773 BasicBlock::iterator BI = SS;
8775 return EraseInstFromFunction(CI);
8779 // Scan down this block to see if there is another stack restore in the
8780 // same block without an intervening call/alloca.
8781 BasicBlock::iterator BI = II;
8782 TerminatorInst *TI = II->getParent()->getTerminator();
8783 bool CannotRemove = false;
8784 for (++BI; &*BI != TI; ++BI) {
8785 if (isa<AllocaInst>(BI)) {
8786 CannotRemove = true;
8789 if (isa<CallInst>(BI)) {
8790 if (!isa<IntrinsicInst>(BI)) {
8791 CannotRemove = true;
8794 // If there is a stackrestore below this one, remove this one.
8795 return EraseInstFromFunction(CI);
8799 // If the stack restore is in a return/unwind block and if there are no
8800 // allocas or calls between the restore and the return, nuke the restore.
8801 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8802 return EraseInstFromFunction(CI);
8808 return visitCallSite(II);
8811 // InvokeInst simplification
8813 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8814 return visitCallSite(&II);
8817 /// isSafeToEliminateVarargsCast - If this cast does not affect the value
8818 /// passed through the varargs area, we can eliminate the use of the cast.
8819 static bool isSafeToEliminateVarargsCast(const CallSite CS,
8820 const CastInst * const CI,
8821 const TargetData * const TD,
8823 if (!CI->isLosslessCast())
8826 // The size of ByVal arguments is derived from the type, so we
8827 // can't change to a type with a different size. If the size were
8828 // passed explicitly we could avoid this check.
8829 if (!CS.paramHasAttr(ix, ParamAttr::ByVal))
8833 cast<PointerType>(CI->getOperand(0)->getType())->getElementType();
8834 const Type* DstTy = cast<PointerType>(CI->getType())->getElementType();
8835 if (!SrcTy->isSized() || !DstTy->isSized())
8837 if (TD->getABITypeSize(SrcTy) != TD->getABITypeSize(DstTy))
8842 // visitCallSite - Improvements for call and invoke instructions.
8844 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8845 bool Changed = false;
8847 // If the callee is a constexpr cast of a function, attempt to move the cast
8848 // to the arguments of the call/invoke.
8849 if (transformConstExprCastCall(CS)) return 0;
8851 Value *Callee = CS.getCalledValue();
8853 if (Function *CalleeF = dyn_cast<Function>(Callee))
8854 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8855 Instruction *OldCall = CS.getInstruction();
8856 // If the call and callee calling conventions don't match, this call must
8857 // be unreachable, as the call is undefined.
8858 new StoreInst(ConstantInt::getTrue(),
8859 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8861 if (!OldCall->use_empty())
8862 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8863 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8864 return EraseInstFromFunction(*OldCall);
8868 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8869 // This instruction is not reachable, just remove it. We insert a store to
8870 // undef so that we know that this code is not reachable, despite the fact
8871 // that we can't modify the CFG here.
8872 new StoreInst(ConstantInt::getTrue(),
8873 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8874 CS.getInstruction());
8876 if (!CS.getInstruction()->use_empty())
8877 CS.getInstruction()->
8878 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8880 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8881 // Don't break the CFG, insert a dummy cond branch.
8882 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
8883 ConstantInt::getTrue(), II);
8885 return EraseInstFromFunction(*CS.getInstruction());
8888 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8889 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8890 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8891 return transformCallThroughTrampoline(CS);
8893 const PointerType *PTy = cast<PointerType>(Callee->getType());
8894 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8895 if (FTy->isVarArg()) {
8896 int ix = FTy->getNumParams() + (isa<InvokeInst>(Callee) ? 3 : 1);
8897 // See if we can optimize any arguments passed through the varargs area of
8899 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8900 E = CS.arg_end(); I != E; ++I, ++ix) {
8901 CastInst *CI = dyn_cast<CastInst>(*I);
8902 if (CI && isSafeToEliminateVarargsCast(CS, CI, TD, ix)) {
8903 *I = CI->getOperand(0);
8909 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8910 // Inline asm calls cannot throw - mark them 'nounwind'.
8911 CS.setDoesNotThrow();
8915 return Changed ? CS.getInstruction() : 0;
8918 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8919 // attempt to move the cast to the arguments of the call/invoke.
8921 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8922 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8923 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8924 if (CE->getOpcode() != Instruction::BitCast ||
8925 !isa<Function>(CE->getOperand(0)))
8927 Function *Callee = cast<Function>(CE->getOperand(0));
8928 Instruction *Caller = CS.getInstruction();
8929 const PAListPtr &CallerPAL = CS.getParamAttrs();
8931 // Okay, this is a cast from a function to a different type. Unless doing so
8932 // would cause a type conversion of one of our arguments, change this call to
8933 // be a direct call with arguments casted to the appropriate types.
8935 const FunctionType *FT = Callee->getFunctionType();
8936 const Type *OldRetTy = Caller->getType();
8938 if (isa<StructType>(FT->getReturnType()))
8939 return false; // TODO: Handle multiple return values.
8941 // Check to see if we are changing the return type...
8942 if (OldRetTy != FT->getReturnType()) {
8943 if (Callee->isDeclaration() &&
8944 // Conversion is ok if changing from pointer to int of same size.
8945 !(isa<PointerType>(FT->getReturnType()) &&
8946 TD->getIntPtrType() == OldRetTy))
8947 return false; // Cannot transform this return value.
8949 if (!Caller->use_empty() &&
8950 // void -> non-void is handled specially
8951 FT->getReturnType() != Type::VoidTy &&
8952 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8953 return false; // Cannot transform this return value.
8955 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
8956 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8957 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8958 return false; // Attribute not compatible with transformed value.
8961 // If the callsite is an invoke instruction, and the return value is used by
8962 // a PHI node in a successor, we cannot change the return type of the call
8963 // because there is no place to put the cast instruction (without breaking
8964 // the critical edge). Bail out in this case.
8965 if (!Caller->use_empty())
8966 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8967 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8969 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8970 if (PN->getParent() == II->getNormalDest() ||
8971 PN->getParent() == II->getUnwindDest())
8975 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8976 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8978 CallSite::arg_iterator AI = CS.arg_begin();
8979 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8980 const Type *ParamTy = FT->getParamType(i);
8981 const Type *ActTy = (*AI)->getType();
8983 if (!CastInst::isCastable(ActTy, ParamTy))
8984 return false; // Cannot transform this parameter value.
8986 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
8987 return false; // Attribute not compatible with transformed value.
8989 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8990 // Some conversions are safe even if we do not have a body.
8991 // Either we can cast directly, or we can upconvert the argument
8992 bool isConvertible = ActTy == ParamTy ||
8993 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8994 (ParamTy->isInteger() && ActTy->isInteger() &&
8995 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8996 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8997 && c->getValue().isStrictlyPositive());
8998 if (Callee->isDeclaration() && !isConvertible) return false;
9001 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
9002 Callee->isDeclaration())
9003 return false; // Do not delete arguments unless we have a function body.
9005 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
9006 !CallerPAL.isEmpty())
9007 // In this case we have more arguments than the new function type, but we
9008 // won't be dropping them. Check that these extra arguments have attributes
9009 // that are compatible with being a vararg call argument.
9010 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
9011 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
9013 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
9014 if (PAttrs & ParamAttr::VarArgsIncompatible)
9018 // Okay, we decided that this is a safe thing to do: go ahead and start
9019 // inserting cast instructions as necessary...
9020 std::vector<Value*> Args;
9021 Args.reserve(NumActualArgs);
9022 SmallVector<ParamAttrsWithIndex, 8> attrVec;
9023 attrVec.reserve(NumCommonArgs);
9025 // Get any return attributes.
9026 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
9028 // If the return value is not being used, the type may not be compatible
9029 // with the existing attributes. Wipe out any problematic attributes.
9030 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
9032 // Add the new return attributes.
9034 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
9036 AI = CS.arg_begin();
9037 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
9038 const Type *ParamTy = FT->getParamType(i);
9039 if ((*AI)->getType() == ParamTy) {
9040 Args.push_back(*AI);
9042 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
9043 false, ParamTy, false);
9044 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
9045 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
9048 // Add any parameter attributes.
9049 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9050 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9053 // If the function takes more arguments than the call was taking, add them
9055 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
9056 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
9058 // If we are removing arguments to the function, emit an obnoxious warning...
9059 if (FT->getNumParams() < NumActualArgs) {
9060 if (!FT->isVarArg()) {
9061 cerr << "WARNING: While resolving call to function '"
9062 << Callee->getName() << "' arguments were dropped!\n";
9064 // Add all of the arguments in their promoted form to the arg list...
9065 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
9066 const Type *PTy = getPromotedType((*AI)->getType());
9067 if (PTy != (*AI)->getType()) {
9068 // Must promote to pass through va_arg area!
9069 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
9071 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
9072 InsertNewInstBefore(Cast, *Caller);
9073 Args.push_back(Cast);
9075 Args.push_back(*AI);
9078 // Add any parameter attributes.
9079 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
9080 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
9085 if (FT->getReturnType() == Type::VoidTy)
9086 Caller->setName(""); // Void type should not have a name.
9088 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
9091 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9092 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
9093 Args.begin(), Args.end(), Caller->getName(), Caller);
9094 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
9095 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
9097 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
9098 Caller->getName(), Caller);
9099 CallInst *CI = cast<CallInst>(Caller);
9100 if (CI->isTailCall())
9101 cast<CallInst>(NC)->setTailCall();
9102 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
9103 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
9106 // Insert a cast of the return type as necessary.
9108 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
9109 if (NV->getType() != Type::VoidTy) {
9110 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
9112 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
9114 // If this is an invoke instruction, we should insert it after the first
9115 // non-phi, instruction in the normal successor block.
9116 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9117 BasicBlock::iterator I = II->getNormalDest()->begin();
9118 while (isa<PHINode>(I)) ++I;
9119 InsertNewInstBefore(NC, *I);
9121 // Otherwise, it's a call, just insert cast right after the call instr
9122 InsertNewInstBefore(NC, *Caller);
9124 AddUsersToWorkList(*Caller);
9126 NV = UndefValue::get(Caller->getType());
9130 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9131 Caller->replaceAllUsesWith(NV);
9132 Caller->eraseFromParent();
9133 RemoveFromWorkList(Caller);
9137 // transformCallThroughTrampoline - Turn a call to a function created by the
9138 // init_trampoline intrinsic into a direct call to the underlying function.
9140 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
9141 Value *Callee = CS.getCalledValue();
9142 const PointerType *PTy = cast<PointerType>(Callee->getType());
9143 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9144 const PAListPtr &Attrs = CS.getParamAttrs();
9146 // If the call already has the 'nest' attribute somewhere then give up -
9147 // otherwise 'nest' would occur twice after splicing in the chain.
9148 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9151 IntrinsicInst *Tramp =
9152 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9154 Function *NestF = cast<Function>(Tramp->getOperand(2)->stripPointerCasts());
9155 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9156 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9158 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9159 if (!NestAttrs.isEmpty()) {
9160 unsigned NestIdx = 1;
9161 const Type *NestTy = 0;
9162 ParameterAttributes NestAttr = ParamAttr::None;
9164 // Look for a parameter marked with the 'nest' attribute.
9165 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9166 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9167 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9168 // Record the parameter type and any other attributes.
9170 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9175 Instruction *Caller = CS.getInstruction();
9176 std::vector<Value*> NewArgs;
9177 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9179 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9180 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9182 // Insert the nest argument into the call argument list, which may
9183 // mean appending it. Likewise for attributes.
9185 // Add any function result attributes.
9186 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9187 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9191 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9193 if (Idx == NestIdx) {
9194 // Add the chain argument and attributes.
9195 Value *NestVal = Tramp->getOperand(3);
9196 if (NestVal->getType() != NestTy)
9197 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9198 NewArgs.push_back(NestVal);
9199 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9205 // Add the original argument and attributes.
9206 NewArgs.push_back(*I);
9207 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9209 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9215 // The trampoline may have been bitcast to a bogus type (FTy).
9216 // Handle this by synthesizing a new function type, equal to FTy
9217 // with the chain parameter inserted.
9219 std::vector<const Type*> NewTypes;
9220 NewTypes.reserve(FTy->getNumParams()+1);
9222 // Insert the chain's type into the list of parameter types, which may
9223 // mean appending it.
9226 FunctionType::param_iterator I = FTy->param_begin(),
9227 E = FTy->param_end();
9231 // Add the chain's type.
9232 NewTypes.push_back(NestTy);
9237 // Add the original type.
9238 NewTypes.push_back(*I);
9244 // Replace the trampoline call with a direct call. Let the generic
9245 // code sort out any function type mismatches.
9246 FunctionType *NewFTy =
9247 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9248 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9249 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9250 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9252 Instruction *NewCaller;
9253 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9254 NewCaller = InvokeInst::Create(NewCallee,
9255 II->getNormalDest(), II->getUnwindDest(),
9256 NewArgs.begin(), NewArgs.end(),
9257 Caller->getName(), Caller);
9258 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9259 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9261 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9262 Caller->getName(), Caller);
9263 if (cast<CallInst>(Caller)->isTailCall())
9264 cast<CallInst>(NewCaller)->setTailCall();
9265 cast<CallInst>(NewCaller)->
9266 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9267 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9269 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9270 Caller->replaceAllUsesWith(NewCaller);
9271 Caller->eraseFromParent();
9272 RemoveFromWorkList(Caller);
9277 // Replace the trampoline call with a direct call. Since there is no 'nest'
9278 // parameter, there is no need to adjust the argument list. Let the generic
9279 // code sort out any function type mismatches.
9280 Constant *NewCallee =
9281 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9282 CS.setCalledFunction(NewCallee);
9283 return CS.getInstruction();
9286 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9287 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9288 /// and a single binop.
9289 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9290 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9291 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9292 isa<CmpInst>(FirstInst));
9293 unsigned Opc = FirstInst->getOpcode();
9294 Value *LHSVal = FirstInst->getOperand(0);
9295 Value *RHSVal = FirstInst->getOperand(1);
9297 const Type *LHSType = LHSVal->getType();
9298 const Type *RHSType = RHSVal->getType();
9300 // Scan to see if all operands are the same opcode, all have one use, and all
9301 // kill their operands (i.e. the operands have one use).
9302 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9303 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9304 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9305 // Verify type of the LHS matches so we don't fold cmp's of different
9306 // types or GEP's with different index types.
9307 I->getOperand(0)->getType() != LHSType ||
9308 I->getOperand(1)->getType() != RHSType)
9311 // If they are CmpInst instructions, check their predicates
9312 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9313 if (cast<CmpInst>(I)->getPredicate() !=
9314 cast<CmpInst>(FirstInst)->getPredicate())
9317 // Keep track of which operand needs a phi node.
9318 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9319 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9322 // Otherwise, this is safe to transform, determine if it is profitable.
9324 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9325 // Indexes are often folded into load/store instructions, so we don't want to
9326 // hide them behind a phi.
9327 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9330 Value *InLHS = FirstInst->getOperand(0);
9331 Value *InRHS = FirstInst->getOperand(1);
9332 PHINode *NewLHS = 0, *NewRHS = 0;
9334 NewLHS = PHINode::Create(LHSType, FirstInst->getOperand(0)->getName()+".pn");
9335 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9336 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9337 InsertNewInstBefore(NewLHS, PN);
9342 NewRHS = PHINode::Create(RHSType, FirstInst->getOperand(1)->getName()+".pn");
9343 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9344 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9345 InsertNewInstBefore(NewRHS, PN);
9349 // Add all operands to the new PHIs.
9350 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9352 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9353 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9356 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9357 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9361 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9362 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
9363 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9364 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9367 assert(isa<GetElementPtrInst>(FirstInst));
9368 return GetElementPtrInst::Create(LHSVal, RHSVal);
9372 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9373 /// of the block that defines it. This means that it must be obvious the value
9374 /// of the load is not changed from the point of the load to the end of the
9377 /// Finally, it is safe, but not profitable, to sink a load targetting a
9378 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9380 static bool isSafeToSinkLoad(LoadInst *L) {
9381 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9383 for (++BBI; BBI != E; ++BBI)
9384 if (BBI->mayWriteToMemory())
9387 // Check for non-address taken alloca. If not address-taken already, it isn't
9388 // profitable to do this xform.
9389 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9390 bool isAddressTaken = false;
9391 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9393 if (isa<LoadInst>(UI)) continue;
9394 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9395 // If storing TO the alloca, then the address isn't taken.
9396 if (SI->getOperand(1) == AI) continue;
9398 isAddressTaken = true;
9402 if (!isAddressTaken)
9410 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9411 // operator and they all are only used by the PHI, PHI together their
9412 // inputs, and do the operation once, to the result of the PHI.
9413 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9414 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9416 // Scan the instruction, looking for input operations that can be folded away.
9417 // If all input operands to the phi are the same instruction (e.g. a cast from
9418 // the same type or "+42") we can pull the operation through the PHI, reducing
9419 // code size and simplifying code.
9420 Constant *ConstantOp = 0;
9421 const Type *CastSrcTy = 0;
9422 bool isVolatile = false;
9423 if (isa<CastInst>(FirstInst)) {
9424 CastSrcTy = FirstInst->getOperand(0)->getType();
9425 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9426 // Can fold binop, compare or shift here if the RHS is a constant,
9427 // otherwise call FoldPHIArgBinOpIntoPHI.
9428 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9429 if (ConstantOp == 0)
9430 return FoldPHIArgBinOpIntoPHI(PN);
9431 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9432 isVolatile = LI->isVolatile();
9433 // We can't sink the load if the loaded value could be modified between the
9434 // load and the PHI.
9435 if (LI->getParent() != PN.getIncomingBlock(0) ||
9436 !isSafeToSinkLoad(LI))
9438 } else if (isa<GetElementPtrInst>(FirstInst)) {
9439 if (FirstInst->getNumOperands() == 2)
9440 return FoldPHIArgBinOpIntoPHI(PN);
9441 // Can't handle general GEPs yet.
9444 return 0; // Cannot fold this operation.
9447 // Check to see if all arguments are the same operation.
9448 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9449 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9450 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9451 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9454 if (I->getOperand(0)->getType() != CastSrcTy)
9455 return 0; // Cast operation must match.
9456 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9457 // We can't sink the load if the loaded value could be modified between
9458 // the load and the PHI.
9459 if (LI->isVolatile() != isVolatile ||
9460 LI->getParent() != PN.getIncomingBlock(i) ||
9461 !isSafeToSinkLoad(LI))
9464 // If the PHI is volatile and its block has multiple successors, sinking
9465 // it would remove a load of the volatile value from the path through the
9468 LI->getParent()->getTerminator()->getNumSuccessors() != 1)
9472 } else if (I->getOperand(1) != ConstantOp) {
9477 // Okay, they are all the same operation. Create a new PHI node of the
9478 // correct type, and PHI together all of the LHS's of the instructions.
9479 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9480 PN.getName()+".in");
9481 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9483 Value *InVal = FirstInst->getOperand(0);
9484 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9486 // Add all operands to the new PHI.
9487 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9488 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9489 if (NewInVal != InVal)
9491 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9496 // The new PHI unions all of the same values together. This is really
9497 // common, so we handle it intelligently here for compile-time speed.
9501 InsertNewInstBefore(NewPN, PN);
9505 // Insert and return the new operation.
9506 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9507 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
9508 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9509 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
9510 if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9511 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
9512 PhiVal, ConstantOp);
9513 assert(isa<LoadInst>(FirstInst) && "Unknown operation");
9515 // If this was a volatile load that we are merging, make sure to loop through
9516 // and mark all the input loads as non-volatile. If we don't do this, we will
9517 // insert a new volatile load and the old ones will not be deletable.
9519 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i)
9520 cast<LoadInst>(PN.getIncomingValue(i))->setVolatile(false);
9522 return new LoadInst(PhiVal, "", isVolatile);
9525 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9527 static bool DeadPHICycle(PHINode *PN,
9528 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9529 if (PN->use_empty()) return true;
9530 if (!PN->hasOneUse()) return false;
9532 // Remember this node, and if we find the cycle, return.
9533 if (!PotentiallyDeadPHIs.insert(PN))
9536 // Don't scan crazily complex things.
9537 if (PotentiallyDeadPHIs.size() == 16)
9540 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9541 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9546 /// PHIsEqualValue - Return true if this phi node is always equal to
9547 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9548 /// z = some value; x = phi (y, z); y = phi (x, z)
9549 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9550 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9551 // See if we already saw this PHI node.
9552 if (!ValueEqualPHIs.insert(PN))
9555 // Don't scan crazily complex things.
9556 if (ValueEqualPHIs.size() == 16)
9559 // Scan the operands to see if they are either phi nodes or are equal to
9561 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9562 Value *Op = PN->getIncomingValue(i);
9563 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9564 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9566 } else if (Op != NonPhiInVal)
9574 // PHINode simplification
9576 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9577 // If LCSSA is around, don't mess with Phi nodes
9578 if (MustPreserveLCSSA) return 0;
9580 if (Value *V = PN.hasConstantValue())
9581 return ReplaceInstUsesWith(PN, V);
9583 // If all PHI operands are the same operation, pull them through the PHI,
9584 // reducing code size.
9585 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9586 PN.getIncomingValue(0)->hasOneUse())
9587 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9590 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9591 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9592 // PHI)... break the cycle.
9593 if (PN.hasOneUse()) {
9594 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9595 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9596 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9597 PotentiallyDeadPHIs.insert(&PN);
9598 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9599 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9602 // If this phi has a single use, and if that use just computes a value for
9603 // the next iteration of a loop, delete the phi. This occurs with unused
9604 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9605 // common case here is good because the only other things that catch this
9606 // are induction variable analysis (sometimes) and ADCE, which is only run
9608 if (PHIUser->hasOneUse() &&
9609 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9610 PHIUser->use_back() == &PN) {
9611 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9615 // We sometimes end up with phi cycles that non-obviously end up being the
9616 // same value, for example:
9617 // z = some value; x = phi (y, z); y = phi (x, z)
9618 // where the phi nodes don't necessarily need to be in the same block. Do a
9619 // quick check to see if the PHI node only contains a single non-phi value, if
9620 // so, scan to see if the phi cycle is actually equal to that value.
9622 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9623 // Scan for the first non-phi operand.
9624 while (InValNo != NumOperandVals &&
9625 isa<PHINode>(PN.getIncomingValue(InValNo)))
9628 if (InValNo != NumOperandVals) {
9629 Value *NonPhiInVal = PN.getOperand(InValNo);
9631 // Scan the rest of the operands to see if there are any conflicts, if so
9632 // there is no need to recursively scan other phis.
9633 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9634 Value *OpVal = PN.getIncomingValue(InValNo);
9635 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9639 // If we scanned over all operands, then we have one unique value plus
9640 // phi values. Scan PHI nodes to see if they all merge in each other or
9642 if (InValNo == NumOperandVals) {
9643 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9644 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9645 return ReplaceInstUsesWith(PN, NonPhiInVal);
9652 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9653 Instruction *InsertPoint,
9655 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9656 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9657 // We must cast correctly to the pointer type. Ensure that we
9658 // sign extend the integer value if it is smaller as this is
9659 // used for address computation.
9660 Instruction::CastOps opcode =
9661 (VTySize < PtrSize ? Instruction::SExt :
9662 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9663 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9667 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9668 Value *PtrOp = GEP.getOperand(0);
9669 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9670 // If so, eliminate the noop.
9671 if (GEP.getNumOperands() == 1)
9672 return ReplaceInstUsesWith(GEP, PtrOp);
9674 if (isa<UndefValue>(GEP.getOperand(0)))
9675 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9677 bool HasZeroPointerIndex = false;
9678 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9679 HasZeroPointerIndex = C->isNullValue();
9681 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9682 return ReplaceInstUsesWith(GEP, PtrOp);
9684 // Eliminate unneeded casts for indices.
9685 bool MadeChange = false;
9687 gep_type_iterator GTI = gep_type_begin(GEP);
9688 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9689 if (isa<SequentialType>(*GTI)) {
9690 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9691 if (CI->getOpcode() == Instruction::ZExt ||
9692 CI->getOpcode() == Instruction::SExt) {
9693 const Type *SrcTy = CI->getOperand(0)->getType();
9694 // We can eliminate a cast from i32 to i64 iff the target
9695 // is a 32-bit pointer target.
9696 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9698 GEP.setOperand(i, CI->getOperand(0));
9702 // If we are using a wider index than needed for this platform, shrink it
9703 // to what we need. If the incoming value needs a cast instruction,
9704 // insert it. This explicit cast can make subsequent optimizations more
9706 Value *Op = GEP.getOperand(i);
9707 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9708 if (Constant *C = dyn_cast<Constant>(Op)) {
9709 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9712 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9714 GEP.setOperand(i, Op);
9720 if (MadeChange) return &GEP;
9722 // If this GEP instruction doesn't move the pointer, and if the input operand
9723 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9724 // real input to the dest type.
9725 if (GEP.hasAllZeroIndices()) {
9726 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9727 // If the bitcast is of an allocation, and the allocation will be
9728 // converted to match the type of the cast, don't touch this.
9729 if (isa<AllocationInst>(BCI->getOperand(0))) {
9730 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9731 if (Instruction *I = visitBitCast(*BCI)) {
9734 BCI->getParent()->getInstList().insert(BCI, I);
9735 ReplaceInstUsesWith(*BCI, I);
9740 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9744 // Combine Indices - If the source pointer to this getelementptr instruction
9745 // is a getelementptr instruction, combine the indices of the two
9746 // getelementptr instructions into a single instruction.
9748 SmallVector<Value*, 8> SrcGEPOperands;
9749 if (User *Src = dyn_castGetElementPtr(PtrOp))
9750 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9752 if (!SrcGEPOperands.empty()) {
9753 // Note that if our source is a gep chain itself that we wait for that
9754 // chain to be resolved before we perform this transformation. This
9755 // avoids us creating a TON of code in some cases.
9757 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9758 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9759 return 0; // Wait until our source is folded to completion.
9761 SmallVector<Value*, 8> Indices;
9763 // Find out whether the last index in the source GEP is a sequential idx.
9764 bool EndsWithSequential = false;
9765 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9766 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9767 EndsWithSequential = !isa<StructType>(*I);
9769 // Can we combine the two pointer arithmetics offsets?
9770 if (EndsWithSequential) {
9771 // Replace: gep (gep %P, long B), long A, ...
9772 // With: T = long A+B; gep %P, T, ...
9774 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9775 if (SO1 == Constant::getNullValue(SO1->getType())) {
9777 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9780 // If they aren't the same type, convert both to an integer of the
9781 // target's pointer size.
9782 if (SO1->getType() != GO1->getType()) {
9783 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9784 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9785 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9786 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9788 unsigned PS = TD->getPointerSizeInBits();
9789 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9790 // Convert GO1 to SO1's type.
9791 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9793 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9794 // Convert SO1 to GO1's type.
9795 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9797 const Type *PT = TD->getIntPtrType();
9798 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9799 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9803 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9804 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9806 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9807 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9811 // Recycle the GEP we already have if possible.
9812 if (SrcGEPOperands.size() == 2) {
9813 GEP.setOperand(0, SrcGEPOperands[0]);
9814 GEP.setOperand(1, Sum);
9817 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9818 SrcGEPOperands.end()-1);
9819 Indices.push_back(Sum);
9820 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9822 } else if (isa<Constant>(*GEP.idx_begin()) &&
9823 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9824 SrcGEPOperands.size() != 1) {
9825 // Otherwise we can do the fold if the first index of the GEP is a zero
9826 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9827 SrcGEPOperands.end());
9828 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9831 if (!Indices.empty())
9832 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
9833 Indices.end(), GEP.getName());
9835 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9836 // GEP of global variable. If all of the indices for this GEP are
9837 // constants, we can promote this to a constexpr instead of an instruction.
9839 // Scan for nonconstants...
9840 SmallVector<Constant*, 8> Indices;
9841 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9842 for (; I != E && isa<Constant>(*I); ++I)
9843 Indices.push_back(cast<Constant>(*I));
9845 if (I == E) { // If they are all constants...
9846 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9847 &Indices[0],Indices.size());
9849 // Replace all uses of the GEP with the new constexpr...
9850 return ReplaceInstUsesWith(GEP, CE);
9852 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9853 if (!isa<PointerType>(X->getType())) {
9854 // Not interesting. Source pointer must be a cast from pointer.
9855 } else if (HasZeroPointerIndex) {
9856 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9857 // into : GEP [10 x i8]* X, i32 0, ...
9859 // This occurs when the program declares an array extern like "int X[];"
9861 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9862 const PointerType *XTy = cast<PointerType>(X->getType());
9863 if (const ArrayType *XATy =
9864 dyn_cast<ArrayType>(XTy->getElementType()))
9865 if (const ArrayType *CATy =
9866 dyn_cast<ArrayType>(CPTy->getElementType()))
9867 if (CATy->getElementType() == XATy->getElementType()) {
9868 // At this point, we know that the cast source type is a pointer
9869 // to an array of the same type as the destination pointer
9870 // array. Because the array type is never stepped over (there
9871 // is a leading zero) we can fold the cast into this GEP.
9872 GEP.setOperand(0, X);
9875 } else if (GEP.getNumOperands() == 2) {
9876 // Transform things like:
9877 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9878 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9879 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9880 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9881 if (isa<ArrayType>(SrcElTy) &&
9882 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9883 TD->getABITypeSize(ResElTy)) {
9885 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9886 Idx[1] = GEP.getOperand(1);
9887 Value *V = InsertNewInstBefore(
9888 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
9889 // V and GEP are both pointer types --> BitCast
9890 return new BitCastInst(V, GEP.getType());
9893 // Transform things like:
9894 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9895 // (where tmp = 8*tmp2) into:
9896 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9898 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9899 uint64_t ArrayEltSize =
9900 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9902 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9903 // allow either a mul, shift, or constant here.
9905 ConstantInt *Scale = 0;
9906 if (ArrayEltSize == 1) {
9907 NewIdx = GEP.getOperand(1);
9908 Scale = ConstantInt::get(NewIdx->getType(), 1);
9909 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9910 NewIdx = ConstantInt::get(CI->getType(), 1);
9912 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9913 if (Inst->getOpcode() == Instruction::Shl &&
9914 isa<ConstantInt>(Inst->getOperand(1))) {
9915 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9916 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9917 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9918 NewIdx = Inst->getOperand(0);
9919 } else if (Inst->getOpcode() == Instruction::Mul &&
9920 isa<ConstantInt>(Inst->getOperand(1))) {
9921 Scale = cast<ConstantInt>(Inst->getOperand(1));
9922 NewIdx = Inst->getOperand(0);
9926 // If the index will be to exactly the right offset with the scale taken
9927 // out, perform the transformation. Note, we don't know whether Scale is
9928 // signed or not. We'll use unsigned version of division/modulo
9929 // operation after making sure Scale doesn't have the sign bit set.
9930 if (Scale && Scale->getSExtValue() >= 0LL &&
9931 Scale->getZExtValue() % ArrayEltSize == 0) {
9932 Scale = ConstantInt::get(Scale->getType(),
9933 Scale->getZExtValue() / ArrayEltSize);
9934 if (Scale->getZExtValue() != 1) {
9935 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9937 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9938 NewIdx = InsertNewInstBefore(Sc, GEP);
9941 // Insert the new GEP instruction.
9943 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9945 Instruction *NewGEP =
9946 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
9947 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9948 // The NewGEP must be pointer typed, so must the old one -> BitCast
9949 return new BitCastInst(NewGEP, GEP.getType());
9958 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9959 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9960 if (AI.isArrayAllocation()) { // Check C != 1
9961 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9963 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9964 AllocationInst *New = 0;
9966 // Create and insert the replacement instruction...
9967 if (isa<MallocInst>(AI))
9968 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9970 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9971 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9974 InsertNewInstBefore(New, AI);
9976 // Scan to the end of the allocation instructions, to skip over a block of
9977 // allocas if possible...
9979 BasicBlock::iterator It = New;
9980 while (isa<AllocationInst>(*It)) ++It;
9982 // Now that I is pointing to the first non-allocation-inst in the block,
9983 // insert our getelementptr instruction...
9985 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9989 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
9990 New->getName()+".sub", It);
9992 // Now make everything use the getelementptr instead of the original
9994 return ReplaceInstUsesWith(AI, V);
9995 } else if (isa<UndefValue>(AI.getArraySize())) {
9996 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10000 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
10001 // Note that we only do this for alloca's, because malloc should allocate and
10002 // return a unique pointer, even for a zero byte allocation.
10003 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
10004 TD->getABITypeSize(AI.getAllocatedType()) == 0)
10005 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
10010 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
10011 Value *Op = FI.getOperand(0);
10013 // free undef -> unreachable.
10014 if (isa<UndefValue>(Op)) {
10015 // Insert a new store to null because we cannot modify the CFG here.
10016 new StoreInst(ConstantInt::getTrue(),
10017 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
10018 return EraseInstFromFunction(FI);
10021 // If we have 'free null' delete the instruction. This can happen in stl code
10022 // when lots of inlining happens.
10023 if (isa<ConstantPointerNull>(Op))
10024 return EraseInstFromFunction(FI);
10026 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
10027 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
10028 FI.setOperand(0, CI->getOperand(0));
10032 // Change free (gep X, 0,0,0,0) into free(X)
10033 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10034 if (GEPI->hasAllZeroIndices()) {
10035 AddToWorkList(GEPI);
10036 FI.setOperand(0, GEPI->getOperand(0));
10041 // Change free(malloc) into nothing, if the malloc has a single use.
10042 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
10043 if (MI->hasOneUse()) {
10044 EraseInstFromFunction(FI);
10045 return EraseInstFromFunction(*MI);
10052 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
10053 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
10054 const TargetData *TD) {
10055 User *CI = cast<User>(LI.getOperand(0));
10056 Value *CastOp = CI->getOperand(0);
10058 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
10059 // Instead of loading constant c string, use corresponding integer value
10060 // directly if string length is small enough.
10061 const std::string &Str = CE->getOperand(0)->getStringValue();
10062 if (!Str.empty()) {
10063 unsigned len = Str.length();
10064 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
10065 unsigned numBits = Ty->getPrimitiveSizeInBits();
10066 // Replace LI with immediate integer store.
10067 if ((numBits >> 3) == len + 1) {
10068 APInt StrVal(numBits, 0);
10069 APInt SingleChar(numBits, 0);
10070 if (TD->isLittleEndian()) {
10071 for (signed i = len-1; i >= 0; i--) {
10072 SingleChar = (uint64_t) Str[i];
10073 StrVal = (StrVal << 8) | SingleChar;
10076 for (unsigned i = 0; i < len; i++) {
10077 SingleChar = (uint64_t) Str[i];
10078 StrVal = (StrVal << 8) | SingleChar;
10080 // Append NULL at the end.
10082 StrVal = (StrVal << 8) | SingleChar;
10084 Value *NL = ConstantInt::get(StrVal);
10085 return IC.ReplaceInstUsesWith(LI, NL);
10090 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10091 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10092 const Type *SrcPTy = SrcTy->getElementType();
10094 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
10095 isa<VectorType>(DestPTy)) {
10096 // If the source is an array, the code below will not succeed. Check to
10097 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10099 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10100 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10101 if (ASrcTy->getNumElements() != 0) {
10103 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10104 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10105 SrcTy = cast<PointerType>(CastOp->getType());
10106 SrcPTy = SrcTy->getElementType();
10109 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
10110 isa<VectorType>(SrcPTy)) &&
10111 // Do not allow turning this into a load of an integer, which is then
10112 // casted to a pointer, this pessimizes pointer analysis a lot.
10113 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
10114 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10115 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10117 // Okay, we are casting from one integer or pointer type to another of
10118 // the same size. Instead of casting the pointer before the load, cast
10119 // the result of the loaded value.
10120 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
10122 LI.isVolatile()),LI);
10123 // Now cast the result of the load.
10124 return new BitCastInst(NewLoad, LI.getType());
10131 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
10132 /// from this value cannot trap. If it is not obviously safe to load from the
10133 /// specified pointer, we do a quick local scan of the basic block containing
10134 /// ScanFrom, to determine if the address is already accessed.
10135 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
10136 // If it is an alloca it is always safe to load from.
10137 if (isa<AllocaInst>(V)) return true;
10139 // If it is a global variable it is mostly safe to load from.
10140 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
10141 // Don't try to evaluate aliases. External weak GV can be null.
10142 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
10144 // Otherwise, be a little bit agressive by scanning the local block where we
10145 // want to check to see if the pointer is already being loaded or stored
10146 // from/to. If so, the previous load or store would have already trapped,
10147 // so there is no harm doing an extra load (also, CSE will later eliminate
10148 // the load entirely).
10149 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
10154 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10155 if (LI->getOperand(0) == V) return true;
10156 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10157 if (SI->getOperand(1) == V) return true;
10163 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10164 /// until we find the underlying object a pointer is referring to or something
10165 /// we don't understand. Note that the returned pointer may be offset from the
10166 /// input, because we ignore GEP indices.
10167 static Value *GetUnderlyingObject(Value *Ptr) {
10169 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10170 if (CE->getOpcode() == Instruction::BitCast ||
10171 CE->getOpcode() == Instruction::GetElementPtr)
10172 Ptr = CE->getOperand(0);
10175 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10176 Ptr = BCI->getOperand(0);
10177 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10178 Ptr = GEP->getOperand(0);
10185 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10186 Value *Op = LI.getOperand(0);
10188 // Attempt to improve the alignment.
10189 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10191 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10192 LI.getAlignment()))
10193 LI.setAlignment(KnownAlign);
10195 // load (cast X) --> cast (load X) iff safe
10196 if (isa<CastInst>(Op))
10197 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10200 // None of the following transforms are legal for volatile loads.
10201 if (LI.isVolatile()) return 0;
10203 if (&LI.getParent()->front() != &LI) {
10204 BasicBlock::iterator BBI = &LI; --BBI;
10205 // If the instruction immediately before this is a store to the same
10206 // address, do a simple form of store->load forwarding.
10207 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10208 if (SI->getOperand(1) == LI.getOperand(0))
10209 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10210 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10211 if (LIB->getOperand(0) == LI.getOperand(0))
10212 return ReplaceInstUsesWith(LI, LIB);
10215 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10216 const Value *GEPI0 = GEPI->getOperand(0);
10217 // TODO: Consider a target hook for valid address spaces for this xform.
10218 if (isa<ConstantPointerNull>(GEPI0) &&
10219 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10220 // Insert a new store to null instruction before the load to indicate
10221 // that this code is not reachable. We do this instead of inserting
10222 // an unreachable instruction directly because we cannot modify the
10224 new StoreInst(UndefValue::get(LI.getType()),
10225 Constant::getNullValue(Op->getType()), &LI);
10226 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10230 if (Constant *C = dyn_cast<Constant>(Op)) {
10231 // load null/undef -> undef
10232 // TODO: Consider a target hook for valid address spaces for this xform.
10233 if (isa<UndefValue>(C) || (C->isNullValue() &&
10234 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10235 // Insert a new store to null instruction before the load to indicate that
10236 // this code is not reachable. We do this instead of inserting an
10237 // unreachable instruction directly because we cannot modify the CFG.
10238 new StoreInst(UndefValue::get(LI.getType()),
10239 Constant::getNullValue(Op->getType()), &LI);
10240 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10243 // Instcombine load (constant global) into the value loaded.
10244 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10245 if (GV->isConstant() && !GV->isDeclaration())
10246 return ReplaceInstUsesWith(LI, GV->getInitializer());
10248 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10249 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10250 if (CE->getOpcode() == Instruction::GetElementPtr) {
10251 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10252 if (GV->isConstant() && !GV->isDeclaration())
10254 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10255 return ReplaceInstUsesWith(LI, V);
10256 if (CE->getOperand(0)->isNullValue()) {
10257 // Insert a new store to null instruction before the load to indicate
10258 // that this code is not reachable. We do this instead of inserting
10259 // an unreachable instruction directly because we cannot modify the
10261 new StoreInst(UndefValue::get(LI.getType()),
10262 Constant::getNullValue(Op->getType()), &LI);
10263 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10266 } else if (CE->isCast()) {
10267 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10273 // If this load comes from anywhere in a constant global, and if the global
10274 // is all undef or zero, we know what it loads.
10275 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10276 if (GV->isConstant() && GV->hasInitializer()) {
10277 if (GV->getInitializer()->isNullValue())
10278 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10279 else if (isa<UndefValue>(GV->getInitializer()))
10280 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10284 if (Op->hasOneUse()) {
10285 // Change select and PHI nodes to select values instead of addresses: this
10286 // helps alias analysis out a lot, allows many others simplifications, and
10287 // exposes redundancy in the code.
10289 // Note that we cannot do the transformation unless we know that the
10290 // introduced loads cannot trap! Something like this is valid as long as
10291 // the condition is always false: load (select bool %C, int* null, int* %G),
10292 // but it would not be valid if we transformed it to load from null
10293 // unconditionally.
10295 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10296 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10297 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10298 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10299 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10300 SI->getOperand(1)->getName()+".val"), LI);
10301 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10302 SI->getOperand(2)->getName()+".val"), LI);
10303 return SelectInst::Create(SI->getCondition(), V1, V2);
10306 // load (select (cond, null, P)) -> load P
10307 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10308 if (C->isNullValue()) {
10309 LI.setOperand(0, SI->getOperand(2));
10313 // load (select (cond, P, null)) -> load P
10314 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10315 if (C->isNullValue()) {
10316 LI.setOperand(0, SI->getOperand(1));
10324 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10326 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10327 User *CI = cast<User>(SI.getOperand(1));
10328 Value *CastOp = CI->getOperand(0);
10330 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10331 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10332 const Type *SrcPTy = SrcTy->getElementType();
10334 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10335 // If the source is an array, the code below will not succeed. Check to
10336 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10338 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10339 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10340 if (ASrcTy->getNumElements() != 0) {
10342 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10343 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10344 SrcTy = cast<PointerType>(CastOp->getType());
10345 SrcPTy = SrcTy->getElementType();
10348 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10349 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10350 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10352 // Okay, we are casting from one integer or pointer type to another of
10353 // the same size. Instead of casting the pointer before
10354 // the store, cast the value to be stored.
10356 Value *SIOp0 = SI.getOperand(0);
10357 Instruction::CastOps opcode = Instruction::BitCast;
10358 const Type* CastSrcTy = SIOp0->getType();
10359 const Type* CastDstTy = SrcPTy;
10360 if (isa<PointerType>(CastDstTy)) {
10361 if (CastSrcTy->isInteger())
10362 opcode = Instruction::IntToPtr;
10363 } else if (isa<IntegerType>(CastDstTy)) {
10364 if (isa<PointerType>(SIOp0->getType()))
10365 opcode = Instruction::PtrToInt;
10367 if (Constant *C = dyn_cast<Constant>(SIOp0))
10368 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10370 NewCast = IC.InsertNewInstBefore(
10371 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10373 return new StoreInst(NewCast, CastOp);
10380 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10381 Value *Val = SI.getOperand(0);
10382 Value *Ptr = SI.getOperand(1);
10384 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10385 EraseInstFromFunction(SI);
10390 // If the RHS is an alloca with a single use, zapify the store, making the
10392 if (Ptr->hasOneUse() && !SI.isVolatile()) {
10393 if (isa<AllocaInst>(Ptr)) {
10394 EraseInstFromFunction(SI);
10399 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10400 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10401 GEP->getOperand(0)->hasOneUse()) {
10402 EraseInstFromFunction(SI);
10408 // Attempt to improve the alignment.
10409 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10411 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10412 SI.getAlignment()))
10413 SI.setAlignment(KnownAlign);
10415 // Do really simple DSE, to catch cases where there are several consequtive
10416 // stores to the same location, separated by a few arithmetic operations. This
10417 // situation often occurs with bitfield accesses.
10418 BasicBlock::iterator BBI = &SI;
10419 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10423 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10424 // Prev store isn't volatile, and stores to the same location?
10425 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10428 EraseInstFromFunction(*PrevSI);
10434 // If this is a load, we have to stop. However, if the loaded value is from
10435 // the pointer we're loading and is producing the pointer we're storing,
10436 // then *this* store is dead (X = load P; store X -> P).
10437 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10438 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10439 EraseInstFromFunction(SI);
10443 // Otherwise, this is a load from some other location. Stores before it
10444 // may not be dead.
10448 // Don't skip over loads or things that can modify memory.
10449 if (BBI->mayWriteToMemory() || BBI->mayReadFromMemory())
10454 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10456 // store X, null -> turns into 'unreachable' in SimplifyCFG
10457 if (isa<ConstantPointerNull>(Ptr)) {
10458 if (!isa<UndefValue>(Val)) {
10459 SI.setOperand(0, UndefValue::get(Val->getType()));
10460 if (Instruction *U = dyn_cast<Instruction>(Val))
10461 AddToWorkList(U); // Dropped a use.
10464 return 0; // Do not modify these!
10467 // store undef, Ptr -> noop
10468 if (isa<UndefValue>(Val)) {
10469 EraseInstFromFunction(SI);
10474 // If the pointer destination is a cast, see if we can fold the cast into the
10476 if (isa<CastInst>(Ptr))
10477 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10479 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10481 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10485 // If this store is the last instruction in the basic block, and if the block
10486 // ends with an unconditional branch, try to move it to the successor block.
10488 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10489 if (BI->isUnconditional())
10490 if (SimplifyStoreAtEndOfBlock(SI))
10491 return 0; // xform done!
10496 /// SimplifyStoreAtEndOfBlock - Turn things like:
10497 /// if () { *P = v1; } else { *P = v2 }
10498 /// into a phi node with a store in the successor.
10500 /// Simplify things like:
10501 /// *P = v1; if () { *P = v2; }
10502 /// into a phi node with a store in the successor.
10504 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10505 BasicBlock *StoreBB = SI.getParent();
10507 // Check to see if the successor block has exactly two incoming edges. If
10508 // so, see if the other predecessor contains a store to the same location.
10509 // if so, insert a PHI node (if needed) and move the stores down.
10510 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10512 // Determine whether Dest has exactly two predecessors and, if so, compute
10513 // the other predecessor.
10514 pred_iterator PI = pred_begin(DestBB);
10515 BasicBlock *OtherBB = 0;
10516 if (*PI != StoreBB)
10519 if (PI == pred_end(DestBB))
10522 if (*PI != StoreBB) {
10527 if (++PI != pred_end(DestBB))
10531 // Verify that the other block ends in a branch and is not otherwise empty.
10532 BasicBlock::iterator BBI = OtherBB->getTerminator();
10533 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10534 if (!OtherBr || BBI == OtherBB->begin())
10537 // If the other block ends in an unconditional branch, check for the 'if then
10538 // else' case. there is an instruction before the branch.
10539 StoreInst *OtherStore = 0;
10540 if (OtherBr->isUnconditional()) {
10541 // If this isn't a store, or isn't a store to the same location, bail out.
10543 OtherStore = dyn_cast<StoreInst>(BBI);
10544 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10547 // Otherwise, the other block ended with a conditional branch. If one of the
10548 // destinations is StoreBB, then we have the if/then case.
10549 if (OtherBr->getSuccessor(0) != StoreBB &&
10550 OtherBr->getSuccessor(1) != StoreBB)
10553 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10554 // if/then triangle. See if there is a store to the same ptr as SI that
10555 // lives in OtherBB.
10557 // Check to see if we find the matching store.
10558 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10559 if (OtherStore->getOperand(1) != SI.getOperand(1))
10563 // If we find something that may be using the stored value, or if we run
10564 // out of instructions, we can't do the xform.
10565 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
10566 BBI == OtherBB->begin())
10570 // In order to eliminate the store in OtherBr, we have to
10571 // make sure nothing reads the stored value in StoreBB.
10572 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10573 // FIXME: This should really be AA driven.
10574 if (isa<LoadInst>(I) || I->mayWriteToMemory())
10579 // Insert a PHI node now if we need it.
10580 Value *MergedVal = OtherStore->getOperand(0);
10581 if (MergedVal != SI.getOperand(0)) {
10582 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10583 PN->reserveOperandSpace(2);
10584 PN->addIncoming(SI.getOperand(0), SI.getParent());
10585 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10586 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10589 // Advance to a place where it is safe to insert the new store and
10591 BBI = DestBB->begin();
10592 while (isa<PHINode>(BBI)) ++BBI;
10593 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10594 OtherStore->isVolatile()), *BBI);
10596 // Nuke the old stores.
10597 EraseInstFromFunction(SI);
10598 EraseInstFromFunction(*OtherStore);
10604 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10605 // Change br (not X), label True, label False to: br X, label False, True
10607 BasicBlock *TrueDest;
10608 BasicBlock *FalseDest;
10609 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10610 !isa<Constant>(X)) {
10611 // Swap Destinations and condition...
10612 BI.setCondition(X);
10613 BI.setSuccessor(0, FalseDest);
10614 BI.setSuccessor(1, TrueDest);
10618 // Cannonicalize fcmp_one -> fcmp_oeq
10619 FCmpInst::Predicate FPred; Value *Y;
10620 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10621 TrueDest, FalseDest)))
10622 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10623 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10624 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10625 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10626 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10627 NewSCC->takeName(I);
10628 // Swap Destinations and condition...
10629 BI.setCondition(NewSCC);
10630 BI.setSuccessor(0, FalseDest);
10631 BI.setSuccessor(1, TrueDest);
10632 RemoveFromWorkList(I);
10633 I->eraseFromParent();
10634 AddToWorkList(NewSCC);
10638 // Cannonicalize icmp_ne -> icmp_eq
10639 ICmpInst::Predicate IPred;
10640 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10641 TrueDest, FalseDest)))
10642 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10643 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10644 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10645 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10646 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10647 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10648 NewSCC->takeName(I);
10649 // Swap Destinations and condition...
10650 BI.setCondition(NewSCC);
10651 BI.setSuccessor(0, FalseDest);
10652 BI.setSuccessor(1, TrueDest);
10653 RemoveFromWorkList(I);
10654 I->eraseFromParent();;
10655 AddToWorkList(NewSCC);
10662 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10663 Value *Cond = SI.getCondition();
10664 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10665 if (I->getOpcode() == Instruction::Add)
10666 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10667 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10668 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10669 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10671 SI.setOperand(0, I->getOperand(0));
10679 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10680 /// is to leave as a vector operation.
10681 static bool CheapToScalarize(Value *V, bool isConstant) {
10682 if (isa<ConstantAggregateZero>(V))
10684 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10685 if (isConstant) return true;
10686 // If all elts are the same, we can extract.
10687 Constant *Op0 = C->getOperand(0);
10688 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10689 if (C->getOperand(i) != Op0)
10693 Instruction *I = dyn_cast<Instruction>(V);
10694 if (!I) return false;
10696 // Insert element gets simplified to the inserted element or is deleted if
10697 // this is constant idx extract element and its a constant idx insertelt.
10698 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10699 isa<ConstantInt>(I->getOperand(2)))
10701 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10703 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10704 if (BO->hasOneUse() &&
10705 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10706 CheapToScalarize(BO->getOperand(1), isConstant)))
10708 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10709 if (CI->hasOneUse() &&
10710 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10711 CheapToScalarize(CI->getOperand(1), isConstant)))
10717 /// Read and decode a shufflevector mask.
10719 /// It turns undef elements into values that are larger than the number of
10720 /// elements in the input.
10721 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10722 unsigned NElts = SVI->getType()->getNumElements();
10723 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10724 return std::vector<unsigned>(NElts, 0);
10725 if (isa<UndefValue>(SVI->getOperand(2)))
10726 return std::vector<unsigned>(NElts, 2*NElts);
10728 std::vector<unsigned> Result;
10729 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10730 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10731 if (isa<UndefValue>(CP->getOperand(i)))
10732 Result.push_back(NElts*2); // undef -> 8
10734 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10738 /// FindScalarElement - Given a vector and an element number, see if the scalar
10739 /// value is already around as a register, for example if it were inserted then
10740 /// extracted from the vector.
10741 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10742 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10743 const VectorType *PTy = cast<VectorType>(V->getType());
10744 unsigned Width = PTy->getNumElements();
10745 if (EltNo >= Width) // Out of range access.
10746 return UndefValue::get(PTy->getElementType());
10748 if (isa<UndefValue>(V))
10749 return UndefValue::get(PTy->getElementType());
10750 else if (isa<ConstantAggregateZero>(V))
10751 return Constant::getNullValue(PTy->getElementType());
10752 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10753 return CP->getOperand(EltNo);
10754 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10755 // If this is an insert to a variable element, we don't know what it is.
10756 if (!isa<ConstantInt>(III->getOperand(2)))
10758 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10760 // If this is an insert to the element we are looking for, return the
10762 if (EltNo == IIElt)
10763 return III->getOperand(1);
10765 // Otherwise, the insertelement doesn't modify the value, recurse on its
10767 return FindScalarElement(III->getOperand(0), EltNo);
10768 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10769 unsigned InEl = getShuffleMask(SVI)[EltNo];
10771 return FindScalarElement(SVI->getOperand(0), InEl);
10772 else if (InEl < Width*2)
10773 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10775 return UndefValue::get(PTy->getElementType());
10778 // Otherwise, we don't know.
10782 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10784 // If vector val is undef, replace extract with scalar undef.
10785 if (isa<UndefValue>(EI.getOperand(0)))
10786 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10788 // If vector val is constant 0, replace extract with scalar 0.
10789 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10790 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10792 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10793 // If vector val is constant with uniform operands, replace EI
10794 // with that operand
10795 Constant *op0 = C->getOperand(0);
10796 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10797 if (C->getOperand(i) != op0) {
10802 return ReplaceInstUsesWith(EI, op0);
10805 // If extracting a specified index from the vector, see if we can recursively
10806 // find a previously computed scalar that was inserted into the vector.
10807 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10808 unsigned IndexVal = IdxC->getZExtValue();
10809 unsigned VectorWidth =
10810 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10812 // If this is extracting an invalid index, turn this into undef, to avoid
10813 // crashing the code below.
10814 if (IndexVal >= VectorWidth)
10815 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10817 // This instruction only demands the single element from the input vector.
10818 // If the input vector has a single use, simplify it based on this use
10820 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10821 uint64_t UndefElts;
10822 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10825 EI.setOperand(0, V);
10830 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10831 return ReplaceInstUsesWith(EI, Elt);
10833 // If the this extractelement is directly using a bitcast from a vector of
10834 // the same number of elements, see if we can find the source element from
10835 // it. In this case, we will end up needing to bitcast the scalars.
10836 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10837 if (const VectorType *VT =
10838 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10839 if (VT->getNumElements() == VectorWidth)
10840 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10841 return new BitCastInst(Elt, EI.getType());
10845 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10846 if (I->hasOneUse()) {
10847 // Push extractelement into predecessor operation if legal and
10848 // profitable to do so
10849 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10850 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10851 if (CheapToScalarize(BO, isConstantElt)) {
10852 ExtractElementInst *newEI0 =
10853 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10854 EI.getName()+".lhs");
10855 ExtractElementInst *newEI1 =
10856 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10857 EI.getName()+".rhs");
10858 InsertNewInstBefore(newEI0, EI);
10859 InsertNewInstBefore(newEI1, EI);
10860 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10862 } else if (isa<LoadInst>(I)) {
10864 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10865 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10866 PointerType::get(EI.getType(), AS),EI);
10867 GetElementPtrInst *GEP =
10868 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName() + ".gep");
10869 InsertNewInstBefore(GEP, EI);
10870 return new LoadInst(GEP);
10873 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10874 // Extracting the inserted element?
10875 if (IE->getOperand(2) == EI.getOperand(1))
10876 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10877 // If the inserted and extracted elements are constants, they must not
10878 // be the same value, extract from the pre-inserted value instead.
10879 if (isa<Constant>(IE->getOperand(2)) &&
10880 isa<Constant>(EI.getOperand(1))) {
10881 AddUsesToWorkList(EI);
10882 EI.setOperand(0, IE->getOperand(0));
10885 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10886 // If this is extracting an element from a shufflevector, figure out where
10887 // it came from and extract from the appropriate input element instead.
10888 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10889 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10891 if (SrcIdx < SVI->getType()->getNumElements())
10892 Src = SVI->getOperand(0);
10893 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10894 SrcIdx -= SVI->getType()->getNumElements();
10895 Src = SVI->getOperand(1);
10897 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10899 return new ExtractElementInst(Src, SrcIdx);
10906 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10907 /// elements from either LHS or RHS, return the shuffle mask and true.
10908 /// Otherwise, return false.
10909 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10910 std::vector<Constant*> &Mask) {
10911 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10912 "Invalid CollectSingleShuffleElements");
10913 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10915 if (isa<UndefValue>(V)) {
10916 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10918 } else if (V == LHS) {
10919 for (unsigned i = 0; i != NumElts; ++i)
10920 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10922 } else if (V == RHS) {
10923 for (unsigned i = 0; i != NumElts; ++i)
10924 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10926 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10927 // If this is an insert of an extract from some other vector, include it.
10928 Value *VecOp = IEI->getOperand(0);
10929 Value *ScalarOp = IEI->getOperand(1);
10930 Value *IdxOp = IEI->getOperand(2);
10932 if (!isa<ConstantInt>(IdxOp))
10934 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10936 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10937 // Okay, we can handle this if the vector we are insertinting into is
10938 // transitively ok.
10939 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10940 // If so, update the mask to reflect the inserted undef.
10941 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10944 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10945 if (isa<ConstantInt>(EI->getOperand(1)) &&
10946 EI->getOperand(0)->getType() == V->getType()) {
10947 unsigned ExtractedIdx =
10948 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10950 // This must be extracting from either LHS or RHS.
10951 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10952 // Okay, we can handle this if the vector we are insertinting into is
10953 // transitively ok.
10954 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10955 // If so, update the mask to reflect the inserted value.
10956 if (EI->getOperand(0) == LHS) {
10957 Mask[InsertedIdx & (NumElts-1)] =
10958 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10960 assert(EI->getOperand(0) == RHS);
10961 Mask[InsertedIdx & (NumElts-1)] =
10962 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10971 // TODO: Handle shufflevector here!
10976 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10977 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10978 /// that computes V and the LHS value of the shuffle.
10979 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10981 assert(isa<VectorType>(V->getType()) &&
10982 (RHS == 0 || V->getType() == RHS->getType()) &&
10983 "Invalid shuffle!");
10984 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10986 if (isa<UndefValue>(V)) {
10987 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10989 } else if (isa<ConstantAggregateZero>(V)) {
10990 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10992 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10993 // If this is an insert of an extract from some other vector, include it.
10994 Value *VecOp = IEI->getOperand(0);
10995 Value *ScalarOp = IEI->getOperand(1);
10996 Value *IdxOp = IEI->getOperand(2);
10998 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10999 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11000 EI->getOperand(0)->getType() == V->getType()) {
11001 unsigned ExtractedIdx =
11002 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11003 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11005 // Either the extracted from or inserted into vector must be RHSVec,
11006 // otherwise we'd end up with a shuffle of three inputs.
11007 if (EI->getOperand(0) == RHS || RHS == 0) {
11008 RHS = EI->getOperand(0);
11009 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
11010 Mask[InsertedIdx & (NumElts-1)] =
11011 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
11015 if (VecOp == RHS) {
11016 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
11017 // Everything but the extracted element is replaced with the RHS.
11018 for (unsigned i = 0; i != NumElts; ++i) {
11019 if (i != InsertedIdx)
11020 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
11025 // If this insertelement is a chain that comes from exactly these two
11026 // vectors, return the vector and the effective shuffle.
11027 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
11028 return EI->getOperand(0);
11033 // TODO: Handle shufflevector here!
11035 // Otherwise, can't do anything fancy. Return an identity vector.
11036 for (unsigned i = 0; i != NumElts; ++i)
11037 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
11041 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
11042 Value *VecOp = IE.getOperand(0);
11043 Value *ScalarOp = IE.getOperand(1);
11044 Value *IdxOp = IE.getOperand(2);
11046 // Inserting an undef or into an undefined place, remove this.
11047 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
11048 ReplaceInstUsesWith(IE, VecOp);
11050 // If the inserted element was extracted from some other vector, and if the
11051 // indexes are constant, try to turn this into a shufflevector operation.
11052 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
11053 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
11054 EI->getOperand(0)->getType() == IE.getType()) {
11055 unsigned NumVectorElts = IE.getType()->getNumElements();
11056 unsigned ExtractedIdx =
11057 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
11058 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
11060 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
11061 return ReplaceInstUsesWith(IE, VecOp);
11063 if (InsertedIdx >= NumVectorElts) // Out of range insert.
11064 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
11066 // If we are extracting a value from a vector, then inserting it right
11067 // back into the same place, just use the input vector.
11068 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
11069 return ReplaceInstUsesWith(IE, VecOp);
11071 // We could theoretically do this for ANY input. However, doing so could
11072 // turn chains of insertelement instructions into a chain of shufflevector
11073 // instructions, and right now we do not merge shufflevectors. As such,
11074 // only do this in a situation where it is clear that there is benefit.
11075 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
11076 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
11077 // the values of VecOp, except then one read from EIOp0.
11078 // Build a new shuffle mask.
11079 std::vector<Constant*> Mask;
11080 if (isa<UndefValue>(VecOp))
11081 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
11083 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
11084 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
11087 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
11088 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
11089 ConstantVector::get(Mask));
11092 // If this insertelement isn't used by some other insertelement, turn it
11093 // (and any insertelements it points to), into one big shuffle.
11094 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
11095 std::vector<Constant*> Mask;
11097 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
11098 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
11099 // We now have a shuffle of LHS, RHS, Mask.
11100 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
11109 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
11110 Value *LHS = SVI.getOperand(0);
11111 Value *RHS = SVI.getOperand(1);
11112 std::vector<unsigned> Mask = getShuffleMask(&SVI);
11114 bool MadeChange = false;
11116 // Undefined shuffle mask -> undefined value.
11117 if (isa<UndefValue>(SVI.getOperand(2)))
11118 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
11120 // If we have shuffle(x, undef, mask) and any elements of mask refer to
11121 // the undef, change them to undefs.
11122 if (isa<UndefValue>(SVI.getOperand(1))) {
11123 // Scan to see if there are any references to the RHS. If so, replace them
11124 // with undef element refs and set MadeChange to true.
11125 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11126 if (Mask[i] >= e && Mask[i] != 2*e) {
11133 // Remap any references to RHS to use LHS.
11134 std::vector<Constant*> Elts;
11135 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11136 if (Mask[i] == 2*e)
11137 Elts.push_back(UndefValue::get(Type::Int32Ty));
11139 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11141 SVI.setOperand(2, ConstantVector::get(Elts));
11145 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
11146 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
11147 if (LHS == RHS || isa<UndefValue>(LHS)) {
11148 if (isa<UndefValue>(LHS) && LHS == RHS) {
11149 // shuffle(undef,undef,mask) -> undef.
11150 return ReplaceInstUsesWith(SVI, LHS);
11153 // Remap any references to RHS to use LHS.
11154 std::vector<Constant*> Elts;
11155 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11156 if (Mask[i] >= 2*e)
11157 Elts.push_back(UndefValue::get(Type::Int32Ty));
11159 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11160 (Mask[i] < e && isa<UndefValue>(LHS)))
11161 Mask[i] = 2*e; // Turn into undef.
11163 Mask[i] &= (e-1); // Force to LHS.
11164 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11167 SVI.setOperand(0, SVI.getOperand(1));
11168 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11169 SVI.setOperand(2, ConstantVector::get(Elts));
11170 LHS = SVI.getOperand(0);
11171 RHS = SVI.getOperand(1);
11175 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11176 bool isLHSID = true, isRHSID = true;
11178 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11179 if (Mask[i] >= e*2) continue; // Ignore undef values.
11180 // Is this an identity shuffle of the LHS value?
11181 isLHSID &= (Mask[i] == i);
11183 // Is this an identity shuffle of the RHS value?
11184 isRHSID &= (Mask[i]-e == i);
11187 // Eliminate identity shuffles.
11188 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11189 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11191 // If the LHS is a shufflevector itself, see if we can combine it with this
11192 // one without producing an unusual shuffle. Here we are really conservative:
11193 // we are absolutely afraid of producing a shuffle mask not in the input
11194 // program, because the code gen may not be smart enough to turn a merged
11195 // shuffle into two specific shuffles: it may produce worse code. As such,
11196 // we only merge two shuffles if the result is one of the two input shuffle
11197 // masks. In this case, merging the shuffles just removes one instruction,
11198 // which we know is safe. This is good for things like turning:
11199 // (splat(splat)) -> splat.
11200 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11201 if (isa<UndefValue>(RHS)) {
11202 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11204 std::vector<unsigned> NewMask;
11205 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11206 if (Mask[i] >= 2*e)
11207 NewMask.push_back(2*e);
11209 NewMask.push_back(LHSMask[Mask[i]]);
11211 // If the result mask is equal to the src shuffle or this shuffle mask, do
11212 // the replacement.
11213 if (NewMask == LHSMask || NewMask == Mask) {
11214 std::vector<Constant*> Elts;
11215 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11216 if (NewMask[i] >= e*2) {
11217 Elts.push_back(UndefValue::get(Type::Int32Ty));
11219 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11222 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11223 LHSSVI->getOperand(1),
11224 ConstantVector::get(Elts));
11229 return MadeChange ? &SVI : 0;
11235 /// TryToSinkInstruction - Try to move the specified instruction from its
11236 /// current block into the beginning of DestBlock, which can only happen if it's
11237 /// safe to move the instruction past all of the instructions between it and the
11238 /// end of its block.
11239 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11240 assert(I->hasOneUse() && "Invariants didn't hold!");
11242 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11243 if (isa<PHINode>(I) || I->mayWriteToMemory() || isa<TerminatorInst>(I))
11246 // Do not sink alloca instructions out of the entry block.
11247 if (isa<AllocaInst>(I) && I->getParent() ==
11248 &DestBlock->getParent()->getEntryBlock())
11251 // We can only sink load instructions if there is nothing between the load and
11252 // the end of block that could change the value.
11253 if (I->mayReadFromMemory()) {
11254 for (BasicBlock::iterator Scan = I, E = I->getParent()->end();
11256 if (Scan->mayWriteToMemory())
11260 BasicBlock::iterator InsertPos = DestBlock->begin();
11261 while (isa<PHINode>(InsertPos)) ++InsertPos;
11263 I->moveBefore(InsertPos);
11269 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11270 /// all reachable code to the worklist.
11272 /// This has a couple of tricks to make the code faster and more powerful. In
11273 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11274 /// them to the worklist (this significantly speeds up instcombine on code where
11275 /// many instructions are dead or constant). Additionally, if we find a branch
11276 /// whose condition is a known constant, we only visit the reachable successors.
11278 static void AddReachableCodeToWorklist(BasicBlock *BB,
11279 SmallPtrSet<BasicBlock*, 64> &Visited,
11281 const TargetData *TD) {
11282 std::vector<BasicBlock*> Worklist;
11283 Worklist.push_back(BB);
11285 while (!Worklist.empty()) {
11286 BB = Worklist.back();
11287 Worklist.pop_back();
11289 // We have now visited this block! If we've already been here, ignore it.
11290 if (!Visited.insert(BB)) continue;
11292 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11293 Instruction *Inst = BBI++;
11295 // DCE instruction if trivially dead.
11296 if (isInstructionTriviallyDead(Inst)) {
11298 DOUT << "IC: DCE: " << *Inst;
11299 Inst->eraseFromParent();
11303 // ConstantProp instruction if trivially constant.
11304 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11305 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11306 Inst->replaceAllUsesWith(C);
11308 Inst->eraseFromParent();
11312 IC.AddToWorkList(Inst);
11315 // Recursively visit successors. If this is a branch or switch on a
11316 // constant, only visit the reachable successor.
11317 TerminatorInst *TI = BB->getTerminator();
11318 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11319 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11320 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11321 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11322 Worklist.push_back(ReachableBB);
11325 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11326 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11327 // See if this is an explicit destination.
11328 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11329 if (SI->getCaseValue(i) == Cond) {
11330 BasicBlock *ReachableBB = SI->getSuccessor(i);
11331 Worklist.push_back(ReachableBB);
11335 // Otherwise it is the default destination.
11336 Worklist.push_back(SI->getSuccessor(0));
11341 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11342 Worklist.push_back(TI->getSuccessor(i));
11346 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11347 bool Changed = false;
11348 TD = &getAnalysis<TargetData>();
11350 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11351 << F.getNameStr() << "\n");
11354 // Do a depth-first traversal of the function, populate the worklist with
11355 // the reachable instructions. Ignore blocks that are not reachable. Keep
11356 // track of which blocks we visit.
11357 SmallPtrSet<BasicBlock*, 64> Visited;
11358 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11360 // Do a quick scan over the function. If we find any blocks that are
11361 // unreachable, remove any instructions inside of them. This prevents
11362 // the instcombine code from having to deal with some bad special cases.
11363 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11364 if (!Visited.count(BB)) {
11365 Instruction *Term = BB->getTerminator();
11366 while (Term != BB->begin()) { // Remove instrs bottom-up
11367 BasicBlock::iterator I = Term; --I;
11369 DOUT << "IC: DCE: " << *I;
11372 if (!I->use_empty())
11373 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11374 I->eraseFromParent();
11379 while (!Worklist.empty()) {
11380 Instruction *I = RemoveOneFromWorkList();
11381 if (I == 0) continue; // skip null values.
11383 // Check to see if we can DCE the instruction.
11384 if (isInstructionTriviallyDead(I)) {
11385 // Add operands to the worklist.
11386 if (I->getNumOperands() < 4)
11387 AddUsesToWorkList(*I);
11390 DOUT << "IC: DCE: " << *I;
11392 I->eraseFromParent();
11393 RemoveFromWorkList(I);
11397 // Instruction isn't dead, see if we can constant propagate it.
11398 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11399 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11401 // Add operands to the worklist.
11402 AddUsesToWorkList(*I);
11403 ReplaceInstUsesWith(*I, C);
11406 I->eraseFromParent();
11407 RemoveFromWorkList(I);
11411 // See if we can trivially sink this instruction to a successor basic block.
11412 // FIXME: Remove GetResultInst test when first class support for aggregates
11414 if (I->hasOneUse() && !isa<GetResultInst>(I)) {
11415 BasicBlock *BB = I->getParent();
11416 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11417 if (UserParent != BB) {
11418 bool UserIsSuccessor = false;
11419 // See if the user is one of our successors.
11420 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11421 if (*SI == UserParent) {
11422 UserIsSuccessor = true;
11426 // If the user is one of our immediate successors, and if that successor
11427 // only has us as a predecessors (we'd have to split the critical edge
11428 // otherwise), we can keep going.
11429 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11430 next(pred_begin(UserParent)) == pred_end(UserParent))
11431 // Okay, the CFG is simple enough, try to sink this instruction.
11432 Changed |= TryToSinkInstruction(I, UserParent);
11436 // Now that we have an instruction, try combining it to simplify it...
11440 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11441 if (Instruction *Result = visit(*I)) {
11443 // Should we replace the old instruction with a new one?
11445 DOUT << "IC: Old = " << *I
11446 << " New = " << *Result;
11448 // Everything uses the new instruction now.
11449 I->replaceAllUsesWith(Result);
11451 // Push the new instruction and any users onto the worklist.
11452 AddToWorkList(Result);
11453 AddUsersToWorkList(*Result);
11455 // Move the name to the new instruction first.
11456 Result->takeName(I);
11458 // Insert the new instruction into the basic block...
11459 BasicBlock *InstParent = I->getParent();
11460 BasicBlock::iterator InsertPos = I;
11462 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11463 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11466 InstParent->getInstList().insert(InsertPos, Result);
11468 // Make sure that we reprocess all operands now that we reduced their
11470 AddUsesToWorkList(*I);
11472 // Instructions can end up on the worklist more than once. Make sure
11473 // we do not process an instruction that has been deleted.
11474 RemoveFromWorkList(I);
11476 // Erase the old instruction.
11477 InstParent->getInstList().erase(I);
11480 DOUT << "IC: Mod = " << OrigI
11481 << " New = " << *I;
11484 // If the instruction was modified, it's possible that it is now dead.
11485 // if so, remove it.
11486 if (isInstructionTriviallyDead(I)) {
11487 // Make sure we process all operands now that we are reducing their
11489 AddUsesToWorkList(*I);
11491 // Instructions may end up in the worklist more than once. Erase all
11492 // occurrences of this instruction.
11493 RemoveFromWorkList(I);
11494 I->eraseFromParent();
11497 AddUsersToWorkList(*I);
11504 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11506 // Do an explicit clear, this shrinks the map if needed.
11507 WorklistMap.clear();
11512 bool InstCombiner::runOnFunction(Function &F) {
11513 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11515 bool EverMadeChange = false;
11517 // Iterate while there is work to do.
11518 unsigned Iteration = 0;
11519 while (DoOneIteration(F, Iteration++))
11520 EverMadeChange = true;
11521 return EverMadeChange;
11524 FunctionPass *llvm::createInstructionCombiningPass() {
11525 return new InstCombiner();