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 algebraic
12 // 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);
375 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
377 void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
378 APInt& KnownOne, unsigned Depth = 0);
379 bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0);
380 bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
382 int &NumCastsRemoved);
383 unsigned GetOrEnforceKnownAlignment(Value *V,
384 unsigned PrefAlign = 0);
387 char InstCombiner::ID = 0;
388 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
391 // getComplexity: Assign a complexity or rank value to LLVM Values...
392 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
393 static unsigned getComplexity(Value *V) {
394 if (isa<Instruction>(V)) {
395 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
399 if (isa<Argument>(V)) return 3;
400 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
403 // isOnlyUse - Return true if this instruction will be deleted if we stop using
405 static bool isOnlyUse(Value *V) {
406 return V->hasOneUse() || isa<Constant>(V);
409 // getPromotedType - Return the specified type promoted as it would be to pass
410 // though a va_arg area...
411 static const Type *getPromotedType(const Type *Ty) {
412 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
413 if (ITy->getBitWidth() < 32)
414 return Type::Int32Ty;
419 /// getBitCastOperand - If the specified operand is a CastInst or a constant
420 /// expression bitcast, return the operand value, otherwise return null.
421 static Value *getBitCastOperand(Value *V) {
422 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
423 return I->getOperand(0);
424 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
425 if (CE->getOpcode() == Instruction::BitCast)
426 return CE->getOperand(0);
430 /// This function is a wrapper around CastInst::isEliminableCastPair. It
431 /// simply extracts arguments and returns what that function returns.
432 static Instruction::CastOps
433 isEliminableCastPair(
434 const CastInst *CI, ///< The first cast instruction
435 unsigned opcode, ///< The opcode of the second cast instruction
436 const Type *DstTy, ///< The target type for the second cast instruction
437 TargetData *TD ///< The target data for pointer size
440 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
441 const Type *MidTy = CI->getType(); // B from above
443 // Get the opcodes of the two Cast instructions
444 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
445 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
447 return Instruction::CastOps(
448 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
449 DstTy, TD->getIntPtrType()));
452 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
453 /// in any code being generated. It does not require codegen if V is simple
454 /// enough or if the cast can be folded into other casts.
455 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
456 const Type *Ty, TargetData *TD) {
457 if (V->getType() == Ty || isa<Constant>(V)) return false;
459 // If this is another cast that can be eliminated, it isn't codegen either.
460 if (const CastInst *CI = dyn_cast<CastInst>(V))
461 if (isEliminableCastPair(CI, opcode, Ty, TD))
466 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
467 /// InsertBefore instruction. This is specialized a bit to avoid inserting
468 /// casts that are known to not do anything...
470 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
471 Value *V, const Type *DestTy,
472 Instruction *InsertBefore) {
473 if (V->getType() == DestTy) return V;
474 if (Constant *C = dyn_cast<Constant>(V))
475 return ConstantExpr::getCast(opcode, C, DestTy);
477 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
480 // SimplifyCommutative - This performs a few simplifications for commutative
483 // 1. Order operands such that they are listed from right (least complex) to
484 // left (most complex). This puts constants before unary operators before
487 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
488 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
490 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
491 bool Changed = false;
492 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
493 Changed = !I.swapOperands();
495 if (!I.isAssociative()) return Changed;
496 Instruction::BinaryOps Opcode = I.getOpcode();
497 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
498 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
499 if (isa<Constant>(I.getOperand(1))) {
500 Constant *Folded = ConstantExpr::get(I.getOpcode(),
501 cast<Constant>(I.getOperand(1)),
502 cast<Constant>(Op->getOperand(1)));
503 I.setOperand(0, Op->getOperand(0));
504 I.setOperand(1, Folded);
506 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
507 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
508 isOnlyUse(Op) && isOnlyUse(Op1)) {
509 Constant *C1 = cast<Constant>(Op->getOperand(1));
510 Constant *C2 = cast<Constant>(Op1->getOperand(1));
512 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
513 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
514 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
518 I.setOperand(0, New);
519 I.setOperand(1, Folded);
526 /// SimplifyCompare - For a CmpInst this function just orders the operands
527 /// so that theyare listed from right (least complex) to left (most complex).
528 /// This puts constants before unary operators before binary operators.
529 bool InstCombiner::SimplifyCompare(CmpInst &I) {
530 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
533 // Compare instructions are not associative so there's nothing else we can do.
537 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
538 // if the LHS is a constant zero (which is the 'negate' form).
540 static inline Value *dyn_castNegVal(Value *V) {
541 if (BinaryOperator::isNeg(V))
542 return BinaryOperator::getNegArgument(V);
544 // Constants can be considered to be negated values if they can be folded.
545 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
546 return ConstantExpr::getNeg(C);
550 static inline Value *dyn_castNotVal(Value *V) {
551 if (BinaryOperator::isNot(V))
552 return BinaryOperator::getNotArgument(V);
554 // Constants can be considered to be not'ed values...
555 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
556 return ConstantInt::get(~C->getValue());
560 // dyn_castFoldableMul - If this value is a multiply that can be folded into
561 // other computations (because it has a constant operand), return the
562 // non-constant operand of the multiply, and set CST to point to the multiplier.
563 // Otherwise, return null.
565 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
566 if (V->hasOneUse() && V->getType()->isInteger())
567 if (Instruction *I = dyn_cast<Instruction>(V)) {
568 if (I->getOpcode() == Instruction::Mul)
569 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
570 return I->getOperand(0);
571 if (I->getOpcode() == Instruction::Shl)
572 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
573 // The multiplier is really 1 << CST.
574 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
575 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
576 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
577 return I->getOperand(0);
583 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
584 /// expression, return it.
585 static User *dyn_castGetElementPtr(Value *V) {
586 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
587 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
588 if (CE->getOpcode() == Instruction::GetElementPtr)
589 return cast<User>(V);
593 /// getOpcode - If this is an Instruction or a ConstantExpr, return the
594 /// opcode value. Otherwise return UserOp1.
595 static unsigned getOpcode(User *U) {
596 if (Instruction *I = dyn_cast<Instruction>(U))
597 return I->getOpcode();
598 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U))
599 return CE->getOpcode();
600 // Use UserOp1 to mean there's no opcode.
601 return Instruction::UserOp1;
604 /// AddOne - Add one to a ConstantInt
605 static ConstantInt *AddOne(ConstantInt *C) {
606 APInt Val(C->getValue());
607 return ConstantInt::get(++Val);
609 /// SubOne - Subtract one from a ConstantInt
610 static ConstantInt *SubOne(ConstantInt *C) {
611 APInt Val(C->getValue());
612 return ConstantInt::get(--Val);
614 /// Add - Add two ConstantInts together
615 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
616 return ConstantInt::get(C1->getValue() + C2->getValue());
618 /// And - Bitwise AND two ConstantInts together
619 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
620 return ConstantInt::get(C1->getValue() & C2->getValue());
622 /// Subtract - Subtract one ConstantInt from another
623 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
624 return ConstantInt::get(C1->getValue() - C2->getValue());
626 /// Multiply - Multiply two ConstantInts together
627 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
628 return ConstantInt::get(C1->getValue() * C2->getValue());
630 /// MultiplyOverflows - True if the multiply can not be expressed in an int
632 static bool MultiplyOverflows(ConstantInt *C1, ConstantInt *C2, bool sign) {
633 uint32_t W = C1->getBitWidth();
634 APInt LHSExt = C1->getValue(), RHSExt = C2->getValue();
643 APInt MulExt = LHSExt * RHSExt;
646 APInt Min = APInt::getSignedMinValue(W).sext(W * 2);
647 APInt Max = APInt::getSignedMaxValue(W).sext(W * 2);
648 return MulExt.slt(Min) || MulExt.sgt(Max);
650 return MulExt.ugt(APInt::getLowBitsSet(W * 2, W));
653 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
654 /// known to be either zero or one and return them in the KnownZero/KnownOne
655 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
657 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
658 /// we cannot optimize based on the assumption that it is zero without changing
659 /// it to be an explicit zero. If we don't change it to zero, other code could
660 /// optimized based on the contradictory assumption that it is non-zero.
661 /// Because instcombine aggressively folds operations with undef args anyway,
662 /// this won't lose us code quality.
663 void InstCombiner::ComputeMaskedBits(Value *V, const APInt &Mask,
664 APInt& KnownZero, APInt& KnownOne,
666 assert(V && "No Value?");
667 assert(Depth <= 6 && "Limit Search Depth");
668 uint32_t BitWidth = Mask.getBitWidth();
669 assert((V->getType()->isInteger() || isa<PointerType>(V->getType())) &&
670 "Not integer or pointer type!");
671 assert((!TD || TD->getTypeSizeInBits(V->getType()) == BitWidth) &&
672 (!isa<IntegerType>(V->getType()) ||
673 V->getType()->getPrimitiveSizeInBits() == BitWidth) &&
674 KnownZero.getBitWidth() == BitWidth &&
675 KnownOne.getBitWidth() == BitWidth &&
676 "V, Mask, KnownOne and KnownZero should have same BitWidth");
677 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
678 // We know all of the bits for a constant!
679 KnownOne = CI->getValue() & Mask;
680 KnownZero = ~KnownOne & Mask;
683 // Null is all-zeros.
684 if (isa<ConstantPointerNull>(V)) {
689 // The address of an aligned GlobalValue has trailing zeros.
690 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
691 unsigned Align = GV->getAlignment();
692 if (Align == 0 && TD && GV->getType()->getElementType()->isSized())
693 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
695 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
696 CountTrailingZeros_32(Align));
703 if (Depth == 6 || Mask == 0)
704 return; // Limit search depth.
706 User *I = dyn_cast<User>(V);
709 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
710 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
712 switch (getOpcode(I)) {
714 case Instruction::And: {
715 // If either the LHS or the RHS are Zero, the result is zero.
716 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
717 APInt Mask2(Mask & ~KnownZero);
718 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
719 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
720 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
722 // Output known-1 bits are only known if set in both the LHS & RHS.
723 KnownOne &= KnownOne2;
724 // Output known-0 are known to be clear if zero in either the LHS | RHS.
725 KnownZero |= KnownZero2;
728 case Instruction::Or: {
729 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
730 APInt Mask2(Mask & ~KnownOne);
731 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
732 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
733 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
735 // Output known-0 bits are only known if clear in both the LHS & RHS.
736 KnownZero &= KnownZero2;
737 // Output known-1 are known to be set if set in either the LHS | RHS.
738 KnownOne |= KnownOne2;
741 case Instruction::Xor: {
742 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
743 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
744 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
745 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
747 // Output known-0 bits are known if clear or set in both the LHS & RHS.
748 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
749 // Output known-1 are known to be set if set in only one of the LHS, RHS.
750 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
751 KnownZero = KnownZeroOut;
754 case Instruction::Mul: {
755 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
756 ComputeMaskedBits(I->getOperand(1), Mask2, KnownZero, KnownOne, Depth+1);
757 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
758 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
759 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
761 // If low bits are zero in either operand, output low known-0 bits.
762 // More trickiness is possible, but this is sufficient for the
763 // interesting case of alignment computation.
765 unsigned TrailZ = KnownZero.countTrailingOnes() +
766 KnownZero2.countTrailingOnes();
767 TrailZ = std::min(TrailZ, BitWidth);
768 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ);
772 case Instruction::Select:
773 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
774 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
775 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
776 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
778 // Only known if known in both the LHS and RHS.
779 KnownOne &= KnownOne2;
780 KnownZero &= KnownZero2;
782 case Instruction::FPTrunc:
783 case Instruction::FPExt:
784 case Instruction::FPToUI:
785 case Instruction::FPToSI:
786 case Instruction::SIToFP:
787 case Instruction::UIToFP:
788 return; // Can't work with floating point.
789 case Instruction::PtrToInt:
790 case Instruction::IntToPtr:
791 // We can't handle these if we don't know the pointer size.
793 // Fall through and handle them the same as zext/trunc.
794 case Instruction::ZExt:
795 case Instruction::Trunc: {
796 // All these have integer operands
797 const Type *SrcTy = I->getOperand(0)->getType();
798 uint32_t SrcBitWidth = TD ?
799 TD->getTypeSizeInBits(SrcTy) :
800 SrcTy->getPrimitiveSizeInBits();
802 MaskIn.zextOrTrunc(SrcBitWidth);
803 KnownZero.zextOrTrunc(SrcBitWidth);
804 KnownOne.zextOrTrunc(SrcBitWidth);
805 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
806 KnownZero.zextOrTrunc(BitWidth);
807 KnownOne.zextOrTrunc(BitWidth);
808 // Any top bits are known to be zero.
809 if (BitWidth > SrcBitWidth)
810 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
813 case Instruction::BitCast: {
814 const Type *SrcTy = I->getOperand(0)->getType();
815 if (SrcTy->isInteger() || isa<PointerType>(SrcTy)) {
816 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
821 case Instruction::SExt: {
822 // Compute the bits in the result that are not present in the input.
823 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
824 uint32_t SrcBitWidth = SrcTy->getBitWidth();
827 MaskIn.trunc(SrcBitWidth);
828 KnownZero.trunc(SrcBitWidth);
829 KnownOne.trunc(SrcBitWidth);
830 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
831 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
832 KnownZero.zext(BitWidth);
833 KnownOne.zext(BitWidth);
835 // If the sign bit of the input is known set or clear, then we know the
836 // top bits of the result.
837 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
838 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
839 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
840 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
843 case Instruction::Shl:
844 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
845 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
846 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
847 APInt Mask2(Mask.lshr(ShiftAmt));
848 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
849 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
850 KnownZero <<= ShiftAmt;
851 KnownOne <<= ShiftAmt;
852 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
856 case Instruction::LShr:
857 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
858 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
859 // Compute the new bits that are at the top now.
860 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
862 // Unsigned shift right.
863 APInt Mask2(Mask.shl(ShiftAmt));
864 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
865 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
866 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
867 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
868 // high bits known zero.
869 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
873 case Instruction::AShr:
874 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
875 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
876 // Compute the new bits that are at the top now.
877 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
879 // Signed shift right.
880 APInt Mask2(Mask.shl(ShiftAmt));
881 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
882 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
883 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
884 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
886 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
887 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
888 KnownZero |= HighBits;
889 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
890 KnownOne |= HighBits;
894 case Instruction::Sub: {
895 if (ConstantInt *CLHS = dyn_cast<ConstantInt>(I->getOperand(0))) {
896 // We know that the top bits of C-X are clear if X contains less bits
897 // than C (i.e. no wrap-around can happen). For example, 20-X is
898 // positive if we can prove that X is >= 0 and < 16.
899 if (!CLHS->getValue().isNegative()) {
900 unsigned NLZ = (CLHS->getValue()+1).countLeadingZeros();
901 // NLZ can't be BitWidth with no sign bit
902 APInt MaskV = APInt::getHighBitsSet(BitWidth, NLZ+1);
903 ComputeMaskedBits(I->getOperand(1), MaskV, KnownZero, KnownOne, Depth+1);
905 // If all of the MaskV bits are known to be zero, then we know the output
906 // top bits are zero, because we now know that the output is from [0-C].
907 if ((KnownZero & MaskV) == MaskV) {
908 unsigned NLZ2 = CLHS->getValue().countLeadingZeros();
909 // Top bits known zero.
910 KnownZero = APInt::getHighBitsSet(BitWidth, NLZ2) & Mask;
911 KnownOne = APInt(BitWidth, 0); // No one bits known.
913 KnownZero = KnownOne = APInt(BitWidth, 0); // Otherwise, nothing known.
920 case Instruction::Add: {
921 // If either the LHS or the RHS are Zero, the result is zero.
922 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
923 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
924 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
925 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
927 // Output known-0 bits are known if clear or set in both the low clear bits
928 // common to both LHS & RHS. For example, 8+(X<<3) is known to have the
930 unsigned KnownZeroOut = std::min(KnownZero.countTrailingOnes(),
931 KnownZero2.countTrailingOnes());
933 KnownZero = APInt::getLowBitsSet(BitWidth, KnownZeroOut);
934 KnownOne = APInt(BitWidth, 0);
937 case Instruction::SRem:
938 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
939 APInt RA = Rem->getValue();
940 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
941 APInt LowBits = RA.isStrictlyPositive() ? ((RA - 1) | RA) : ~RA;
942 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
943 ComputeMaskedBits(I->getOperand(0), Mask2,KnownZero2,KnownOne2,Depth+1);
945 // The sign of a remainder is equal to the sign of the first
946 // operand (zero being positive).
947 if (KnownZero2[BitWidth-1] || ((KnownZero2 & LowBits) == LowBits))
948 KnownZero2 |= ~LowBits;
949 else if (KnownOne2[BitWidth-1])
950 KnownOne2 |= ~LowBits;
952 KnownZero |= KnownZero2 & Mask;
953 KnownOne |= KnownOne2 & Mask;
955 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
959 case Instruction::URem:
960 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
961 APInt RA = Rem->getValue();
962 if (RA.isStrictlyPositive() && RA.isPowerOf2()) {
963 APInt LowBits = (RA - 1) | RA;
964 APInt Mask2 = LowBits & Mask;
965 KnownZero |= ~LowBits & Mask;
966 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne,Depth+1);
967 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
970 // Since the result is less than or equal to RHS, any leading zero bits
971 // in RHS must also exist in the result.
972 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
973 ComputeMaskedBits(I->getOperand(1), AllOnes, KnownZero2, KnownOne2,
976 uint32_t Leaders = KnownZero2.countLeadingOnes();
977 KnownZero |= APInt::getHighBitsSet(BitWidth, Leaders) & Mask;
978 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
982 case Instruction::Alloca:
983 case Instruction::Malloc: {
984 AllocationInst *AI = cast<AllocationInst>(V);
985 unsigned Align = AI->getAlignment();
986 if (Align == 0 && TD) {
987 if (isa<AllocaInst>(AI))
988 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
989 else if (isa<MallocInst>(AI)) {
990 // Malloc returns maximally aligned memory.
991 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
994 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
997 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
1002 KnownZero = Mask & APInt::getLowBitsSet(BitWidth,
1003 CountTrailingZeros_32(Align));
1006 case Instruction::GetElementPtr: {
1007 // Analyze all of the subscripts of this getelementptr instruction
1008 // to determine if we can prove known low zero bits.
1009 APInt LocalMask = APInt::getAllOnesValue(BitWidth);
1010 APInt LocalKnownZero(BitWidth, 0), LocalKnownOne(BitWidth, 0);
1011 ComputeMaskedBits(I->getOperand(0), LocalMask,
1012 LocalKnownZero, LocalKnownOne, Depth+1);
1013 unsigned TrailZ = LocalKnownZero.countTrailingOnes();
1015 gep_type_iterator GTI = gep_type_begin(I);
1016 for (unsigned i = 1, e = I->getNumOperands(); i != e; ++i, ++GTI) {
1017 Value *Index = I->getOperand(i);
1018 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
1019 // Handle struct member offset arithmetic.
1021 const StructLayout *SL = TD->getStructLayout(STy);
1022 unsigned Idx = cast<ConstantInt>(Index)->getZExtValue();
1023 uint64_t Offset = SL->getElementOffset(Idx);
1024 TrailZ = std::min(TrailZ,
1025 CountTrailingZeros_64(Offset));
1027 // Handle array index arithmetic.
1028 const Type *IndexedTy = GTI.getIndexedType();
1029 if (!IndexedTy->isSized()) return;
1030 unsigned GEPOpiBits = Index->getType()->getPrimitiveSizeInBits();
1031 uint64_t TypeSize = TD ? TD->getABITypeSize(IndexedTy) : 1;
1032 LocalMask = APInt::getAllOnesValue(GEPOpiBits);
1033 LocalKnownZero = LocalKnownOne = APInt(GEPOpiBits, 0);
1034 ComputeMaskedBits(Index, LocalMask,
1035 LocalKnownZero, LocalKnownOne, Depth+1);
1036 TrailZ = std::min(TrailZ,
1037 CountTrailingZeros_64(TypeSize) +
1038 LocalKnownZero.countTrailingOnes());
1042 KnownZero = APInt::getLowBitsSet(BitWidth, TrailZ) & Mask;
1045 case Instruction::PHI: {
1046 PHINode *P = cast<PHINode>(I);
1047 // Handle the case of a simple two-predecessor recurrence PHI.
1048 // There's a lot more that could theoretically be done here, but
1049 // this is sufficient to catch some interesting cases.
1050 if (P->getNumIncomingValues() == 2) {
1051 for (unsigned i = 0; i != 2; ++i) {
1052 Value *L = P->getIncomingValue(i);
1053 Value *R = P->getIncomingValue(!i);
1054 User *LU = dyn_cast<User>(L);
1055 unsigned Opcode = LU ? getOpcode(LU) : (unsigned)Instruction::UserOp1;
1056 // Check for operations that have the property that if
1057 // both their operands have low zero bits, the result
1058 // will have low zero bits.
1059 if (Opcode == Instruction::Add ||
1060 Opcode == Instruction::Sub ||
1061 Opcode == Instruction::And ||
1062 Opcode == Instruction::Or ||
1063 Opcode == Instruction::Mul) {
1064 Value *LL = LU->getOperand(0);
1065 Value *LR = LU->getOperand(1);
1066 // Find a recurrence.
1073 // Ok, we have a PHI of the form L op= R. Check for low
1075 APInt Mask2 = APInt::getAllOnesValue(BitWidth);
1076 ComputeMaskedBits(R, Mask2, KnownZero2, KnownOne2, Depth+1);
1077 Mask2 = APInt::getLowBitsSet(BitWidth,
1078 KnownZero2.countTrailingOnes());
1081 ComputeMaskedBits(L, Mask2, KnownZero2, KnownOne2, Depth+1);
1083 APInt::getLowBitsSet(BitWidth,
1084 KnownZero2.countTrailingOnes());
1094 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
1095 /// this predicate to simplify operations downstream. Mask is known to be zero
1096 /// for bits that V cannot have.
1097 bool InstCombiner::MaskedValueIsZero(Value *V, const APInt& Mask,
1099 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
1100 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
1101 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1102 return (KnownZero & Mask) == Mask;
1105 /// ShrinkDemandedConstant - Check to see if the specified operand of the
1106 /// specified instruction is a constant integer. If so, check to see if there
1107 /// are any bits set in the constant that are not demanded. If so, shrink the
1108 /// constant and return true.
1109 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
1111 assert(I && "No instruction?");
1112 assert(OpNo < I->getNumOperands() && "Operand index too large");
1114 // If the operand is not a constant integer, nothing to do.
1115 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
1116 if (!OpC) return false;
1118 // If there are no bits set that aren't demanded, nothing to do.
1119 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
1120 if ((~Demanded & OpC->getValue()) == 0)
1123 // This instruction is producing bits that are not demanded. Shrink the RHS.
1124 Demanded &= OpC->getValue();
1125 I->setOperand(OpNo, ConstantInt::get(Demanded));
1129 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
1130 // set of known zero and one bits, compute the maximum and minimum values that
1131 // could have the specified known zero and known one bits, returning them in
1133 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
1134 const APInt& KnownZero,
1135 const APInt& KnownOne,
1136 APInt& Min, APInt& Max) {
1137 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
1138 assert(KnownZero.getBitWidth() == BitWidth &&
1139 KnownOne.getBitWidth() == BitWidth &&
1140 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
1141 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1142 APInt UnknownBits = ~(KnownZero|KnownOne);
1144 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
1145 // bit if it is unknown.
1147 Max = KnownOne|UnknownBits;
1149 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
1150 Min.set(BitWidth-1);
1151 Max.clear(BitWidth-1);
1155 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
1156 // a set of known zero and one bits, compute the maximum and minimum values that
1157 // could have the specified known zero and known one bits, returning them in
1159 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
1160 const APInt &KnownZero,
1161 const APInt &KnownOne,
1162 APInt &Min, APInt &Max) {
1163 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
1164 assert(KnownZero.getBitWidth() == BitWidth &&
1165 KnownOne.getBitWidth() == BitWidth &&
1166 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
1167 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
1168 APInt UnknownBits = ~(KnownZero|KnownOne);
1170 // The minimum value is when the unknown bits are all zeros.
1172 // The maximum value is when the unknown bits are all ones.
1173 Max = KnownOne|UnknownBits;
1176 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
1177 /// value based on the demanded bits. When this function is called, it is known
1178 /// that only the bits set in DemandedMask of the result of V are ever used
1179 /// downstream. Consequently, depending on the mask and V, it may be possible
1180 /// to replace V with a constant or one of its operands. In such cases, this
1181 /// function does the replacement and returns true. In all other cases, it
1182 /// returns false after analyzing the expression and setting KnownOne and known
1183 /// to be one in the expression. KnownZero contains all the bits that are known
1184 /// to be zero in the expression. These are provided to potentially allow the
1185 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
1186 /// the expression. KnownOne and KnownZero always follow the invariant that
1187 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
1188 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
1189 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
1190 /// and KnownOne must all be the same.
1191 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
1192 APInt& KnownZero, APInt& KnownOne,
1194 assert(V != 0 && "Null pointer of Value???");
1195 assert(Depth <= 6 && "Limit Search Depth");
1196 uint32_t BitWidth = DemandedMask.getBitWidth();
1197 const IntegerType *VTy = cast<IntegerType>(V->getType());
1198 assert(VTy->getBitWidth() == BitWidth &&
1199 KnownZero.getBitWidth() == BitWidth &&
1200 KnownOne.getBitWidth() == BitWidth &&
1201 "Value *V, DemandedMask, KnownZero and KnownOne \
1202 must have same BitWidth");
1203 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
1204 // We know all of the bits for a constant!
1205 KnownOne = CI->getValue() & DemandedMask;
1206 KnownZero = ~KnownOne & DemandedMask;
1212 if (!V->hasOneUse()) { // Other users may use these bits.
1213 if (Depth != 0) { // Not at the root.
1214 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
1215 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
1218 // If this is the root being simplified, allow it to have multiple uses,
1219 // just set the DemandedMask to all bits.
1220 DemandedMask = APInt::getAllOnesValue(BitWidth);
1221 } else if (DemandedMask == 0) { // Not demanding any bits from V.
1222 if (V != UndefValue::get(VTy))
1223 return UpdateValueUsesWith(V, UndefValue::get(VTy));
1225 } else if (Depth == 6) { // Limit search depth.
1229 Instruction *I = dyn_cast<Instruction>(V);
1230 if (!I) return false; // Only analyze instructions.
1232 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
1233 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
1234 switch (I->getOpcode()) {
1236 case Instruction::And:
1237 // If either the LHS or the RHS are Zero, the result is zero.
1238 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1239 RHSKnownZero, RHSKnownOne, Depth+1))
1241 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1242 "Bits known to be one AND zero?");
1244 // If something is known zero on the RHS, the bits aren't demanded on the
1246 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
1247 LHSKnownZero, LHSKnownOne, Depth+1))
1249 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1250 "Bits known to be one AND zero?");
1252 // If all of the demanded bits are known 1 on one side, return the other.
1253 // These bits cannot contribute to the result of the 'and'.
1254 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
1255 (DemandedMask & ~LHSKnownZero))
1256 return UpdateValueUsesWith(I, I->getOperand(0));
1257 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
1258 (DemandedMask & ~RHSKnownZero))
1259 return UpdateValueUsesWith(I, I->getOperand(1));
1261 // If all of the demanded bits in the inputs are known zeros, return zero.
1262 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
1263 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
1265 // If the RHS is a constant, see if we can simplify it.
1266 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
1267 return UpdateValueUsesWith(I, I);
1269 // Output known-1 bits are only known if set in both the LHS & RHS.
1270 RHSKnownOne &= LHSKnownOne;
1271 // Output known-0 are known to be clear if zero in either the LHS | RHS.
1272 RHSKnownZero |= LHSKnownZero;
1274 case Instruction::Or:
1275 // If either the LHS or the RHS are One, the result is One.
1276 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1277 RHSKnownZero, RHSKnownOne, Depth+1))
1279 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1280 "Bits known to be one AND zero?");
1281 // If something is known one on the RHS, the bits aren't demanded on the
1283 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
1284 LHSKnownZero, LHSKnownOne, Depth+1))
1286 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1287 "Bits known to be one AND zero?");
1289 // If all of the demanded bits are known zero on one side, return the other.
1290 // These bits cannot contribute to the result of the 'or'.
1291 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1292 (DemandedMask & ~LHSKnownOne))
1293 return UpdateValueUsesWith(I, I->getOperand(0));
1294 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1295 (DemandedMask & ~RHSKnownOne))
1296 return UpdateValueUsesWith(I, I->getOperand(1));
1298 // If all of the potentially set bits on one side are known to be set on
1299 // the other side, just use the 'other' side.
1300 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1301 (DemandedMask & (~RHSKnownZero)))
1302 return UpdateValueUsesWith(I, I->getOperand(0));
1303 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1304 (DemandedMask & (~LHSKnownZero)))
1305 return UpdateValueUsesWith(I, I->getOperand(1));
1307 // If the RHS is a constant, see if we can simplify it.
1308 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1309 return UpdateValueUsesWith(I, I);
1311 // Output known-0 bits are only known if clear in both the LHS & RHS.
1312 RHSKnownZero &= LHSKnownZero;
1313 // Output known-1 are known to be set if set in either the LHS | RHS.
1314 RHSKnownOne |= LHSKnownOne;
1316 case Instruction::Xor: {
1317 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1318 RHSKnownZero, RHSKnownOne, Depth+1))
1320 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1321 "Bits known to be one AND zero?");
1322 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1323 LHSKnownZero, LHSKnownOne, Depth+1))
1325 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1326 "Bits known to be one AND zero?");
1328 // If all of the demanded bits are known zero on one side, return the other.
1329 // These bits cannot contribute to the result of the 'xor'.
1330 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1331 return UpdateValueUsesWith(I, I->getOperand(0));
1332 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1333 return UpdateValueUsesWith(I, I->getOperand(1));
1335 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1336 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1337 (RHSKnownOne & LHSKnownOne);
1338 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1339 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1340 (RHSKnownOne & LHSKnownZero);
1342 // If all of the demanded bits are known to be zero on one side or the
1343 // other, turn this into an *inclusive* or.
1344 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1345 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1347 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1349 InsertNewInstBefore(Or, *I);
1350 return UpdateValueUsesWith(I, Or);
1353 // If all of the demanded bits on one side are known, and all of the set
1354 // bits on that side are also known to be set on the other side, turn this
1355 // into an AND, as we know the bits will be cleared.
1356 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1357 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1359 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1360 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1362 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1363 InsertNewInstBefore(And, *I);
1364 return UpdateValueUsesWith(I, And);
1368 // If the RHS is a constant, see if we can simplify it.
1369 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1370 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1371 return UpdateValueUsesWith(I, I);
1373 RHSKnownZero = KnownZeroOut;
1374 RHSKnownOne = KnownOneOut;
1377 case Instruction::Select:
1378 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1379 RHSKnownZero, RHSKnownOne, Depth+1))
1381 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1382 LHSKnownZero, LHSKnownOne, Depth+1))
1384 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1385 "Bits known to be one AND zero?");
1386 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1387 "Bits known to be one AND zero?");
1389 // If the operands are constants, see if we can simplify them.
1390 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1391 return UpdateValueUsesWith(I, I);
1392 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1393 return UpdateValueUsesWith(I, I);
1395 // Only known if known in both the LHS and RHS.
1396 RHSKnownOne &= LHSKnownOne;
1397 RHSKnownZero &= LHSKnownZero;
1399 case Instruction::Trunc: {
1401 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1402 DemandedMask.zext(truncBf);
1403 RHSKnownZero.zext(truncBf);
1404 RHSKnownOne.zext(truncBf);
1405 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1406 RHSKnownZero, RHSKnownOne, Depth+1))
1408 DemandedMask.trunc(BitWidth);
1409 RHSKnownZero.trunc(BitWidth);
1410 RHSKnownOne.trunc(BitWidth);
1411 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1412 "Bits known to be one AND zero?");
1415 case Instruction::BitCast:
1416 if (!I->getOperand(0)->getType()->isInteger())
1419 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1420 RHSKnownZero, RHSKnownOne, Depth+1))
1422 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1423 "Bits known to be one AND zero?");
1425 case Instruction::ZExt: {
1426 // Compute the bits in the result that are not present in the input.
1427 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1428 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1430 DemandedMask.trunc(SrcBitWidth);
1431 RHSKnownZero.trunc(SrcBitWidth);
1432 RHSKnownOne.trunc(SrcBitWidth);
1433 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1434 RHSKnownZero, RHSKnownOne, Depth+1))
1436 DemandedMask.zext(BitWidth);
1437 RHSKnownZero.zext(BitWidth);
1438 RHSKnownOne.zext(BitWidth);
1439 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1440 "Bits known to be one AND zero?");
1441 // The top bits are known to be zero.
1442 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1445 case Instruction::SExt: {
1446 // Compute the bits in the result that are not present in the input.
1447 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1448 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1450 APInt InputDemandedBits = DemandedMask &
1451 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1453 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1454 // If any of the sign extended bits are demanded, we know that the sign
1456 if ((NewBits & DemandedMask) != 0)
1457 InputDemandedBits.set(SrcBitWidth-1);
1459 InputDemandedBits.trunc(SrcBitWidth);
1460 RHSKnownZero.trunc(SrcBitWidth);
1461 RHSKnownOne.trunc(SrcBitWidth);
1462 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1463 RHSKnownZero, RHSKnownOne, Depth+1))
1465 InputDemandedBits.zext(BitWidth);
1466 RHSKnownZero.zext(BitWidth);
1467 RHSKnownOne.zext(BitWidth);
1468 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1469 "Bits known to be one AND zero?");
1471 // If the sign bit of the input is known set or clear, then we know the
1472 // top bits of the result.
1474 // If the input sign bit is known zero, or if the NewBits are not demanded
1475 // convert this into a zero extension.
1476 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1478 // Convert to ZExt cast
1479 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1480 return UpdateValueUsesWith(I, NewCast);
1481 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1482 RHSKnownOne |= NewBits;
1486 case Instruction::Add: {
1487 // Figure out what the input bits are. If the top bits of the and result
1488 // are not demanded, then the add doesn't demand them from its input
1490 uint32_t NLZ = DemandedMask.countLeadingZeros();
1492 // If there is a constant on the RHS, there are a variety of xformations
1494 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1495 // If null, this should be simplified elsewhere. Some of the xforms here
1496 // won't work if the RHS is zero.
1500 // If the top bit of the output is demanded, demand everything from the
1501 // input. Otherwise, we demand all the input bits except NLZ top bits.
1502 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1504 // Find information about known zero/one bits in the input.
1505 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1506 LHSKnownZero, LHSKnownOne, Depth+1))
1509 // If the RHS of the add has bits set that can't affect the input, reduce
1511 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1512 return UpdateValueUsesWith(I, I);
1514 // Avoid excess work.
1515 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1518 // Turn it into OR if input bits are zero.
1519 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1521 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1523 InsertNewInstBefore(Or, *I);
1524 return UpdateValueUsesWith(I, Or);
1527 // We can say something about the output known-zero and known-one bits,
1528 // depending on potential carries from the input constant and the
1529 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1530 // bits set and the RHS constant is 0x01001, then we know we have a known
1531 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1533 // To compute this, we first compute the potential carry bits. These are
1534 // the bits which may be modified. I'm not aware of a better way to do
1536 const APInt& RHSVal = RHS->getValue();
1537 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1539 // Now that we know which bits have carries, compute the known-1/0 sets.
1541 // Bits are known one if they are known zero in one operand and one in the
1542 // other, and there is no input carry.
1543 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1544 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1546 // Bits are known zero if they are known zero in both operands and there
1547 // is no input carry.
1548 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1550 // If the high-bits of this ADD are not demanded, then it does not demand
1551 // the high bits of its LHS or RHS.
1552 if (DemandedMask[BitWidth-1] == 0) {
1553 // Right fill the mask of bits for this ADD to demand the most
1554 // significant bit and all those below it.
1555 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1556 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1557 LHSKnownZero, LHSKnownOne, Depth+1))
1559 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1560 LHSKnownZero, LHSKnownOne, Depth+1))
1566 case Instruction::Sub:
1567 // If the high-bits of this SUB are not demanded, then it does not demand
1568 // the high bits of its LHS or RHS.
1569 if (DemandedMask[BitWidth-1] == 0) {
1570 // Right fill the mask of bits for this SUB to demand the most
1571 // significant bit and all those below it.
1572 uint32_t NLZ = DemandedMask.countLeadingZeros();
1573 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1574 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1575 LHSKnownZero, LHSKnownOne, Depth+1))
1577 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1578 LHSKnownZero, LHSKnownOne, Depth+1))
1582 case Instruction::Shl:
1583 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1584 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1585 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1586 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1587 RHSKnownZero, RHSKnownOne, Depth+1))
1589 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1590 "Bits known to be one AND zero?");
1591 RHSKnownZero <<= ShiftAmt;
1592 RHSKnownOne <<= ShiftAmt;
1593 // low bits known zero.
1595 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1598 case Instruction::LShr:
1599 // For a logical shift right
1600 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1601 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1603 // Unsigned shift right.
1604 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1605 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1606 RHSKnownZero, RHSKnownOne, Depth+1))
1608 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1609 "Bits known to be one AND zero?");
1610 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1611 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1613 // Compute the new bits that are at the top now.
1614 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1615 RHSKnownZero |= HighBits; // high bits known zero.
1619 case Instruction::AShr:
1620 // If this is an arithmetic shift right and only the low-bit is set, we can
1621 // always convert this into a logical shr, even if the shift amount is
1622 // variable. The low bit of the shift cannot be an input sign bit unless
1623 // the shift amount is >= the size of the datatype, which is undefined.
1624 if (DemandedMask == 1) {
1625 // Perform the logical shift right.
1626 Value *NewVal = BinaryOperator::createLShr(
1627 I->getOperand(0), I->getOperand(1), I->getName());
1628 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1629 return UpdateValueUsesWith(I, NewVal);
1632 // If the sign bit is the only bit demanded by this ashr, then there is no
1633 // need to do it, the shift doesn't change the high bit.
1634 if (DemandedMask.isSignBit())
1635 return UpdateValueUsesWith(I, I->getOperand(0));
1637 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1638 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1640 // Signed shift right.
1641 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1642 // If any of the "high bits" are demanded, we should set the sign bit as
1644 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1645 DemandedMaskIn.set(BitWidth-1);
1646 if (SimplifyDemandedBits(I->getOperand(0),
1648 RHSKnownZero, RHSKnownOne, Depth+1))
1650 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1651 "Bits known to be one AND zero?");
1652 // Compute the new bits that are at the top now.
1653 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1654 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1655 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1657 // Handle the sign bits.
1658 APInt SignBit(APInt::getSignBit(BitWidth));
1659 // Adjust to where it is now in the mask.
1660 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1662 // If the input sign bit is known to be zero, or if none of the top bits
1663 // are demanded, turn this into an unsigned shift right.
1664 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1665 (HighBits & ~DemandedMask) == HighBits) {
1666 // Perform the logical shift right.
1667 Value *NewVal = BinaryOperator::createLShr(
1668 I->getOperand(0), SA, I->getName());
1669 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1670 return UpdateValueUsesWith(I, NewVal);
1671 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1672 RHSKnownOne |= HighBits;
1676 case Instruction::SRem:
1677 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1678 APInt RA = Rem->getValue();
1679 if (RA.isPowerOf2() || (-RA).isPowerOf2()) {
1680 APInt LowBits = RA.isStrictlyPositive() ? (RA - 1) | RA : ~RA;
1681 APInt Mask2 = LowBits | APInt::getSignBit(BitWidth);
1682 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1683 LHSKnownZero, LHSKnownOne, Depth+1))
1686 if (LHSKnownZero[BitWidth-1] || ((LHSKnownZero & LowBits) == LowBits))
1687 LHSKnownZero |= ~LowBits;
1688 else if (LHSKnownOne[BitWidth-1])
1689 LHSKnownOne |= ~LowBits;
1691 KnownZero |= LHSKnownZero & DemandedMask;
1692 KnownOne |= LHSKnownOne & DemandedMask;
1694 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1698 case Instruction::URem:
1699 if (ConstantInt *Rem = dyn_cast<ConstantInt>(I->getOperand(1))) {
1700 APInt RA = Rem->getValue();
1701 if (RA.isPowerOf2()) {
1702 APInt LowBits = (RA - 1) | RA;
1703 APInt Mask2 = LowBits & DemandedMask;
1704 KnownZero |= ~LowBits & DemandedMask;
1705 if (SimplifyDemandedBits(I->getOperand(0), Mask2,
1706 KnownZero, KnownOne, Depth+1))
1709 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
1712 APInt KnownZero2(BitWidth, 0), KnownOne2(BitWidth, 0);
1713 APInt AllOnes = APInt::getAllOnesValue(BitWidth);
1714 if (SimplifyDemandedBits(I->getOperand(1), AllOnes,
1715 KnownZero2, KnownOne2, Depth+1))
1718 uint32_t Leaders = KnownZero2.countLeadingOnes();
1719 KnownZero |= APInt::getHighBitsSet(BitWidth, Leaders) & DemandedMask;
1724 // If the client is only demanding bits that we know, return the known
1726 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1727 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1732 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1733 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1734 /// actually used by the caller. This method analyzes which elements of the
1735 /// operand are undef and returns that information in UndefElts.
1737 /// If the information about demanded elements can be used to simplify the
1738 /// operation, the operation is simplified, then the resultant value is
1739 /// returned. This returns null if no change was made.
1740 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1741 uint64_t &UndefElts,
1743 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1744 assert(VWidth <= 64 && "Vector too wide to analyze!");
1745 uint64_t EltMask = ~0ULL >> (64-VWidth);
1746 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1747 "Invalid DemandedElts!");
1749 if (isa<UndefValue>(V)) {
1750 // If the entire vector is undefined, just return this info.
1751 UndefElts = EltMask;
1753 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1754 UndefElts = EltMask;
1755 return UndefValue::get(V->getType());
1759 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1760 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1761 Constant *Undef = UndefValue::get(EltTy);
1763 std::vector<Constant*> Elts;
1764 for (unsigned i = 0; i != VWidth; ++i)
1765 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1766 Elts.push_back(Undef);
1767 UndefElts |= (1ULL << i);
1768 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1769 Elts.push_back(Undef);
1770 UndefElts |= (1ULL << i);
1771 } else { // Otherwise, defined.
1772 Elts.push_back(CP->getOperand(i));
1775 // If we changed the constant, return it.
1776 Constant *NewCP = ConstantVector::get(Elts);
1777 return NewCP != CP ? NewCP : 0;
1778 } else if (isa<ConstantAggregateZero>(V)) {
1779 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1781 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1782 Constant *Zero = Constant::getNullValue(EltTy);
1783 Constant *Undef = UndefValue::get(EltTy);
1784 std::vector<Constant*> Elts;
1785 for (unsigned i = 0; i != VWidth; ++i)
1786 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1787 UndefElts = DemandedElts ^ EltMask;
1788 return ConstantVector::get(Elts);
1791 if (!V->hasOneUse()) { // Other users may use these bits.
1792 if (Depth != 0) { // Not at the root.
1793 // TODO: Just compute the UndefElts information recursively.
1797 } else if (Depth == 10) { // Limit search depth.
1801 Instruction *I = dyn_cast<Instruction>(V);
1802 if (!I) return false; // Only analyze instructions.
1804 bool MadeChange = false;
1805 uint64_t UndefElts2;
1807 switch (I->getOpcode()) {
1810 case Instruction::InsertElement: {
1811 // If this is a variable index, we don't know which element it overwrites.
1812 // demand exactly the same input as we produce.
1813 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1815 // Note that we can't propagate undef elt info, because we don't know
1816 // which elt is getting updated.
1817 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1818 UndefElts2, Depth+1);
1819 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1823 // If this is inserting an element that isn't demanded, remove this
1825 unsigned IdxNo = Idx->getZExtValue();
1826 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1827 return AddSoonDeadInstToWorklist(*I, 0);
1829 // Otherwise, the element inserted overwrites whatever was there, so the
1830 // input demanded set is simpler than the output set.
1831 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1832 DemandedElts & ~(1ULL << IdxNo),
1833 UndefElts, Depth+1);
1834 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1836 // The inserted element is defined.
1837 UndefElts |= 1ULL << IdxNo;
1840 case Instruction::BitCast: {
1841 // Vector->vector casts only.
1842 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1844 unsigned InVWidth = VTy->getNumElements();
1845 uint64_t InputDemandedElts = 0;
1848 if (VWidth == InVWidth) {
1849 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1850 // elements as are demanded of us.
1852 InputDemandedElts = DemandedElts;
1853 } else if (VWidth > InVWidth) {
1857 // If there are more elements in the result than there are in the source,
1858 // then an input element is live if any of the corresponding output
1859 // elements are live.
1860 Ratio = VWidth/InVWidth;
1861 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1862 if (DemandedElts & (1ULL << OutIdx))
1863 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1869 // If there are more elements in the source than there are in the result,
1870 // then an input element is live if the corresponding output element is
1872 Ratio = InVWidth/VWidth;
1873 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1874 if (DemandedElts & (1ULL << InIdx/Ratio))
1875 InputDemandedElts |= 1ULL << InIdx;
1878 // div/rem demand all inputs, because they don't want divide by zero.
1879 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1880 UndefElts2, Depth+1);
1882 I->setOperand(0, TmpV);
1886 UndefElts = UndefElts2;
1887 if (VWidth > InVWidth) {
1888 assert(0 && "Unimp");
1889 // If there are more elements in the result than there are in the source,
1890 // then an output element is undef if the corresponding input element is
1892 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1893 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1894 UndefElts |= 1ULL << OutIdx;
1895 } else if (VWidth < InVWidth) {
1896 assert(0 && "Unimp");
1897 // If there are more elements in the source than there are in the result,
1898 // then a result element is undef if all of the corresponding input
1899 // elements are undef.
1900 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1901 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1902 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1903 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1907 case Instruction::And:
1908 case Instruction::Or:
1909 case Instruction::Xor:
1910 case Instruction::Add:
1911 case Instruction::Sub:
1912 case Instruction::Mul:
1913 // div/rem demand all inputs, because they don't want divide by zero.
1914 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1915 UndefElts, Depth+1);
1916 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1917 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1918 UndefElts2, Depth+1);
1919 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1921 // Output elements are undefined if both are undefined. Consider things
1922 // like undef&0. The result is known zero, not undef.
1923 UndefElts &= UndefElts2;
1926 case Instruction::Call: {
1927 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1929 switch (II->getIntrinsicID()) {
1932 // Binary vector operations that work column-wise. A dest element is a
1933 // function of the corresponding input elements from the two inputs.
1934 case Intrinsic::x86_sse_sub_ss:
1935 case Intrinsic::x86_sse_mul_ss:
1936 case Intrinsic::x86_sse_min_ss:
1937 case Intrinsic::x86_sse_max_ss:
1938 case Intrinsic::x86_sse2_sub_sd:
1939 case Intrinsic::x86_sse2_mul_sd:
1940 case Intrinsic::x86_sse2_min_sd:
1941 case Intrinsic::x86_sse2_max_sd:
1942 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1943 UndefElts, Depth+1);
1944 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1945 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1946 UndefElts2, Depth+1);
1947 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1949 // If only the low elt is demanded and this is a scalarizable intrinsic,
1950 // scalarize it now.
1951 if (DemandedElts == 1) {
1952 switch (II->getIntrinsicID()) {
1954 case Intrinsic::x86_sse_sub_ss:
1955 case Intrinsic::x86_sse_mul_ss:
1956 case Intrinsic::x86_sse2_sub_sd:
1957 case Intrinsic::x86_sse2_mul_sd:
1958 // TODO: Lower MIN/MAX/ABS/etc
1959 Value *LHS = II->getOperand(1);
1960 Value *RHS = II->getOperand(2);
1961 // Extract the element as scalars.
1962 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1963 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1965 switch (II->getIntrinsicID()) {
1966 default: assert(0 && "Case stmts out of sync!");
1967 case Intrinsic::x86_sse_sub_ss:
1968 case Intrinsic::x86_sse2_sub_sd:
1969 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1970 II->getName()), *II);
1972 case Intrinsic::x86_sse_mul_ss:
1973 case Intrinsic::x86_sse2_mul_sd:
1974 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1975 II->getName()), *II);
1980 InsertElementInst::Create(UndefValue::get(II->getType()), TmpV, 0U,
1982 InsertNewInstBefore(New, *II);
1983 AddSoonDeadInstToWorklist(*II, 0);
1988 // Output elements are undefined if both are undefined. Consider things
1989 // like undef&0. The result is known zero, not undef.
1990 UndefElts &= UndefElts2;
1996 return MadeChange ? I : 0;
1999 /// @returns true if the specified compare predicate is
2000 /// true when both operands are equal...
2001 /// @brief Determine if the icmp Predicate is true when both operands are equal
2002 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
2003 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
2004 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
2005 pred == ICmpInst::ICMP_SLE;
2008 /// @returns true if the specified compare instruction is
2009 /// true when both operands are equal...
2010 /// @brief Determine if the ICmpInst returns true when both operands are equal
2011 static bool isTrueWhenEqual(ICmpInst &ICI) {
2012 return isTrueWhenEqual(ICI.getPredicate());
2015 /// AssociativeOpt - Perform an optimization on an associative operator. This
2016 /// function is designed to check a chain of associative operators for a
2017 /// potential to apply a certain optimization. Since the optimization may be
2018 /// applicable if the expression was reassociated, this checks the chain, then
2019 /// reassociates the expression as necessary to expose the optimization
2020 /// opportunity. This makes use of a special Functor, which must define
2021 /// 'shouldApply' and 'apply' methods.
2023 template<typename Functor>
2024 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
2025 unsigned Opcode = Root.getOpcode();
2026 Value *LHS = Root.getOperand(0);
2028 // Quick check, see if the immediate LHS matches...
2029 if (F.shouldApply(LHS))
2030 return F.apply(Root);
2032 // Otherwise, if the LHS is not of the same opcode as the root, return.
2033 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2034 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
2035 // Should we apply this transform to the RHS?
2036 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
2038 // If not to the RHS, check to see if we should apply to the LHS...
2039 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
2040 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
2044 // If the functor wants to apply the optimization to the RHS of LHSI,
2045 // reassociate the expression from ((? op A) op B) to (? op (A op B))
2047 BasicBlock *BB = Root.getParent();
2049 // Now all of the instructions are in the current basic block, go ahead
2050 // and perform the reassociation.
2051 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
2053 // First move the selected RHS to the LHS of the root...
2054 Root.setOperand(0, LHSI->getOperand(1));
2056 // Make what used to be the LHS of the root be the user of the root...
2057 Value *ExtraOperand = TmpLHSI->getOperand(1);
2058 if (&Root == TmpLHSI) {
2059 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
2062 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
2063 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
2064 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
2065 BasicBlock::iterator ARI = &Root; ++ARI;
2066 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
2069 // Now propagate the ExtraOperand down the chain of instructions until we
2071 while (TmpLHSI != LHSI) {
2072 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
2073 // Move the instruction to immediately before the chain we are
2074 // constructing to avoid breaking dominance properties.
2075 NextLHSI->getParent()->getInstList().remove(NextLHSI);
2076 BB->getInstList().insert(ARI, NextLHSI);
2079 Value *NextOp = NextLHSI->getOperand(1);
2080 NextLHSI->setOperand(1, ExtraOperand);
2082 ExtraOperand = NextOp;
2085 // Now that the instructions are reassociated, have the functor perform
2086 // the transformation...
2087 return F.apply(Root);
2090 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
2096 // AddRHS - Implements: X + X --> X << 1
2099 AddRHS(Value *rhs) : RHS(rhs) {}
2100 bool shouldApply(Value *LHS) const { return LHS == RHS; }
2101 Instruction *apply(BinaryOperator &Add) const {
2102 return BinaryOperator::createShl(Add.getOperand(0),
2103 ConstantInt::get(Add.getType(), 1));
2107 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
2109 struct AddMaskingAnd {
2111 AddMaskingAnd(Constant *c) : C2(c) {}
2112 bool shouldApply(Value *LHS) const {
2114 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
2115 ConstantExpr::getAnd(C1, C2)->isNullValue();
2117 Instruction *apply(BinaryOperator &Add) const {
2118 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
2122 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
2124 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
2125 if (Constant *SOC = dyn_cast<Constant>(SO))
2126 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
2128 return IC->InsertNewInstBefore(CastInst::create(
2129 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
2132 // Figure out if the constant is the left or the right argument.
2133 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
2134 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
2136 if (Constant *SOC = dyn_cast<Constant>(SO)) {
2138 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
2139 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
2142 Value *Op0 = SO, *Op1 = ConstOperand;
2144 std::swap(Op0, Op1);
2146 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2147 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
2148 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2149 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
2150 SO->getName()+".cmp");
2152 assert(0 && "Unknown binary instruction type!");
2155 return IC->InsertNewInstBefore(New, I);
2158 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
2159 // constant as the other operand, try to fold the binary operator into the
2160 // select arguments. This also works for Cast instructions, which obviously do
2161 // not have a second operand.
2162 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
2164 // Don't modify shared select instructions
2165 if (!SI->hasOneUse()) return 0;
2166 Value *TV = SI->getOperand(1);
2167 Value *FV = SI->getOperand(2);
2169 if (isa<Constant>(TV) || isa<Constant>(FV)) {
2170 // Bool selects with constant operands can be folded to logical ops.
2171 if (SI->getType() == Type::Int1Ty) return 0;
2173 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
2174 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
2176 return SelectInst::Create(SI->getCondition(), SelectTrueVal,
2183 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
2184 /// node as operand #0, see if we can fold the instruction into the PHI (which
2185 /// is only possible if all operands to the PHI are constants).
2186 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
2187 PHINode *PN = cast<PHINode>(I.getOperand(0));
2188 unsigned NumPHIValues = PN->getNumIncomingValues();
2189 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
2191 // Check to see if all of the operands of the PHI are constants. If there is
2192 // one non-constant value, remember the BB it is. If there is more than one
2193 // or if *it* is a PHI, bail out.
2194 BasicBlock *NonConstBB = 0;
2195 for (unsigned i = 0; i != NumPHIValues; ++i)
2196 if (!isa<Constant>(PN->getIncomingValue(i))) {
2197 if (NonConstBB) return 0; // More than one non-const value.
2198 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
2199 NonConstBB = PN->getIncomingBlock(i);
2201 // If the incoming non-constant value is in I's block, we have an infinite
2203 if (NonConstBB == I.getParent())
2207 // If there is exactly one non-constant value, we can insert a copy of the
2208 // operation in that block. However, if this is a critical edge, we would be
2209 // inserting the computation one some other paths (e.g. inside a loop). Only
2210 // do this if the pred block is unconditionally branching into the phi block.
2212 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
2213 if (!BI || !BI->isUnconditional()) return 0;
2216 // Okay, we can do the transformation: create the new PHI node.
2217 PHINode *NewPN = PHINode::Create(I.getType(), "");
2218 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
2219 InsertNewInstBefore(NewPN, *PN);
2220 NewPN->takeName(PN);
2222 // Next, add all of the operands to the PHI.
2223 if (I.getNumOperands() == 2) {
2224 Constant *C = cast<Constant>(I.getOperand(1));
2225 for (unsigned i = 0; i != NumPHIValues; ++i) {
2227 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2228 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2229 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
2231 InV = ConstantExpr::get(I.getOpcode(), InC, C);
2233 assert(PN->getIncomingBlock(i) == NonConstBB);
2234 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
2235 InV = BinaryOperator::create(BO->getOpcode(),
2236 PN->getIncomingValue(i), C, "phitmp",
2237 NonConstBB->getTerminator());
2238 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
2239 InV = CmpInst::create(CI->getOpcode(),
2241 PN->getIncomingValue(i), C, "phitmp",
2242 NonConstBB->getTerminator());
2244 assert(0 && "Unknown binop!");
2246 AddToWorkList(cast<Instruction>(InV));
2248 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2251 CastInst *CI = cast<CastInst>(&I);
2252 const Type *RetTy = CI->getType();
2253 for (unsigned i = 0; i != NumPHIValues; ++i) {
2255 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
2256 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
2258 assert(PN->getIncomingBlock(i) == NonConstBB);
2259 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
2260 I.getType(), "phitmp",
2261 NonConstBB->getTerminator());
2262 AddToWorkList(cast<Instruction>(InV));
2264 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
2267 return ReplaceInstUsesWith(I, NewPN);
2271 /// CannotBeNegativeZero - Return true if we can prove that the specified FP
2272 /// value is never equal to -0.0.
2274 /// Note that this function will need to be revisited when we support nondefault
2277 static bool CannotBeNegativeZero(const Value *V) {
2278 if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V))
2279 return !CFP->getValueAPF().isNegZero();
2281 // (add x, 0.0) is guaranteed to return +0.0, not -0.0.
2282 if (const Instruction *I = dyn_cast<Instruction>(V)) {
2283 if (I->getOpcode() == Instruction::Add &&
2284 isa<ConstantFP>(I->getOperand(1)) &&
2285 cast<ConstantFP>(I->getOperand(1))->isNullValue())
2288 if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(I))
2289 if (II->getIntrinsicID() == Intrinsic::sqrt)
2290 return CannotBeNegativeZero(II->getOperand(1));
2292 if (const CallInst *CI = dyn_cast<CallInst>(I))
2293 if (const Function *F = CI->getCalledFunction()) {
2294 if (F->isDeclaration()) {
2295 switch (F->getNameLen()) {
2296 case 3: // abs(x) != -0.0
2297 if (!strcmp(F->getNameStart(), "abs")) return true;
2299 case 4: // abs[lf](x) != -0.0
2300 if (!strcmp(F->getNameStart(), "absf")) return true;
2301 if (!strcmp(F->getNameStart(), "absl")) return true;
2312 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
2313 bool Changed = SimplifyCommutative(I);
2314 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
2316 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
2317 // X + undef -> undef
2318 if (isa<UndefValue>(RHS))
2319 return ReplaceInstUsesWith(I, RHS);
2322 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
2323 if (RHSC->isNullValue())
2324 return ReplaceInstUsesWith(I, LHS);
2325 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
2326 if (CFP->isExactlyValue(ConstantFP::getNegativeZero
2327 (I.getType())->getValueAPF()))
2328 return ReplaceInstUsesWith(I, LHS);
2331 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
2332 // X + (signbit) --> X ^ signbit
2333 const APInt& Val = CI->getValue();
2334 uint32_t BitWidth = Val.getBitWidth();
2335 if (Val == APInt::getSignBit(BitWidth))
2336 return BinaryOperator::createXor(LHS, RHS);
2338 // See if SimplifyDemandedBits can simplify this. This handles stuff like
2339 // (X & 254)+1 -> (X&254)|1
2340 if (!isa<VectorType>(I.getType())) {
2341 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
2342 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
2343 KnownZero, KnownOne))
2348 if (isa<PHINode>(LHS))
2349 if (Instruction *NV = FoldOpIntoPhi(I))
2352 ConstantInt *XorRHS = 0;
2354 if (isa<ConstantInt>(RHSC) &&
2355 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
2356 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
2357 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
2359 uint32_t Size = TySizeBits / 2;
2360 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
2361 APInt CFF80Val(-C0080Val);
2363 if (TySizeBits > Size) {
2364 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
2365 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
2366 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
2367 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
2368 // This is a sign extend if the top bits are known zero.
2369 if (!MaskedValueIsZero(XorLHS,
2370 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
2371 Size = 0; // Not a sign ext, but can't be any others either.
2376 C0080Val = APIntOps::lshr(C0080Val, Size);
2377 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2378 } while (Size >= 1);
2380 // FIXME: This shouldn't be necessary. When the backends can handle types
2381 // with funny bit widths then this whole cascade of if statements should
2382 // be removed. It is just here to get the size of the "middle" type back
2383 // up to something that the back ends can handle.
2384 const Type *MiddleType = 0;
2387 case 32: MiddleType = Type::Int32Ty; break;
2388 case 16: MiddleType = Type::Int16Ty; break;
2389 case 8: MiddleType = Type::Int8Ty; break;
2392 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2393 InsertNewInstBefore(NewTrunc, I);
2394 return new SExtInst(NewTrunc, I.getType(), I.getName());
2400 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2401 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2403 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2404 if (RHSI->getOpcode() == Instruction::Sub)
2405 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2406 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2408 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2409 if (LHSI->getOpcode() == Instruction::Sub)
2410 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2411 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2416 // -A + -B --> -(A + B)
2417 if (Value *LHSV = dyn_castNegVal(LHS)) {
2418 if (LHS->getType()->isIntOrIntVector()) {
2419 if (Value *RHSV = dyn_castNegVal(RHS)) {
2420 Instruction *NewAdd = BinaryOperator::createAdd(LHSV, RHSV, "sum");
2421 InsertNewInstBefore(NewAdd, I);
2422 return BinaryOperator::createNeg(NewAdd);
2426 return BinaryOperator::createSub(RHS, LHSV);
2430 if (!isa<Constant>(RHS))
2431 if (Value *V = dyn_castNegVal(RHS))
2432 return BinaryOperator::createSub(LHS, V);
2436 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2437 if (X == RHS) // X*C + X --> X * (C+1)
2438 return BinaryOperator::createMul(RHS, AddOne(C2));
2440 // X*C1 + X*C2 --> X * (C1+C2)
2442 if (X == dyn_castFoldableMul(RHS, C1))
2443 return BinaryOperator::createMul(X, Add(C1, C2));
2446 // X + X*C --> X * (C+1)
2447 if (dyn_castFoldableMul(RHS, C2) == LHS)
2448 return BinaryOperator::createMul(LHS, AddOne(C2));
2450 // X + ~X --> -1 since ~X = -X-1
2451 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2452 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2455 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2456 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2457 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2460 // W*X + Y*Z --> W * (X+Z) iff W == Y
2461 if (I.getType()->isIntOrIntVector()) {
2462 Value *W, *X, *Y, *Z;
2463 if (match(LHS, m_Mul(m_Value(W), m_Value(X))) &&
2464 match(RHS, m_Mul(m_Value(Y), m_Value(Z)))) {
2468 } else if (Y == X) {
2470 } else if (X == Z) {
2477 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, Z,
2478 LHS->getName()), I);
2479 return BinaryOperator::createMul(W, NewAdd);
2484 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2486 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2487 return BinaryOperator::createSub(SubOne(CRHS), X);
2489 // (X & FF00) + xx00 -> (X+xx00) & FF00
2490 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2491 Constant *Anded = And(CRHS, C2);
2492 if (Anded == CRHS) {
2493 // See if all bits from the first bit set in the Add RHS up are included
2494 // in the mask. First, get the rightmost bit.
2495 const APInt& AddRHSV = CRHS->getValue();
2497 // Form a mask of all bits from the lowest bit added through the top.
2498 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2500 // See if the and mask includes all of these bits.
2501 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2503 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2504 // Okay, the xform is safe. Insert the new add pronto.
2505 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2506 LHS->getName()), I);
2507 return BinaryOperator::createAnd(NewAdd, C2);
2512 // Try to fold constant add into select arguments.
2513 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2514 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2518 // add (cast *A to intptrtype) B ->
2519 // cast (GEP (cast *A to sbyte*) B) --> intptrtype
2521 CastInst *CI = dyn_cast<CastInst>(LHS);
2524 CI = dyn_cast<CastInst>(RHS);
2527 if (CI && CI->getType()->isSized() &&
2528 (CI->getType()->getPrimitiveSizeInBits() ==
2529 TD->getIntPtrType()->getPrimitiveSizeInBits())
2530 && isa<PointerType>(CI->getOperand(0)->getType())) {
2532 cast<PointerType>(CI->getOperand(0)->getType())->getAddressSpace();
2533 Value *I2 = InsertBitCastBefore(CI->getOperand(0),
2534 PointerType::get(Type::Int8Ty, AS), I);
2535 I2 = InsertNewInstBefore(GetElementPtrInst::Create(I2, Other, "ctg2"), I);
2536 return new PtrToIntInst(I2, CI->getType());
2540 // add (select X 0 (sub n A)) A --> select X A n
2542 SelectInst *SI = dyn_cast<SelectInst>(LHS);
2545 SI = dyn_cast<SelectInst>(RHS);
2548 if (SI && SI->hasOneUse()) {
2549 Value *TV = SI->getTrueValue();
2550 Value *FV = SI->getFalseValue();
2553 // Can we fold the add into the argument of the select?
2554 // We check both true and false select arguments for a matching subtract.
2555 if (match(FV, m_Zero()) && match(TV, m_Sub(m_Value(N), m_Value(A))) &&
2556 A == Other) // Fold the add into the true select value.
2557 return SelectInst::Create(SI->getCondition(), N, A);
2558 if (match(TV, m_Zero()) && match(FV, m_Sub(m_Value(N), m_Value(A))) &&
2559 A == Other) // Fold the add into the false select value.
2560 return SelectInst::Create(SI->getCondition(), A, N);
2564 // Check for X+0.0. Simplify it to X if we know X is not -0.0.
2565 if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHS))
2566 if (CFP->getValueAPF().isPosZero() && CannotBeNegativeZero(LHS))
2567 return ReplaceInstUsesWith(I, LHS);
2569 return Changed ? &I : 0;
2572 // isSignBit - Return true if the value represented by the constant only has the
2573 // highest order bit set.
2574 static bool isSignBit(ConstantInt *CI) {
2575 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2576 return CI->getValue() == APInt::getSignBit(NumBits);
2579 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2580 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2582 if (Op0 == Op1) // sub X, X -> 0
2583 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2585 // If this is a 'B = x-(-A)', change to B = x+A...
2586 if (Value *V = dyn_castNegVal(Op1))
2587 return BinaryOperator::createAdd(Op0, V);
2589 if (isa<UndefValue>(Op0))
2590 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2591 if (isa<UndefValue>(Op1))
2592 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2594 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2595 // Replace (-1 - A) with (~A)...
2596 if (C->isAllOnesValue())
2597 return BinaryOperator::createNot(Op1);
2599 // C - ~X == X + (1+C)
2601 if (match(Op1, m_Not(m_Value(X))))
2602 return BinaryOperator::createAdd(X, AddOne(C));
2604 // -(X >>u 31) -> (X >>s 31)
2605 // -(X >>s 31) -> (X >>u 31)
2607 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1)) {
2608 if (SI->getOpcode() == Instruction::LShr) {
2609 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2610 // Check to see if we are shifting out everything but the sign bit.
2611 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2612 SI->getType()->getPrimitiveSizeInBits()-1) {
2613 // Ok, the transformation is safe. Insert AShr.
2614 return BinaryOperator::create(Instruction::AShr,
2615 SI->getOperand(0), CU, SI->getName());
2619 else if (SI->getOpcode() == Instruction::AShr) {
2620 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2621 // Check to see if we are shifting out everything but the sign bit.
2622 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2623 SI->getType()->getPrimitiveSizeInBits()-1) {
2624 // Ok, the transformation is safe. Insert LShr.
2625 return BinaryOperator::createLShr(
2626 SI->getOperand(0), CU, SI->getName());
2633 // Try to fold constant sub into select arguments.
2634 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2635 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2638 if (isa<PHINode>(Op0))
2639 if (Instruction *NV = FoldOpIntoPhi(I))
2643 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2644 if (Op1I->getOpcode() == Instruction::Add &&
2645 !Op0->getType()->isFPOrFPVector()) {
2646 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2647 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2648 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2649 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2650 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2651 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2652 // C1-(X+C2) --> (C1-C2)-X
2653 return BinaryOperator::createSub(Subtract(CI1, CI2),
2654 Op1I->getOperand(0));
2658 if (Op1I->hasOneUse()) {
2659 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2660 // is not used by anyone else...
2662 if (Op1I->getOpcode() == Instruction::Sub &&
2663 !Op1I->getType()->isFPOrFPVector()) {
2664 // Swap the two operands of the subexpr...
2665 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2666 Op1I->setOperand(0, IIOp1);
2667 Op1I->setOperand(1, IIOp0);
2669 // Create the new top level add instruction...
2670 return BinaryOperator::createAdd(Op0, Op1);
2673 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2675 if (Op1I->getOpcode() == Instruction::And &&
2676 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2677 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2680 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2681 return BinaryOperator::createAnd(Op0, NewNot);
2684 // 0 - (X sdiv C) -> (X sdiv -C)
2685 if (Op1I->getOpcode() == Instruction::SDiv)
2686 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2688 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2689 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2690 ConstantExpr::getNeg(DivRHS));
2692 // X - X*C --> X * (1-C)
2693 ConstantInt *C2 = 0;
2694 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2695 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2696 return BinaryOperator::createMul(Op0, CP1);
2699 // X - ((X / Y) * Y) --> X % Y
2700 if (Op1I->getOpcode() == Instruction::Mul)
2701 if (Instruction *I = dyn_cast<Instruction>(Op1I->getOperand(0)))
2702 if (Op0 == I->getOperand(0) &&
2703 Op1I->getOperand(1) == I->getOperand(1)) {
2704 if (I->getOpcode() == Instruction::SDiv)
2705 return BinaryOperator::createSRem(Op0, Op1I->getOperand(1));
2706 if (I->getOpcode() == Instruction::UDiv)
2707 return BinaryOperator::createURem(Op0, Op1I->getOperand(1));
2712 if (!Op0->getType()->isFPOrFPVector())
2713 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
2714 if (Op0I->getOpcode() == Instruction::Add) {
2715 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2716 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2717 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2718 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2719 } else if (Op0I->getOpcode() == Instruction::Sub) {
2720 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2721 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2726 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2727 if (X == Op1) // X*C - X --> X * (C-1)
2728 return BinaryOperator::createMul(Op1, SubOne(C1));
2730 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2731 if (X == dyn_castFoldableMul(Op1, C2))
2732 return BinaryOperator::createMul(X, Subtract(C1, C2));
2737 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2738 /// comparison only checks the sign bit. If it only checks the sign bit, set
2739 /// TrueIfSigned if the result of the comparison is true when the input value is
2741 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2742 bool &TrueIfSigned) {
2744 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2745 TrueIfSigned = true;
2746 return RHS->isZero();
2747 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2748 TrueIfSigned = true;
2749 return RHS->isAllOnesValue();
2750 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2751 TrueIfSigned = false;
2752 return RHS->isAllOnesValue();
2753 case ICmpInst::ICMP_UGT:
2754 // True if LHS u> RHS and RHS == high-bit-mask - 1
2755 TrueIfSigned = true;
2756 return RHS->getValue() ==
2757 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2758 case ICmpInst::ICMP_UGE:
2759 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2760 TrueIfSigned = true;
2761 return RHS->getValue() ==
2762 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2768 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2769 bool Changed = SimplifyCommutative(I);
2770 Value *Op0 = I.getOperand(0);
2772 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2773 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2775 // Simplify mul instructions with a constant RHS...
2776 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2777 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2779 // ((X << C1)*C2) == (X * (C2 << C1))
2780 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2781 if (SI->getOpcode() == Instruction::Shl)
2782 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2783 return BinaryOperator::createMul(SI->getOperand(0),
2784 ConstantExpr::getShl(CI, ShOp));
2787 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2788 if (CI->equalsInt(1)) // X * 1 == X
2789 return ReplaceInstUsesWith(I, Op0);
2790 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2791 return BinaryOperator::createNeg(Op0, I.getName());
2793 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2794 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2795 return BinaryOperator::createShl(Op0,
2796 ConstantInt::get(Op0->getType(), Val.logBase2()));
2798 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2799 if (Op1F->isNullValue())
2800 return ReplaceInstUsesWith(I, Op1);
2802 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2803 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2804 // We need a better interface for long double here.
2805 if (Op1->getType() == Type::FloatTy || Op1->getType() == Type::DoubleTy)
2806 if (Op1F->isExactlyValue(1.0))
2807 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2810 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2811 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2812 isa<ConstantInt>(Op0I->getOperand(1))) {
2813 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2814 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2816 InsertNewInstBefore(Add, I);
2817 Value *C1C2 = ConstantExpr::getMul(Op1,
2818 cast<Constant>(Op0I->getOperand(1)));
2819 return BinaryOperator::createAdd(Add, C1C2);
2823 // Try to fold constant mul into select arguments.
2824 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2825 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2828 if (isa<PHINode>(Op0))
2829 if (Instruction *NV = FoldOpIntoPhi(I))
2833 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2834 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2835 return BinaryOperator::createMul(Op0v, Op1v);
2837 // If one of the operands of the multiply is a cast from a boolean value, then
2838 // we know the bool is either zero or one, so this is a 'masking' multiply.
2839 // See if we can simplify things based on how the boolean was originally
2841 CastInst *BoolCast = 0;
2842 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2843 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2846 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2847 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2850 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2851 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2852 const Type *SCOpTy = SCIOp0->getType();
2855 // If the icmp is true iff the sign bit of X is set, then convert this
2856 // multiply into a shift/and combination.
2857 if (isa<ConstantInt>(SCIOp1) &&
2858 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2860 // Shift the X value right to turn it into "all signbits".
2861 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2862 SCOpTy->getPrimitiveSizeInBits()-1);
2864 InsertNewInstBefore(
2865 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2866 BoolCast->getOperand(0)->getName()+
2869 // If the multiply type is not the same as the source type, sign extend
2870 // or truncate to the multiply type.
2871 if (I.getType() != V->getType()) {
2872 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2873 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2874 Instruction::CastOps opcode =
2875 (SrcBits == DstBits ? Instruction::BitCast :
2876 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2877 V = InsertCastBefore(opcode, V, I.getType(), I);
2880 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2881 return BinaryOperator::createAnd(V, OtherOp);
2886 return Changed ? &I : 0;
2889 /// This function implements the transforms on div instructions that work
2890 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2891 /// used by the visitors to those instructions.
2892 /// @brief Transforms common to all three div instructions
2893 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2894 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2896 // undef / X -> 0 for integer.
2897 // undef / X -> undef for FP (the undef could be a snan).
2898 if (isa<UndefValue>(Op0)) {
2899 if (Op0->getType()->isFPOrFPVector())
2900 return ReplaceInstUsesWith(I, Op0);
2901 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2904 // X / undef -> undef
2905 if (isa<UndefValue>(Op1))
2906 return ReplaceInstUsesWith(I, Op1);
2908 // Handle cases involving: [su]div X, (select Cond, Y, Z)
2909 // This does not apply for fdiv.
2910 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2911 // [su]div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in
2912 // the same basic block, then we replace the select with Y, and the
2913 // condition of the select with false (if the cond value is in the same BB).
2914 // If the select has uses other than the div, this allows them to be
2915 // simplified also. Note that div X, Y is just as good as div X, 0 (undef)
2916 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(1)))
2917 if (ST->isNullValue()) {
2918 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2919 if (CondI && CondI->getParent() == I.getParent())
2920 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2921 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2922 I.setOperand(1, SI->getOperand(2));
2924 UpdateValueUsesWith(SI, SI->getOperand(2));
2928 // Likewise for: [su]div X, (Cond ? Y : 0) -> div X, Y
2929 if (ConstantInt *ST = dyn_cast<ConstantInt>(SI->getOperand(2)))
2930 if (ST->isNullValue()) {
2931 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2932 if (CondI && CondI->getParent() == I.getParent())
2933 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2934 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2935 I.setOperand(1, SI->getOperand(1));
2937 UpdateValueUsesWith(SI, SI->getOperand(1));
2945 /// This function implements the transforms common to both integer division
2946 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2947 /// division instructions.
2948 /// @brief Common integer divide transforms
2949 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2950 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2952 if (Instruction *Common = commonDivTransforms(I))
2955 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2957 if (RHS->equalsInt(1))
2958 return ReplaceInstUsesWith(I, Op0);
2960 // (X / C1) / C2 -> X / (C1*C2)
2961 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2962 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2963 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2964 if (MultiplyOverflows(RHS, LHSRHS, I.getOpcode()==Instruction::SDiv))
2965 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2967 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2968 Multiply(RHS, LHSRHS));
2971 if (!RHS->isZero()) { // avoid X udiv 0
2972 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2973 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2975 if (isa<PHINode>(Op0))
2976 if (Instruction *NV = FoldOpIntoPhi(I))
2981 // 0 / X == 0, we don't need to preserve faults!
2982 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2983 if (LHS->equalsInt(0))
2984 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2989 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2990 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2992 // Handle the integer div common cases
2993 if (Instruction *Common = commonIDivTransforms(I))
2996 // X udiv C^2 -> X >> C
2997 // Check to see if this is an unsigned division with an exact power of 2,
2998 // if so, convert to a right shift.
2999 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
3000 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
3001 return BinaryOperator::createLShr(Op0,
3002 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
3005 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
3006 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
3007 if (RHSI->getOpcode() == Instruction::Shl &&
3008 isa<ConstantInt>(RHSI->getOperand(0))) {
3009 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
3010 if (C1.isPowerOf2()) {
3011 Value *N = RHSI->getOperand(1);
3012 const Type *NTy = N->getType();
3013 if (uint32_t C2 = C1.logBase2()) {
3014 Constant *C2V = ConstantInt::get(NTy, C2);
3015 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
3017 return BinaryOperator::createLShr(Op0, N);
3022 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
3023 // where C1&C2 are powers of two.
3024 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
3025 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3026 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3027 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
3028 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
3029 // Compute the shift amounts
3030 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
3031 // Construct the "on true" case of the select
3032 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
3033 Instruction *TSI = BinaryOperator::createLShr(
3034 Op0, TC, SI->getName()+".t");
3035 TSI = InsertNewInstBefore(TSI, I);
3037 // Construct the "on false" case of the select
3038 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
3039 Instruction *FSI = BinaryOperator::createLShr(
3040 Op0, FC, SI->getName()+".f");
3041 FSI = InsertNewInstBefore(FSI, I);
3043 // construct the select instruction and return it.
3044 return SelectInst::Create(SI->getOperand(0), TSI, FSI, SI->getName());
3050 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
3051 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3053 // Handle the integer div common cases
3054 if (Instruction *Common = commonIDivTransforms(I))
3057 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3059 if (RHS->isAllOnesValue())
3060 return BinaryOperator::createNeg(Op0);
3063 if (Value *LHSNeg = dyn_castNegVal(Op0))
3064 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
3067 // If the sign bits of both operands are zero (i.e. we can prove they are
3068 // unsigned inputs), turn this into a udiv.
3069 if (I.getType()->isInteger()) {
3070 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3071 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3072 // X sdiv Y -> X udiv Y, iff X and Y don't have sign bit set
3073 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
3080 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
3081 return commonDivTransforms(I);
3084 /// This function implements the transforms on rem instructions that work
3085 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
3086 /// is used by the visitors to those instructions.
3087 /// @brief Transforms common to all three rem instructions
3088 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
3089 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3091 // 0 % X == 0 for integer, we don't need to preserve faults!
3092 if (Constant *LHS = dyn_cast<Constant>(Op0))
3093 if (LHS->isNullValue())
3094 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3096 if (isa<UndefValue>(Op0)) { // undef % X -> 0
3097 if (I.getType()->isFPOrFPVector())
3098 return ReplaceInstUsesWith(I, Op0); // X % undef -> undef (could be SNaN)
3099 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3101 if (isa<UndefValue>(Op1))
3102 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
3104 // Handle cases involving: rem X, (select Cond, Y, Z)
3105 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3106 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
3107 // the same basic block, then we replace the select with Y, and the
3108 // condition of the select with false (if the cond value is in the same
3109 // BB). If the select has uses other than the div, this allows them to be
3111 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
3112 if (ST->isNullValue()) {
3113 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3114 if (CondI && CondI->getParent() == I.getParent())
3115 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
3116 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3117 I.setOperand(1, SI->getOperand(2));
3119 UpdateValueUsesWith(SI, SI->getOperand(2));
3122 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
3123 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
3124 if (ST->isNullValue()) {
3125 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
3126 if (CondI && CondI->getParent() == I.getParent())
3127 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
3128 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
3129 I.setOperand(1, SI->getOperand(1));
3131 UpdateValueUsesWith(SI, SI->getOperand(1));
3139 /// This function implements the transforms common to both integer remainder
3140 /// instructions (urem and srem). It is called by the visitors to those integer
3141 /// remainder instructions.
3142 /// @brief Common integer remainder transforms
3143 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
3144 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3146 if (Instruction *common = commonRemTransforms(I))
3149 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3150 // X % 0 == undef, we don't need to preserve faults!
3151 if (RHS->equalsInt(0))
3152 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3154 if (RHS->equalsInt(1)) // X % 1 == 0
3155 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3157 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
3158 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
3159 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3161 } else if (isa<PHINode>(Op0I)) {
3162 if (Instruction *NV = FoldOpIntoPhi(I))
3166 // See if we can fold away this rem instruction.
3167 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3168 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3169 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3170 KnownZero, KnownOne))
3178 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
3179 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3181 if (Instruction *common = commonIRemTransforms(I))
3184 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3185 // X urem C^2 -> X and C
3186 // Check to see if this is an unsigned remainder with an exact power of 2,
3187 // if so, convert to a bitwise and.
3188 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
3189 if (C->getValue().isPowerOf2())
3190 return BinaryOperator::createAnd(Op0, SubOne(C));
3193 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
3194 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
3195 if (RHSI->getOpcode() == Instruction::Shl &&
3196 isa<ConstantInt>(RHSI->getOperand(0))) {
3197 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
3198 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
3199 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
3201 return BinaryOperator::createAnd(Op0, Add);
3206 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
3207 // where C1&C2 are powers of two.
3208 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
3209 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
3210 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
3211 // STO == 0 and SFO == 0 handled above.
3212 if ((STO->getValue().isPowerOf2()) &&
3213 (SFO->getValue().isPowerOf2())) {
3214 Value *TrueAnd = InsertNewInstBefore(
3215 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
3216 Value *FalseAnd = InsertNewInstBefore(
3217 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
3218 return SelectInst::Create(SI->getOperand(0), TrueAnd, FalseAnd);
3226 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
3227 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3229 // Handle the integer rem common cases
3230 if (Instruction *common = commonIRemTransforms(I))
3233 if (Value *RHSNeg = dyn_castNegVal(Op1))
3234 if (!isa<ConstantInt>(RHSNeg) ||
3235 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
3237 AddUsesToWorkList(I);
3238 I.setOperand(1, RHSNeg);
3242 // If the sign bits of both operands are zero (i.e. we can prove they are
3243 // unsigned inputs), turn this into a urem.
3244 if (I.getType()->isInteger()) {
3245 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
3246 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
3247 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
3248 return BinaryOperator::createURem(Op0, Op1, I.getName());
3255 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
3256 return commonRemTransforms(I);
3259 // isMaxValueMinusOne - return true if this is Max-1
3260 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
3261 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3263 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
3264 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
3267 // isMinValuePlusOne - return true if this is Min+1
3268 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
3270 return C->getValue() == 1; // unsigned
3272 // Calculate 1111111111000000000000
3273 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
3274 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
3277 // isOneBitSet - Return true if there is exactly one bit set in the specified
3279 static bool isOneBitSet(const ConstantInt *CI) {
3280 return CI->getValue().isPowerOf2();
3283 // isHighOnes - Return true if the constant is of the form 1+0+.
3284 // This is the same as lowones(~X).
3285 static bool isHighOnes(const ConstantInt *CI) {
3286 return (~CI->getValue() + 1).isPowerOf2();
3289 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
3290 /// are carefully arranged to allow folding of expressions such as:
3292 /// (A < B) | (A > B) --> (A != B)
3294 /// Note that this is only valid if the first and second predicates have the
3295 /// same sign. Is illegal to do: (A u< B) | (A s> B)
3297 /// Three bits are used to represent the condition, as follows:
3302 /// <=> Value Definition
3303 /// 000 0 Always false
3310 /// 111 7 Always true
3312 static unsigned getICmpCode(const ICmpInst *ICI) {
3313 switch (ICI->getPredicate()) {
3315 case ICmpInst::ICMP_UGT: return 1; // 001
3316 case ICmpInst::ICMP_SGT: return 1; // 001
3317 case ICmpInst::ICMP_EQ: return 2; // 010
3318 case ICmpInst::ICMP_UGE: return 3; // 011
3319 case ICmpInst::ICMP_SGE: return 3; // 011
3320 case ICmpInst::ICMP_ULT: return 4; // 100
3321 case ICmpInst::ICMP_SLT: return 4; // 100
3322 case ICmpInst::ICMP_NE: return 5; // 101
3323 case ICmpInst::ICMP_ULE: return 6; // 110
3324 case ICmpInst::ICMP_SLE: return 6; // 110
3327 assert(0 && "Invalid ICmp predicate!");
3332 /// getICmpValue - This is the complement of getICmpCode, which turns an
3333 /// opcode and two operands into either a constant true or false, or a brand
3334 /// new ICmp instruction. The sign is passed in to determine which kind
3335 /// of predicate to use in new icmp instructions.
3336 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
3338 default: assert(0 && "Illegal ICmp code!");
3339 case 0: return ConstantInt::getFalse();
3342 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
3344 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
3345 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
3348 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
3350 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
3353 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
3355 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
3356 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
3359 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
3361 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
3362 case 7: return ConstantInt::getTrue();
3366 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
3367 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
3368 (ICmpInst::isSignedPredicate(p1) &&
3369 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
3370 (ICmpInst::isSignedPredicate(p2) &&
3371 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
3375 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3376 struct FoldICmpLogical {
3379 ICmpInst::Predicate pred;
3380 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
3381 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
3382 pred(ICI->getPredicate()) {}
3383 bool shouldApply(Value *V) const {
3384 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
3385 if (PredicatesFoldable(pred, ICI->getPredicate()))
3386 return ((ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS) ||
3387 (ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS));
3390 Instruction *apply(Instruction &Log) const {
3391 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
3392 if (ICI->getOperand(0) != LHS) {
3393 assert(ICI->getOperand(1) == LHS);
3394 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
3397 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
3398 unsigned LHSCode = getICmpCode(ICI);
3399 unsigned RHSCode = getICmpCode(RHSICI);
3401 switch (Log.getOpcode()) {
3402 case Instruction::And: Code = LHSCode & RHSCode; break;
3403 case Instruction::Or: Code = LHSCode | RHSCode; break;
3404 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
3405 default: assert(0 && "Illegal logical opcode!"); return 0;
3408 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
3409 ICmpInst::isSignedPredicate(ICI->getPredicate());
3411 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
3412 if (Instruction *I = dyn_cast<Instruction>(RV))
3414 // Otherwise, it's a constant boolean value...
3415 return IC.ReplaceInstUsesWith(Log, RV);
3418 } // end anonymous namespace
3420 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
3421 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
3422 // guaranteed to be a binary operator.
3423 Instruction *InstCombiner::OptAndOp(Instruction *Op,
3425 ConstantInt *AndRHS,
3426 BinaryOperator &TheAnd) {
3427 Value *X = Op->getOperand(0);
3428 Constant *Together = 0;
3430 Together = And(AndRHS, OpRHS);
3432 switch (Op->getOpcode()) {
3433 case Instruction::Xor:
3434 if (Op->hasOneUse()) {
3435 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3436 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3437 InsertNewInstBefore(And, TheAnd);
3439 return BinaryOperator::createXor(And, Together);
3442 case Instruction::Or:
3443 if (Together == AndRHS) // (X | C) & C --> C
3444 return ReplaceInstUsesWith(TheAnd, AndRHS);
3446 if (Op->hasOneUse() && Together != OpRHS) {
3447 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3448 Instruction *Or = BinaryOperator::createOr(X, Together);
3449 InsertNewInstBefore(Or, TheAnd);
3451 return BinaryOperator::createAnd(Or, AndRHS);
3454 case Instruction::Add:
3455 if (Op->hasOneUse()) {
3456 // Adding a one to a single bit bit-field should be turned into an XOR
3457 // of the bit. First thing to check is to see if this AND is with a
3458 // single bit constant.
3459 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3461 // If there is only one bit set...
3462 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3463 // Ok, at this point, we know that we are masking the result of the
3464 // ADD down to exactly one bit. If the constant we are adding has
3465 // no bits set below this bit, then we can eliminate the ADD.
3466 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3468 // Check to see if any bits below the one bit set in AndRHSV are set.
3469 if ((AddRHS & (AndRHSV-1)) == 0) {
3470 // If not, the only thing that can effect the output of the AND is
3471 // the bit specified by AndRHSV. If that bit is set, the effect of
3472 // the XOR is to toggle the bit. If it is clear, then the ADD has
3474 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3475 TheAnd.setOperand(0, X);
3478 // Pull the XOR out of the AND.
3479 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3480 InsertNewInstBefore(NewAnd, TheAnd);
3481 NewAnd->takeName(Op);
3482 return BinaryOperator::createXor(NewAnd, AndRHS);
3489 case Instruction::Shl: {
3490 // We know that the AND will not produce any of the bits shifted in, so if
3491 // the anded constant includes them, clear them now!
3493 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3494 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3495 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3496 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3498 if (CI->getValue() == ShlMask) {
3499 // Masking out bits that the shift already masks
3500 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3501 } else if (CI != AndRHS) { // Reducing bits set in and.
3502 TheAnd.setOperand(1, CI);
3507 case Instruction::LShr:
3509 // We know that the AND will not produce any of the bits shifted in, so if
3510 // the anded constant includes them, clear them now! This only applies to
3511 // unsigned shifts, because a signed shr may bring in set bits!
3513 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3514 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3515 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3516 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3518 if (CI->getValue() == ShrMask) {
3519 // Masking out bits that the shift already masks.
3520 return ReplaceInstUsesWith(TheAnd, Op);
3521 } else if (CI != AndRHS) {
3522 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3527 case Instruction::AShr:
3529 // See if this is shifting in some sign extension, then masking it out
3531 if (Op->hasOneUse()) {
3532 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3533 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3534 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3535 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3536 if (C == AndRHS) { // Masking out bits shifted in.
3537 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3538 // Make the argument unsigned.
3539 Value *ShVal = Op->getOperand(0);
3540 ShVal = InsertNewInstBefore(
3541 BinaryOperator::createLShr(ShVal, OpRHS,
3542 Op->getName()), TheAnd);
3543 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3552 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3553 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3554 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3555 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3556 /// insert new instructions.
3557 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3558 bool isSigned, bool Inside,
3560 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3561 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3562 "Lo is not <= Hi in range emission code!");
3565 if (Lo == Hi) // Trivially false.
3566 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3568 // V >= Min && V < Hi --> V < Hi
3569 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3570 ICmpInst::Predicate pred = (isSigned ?
3571 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3572 return new ICmpInst(pred, V, Hi);
3575 // Emit V-Lo <u Hi-Lo
3576 Constant *NegLo = ConstantExpr::getNeg(Lo);
3577 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3578 InsertNewInstBefore(Add, IB);
3579 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3580 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3583 if (Lo == Hi) // Trivially true.
3584 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3586 // V < Min || V >= Hi -> V > Hi-1
3587 Hi = SubOne(cast<ConstantInt>(Hi));
3588 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3589 ICmpInst::Predicate pred = (isSigned ?
3590 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3591 return new ICmpInst(pred, V, Hi);
3594 // Emit V-Lo >u Hi-1-Lo
3595 // Note that Hi has already had one subtracted from it, above.
3596 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3597 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3598 InsertNewInstBefore(Add, IB);
3599 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3600 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3603 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3604 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3605 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3606 // not, since all 1s are not contiguous.
3607 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3608 const APInt& V = Val->getValue();
3609 uint32_t BitWidth = Val->getType()->getBitWidth();
3610 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3612 // look for the first zero bit after the run of ones
3613 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3614 // look for the first non-zero bit
3615 ME = V.getActiveBits();
3619 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3620 /// where isSub determines whether the operator is a sub. If we can fold one of
3621 /// the following xforms:
3623 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3624 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3625 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3627 /// return (A +/- B).
3629 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3630 ConstantInt *Mask, bool isSub,
3632 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3633 if (!LHSI || LHSI->getNumOperands() != 2 ||
3634 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3636 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3638 switch (LHSI->getOpcode()) {
3640 case Instruction::And:
3641 if (And(N, Mask) == Mask) {
3642 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3643 if ((Mask->getValue().countLeadingZeros() +
3644 Mask->getValue().countPopulation()) ==
3645 Mask->getValue().getBitWidth())
3648 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3649 // part, we don't need any explicit masks to take them out of A. If that
3650 // is all N is, ignore it.
3651 uint32_t MB = 0, ME = 0;
3652 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3653 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3654 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3655 if (MaskedValueIsZero(RHS, Mask))
3660 case Instruction::Or:
3661 case Instruction::Xor:
3662 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3663 if ((Mask->getValue().countLeadingZeros() +
3664 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3665 && And(N, Mask)->isZero())
3672 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3674 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3675 return InsertNewInstBefore(New, I);
3678 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3679 bool Changed = SimplifyCommutative(I);
3680 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3682 if (isa<UndefValue>(Op1)) // X & undef -> 0
3683 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3687 return ReplaceInstUsesWith(I, Op1);
3689 // See if we can simplify any instructions used by the instruction whose sole
3690 // purpose is to compute bits we don't care about.
3691 if (!isa<VectorType>(I.getType())) {
3692 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3693 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3694 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3695 KnownZero, KnownOne))
3698 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3699 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3700 return ReplaceInstUsesWith(I, I.getOperand(0));
3701 } else if (isa<ConstantAggregateZero>(Op1)) {
3702 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3706 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3707 const APInt& AndRHSMask = AndRHS->getValue();
3708 APInt NotAndRHS(~AndRHSMask);
3710 // Optimize a variety of ((val OP C1) & C2) combinations...
3711 if (isa<BinaryOperator>(Op0)) {
3712 Instruction *Op0I = cast<Instruction>(Op0);
3713 Value *Op0LHS = Op0I->getOperand(0);
3714 Value *Op0RHS = Op0I->getOperand(1);
3715 switch (Op0I->getOpcode()) {
3716 case Instruction::Xor:
3717 case Instruction::Or:
3718 // If the mask is only needed on one incoming arm, push it up.
3719 if (Op0I->hasOneUse()) {
3720 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3721 // Not masking anything out for the LHS, move to RHS.
3722 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3723 Op0RHS->getName()+".masked");
3724 InsertNewInstBefore(NewRHS, I);
3725 return BinaryOperator::create(
3726 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3728 if (!isa<Constant>(Op0RHS) &&
3729 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3730 // Not masking anything out for the RHS, move to LHS.
3731 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3732 Op0LHS->getName()+".masked");
3733 InsertNewInstBefore(NewLHS, I);
3734 return BinaryOperator::create(
3735 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3740 case Instruction::Add:
3741 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3742 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3743 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3744 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3745 return BinaryOperator::createAnd(V, AndRHS);
3746 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3747 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3750 case Instruction::Sub:
3751 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3752 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3753 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3754 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3755 return BinaryOperator::createAnd(V, AndRHS);
3759 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3760 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3762 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3763 // If this is an integer truncation or change from signed-to-unsigned, and
3764 // if the source is an and/or with immediate, transform it. This
3765 // frequently occurs for bitfield accesses.
3766 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3767 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3768 CastOp->getNumOperands() == 2)
3769 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1))) {
3770 if (CastOp->getOpcode() == Instruction::And) {
3771 // Change: and (cast (and X, C1) to T), C2
3772 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3773 // This will fold the two constants together, which may allow
3774 // other simplifications.
3775 Instruction *NewCast = CastInst::createTruncOrBitCast(
3776 CastOp->getOperand(0), I.getType(),
3777 CastOp->getName()+".shrunk");
3778 NewCast = InsertNewInstBefore(NewCast, I);
3779 // trunc_or_bitcast(C1)&C2
3780 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3781 C3 = ConstantExpr::getAnd(C3, AndRHS);
3782 return BinaryOperator::createAnd(NewCast, C3);
3783 } else if (CastOp->getOpcode() == Instruction::Or) {
3784 // Change: and (cast (or X, C1) to T), C2
3785 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3786 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3787 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3788 return ReplaceInstUsesWith(I, AndRHS);
3794 // Try to fold constant and into select arguments.
3795 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3796 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3798 if (isa<PHINode>(Op0))
3799 if (Instruction *NV = FoldOpIntoPhi(I))
3803 Value *Op0NotVal = dyn_castNotVal(Op0);
3804 Value *Op1NotVal = dyn_castNotVal(Op1);
3806 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3807 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3809 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3810 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3811 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3812 I.getName()+".demorgan");
3813 InsertNewInstBefore(Or, I);
3814 return BinaryOperator::createNot(Or);
3818 Value *A = 0, *B = 0, *C = 0, *D = 0;
3819 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3820 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3821 return ReplaceInstUsesWith(I, Op1);
3823 // (A|B) & ~(A&B) -> A^B
3824 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3825 if ((A == C && B == D) || (A == D && B == C))
3826 return BinaryOperator::createXor(A, B);
3830 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3831 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3832 return ReplaceInstUsesWith(I, Op0);
3834 // ~(A&B) & (A|B) -> A^B
3835 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3836 if ((A == C && B == D) || (A == D && B == C))
3837 return BinaryOperator::createXor(A, B);
3841 if (Op0->hasOneUse() &&
3842 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3843 if (A == Op1) { // (A^B)&A -> A&(A^B)
3844 I.swapOperands(); // Simplify below
3845 std::swap(Op0, Op1);
3846 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3847 cast<BinaryOperator>(Op0)->swapOperands();
3848 I.swapOperands(); // Simplify below
3849 std::swap(Op0, Op1);
3852 if (Op1->hasOneUse() &&
3853 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3854 if (B == Op0) { // B&(A^B) -> B&(B^A)
3855 cast<BinaryOperator>(Op1)->swapOperands();
3858 if (A == Op0) { // A&(A^B) -> A & ~B
3859 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3860 InsertNewInstBefore(NotB, I);
3861 return BinaryOperator::createAnd(A, NotB);
3866 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3867 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3868 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3871 Value *LHSVal, *RHSVal;
3872 ConstantInt *LHSCst, *RHSCst;
3873 ICmpInst::Predicate LHSCC, RHSCC;
3874 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3875 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3876 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3877 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3878 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3879 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3880 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3881 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3883 // Don't try to fold ICMP_SLT + ICMP_ULT.
3884 (ICmpInst::isEquality(LHSCC) || ICmpInst::isEquality(RHSCC) ||
3885 ICmpInst::isSignedPredicate(LHSCC) ==
3886 ICmpInst::isSignedPredicate(RHSCC))) {
3887 // Ensure that the larger constant is on the RHS.
3888 ICmpInst::Predicate GT;
3889 if (ICmpInst::isSignedPredicate(LHSCC) ||
3890 (ICmpInst::isEquality(LHSCC) &&
3891 ICmpInst::isSignedPredicate(RHSCC)))
3892 GT = ICmpInst::ICMP_SGT;
3894 GT = ICmpInst::ICMP_UGT;
3896 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3897 ICmpInst *LHS = cast<ICmpInst>(Op0);
3898 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3899 std::swap(LHS, RHS);
3900 std::swap(LHSCst, RHSCst);
3901 std::swap(LHSCC, RHSCC);
3904 // At this point, we know we have have two icmp instructions
3905 // comparing a value against two constants and and'ing the result
3906 // together. Because of the above check, we know that we only have
3907 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3908 // (from the FoldICmpLogical check above), that the two constants
3909 // are not equal and that the larger constant is on the RHS
3910 assert(LHSCst != RHSCst && "Compares not folded above?");
3913 default: assert(0 && "Unknown integer condition code!");
3914 case ICmpInst::ICMP_EQ:
3916 default: assert(0 && "Unknown integer condition code!");
3917 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3918 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3919 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3920 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3921 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3922 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3923 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3924 return ReplaceInstUsesWith(I, LHS);
3926 case ICmpInst::ICMP_NE:
3928 default: assert(0 && "Unknown integer condition code!");
3929 case ICmpInst::ICMP_ULT:
3930 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3931 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3932 break; // (X != 13 & X u< 15) -> no change
3933 case ICmpInst::ICMP_SLT:
3934 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3935 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3936 break; // (X != 13 & X s< 15) -> no change
3937 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3938 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3939 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3940 return ReplaceInstUsesWith(I, RHS);
3941 case ICmpInst::ICMP_NE:
3942 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3943 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3944 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3945 LHSVal->getName()+".off");
3946 InsertNewInstBefore(Add, I);
3947 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3948 ConstantInt::get(Add->getType(), 1));
3950 break; // (X != 13 & X != 15) -> no change
3953 case ICmpInst::ICMP_ULT:
3955 default: assert(0 && "Unknown integer condition code!");
3956 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3957 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3958 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3959 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3961 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3962 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3963 return ReplaceInstUsesWith(I, LHS);
3964 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3968 case ICmpInst::ICMP_SLT:
3970 default: assert(0 && "Unknown integer condition code!");
3971 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3972 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3973 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3974 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3976 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3977 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3978 return ReplaceInstUsesWith(I, LHS);
3979 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3983 case ICmpInst::ICMP_UGT:
3985 default: assert(0 && "Unknown integer condition code!");
3986 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3987 return ReplaceInstUsesWith(I, LHS);
3988 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3989 return ReplaceInstUsesWith(I, RHS);
3990 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3992 case ICmpInst::ICMP_NE:
3993 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3994 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3995 break; // (X u> 13 & X != 15) -> no change
3996 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3997 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3999 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
4003 case ICmpInst::ICMP_SGT:
4005 default: assert(0 && "Unknown integer condition code!");
4006 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X == 15
4007 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
4008 return ReplaceInstUsesWith(I, RHS);
4009 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
4011 case ICmpInst::ICMP_NE:
4012 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
4013 return new ICmpInst(LHSCC, LHSVal, RHSCst);
4014 break; // (X s> 13 & X != 15) -> no change
4015 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
4016 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
4018 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
4026 // fold (and (cast A), (cast B)) -> (cast (and A, B))
4027 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4028 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4029 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
4030 const Type *SrcTy = Op0C->getOperand(0)->getType();
4031 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4032 // Only do this if the casts both really cause code to be generated.
4033 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4035 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4037 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
4038 Op1C->getOperand(0),
4040 InsertNewInstBefore(NewOp, I);
4041 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4045 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
4046 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4047 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4048 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4049 SI0->getOperand(1) == SI1->getOperand(1) &&
4050 (SI0->hasOneUse() || SI1->hasOneUse())) {
4051 Instruction *NewOp =
4052 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
4054 SI0->getName()), I);
4055 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4056 SI1->getOperand(1));
4060 // (fcmp ord x, c) & (fcmp ord y, c) -> (fcmp ord x, y)
4061 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4062 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4063 if (LHS->getPredicate() == FCmpInst::FCMP_ORD &&
4064 RHS->getPredicate() == FCmpInst::FCMP_ORD)
4065 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4066 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4067 // If either of the constants are nans, then the whole thing returns
4069 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4070 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4071 return new FCmpInst(FCmpInst::FCMP_ORD, LHS->getOperand(0),
4072 RHS->getOperand(0));
4077 return Changed ? &I : 0;
4080 /// CollectBSwapParts - Look to see if the specified value defines a single byte
4081 /// in the result. If it does, and if the specified byte hasn't been filled in
4082 /// yet, fill it in and return false.
4083 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
4084 Instruction *I = dyn_cast<Instruction>(V);
4085 if (I == 0) return true;
4087 // If this is an or instruction, it is an inner node of the bswap.
4088 if (I->getOpcode() == Instruction::Or)
4089 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
4090 CollectBSwapParts(I->getOperand(1), ByteValues);
4092 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
4093 // If this is a shift by a constant int, and it is "24", then its operand
4094 // defines a byte. We only handle unsigned types here.
4095 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
4096 // Not shifting the entire input by N-1 bytes?
4097 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
4098 8*(ByteValues.size()-1))
4102 if (I->getOpcode() == Instruction::Shl) {
4103 // X << 24 defines the top byte with the lowest of the input bytes.
4104 DestNo = ByteValues.size()-1;
4106 // X >>u 24 defines the low byte with the highest of the input bytes.
4110 // If the destination byte value is already defined, the values are or'd
4111 // together, which isn't a bswap (unless it's an or of the same bits).
4112 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
4114 ByteValues[DestNo] = I->getOperand(0);
4118 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
4120 Value *Shift = 0, *ShiftLHS = 0;
4121 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
4122 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
4123 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
4125 Instruction *SI = cast<Instruction>(Shift);
4127 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
4128 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
4129 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
4132 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
4134 if (AndAmt->getValue().getActiveBits() > 64)
4136 uint64_t AndAmtVal = AndAmt->getZExtValue();
4137 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
4138 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
4140 // Unknown mask for bswap.
4141 if (DestByte == ByteValues.size()) return true;
4143 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
4145 if (SI->getOpcode() == Instruction::Shl)
4146 SrcByte = DestByte - ShiftBytes;
4148 SrcByte = DestByte + ShiftBytes;
4150 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
4151 if (SrcByte != ByteValues.size()-DestByte-1)
4154 // If the destination byte value is already defined, the values are or'd
4155 // together, which isn't a bswap (unless it's an or of the same bits).
4156 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
4158 ByteValues[DestByte] = SI->getOperand(0);
4162 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
4163 /// If so, insert the new bswap intrinsic and return it.
4164 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
4165 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
4166 if (!ITy || ITy->getBitWidth() % 16)
4167 return 0; // Can only bswap pairs of bytes. Can't do vectors.
4169 /// ByteValues - For each byte of the result, we keep track of which value
4170 /// defines each byte.
4171 SmallVector<Value*, 8> ByteValues;
4172 ByteValues.resize(ITy->getBitWidth()/8);
4174 // Try to find all the pieces corresponding to the bswap.
4175 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
4176 CollectBSwapParts(I.getOperand(1), ByteValues))
4179 // Check to see if all of the bytes come from the same value.
4180 Value *V = ByteValues[0];
4181 if (V == 0) return 0; // Didn't find a byte? Must be zero.
4183 // Check to make sure that all of the bytes come from the same value.
4184 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
4185 if (ByteValues[i] != V)
4187 const Type *Tys[] = { ITy };
4188 Module *M = I.getParent()->getParent()->getParent();
4189 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
4190 return CallInst::Create(F, V);
4194 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
4195 bool Changed = SimplifyCommutative(I);
4196 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4198 if (isa<UndefValue>(Op1)) // X | undef -> -1
4199 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4203 return ReplaceInstUsesWith(I, Op0);
4205 // See if we can simplify any instructions used by the instruction whose sole
4206 // purpose is to compute bits we don't care about.
4207 if (!isa<VectorType>(I.getType())) {
4208 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4209 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4210 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4211 KnownZero, KnownOne))
4213 } else if (isa<ConstantAggregateZero>(Op1)) {
4214 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
4215 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
4216 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
4217 return ReplaceInstUsesWith(I, I.getOperand(1));
4223 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4224 ConstantInt *C1 = 0; Value *X = 0;
4225 // (X & C1) | C2 --> (X | C2) & (C1|C2)
4226 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4227 Instruction *Or = BinaryOperator::createOr(X, RHS);
4228 InsertNewInstBefore(Or, I);
4230 return BinaryOperator::createAnd(Or,
4231 ConstantInt::get(RHS->getValue() | C1->getValue()));
4234 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
4235 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
4236 Instruction *Or = BinaryOperator::createOr(X, RHS);
4237 InsertNewInstBefore(Or, I);
4239 return BinaryOperator::createXor(Or,
4240 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
4243 // Try to fold constant and into select arguments.
4244 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4245 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4247 if (isa<PHINode>(Op0))
4248 if (Instruction *NV = FoldOpIntoPhi(I))
4252 Value *A = 0, *B = 0;
4253 ConstantInt *C1 = 0, *C2 = 0;
4255 if (match(Op0, m_And(m_Value(A), m_Value(B))))
4256 if (A == Op1 || B == Op1) // (A & ?) | A --> A
4257 return ReplaceInstUsesWith(I, Op1);
4258 if (match(Op1, m_And(m_Value(A), m_Value(B))))
4259 if (A == Op0 || B == Op0) // A | (A & ?) --> A
4260 return ReplaceInstUsesWith(I, Op0);
4262 // (A | B) | C and A | (B | C) -> bswap if possible.
4263 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
4264 if (match(Op0, m_Or(m_Value(), m_Value())) ||
4265 match(Op1, m_Or(m_Value(), m_Value())) ||
4266 (match(Op0, m_Shift(m_Value(), m_Value())) &&
4267 match(Op1, m_Shift(m_Value(), m_Value())))) {
4268 if (Instruction *BSwap = MatchBSwap(I))
4272 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
4273 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4274 MaskedValueIsZero(Op1, C1->getValue())) {
4275 Instruction *NOr = BinaryOperator::createOr(A, Op1);
4276 InsertNewInstBefore(NOr, I);
4278 return BinaryOperator::createXor(NOr, C1);
4281 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
4282 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
4283 MaskedValueIsZero(Op0, C1->getValue())) {
4284 Instruction *NOr = BinaryOperator::createOr(A, Op0);
4285 InsertNewInstBefore(NOr, I);
4287 return BinaryOperator::createXor(NOr, C1);
4291 Value *C = 0, *D = 0;
4292 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
4293 match(Op1, m_And(m_Value(B), m_Value(D)))) {
4294 Value *V1 = 0, *V2 = 0, *V3 = 0;
4295 C1 = dyn_cast<ConstantInt>(C);
4296 C2 = dyn_cast<ConstantInt>(D);
4297 if (C1 && C2) { // (A & C1)|(B & C2)
4298 // If we have: ((V + N) & C1) | (V & C2)
4299 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
4300 // replace with V+N.
4301 if (C1->getValue() == ~C2->getValue()) {
4302 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
4303 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
4304 // Add commutes, try both ways.
4305 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
4306 return ReplaceInstUsesWith(I, A);
4307 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
4308 return ReplaceInstUsesWith(I, A);
4310 // Or commutes, try both ways.
4311 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
4312 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
4313 // Add commutes, try both ways.
4314 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
4315 return ReplaceInstUsesWith(I, B);
4316 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
4317 return ReplaceInstUsesWith(I, B);
4320 V1 = 0; V2 = 0; V3 = 0;
4323 // Check to see if we have any common things being and'ed. If so, find the
4324 // terms for V1 & (V2|V3).
4325 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
4326 if (A == B) // (A & C)|(A & D) == A & (C|D)
4327 V1 = A, V2 = C, V3 = D;
4328 else if (A == D) // (A & C)|(B & A) == A & (B|C)
4329 V1 = A, V2 = B, V3 = C;
4330 else if (C == B) // (A & C)|(C & D) == C & (A|D)
4331 V1 = C, V2 = A, V3 = D;
4332 else if (C == D) // (A & C)|(B & C) == C & (A|B)
4333 V1 = C, V2 = A, V3 = B;
4337 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
4338 return BinaryOperator::createAnd(V1, Or);
4343 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
4344 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
4345 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
4346 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
4347 SI0->getOperand(1) == SI1->getOperand(1) &&
4348 (SI0->hasOneUse() || SI1->hasOneUse())) {
4349 Instruction *NewOp =
4350 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
4352 SI0->getName()), I);
4353 return BinaryOperator::create(SI1->getOpcode(), NewOp,
4354 SI1->getOperand(1));
4358 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
4359 if (A == Op1) // ~A | A == -1
4360 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4364 // Note, A is still live here!
4365 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
4367 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4369 // (~A | ~B) == (~(A & B)) - De Morgan's Law
4370 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
4371 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
4372 I.getName()+".demorgan"), I);
4373 return BinaryOperator::createNot(And);
4377 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
4378 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
4379 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4382 Value *LHSVal, *RHSVal;
4383 ConstantInt *LHSCst, *RHSCst;
4384 ICmpInst::Predicate LHSCC, RHSCC;
4385 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
4386 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
4387 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
4388 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
4389 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
4390 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
4391 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
4392 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
4393 // We can't fold (ugt x, C) | (sgt x, C2).
4394 PredicatesFoldable(LHSCC, RHSCC)) {
4395 // Ensure that the larger constant is on the RHS.
4396 ICmpInst *LHS = cast<ICmpInst>(Op0);
4398 if (ICmpInst::isSignedPredicate(LHSCC))
4399 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
4401 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
4404 std::swap(LHS, RHS);
4405 std::swap(LHSCst, RHSCst);
4406 std::swap(LHSCC, RHSCC);
4409 // At this point, we know we have have two icmp instructions
4410 // comparing a value against two constants and or'ing the result
4411 // together. Because of the above check, we know that we only have
4412 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
4413 // FoldICmpLogical check above), that the two constants are not
4415 assert(LHSCst != RHSCst && "Compares not folded above?");
4418 default: assert(0 && "Unknown integer condition code!");
4419 case ICmpInst::ICMP_EQ:
4421 default: assert(0 && "Unknown integer condition code!");
4422 case ICmpInst::ICMP_EQ:
4423 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
4424 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
4425 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
4426 LHSVal->getName()+".off");
4427 InsertNewInstBefore(Add, I);
4428 AddCST = Subtract(AddOne(RHSCst), LHSCst);
4429 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
4431 break; // (X == 13 | X == 15) -> no change
4432 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4433 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4435 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4436 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4437 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4438 return ReplaceInstUsesWith(I, RHS);
4441 case ICmpInst::ICMP_NE:
4443 default: assert(0 && "Unknown integer condition code!");
4444 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4445 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4446 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4447 return ReplaceInstUsesWith(I, LHS);
4448 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4449 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4450 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4451 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4454 case ICmpInst::ICMP_ULT:
4456 default: assert(0 && "Unknown integer condition code!");
4457 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4459 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4460 // If RHSCst is [us]MAXINT, it is always false. Not handling
4461 // this can cause overflow.
4462 if (RHSCst->isMaxValue(false))
4463 return ReplaceInstUsesWith(I, LHS);
4464 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4466 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4468 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4469 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4470 return ReplaceInstUsesWith(I, RHS);
4471 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4475 case ICmpInst::ICMP_SLT:
4477 default: assert(0 && "Unknown integer condition code!");
4478 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4480 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4481 // If RHSCst is [us]MAXINT, it is always false. Not handling
4482 // this can cause overflow.
4483 if (RHSCst->isMaxValue(true))
4484 return ReplaceInstUsesWith(I, LHS);
4485 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4487 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4489 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4490 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4491 return ReplaceInstUsesWith(I, RHS);
4492 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4496 case ICmpInst::ICMP_UGT:
4498 default: assert(0 && "Unknown integer condition code!");
4499 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4500 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4501 return ReplaceInstUsesWith(I, LHS);
4502 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4504 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4505 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4506 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4507 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4511 case ICmpInst::ICMP_SGT:
4513 default: assert(0 && "Unknown integer condition code!");
4514 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4515 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4516 return ReplaceInstUsesWith(I, LHS);
4517 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4519 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4520 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4521 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4522 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4530 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4531 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4532 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4533 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4534 if (!isa<ICmpInst>(Op0C->getOperand(0)) ||
4535 !isa<ICmpInst>(Op1C->getOperand(0))) {
4536 const Type *SrcTy = Op0C->getOperand(0)->getType();
4537 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4538 // Only do this if the casts both really cause code to be
4540 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4542 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4544 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4545 Op1C->getOperand(0),
4547 InsertNewInstBefore(NewOp, I);
4548 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4555 // (fcmp uno x, c) | (fcmp uno y, c) -> (fcmp uno x, y)
4556 if (FCmpInst *LHS = dyn_cast<FCmpInst>(I.getOperand(0))) {
4557 if (FCmpInst *RHS = dyn_cast<FCmpInst>(I.getOperand(1))) {
4558 if (LHS->getPredicate() == FCmpInst::FCMP_UNO &&
4559 RHS->getPredicate() == FCmpInst::FCMP_UNO &&
4560 LHS->getOperand(0)->getType() == RHS->getOperand(0)->getType())
4561 if (ConstantFP *LHSC = dyn_cast<ConstantFP>(LHS->getOperand(1)))
4562 if (ConstantFP *RHSC = dyn_cast<ConstantFP>(RHS->getOperand(1))) {
4563 // If either of the constants are nans, then the whole thing returns
4565 if (LHSC->getValueAPF().isNaN() || RHSC->getValueAPF().isNaN())
4566 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4568 // Otherwise, no need to compare the two constants, compare the
4570 return new FCmpInst(FCmpInst::FCMP_UNO, LHS->getOperand(0),
4571 RHS->getOperand(0));
4576 return Changed ? &I : 0;
4579 // XorSelf - Implements: X ^ X --> 0
4582 XorSelf(Value *rhs) : RHS(rhs) {}
4583 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4584 Instruction *apply(BinaryOperator &Xor) const {
4590 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4591 bool Changed = SimplifyCommutative(I);
4592 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4594 if (isa<UndefValue>(Op1)) {
4595 if (isa<UndefValue>(Op0))
4596 // Handle undef ^ undef -> 0 special case. This is a common
4598 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4599 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4602 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4603 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4604 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4605 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4608 // See if we can simplify any instructions used by the instruction whose sole
4609 // purpose is to compute bits we don't care about.
4610 if (!isa<VectorType>(I.getType())) {
4611 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4612 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4613 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4614 KnownZero, KnownOne))
4616 } else if (isa<ConstantAggregateZero>(Op1)) {
4617 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4620 // Is this a ~ operation?
4621 if (Value *NotOp = dyn_castNotVal(&I)) {
4622 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4623 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4624 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4625 if (Op0I->getOpcode() == Instruction::And ||
4626 Op0I->getOpcode() == Instruction::Or) {
4627 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4628 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4630 BinaryOperator::createNot(Op0I->getOperand(1),
4631 Op0I->getOperand(1)->getName()+".not");
4632 InsertNewInstBefore(NotY, I);
4633 if (Op0I->getOpcode() == Instruction::And)
4634 return BinaryOperator::createOr(Op0NotVal, NotY);
4636 return BinaryOperator::createAnd(Op0NotVal, NotY);
4643 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4644 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4645 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4646 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4647 return new ICmpInst(ICI->getInversePredicate(),
4648 ICI->getOperand(0), ICI->getOperand(1));
4650 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4651 return new FCmpInst(FCI->getInversePredicate(),
4652 FCI->getOperand(0), FCI->getOperand(1));
4655 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4656 // ~(c-X) == X-c-1 == X+(-c-1)
4657 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4658 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4659 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4660 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4661 ConstantInt::get(I.getType(), 1));
4662 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4665 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1))) {
4666 if (Op0I->getOpcode() == Instruction::Add) {
4667 // ~(X-c) --> (-c-1)-X
4668 if (RHS->isAllOnesValue()) {
4669 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4670 return BinaryOperator::createSub(
4671 ConstantExpr::getSub(NegOp0CI,
4672 ConstantInt::get(I.getType(), 1)),
4673 Op0I->getOperand(0));
4674 } else if (RHS->getValue().isSignBit()) {
4675 // (X + C) ^ signbit -> (X + C + signbit)
4676 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4677 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4680 } else if (Op0I->getOpcode() == Instruction::Or) {
4681 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4682 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4683 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4684 // Anything in both C1 and C2 is known to be zero, remove it from
4686 Constant *CommonBits = And(Op0CI, RHS);
4687 NewRHS = ConstantExpr::getAnd(NewRHS,
4688 ConstantExpr::getNot(CommonBits));
4689 AddToWorkList(Op0I);
4690 I.setOperand(0, Op0I->getOperand(0));
4691 I.setOperand(1, NewRHS);
4698 // Try to fold constant and into select arguments.
4699 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4700 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4702 if (isa<PHINode>(Op0))
4703 if (Instruction *NV = FoldOpIntoPhi(I))
4707 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4709 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4711 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4713 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4716 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4719 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4720 if (A == Op0) { // B^(B|A) == (A|B)^B
4721 Op1I->swapOperands();
4723 std::swap(Op0, Op1);
4724 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4725 I.swapOperands(); // Simplified below.
4726 std::swap(Op0, Op1);
4728 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4729 if (Op0 == A) // A^(A^B) == B
4730 return ReplaceInstUsesWith(I, B);
4731 else if (Op0 == B) // A^(B^A) == B
4732 return ReplaceInstUsesWith(I, A);
4733 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4734 if (A == Op0) { // A^(A&B) -> A^(B&A)
4735 Op1I->swapOperands();
4738 if (B == Op0) { // A^(B&A) -> (B&A)^A
4739 I.swapOperands(); // Simplified below.
4740 std::swap(Op0, Op1);
4745 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4748 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4749 if (A == Op1) // (B|A)^B == (A|B)^B
4751 if (B == Op1) { // (A|B)^B == A & ~B
4753 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4754 return BinaryOperator::createAnd(A, NotB);
4756 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4757 if (Op1 == A) // (A^B)^A == B
4758 return ReplaceInstUsesWith(I, B);
4759 else if (Op1 == B) // (B^A)^A == B
4760 return ReplaceInstUsesWith(I, A);
4761 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4762 if (A == Op1) // (A&B)^A -> (B&A)^A
4764 if (B == Op1 && // (B&A)^A == ~B & A
4765 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4767 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4768 return BinaryOperator::createAnd(N, Op1);
4773 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4774 if (Op0I && Op1I && Op0I->isShift() &&
4775 Op0I->getOpcode() == Op1I->getOpcode() &&
4776 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4777 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4778 Instruction *NewOp =
4779 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4780 Op1I->getOperand(0),
4781 Op0I->getName()), I);
4782 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4783 Op1I->getOperand(1));
4787 Value *A, *B, *C, *D;
4788 // (A & B)^(A | B) -> A ^ B
4789 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4790 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4791 if ((A == C && B == D) || (A == D && B == C))
4792 return BinaryOperator::createXor(A, B);
4794 // (A | B)^(A & B) -> A ^ B
4795 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4796 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4797 if ((A == C && B == D) || (A == D && B == C))
4798 return BinaryOperator::createXor(A, B);
4802 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4803 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4804 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4805 // (X & Y)^(X & Y) -> (Y^Z) & X
4806 Value *X = 0, *Y = 0, *Z = 0;
4808 X = A, Y = B, Z = D;
4810 X = A, Y = B, Z = C;
4812 X = B, Y = A, Z = D;
4814 X = B, Y = A, Z = C;
4817 Instruction *NewOp =
4818 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4819 return BinaryOperator::createAnd(NewOp, X);
4824 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4825 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4826 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4829 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4830 if (CastInst *Op0C = dyn_cast<CastInst>(Op0)) {
4831 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4832 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4833 const Type *SrcTy = Op0C->getOperand(0)->getType();
4834 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4835 // Only do this if the casts both really cause code to be generated.
4836 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4838 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4840 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4841 Op1C->getOperand(0),
4843 InsertNewInstBefore(NewOp, I);
4844 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4848 return Changed ? &I : 0;
4851 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4852 /// overflowed for this type.
4853 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4854 ConstantInt *In2, bool IsSigned = false) {
4855 Result = cast<ConstantInt>(Add(In1, In2));
4858 if (In2->getValue().isNegative())
4859 return Result->getValue().sgt(In1->getValue());
4861 return Result->getValue().slt(In1->getValue());
4863 return Result->getValue().ult(In1->getValue());
4866 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4867 /// code necessary to compute the offset from the base pointer (without adding
4868 /// in the base pointer). Return the result as a signed integer of intptr size.
4869 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4870 TargetData &TD = IC.getTargetData();
4871 gep_type_iterator GTI = gep_type_begin(GEP);
4872 const Type *IntPtrTy = TD.getIntPtrType();
4873 Value *Result = Constant::getNullValue(IntPtrTy);
4875 // Build a mask for high order bits.
4876 unsigned IntPtrWidth = TD.getPointerSizeInBits();
4877 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4879 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4880 Value *Op = GEP->getOperand(i);
4881 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType()) & PtrSizeMask;
4882 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4883 if (OpC->isZero()) continue;
4885 // Handle a struct index, which adds its field offset to the pointer.
4886 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4887 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4889 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4890 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4892 Result = IC.InsertNewInstBefore(
4893 BinaryOperator::createAdd(Result,
4894 ConstantInt::get(IntPtrTy, Size),
4895 GEP->getName()+".offs"), I);
4899 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4900 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4901 Scale = ConstantExpr::getMul(OC, Scale);
4902 if (Constant *RC = dyn_cast<Constant>(Result))
4903 Result = ConstantExpr::getAdd(RC, Scale);
4905 // Emit an add instruction.
4906 Result = IC.InsertNewInstBefore(
4907 BinaryOperator::createAdd(Result, Scale,
4908 GEP->getName()+".offs"), I);
4912 // Convert to correct type.
4913 if (Op->getType() != IntPtrTy) {
4914 if (Constant *OpC = dyn_cast<Constant>(Op))
4915 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4917 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4918 Op->getName()+".c"), I);
4921 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4922 if (Constant *OpC = dyn_cast<Constant>(Op))
4923 Op = ConstantExpr::getMul(OpC, Scale);
4924 else // We'll let instcombine(mul) convert this to a shl if possible.
4925 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4926 GEP->getName()+".idx"), I);
4929 // Emit an add instruction.
4930 if (isa<Constant>(Op) && isa<Constant>(Result))
4931 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4932 cast<Constant>(Result));
4934 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4935 GEP->getName()+".offs"), I);
4941 /// EvaluateGEPOffsetExpression - Return an value that can be used to compare of
4942 /// the *offset* implied by GEP to zero. For example, if we have &A[i], we want
4943 /// to return 'i' for "icmp ne i, 0". Note that, in general, indices can be
4944 /// complex, and scales are involved. The above expression would also be legal
4945 /// to codegen as "icmp ne (i*4), 0" (assuming A is a pointer to i32). This
4946 /// later form is less amenable to optimization though, and we are allowed to
4947 /// generate the first by knowing that pointer arithmetic doesn't overflow.
4949 /// If we can't emit an optimized form for this expression, this returns null.
4951 static Value *EvaluateGEPOffsetExpression(User *GEP, Instruction &I,
4953 TargetData &TD = IC.getTargetData();
4954 gep_type_iterator GTI = gep_type_begin(GEP);
4956 // Check to see if this gep only has a single variable index. If so, and if
4957 // any constant indices are a multiple of its scale, then we can compute this
4958 // in terms of the scale of the variable index. For example, if the GEP
4959 // implies an offset of "12 + i*4", then we can codegen this as "3 + i",
4960 // because the expression will cross zero at the same point.
4961 unsigned i, e = GEP->getNumOperands();
4963 for (i = 1; i != e; ++i, ++GTI) {
4964 if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i))) {
4965 // Compute the aggregate offset of constant indices.
4966 if (CI->isZero()) continue;
4968 // Handle a struct index, which adds its field offset to the pointer.
4969 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4970 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
4972 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
4973 Offset += Size*CI->getSExtValue();
4976 // Found our variable index.
4981 // If there are no variable indices, we must have a constant offset, just
4982 // evaluate it the general way.
4983 if (i == e) return 0;
4985 Value *VariableIdx = GEP->getOperand(i);
4986 // Determine the scale factor of the variable element. For example, this is
4987 // 4 if the variable index is into an array of i32.
4988 uint64_t VariableScale = TD.getABITypeSize(GTI.getIndexedType());
4990 // Verify that there are no other variable indices. If so, emit the hard way.
4991 for (++i, ++GTI; i != e; ++i, ++GTI) {
4992 ConstantInt *CI = dyn_cast<ConstantInt>(GEP->getOperand(i));
4995 // Compute the aggregate offset of constant indices.
4996 if (CI->isZero()) continue;
4998 // Handle a struct index, which adds its field offset to the pointer.
4999 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
5000 Offset += TD.getStructLayout(STy)->getElementOffset(CI->getZExtValue());
5002 uint64_t Size = TD.getABITypeSize(GTI.getIndexedType());
5003 Offset += Size*CI->getSExtValue();
5007 // Okay, we know we have a single variable index, which must be a
5008 // pointer/array/vector index. If there is no offset, life is simple, return
5010 unsigned IntPtrWidth = TD.getPointerSizeInBits();
5012 // Cast to intptrty in case a truncation occurs. If an extension is needed,
5013 // we don't need to bother extending: the extension won't affect where the
5014 // computation crosses zero.
5015 if (VariableIdx->getType()->getPrimitiveSizeInBits() > IntPtrWidth)
5016 VariableIdx = new TruncInst(VariableIdx, TD.getIntPtrType(),
5017 VariableIdx->getNameStart(), &I);
5021 // Otherwise, there is an index. The computation we will do will be modulo
5022 // the pointer size, so get it.
5023 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
5025 Offset &= PtrSizeMask;
5026 VariableScale &= PtrSizeMask;
5028 // To do this transformation, any constant index must be a multiple of the
5029 // variable scale factor. For example, we can evaluate "12 + 4*i" as "3 + i",
5030 // but we can't evaluate "10 + 3*i" in terms of i. Check that the offset is a
5031 // multiple of the variable scale.
5032 int64_t NewOffs = Offset / (int64_t)VariableScale;
5033 if (Offset != NewOffs*(int64_t)VariableScale)
5036 // Okay, we can do this evaluation. Start by converting the index to intptr.
5037 const Type *IntPtrTy = TD.getIntPtrType();
5038 if (VariableIdx->getType() != IntPtrTy)
5039 VariableIdx = CastInst::createIntegerCast(VariableIdx, IntPtrTy,
5041 VariableIdx->getNameStart(), &I);
5042 Constant *OffsetVal = ConstantInt::get(IntPtrTy, NewOffs);
5043 return BinaryOperator::createAdd(VariableIdx, OffsetVal, "offset", &I);
5047 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
5048 /// else. At this point we know that the GEP is on the LHS of the comparison.
5049 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
5050 ICmpInst::Predicate Cond,
5052 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
5054 // Look through bitcasts.
5055 if (BitCastInst *BCI = dyn_cast<BitCastInst>(RHS))
5056 RHS = BCI->getOperand(0);
5058 Value *PtrBase = GEPLHS->getOperand(0);
5059 if (PtrBase == RHS) {
5060 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
5061 // This transformation (ignoring the base and scales) is valid because we
5062 // know pointers can't overflow. See if we can output an optimized form.
5063 Value *Offset = EvaluateGEPOffsetExpression(GEPLHS, I, *this);
5065 // If not, synthesize the offset the hard way.
5067 Offset = EmitGEPOffset(GEPLHS, I, *this);
5068 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
5069 Constant::getNullValue(Offset->getType()));
5070 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
5071 // If the base pointers are different, but the indices are the same, just
5072 // compare the base pointer.
5073 if (PtrBase != GEPRHS->getOperand(0)) {
5074 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
5075 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
5076 GEPRHS->getOperand(0)->getType();
5078 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5079 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5080 IndicesTheSame = false;
5084 // If all indices are the same, just compare the base pointers.
5086 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
5087 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
5089 // Otherwise, the base pointers are different and the indices are
5090 // different, bail out.
5094 // If one of the GEPs has all zero indices, recurse.
5095 bool AllZeros = true;
5096 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
5097 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
5098 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
5103 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
5104 ICmpInst::getSwappedPredicate(Cond), I);
5106 // If the other GEP has all zero indices, recurse.
5108 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5109 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
5110 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
5115 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
5117 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
5118 // If the GEPs only differ by one index, compare it.
5119 unsigned NumDifferences = 0; // Keep track of # differences.
5120 unsigned DiffOperand = 0; // The operand that differs.
5121 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
5122 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
5123 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
5124 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
5125 // Irreconcilable differences.
5129 if (NumDifferences++) break;
5134 if (NumDifferences == 0) // SAME GEP?
5135 return ReplaceInstUsesWith(I, // No comparison is needed here.
5136 ConstantInt::get(Type::Int1Ty,
5137 isTrueWhenEqual(Cond)));
5139 else if (NumDifferences == 1) {
5140 Value *LHSV = GEPLHS->getOperand(DiffOperand);
5141 Value *RHSV = GEPRHS->getOperand(DiffOperand);
5142 // Make sure we do a signed comparison here.
5143 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
5147 // Only lower this if the icmp is the only user of the GEP or if we expect
5148 // the result to fold to a constant!
5149 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
5150 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
5151 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
5152 Value *L = EmitGEPOffset(GEPLHS, I, *this);
5153 Value *R = EmitGEPOffset(GEPRHS, I, *this);
5154 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
5160 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
5161 bool Changed = SimplifyCompare(I);
5162 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5164 // Fold trivial predicates.
5165 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
5166 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
5167 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
5168 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5170 // Simplify 'fcmp pred X, X'
5172 switch (I.getPredicate()) {
5173 default: assert(0 && "Unknown predicate!");
5174 case FCmpInst::FCMP_UEQ: // True if unordered or equal
5175 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
5176 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
5177 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
5178 case FCmpInst::FCMP_OGT: // True if ordered and greater than
5179 case FCmpInst::FCMP_OLT: // True if ordered and less than
5180 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
5181 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
5183 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
5184 case FCmpInst::FCMP_ULT: // True if unordered or less than
5185 case FCmpInst::FCMP_UGT: // True if unordered or greater than
5186 case FCmpInst::FCMP_UNE: // True if unordered or not equal
5187 // Canonicalize these to be 'fcmp uno %X, 0.0'.
5188 I.setPredicate(FCmpInst::FCMP_UNO);
5189 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5192 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
5193 case FCmpInst::FCMP_OEQ: // True if ordered and equal
5194 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
5195 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
5196 // Canonicalize these to be 'fcmp ord %X, 0.0'.
5197 I.setPredicate(FCmpInst::FCMP_ORD);
5198 I.setOperand(1, Constant::getNullValue(Op0->getType()));
5203 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
5204 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5206 // Handle fcmp with constant RHS
5207 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5208 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5209 switch (LHSI->getOpcode()) {
5210 case Instruction::PHI:
5211 if (Instruction *NV = FoldOpIntoPhi(I))
5214 case Instruction::Select:
5215 // If either operand of the select is a constant, we can fold the
5216 // comparison into the select arms, which will cause one to be
5217 // constant folded and the select turned into a bitwise or.
5218 Value *Op1 = 0, *Op2 = 0;
5219 if (LHSI->hasOneUse()) {
5220 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5221 // Fold the known value into the constant operand.
5222 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5223 // Insert a new FCmp of the other select operand.
5224 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5225 LHSI->getOperand(2), RHSC,
5227 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5228 // Fold the known value into the constant operand.
5229 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
5230 // Insert a new FCmp of the other select operand.
5231 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
5232 LHSI->getOperand(1), RHSC,
5238 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5243 return Changed ? &I : 0;
5246 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
5247 bool Changed = SimplifyCompare(I);
5248 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5249 const Type *Ty = Op0->getType();
5253 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5254 isTrueWhenEqual(I)));
5256 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
5257 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
5259 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
5260 // addresses never equal each other! We already know that Op0 != Op1.
5261 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
5262 isa<ConstantPointerNull>(Op0)) &&
5263 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
5264 isa<ConstantPointerNull>(Op1)))
5265 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5266 !isTrueWhenEqual(I)));
5268 // icmp's with boolean values can always be turned into bitwise operations
5269 if (Ty == Type::Int1Ty) {
5270 switch (I.getPredicate()) {
5271 default: assert(0 && "Invalid icmp instruction!");
5272 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
5273 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
5274 InsertNewInstBefore(Xor, I);
5275 return BinaryOperator::createNot(Xor);
5277 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
5278 return BinaryOperator::createXor(Op0, Op1);
5280 case ICmpInst::ICMP_UGT:
5281 case ICmpInst::ICMP_SGT:
5282 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
5284 case ICmpInst::ICMP_ULT:
5285 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
5286 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
5287 InsertNewInstBefore(Not, I);
5288 return BinaryOperator::createAnd(Not, Op1);
5290 case ICmpInst::ICMP_UGE:
5291 case ICmpInst::ICMP_SGE:
5292 std::swap(Op0, Op1); // Change icmp ge -> icmp le
5294 case ICmpInst::ICMP_ULE:
5295 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
5296 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
5297 InsertNewInstBefore(Not, I);
5298 return BinaryOperator::createOr(Not, Op1);
5303 // See if we are doing a comparison between a constant and an instruction that
5304 // can be folded into the comparison.
5305 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
5308 // (icmp ne/eq (sub A B) 0) -> (icmp ne/eq A, B)
5309 if (I.isEquality() && CI->isNullValue() &&
5310 match(Op0, m_Sub(m_Value(A), m_Value(B)))) {
5311 // (icmp cond A B) if cond is equality
5312 return new ICmpInst(I.getPredicate(), A, B);
5315 switch (I.getPredicate()) {
5317 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
5318 if (CI->isMinValue(false))
5319 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5320 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
5321 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
5322 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
5323 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5324 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
5325 if (CI->isMinValue(true))
5326 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
5327 ConstantInt::getAllOnesValue(Op0->getType()));
5331 case ICmpInst::ICMP_SLT:
5332 if (CI->isMinValue(true)) // A <s MIN -> FALSE
5333 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5334 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
5335 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5336 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
5337 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
5340 case ICmpInst::ICMP_UGT:
5341 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
5342 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5343 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
5344 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5345 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
5346 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5348 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
5349 if (CI->isMaxValue(true))
5350 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
5351 ConstantInt::getNullValue(Op0->getType()));
5354 case ICmpInst::ICMP_SGT:
5355 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
5356 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5357 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
5358 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
5359 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
5360 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
5363 case ICmpInst::ICMP_ULE:
5364 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
5365 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5366 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
5367 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5368 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
5369 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5372 case ICmpInst::ICMP_SLE:
5373 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
5374 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5375 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
5376 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5377 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
5378 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
5381 case ICmpInst::ICMP_UGE:
5382 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
5383 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5384 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
5385 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5386 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
5387 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5390 case ICmpInst::ICMP_SGE:
5391 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
5392 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5393 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
5394 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
5395 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
5396 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
5400 // If we still have a icmp le or icmp ge instruction, turn it into the
5401 // appropriate icmp lt or icmp gt instruction. Since the border cases have
5402 // already been handled above, this requires little checking.
5404 switch (I.getPredicate()) {
5406 case ICmpInst::ICMP_ULE:
5407 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
5408 case ICmpInst::ICMP_SLE:
5409 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
5410 case ICmpInst::ICMP_UGE:
5411 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
5412 case ICmpInst::ICMP_SGE:
5413 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
5416 // See if we can fold the comparison based on bits known to be zero or one
5417 // in the input. If this comparison is a normal comparison, it demands all
5418 // bits, if it is a sign bit comparison, it only demands the sign bit.
5421 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
5423 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
5424 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
5425 if (SimplifyDemandedBits(Op0,
5426 isSignBit ? APInt::getSignBit(BitWidth)
5427 : APInt::getAllOnesValue(BitWidth),
5428 KnownZero, KnownOne, 0))
5431 // Given the known and unknown bits, compute a range that the LHS could be
5433 if ((KnownOne | KnownZero) != 0) {
5434 // Compute the Min, Max and RHS values based on the known bits. For the
5435 // EQ and NE we use unsigned values.
5436 APInt Min(BitWidth, 0), Max(BitWidth, 0);
5437 const APInt& RHSVal = CI->getValue();
5438 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
5439 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5442 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
5445 switch (I.getPredicate()) { // LE/GE have been folded already.
5446 default: assert(0 && "Unknown icmp opcode!");
5447 case ICmpInst::ICMP_EQ:
5448 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5449 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5451 case ICmpInst::ICMP_NE:
5452 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
5453 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5455 case ICmpInst::ICMP_ULT:
5456 if (Max.ult(RHSVal))
5457 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5458 if (Min.uge(RHSVal))
5459 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5461 case ICmpInst::ICMP_UGT:
5462 if (Min.ugt(RHSVal))
5463 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5464 if (Max.ule(RHSVal))
5465 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5467 case ICmpInst::ICMP_SLT:
5468 if (Max.slt(RHSVal))
5469 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5470 if (Min.sgt(RHSVal))
5471 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5473 case ICmpInst::ICMP_SGT:
5474 if (Min.sgt(RHSVal))
5475 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
5476 if (Max.sle(RHSVal))
5477 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
5482 // Since the RHS is a ConstantInt (CI), if the left hand side is an
5483 // instruction, see if that instruction also has constants so that the
5484 // instruction can be folded into the icmp
5485 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5486 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
5490 // Handle icmp with constant (but not simple integer constant) RHS
5491 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
5492 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
5493 switch (LHSI->getOpcode()) {
5494 case Instruction::GetElementPtr:
5495 if (RHSC->isNullValue()) {
5496 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
5497 bool isAllZeros = true;
5498 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
5499 if (!isa<Constant>(LHSI->getOperand(i)) ||
5500 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
5505 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
5506 Constant::getNullValue(LHSI->getOperand(0)->getType()));
5510 case Instruction::PHI:
5511 if (Instruction *NV = FoldOpIntoPhi(I))
5514 case Instruction::Select: {
5515 // If either operand of the select is a constant, we can fold the
5516 // comparison into the select arms, which will cause one to be
5517 // constant folded and the select turned into a bitwise or.
5518 Value *Op1 = 0, *Op2 = 0;
5519 if (LHSI->hasOneUse()) {
5520 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
5521 // Fold the known value into the constant operand.
5522 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5523 // Insert a new ICmp of the other select operand.
5524 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5525 LHSI->getOperand(2), RHSC,
5527 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
5528 // Fold the known value into the constant operand.
5529 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
5530 // Insert a new ICmp of the other select operand.
5531 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
5532 LHSI->getOperand(1), RHSC,
5538 return SelectInst::Create(LHSI->getOperand(0), Op1, Op2);
5541 case Instruction::Malloc:
5542 // If we have (malloc != null), and if the malloc has a single use, we
5543 // can assume it is successful and remove the malloc.
5544 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5545 AddToWorkList(LHSI);
5546 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5547 !isTrueWhenEqual(I)));
5553 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5554 if (User *GEP = dyn_castGetElementPtr(Op0))
5555 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5557 if (User *GEP = dyn_castGetElementPtr(Op1))
5558 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5559 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5562 // Test to see if the operands of the icmp are casted versions of other
5563 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5565 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5566 if (isa<PointerType>(Op0->getType()) &&
5567 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5568 // We keep moving the cast from the left operand over to the right
5569 // operand, where it can often be eliminated completely.
5570 Op0 = CI->getOperand(0);
5572 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5573 // so eliminate it as well.
5574 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5575 Op1 = CI2->getOperand(0);
5577 // If Op1 is a constant, we can fold the cast into the constant.
5578 if (Op0->getType() != Op1->getType()) {
5579 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5580 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5582 // Otherwise, cast the RHS right before the icmp
5583 Op1 = InsertBitCastBefore(Op1, Op0->getType(), I);
5586 return new ICmpInst(I.getPredicate(), Op0, Op1);
5590 if (isa<CastInst>(Op0)) {
5591 // Handle the special case of: icmp (cast bool to X), <cst>
5592 // This comes up when you have code like
5595 // For generality, we handle any zero-extension of any operand comparison
5596 // with a constant or another cast from the same type.
5597 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5598 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5602 if (I.isEquality()) {
5603 Value *A, *B, *C, *D;
5604 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5605 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5606 Value *OtherVal = A == Op1 ? B : A;
5607 return new ICmpInst(I.getPredicate(), OtherVal,
5608 Constant::getNullValue(A->getType()));
5611 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5612 // A^c1 == C^c2 --> A == C^(c1^c2)
5613 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5614 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5615 if (Op1->hasOneUse()) {
5616 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5617 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5618 return new ICmpInst(I.getPredicate(), A,
5619 InsertNewInstBefore(Xor, I));
5622 // A^B == A^D -> B == D
5623 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5624 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5625 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5626 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5630 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5631 (A == Op0 || B == Op0)) {
5632 // A == (A^B) -> B == 0
5633 Value *OtherVal = A == Op0 ? B : A;
5634 return new ICmpInst(I.getPredicate(), OtherVal,
5635 Constant::getNullValue(A->getType()));
5637 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5638 // (A-B) == A -> B == 0
5639 return new ICmpInst(I.getPredicate(), B,
5640 Constant::getNullValue(B->getType()));
5642 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5643 // A == (A-B) -> B == 0
5644 return new ICmpInst(I.getPredicate(), B,
5645 Constant::getNullValue(B->getType()));
5648 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5649 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5650 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5651 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5652 Value *X = 0, *Y = 0, *Z = 0;
5655 X = B; Y = D; Z = A;
5656 } else if (A == D) {
5657 X = B; Y = C; Z = A;
5658 } else if (B == C) {
5659 X = A; Y = D; Z = B;
5660 } else if (B == D) {
5661 X = A; Y = C; Z = B;
5664 if (X) { // Build (X^Y) & Z
5665 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5666 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5667 I.setOperand(0, Op1);
5668 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5673 return Changed ? &I : 0;
5677 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5678 /// and CmpRHS are both known to be integer constants.
5679 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5680 ConstantInt *DivRHS) {
5681 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5682 const APInt &CmpRHSV = CmpRHS->getValue();
5684 // FIXME: If the operand types don't match the type of the divide
5685 // then don't attempt this transform. The code below doesn't have the
5686 // logic to deal with a signed divide and an unsigned compare (and
5687 // vice versa). This is because (x /s C1) <s C2 produces different
5688 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5689 // (x /u C1) <u C2. Simply casting the operands and result won't
5690 // work. :( The if statement below tests that condition and bails
5692 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5693 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5695 if (DivRHS->isZero())
5696 return 0; // The ProdOV computation fails on divide by zero.
5698 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5699 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5700 // C2 (CI). By solving for X we can turn this into a range check
5701 // instead of computing a divide.
5702 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5704 // Determine if the product overflows by seeing if the product is
5705 // not equal to the divide. Make sure we do the same kind of divide
5706 // as in the LHS instruction that we're folding.
5707 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5708 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5710 // Get the ICmp opcode
5711 ICmpInst::Predicate Pred = ICI.getPredicate();
5713 // Figure out the interval that is being checked. For example, a comparison
5714 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5715 // Compute this interval based on the constants involved and the signedness of
5716 // the compare/divide. This computes a half-open interval, keeping track of
5717 // whether either value in the interval overflows. After analysis each
5718 // overflow variable is set to 0 if it's corresponding bound variable is valid
5719 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5720 int LoOverflow = 0, HiOverflow = 0;
5721 ConstantInt *LoBound = 0, *HiBound = 0;
5724 if (!DivIsSigned) { // udiv
5725 // e.g. X/5 op 3 --> [15, 20)
5727 HiOverflow = LoOverflow = ProdOV;
5729 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5730 } else if (DivRHS->getValue().isStrictlyPositive()) { // Divisor is > 0.
5731 if (CmpRHSV == 0) { // (X / pos) op 0
5732 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5733 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5735 } else if (CmpRHSV.isStrictlyPositive()) { // (X / pos) op pos
5736 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5737 HiOverflow = LoOverflow = ProdOV;
5739 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5740 } else { // (X / pos) op neg
5741 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5742 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5743 LoOverflow = AddWithOverflow(LoBound, Prod,
5744 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5745 HiBound = AddOne(Prod);
5746 HiOverflow = ProdOV ? -1 : 0;
5748 } else if (DivRHS->getValue().isNegative()) { // Divisor is < 0.
5749 if (CmpRHSV == 0) { // (X / neg) op 0
5750 // e.g. X/-5 op 0 --> [-4, 5)
5751 LoBound = AddOne(DivRHS);
5752 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5753 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5754 HiOverflow = 1; // [INTMIN+1, overflow)
5755 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5757 } else if (CmpRHSV.isStrictlyPositive()) { // (X / neg) op pos
5758 // e.g. X/-5 op 3 --> [-19, -14)
5759 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5761 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5762 HiBound = AddOne(Prod);
5763 } else { // (X / neg) op neg
5764 // e.g. X/-5 op -3 --> [15, 20)
5766 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5767 HiBound = Subtract(Prod, DivRHS);
5770 // Dividing by a negative swaps the condition. LT <-> GT
5771 Pred = ICmpInst::getSwappedPredicate(Pred);
5774 Value *X = DivI->getOperand(0);
5776 default: assert(0 && "Unhandled icmp opcode!");
5777 case ICmpInst::ICMP_EQ:
5778 if (LoOverflow && HiOverflow)
5779 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5780 else if (HiOverflow)
5781 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5782 ICmpInst::ICMP_UGE, X, LoBound);
5783 else if (LoOverflow)
5784 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5785 ICmpInst::ICMP_ULT, X, HiBound);
5787 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5788 case ICmpInst::ICMP_NE:
5789 if (LoOverflow && HiOverflow)
5790 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5791 else if (HiOverflow)
5792 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5793 ICmpInst::ICMP_ULT, X, LoBound);
5794 else if (LoOverflow)
5795 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5796 ICmpInst::ICMP_UGE, X, HiBound);
5798 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5799 case ICmpInst::ICMP_ULT:
5800 case ICmpInst::ICMP_SLT:
5801 if (LoOverflow == +1) // Low bound is greater than input range.
5802 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5803 if (LoOverflow == -1) // Low bound is less than input range.
5804 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5805 return new ICmpInst(Pred, X, LoBound);
5806 case ICmpInst::ICMP_UGT:
5807 case ICmpInst::ICMP_SGT:
5808 if (HiOverflow == +1) // High bound greater than input range.
5809 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5810 else if (HiOverflow == -1) // High bound less than input range.
5811 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5812 if (Pred == ICmpInst::ICMP_UGT)
5813 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5815 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5820 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5822 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5825 const APInt &RHSV = RHS->getValue();
5827 switch (LHSI->getOpcode()) {
5828 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5829 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5830 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5832 if ((ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0) ||
5833 (ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue())) {
5834 Value *CompareVal = LHSI->getOperand(0);
5836 // If the sign bit of the XorCST is not set, there is no change to
5837 // the operation, just stop using the Xor.
5838 if (!XorCST->getValue().isNegative()) {
5839 ICI.setOperand(0, CompareVal);
5840 AddToWorkList(LHSI);
5844 // Was the old condition true if the operand is positive?
5845 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5847 // If so, the new one isn't.
5848 isTrueIfPositive ^= true;
5850 if (isTrueIfPositive)
5851 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5853 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5857 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5858 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5859 LHSI->getOperand(0)->hasOneUse()) {
5860 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5862 // If the LHS is an AND of a truncating cast, we can widen the
5863 // and/compare to be the input width without changing the value
5864 // produced, eliminating a cast.
5865 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5866 // We can do this transformation if either the AND constant does not
5867 // have its sign bit set or if it is an equality comparison.
5868 // Extending a relational comparison when we're checking the sign
5869 // bit would not work.
5870 if (Cast->hasOneUse() &&
5871 (ICI.isEquality() ||
5872 (AndCST->getValue().isNonNegative() && RHSV.isNonNegative()))) {
5874 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5875 APInt NewCST = AndCST->getValue();
5876 NewCST.zext(BitWidth);
5878 NewCI.zext(BitWidth);
5879 Instruction *NewAnd =
5880 BinaryOperator::createAnd(Cast->getOperand(0),
5881 ConstantInt::get(NewCST),LHSI->getName());
5882 InsertNewInstBefore(NewAnd, ICI);
5883 return new ICmpInst(ICI.getPredicate(), NewAnd,
5884 ConstantInt::get(NewCI));
5888 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5889 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5890 // happens a LOT in code produced by the C front-end, for bitfield
5892 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5893 if (Shift && !Shift->isShift())
5897 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5898 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5899 const Type *AndTy = AndCST->getType(); // Type of the and.
5901 // We can fold this as long as we can't shift unknown bits
5902 // into the mask. This can only happen with signed shift
5903 // rights, as they sign-extend.
5905 bool CanFold = Shift->isLogicalShift();
5907 // To test for the bad case of the signed shr, see if any
5908 // of the bits shifted in could be tested after the mask.
5909 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5910 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5912 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5913 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5914 AndCST->getValue()) == 0)
5920 if (Shift->getOpcode() == Instruction::Shl)
5921 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5923 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5925 // Check to see if we are shifting out any of the bits being
5927 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5928 // If we shifted bits out, the fold is not going to work out.
5929 // As a special case, check to see if this means that the
5930 // result is always true or false now.
5931 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5932 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5933 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5934 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5936 ICI.setOperand(1, NewCst);
5937 Constant *NewAndCST;
5938 if (Shift->getOpcode() == Instruction::Shl)
5939 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5941 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5942 LHSI->setOperand(1, NewAndCST);
5943 LHSI->setOperand(0, Shift->getOperand(0));
5944 AddToWorkList(Shift); // Shift is dead.
5945 AddUsesToWorkList(ICI);
5951 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5952 // preferable because it allows the C<<Y expression to be hoisted out
5953 // of a loop if Y is invariant and X is not.
5954 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5955 ICI.isEquality() && !Shift->isArithmeticShift() &&
5956 isa<Instruction>(Shift->getOperand(0))) {
5959 if (Shift->getOpcode() == Instruction::LShr) {
5960 NS = BinaryOperator::createShl(AndCST,
5961 Shift->getOperand(1), "tmp");
5963 // Insert a logical shift.
5964 NS = BinaryOperator::createLShr(AndCST,
5965 Shift->getOperand(1), "tmp");
5967 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5969 // Compute X & (C << Y).
5970 Instruction *NewAnd =
5971 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5972 InsertNewInstBefore(NewAnd, ICI);
5974 ICI.setOperand(0, NewAnd);
5980 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5981 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5984 uint32_t TypeBits = RHSV.getBitWidth();
5986 // Check that the shift amount is in range. If not, don't perform
5987 // undefined shifts. When the shift is visited it will be
5989 if (ShAmt->uge(TypeBits))
5992 if (ICI.isEquality()) {
5993 // If we are comparing against bits always shifted out, the
5994 // comparison cannot succeed.
5996 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5997 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5998 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5999 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6000 return ReplaceInstUsesWith(ICI, Cst);
6003 if (LHSI->hasOneUse()) {
6004 // Otherwise strength reduce the shift into an and.
6005 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6007 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
6010 BinaryOperator::createAnd(LHSI->getOperand(0),
6011 Mask, LHSI->getName()+".mask");
6012 Value *And = InsertNewInstBefore(AndI, ICI);
6013 return new ICmpInst(ICI.getPredicate(), And,
6014 ConstantInt::get(RHSV.lshr(ShAmtVal)));
6018 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
6019 bool TrueIfSigned = false;
6020 if (LHSI->hasOneUse() &&
6021 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
6022 // (X << 31) <s 0 --> (X&1) != 0
6023 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
6024 (TypeBits-ShAmt->getZExtValue()-1));
6026 BinaryOperator::createAnd(LHSI->getOperand(0),
6027 Mask, LHSI->getName()+".mask");
6028 Value *And = InsertNewInstBefore(AndI, ICI);
6030 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
6031 And, Constant::getNullValue(And->getType()));
6036 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
6037 case Instruction::AShr: {
6038 // Only handle equality comparisons of shift-by-constant.
6039 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6040 if (!ShAmt || !ICI.isEquality()) break;
6042 // Check that the shift amount is in range. If not, don't perform
6043 // undefined shifts. When the shift is visited it will be
6045 uint32_t TypeBits = RHSV.getBitWidth();
6046 if (ShAmt->uge(TypeBits))
6049 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
6051 // If we are comparing against bits always shifted out, the
6052 // comparison cannot succeed.
6053 APInt Comp = RHSV << ShAmtVal;
6054 if (LHSI->getOpcode() == Instruction::LShr)
6055 Comp = Comp.lshr(ShAmtVal);
6057 Comp = Comp.ashr(ShAmtVal);
6059 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
6060 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6061 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
6062 return ReplaceInstUsesWith(ICI, Cst);
6065 // Otherwise, check to see if the bits shifted out are known to be zero.
6066 // If so, we can compare against the unshifted value:
6067 // (X & 4) >> 1 == 2 --> (X & 4) == 4.
6068 if (LHSI->hasOneUse() &&
6069 MaskedValueIsZero(LHSI->getOperand(0),
6070 APInt::getLowBitsSet(Comp.getBitWidth(), ShAmtVal))) {
6071 return new ICmpInst(ICI.getPredicate(), LHSI->getOperand(0),
6072 ConstantExpr::getShl(RHS, ShAmt));
6075 if (LHSI->hasOneUse()) {
6076 // Otherwise strength reduce the shift into an and.
6077 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
6078 Constant *Mask = ConstantInt::get(Val);
6081 BinaryOperator::createAnd(LHSI->getOperand(0),
6082 Mask, LHSI->getName()+".mask");
6083 Value *And = InsertNewInstBefore(AndI, ICI);
6084 return new ICmpInst(ICI.getPredicate(), And,
6085 ConstantExpr::getShl(RHS, ShAmt));
6090 case Instruction::SDiv:
6091 case Instruction::UDiv:
6092 // Fold: icmp pred ([us]div X, C1), C2 -> range test
6093 // Fold this div into the comparison, producing a range check.
6094 // Determine, based on the divide type, what the range is being
6095 // checked. If there is an overflow on the low or high side, remember
6096 // it, otherwise compute the range [low, hi) bounding the new value.
6097 // See: InsertRangeTest above for the kinds of replacements possible.
6098 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
6099 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
6104 case Instruction::Add:
6105 // Fold: icmp pred (add, X, C1), C2
6107 if (!ICI.isEquality()) {
6108 ConstantInt *LHSC = dyn_cast<ConstantInt>(LHSI->getOperand(1));
6110 const APInt &LHSV = LHSC->getValue();
6112 ConstantRange CR = ICI.makeConstantRange(ICI.getPredicate(), RHSV)
6115 if (ICI.isSignedPredicate()) {
6116 if (CR.getLower().isSignBit()) {
6117 return new ICmpInst(ICmpInst::ICMP_SLT, LHSI->getOperand(0),
6118 ConstantInt::get(CR.getUpper()));
6119 } else if (CR.getUpper().isSignBit()) {
6120 return new ICmpInst(ICmpInst::ICMP_SGE, LHSI->getOperand(0),
6121 ConstantInt::get(CR.getLower()));
6124 if (CR.getLower().isMinValue()) {
6125 return new ICmpInst(ICmpInst::ICMP_ULT, LHSI->getOperand(0),
6126 ConstantInt::get(CR.getUpper()));
6127 } else if (CR.getUpper().isMinValue()) {
6128 return new ICmpInst(ICmpInst::ICMP_UGE, LHSI->getOperand(0),
6129 ConstantInt::get(CR.getLower()));
6136 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
6137 if (ICI.isEquality()) {
6138 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
6140 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
6141 // the second operand is a constant, simplify a bit.
6142 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
6143 switch (BO->getOpcode()) {
6144 case Instruction::SRem:
6145 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
6146 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
6147 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
6148 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
6149 Instruction *NewRem =
6150 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
6152 InsertNewInstBefore(NewRem, ICI);
6153 return new ICmpInst(ICI.getPredicate(), NewRem,
6154 Constant::getNullValue(BO->getType()));
6158 case Instruction::Add:
6159 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
6160 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6161 if (BO->hasOneUse())
6162 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6163 Subtract(RHS, BOp1C));
6164 } else if (RHSV == 0) {
6165 // Replace ((add A, B) != 0) with (A != -B) if A or B is
6166 // efficiently invertible, or if the add has just this one use.
6167 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
6169 if (Value *NegVal = dyn_castNegVal(BOp1))
6170 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
6171 else if (Value *NegVal = dyn_castNegVal(BOp0))
6172 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
6173 else if (BO->hasOneUse()) {
6174 Instruction *Neg = BinaryOperator::createNeg(BOp1);
6175 InsertNewInstBefore(Neg, ICI);
6177 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
6181 case Instruction::Xor:
6182 // For the xor case, we can xor two constants together, eliminating
6183 // the explicit xor.
6184 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
6185 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6186 ConstantExpr::getXor(RHS, BOC));
6189 case Instruction::Sub:
6190 // Replace (([sub|xor] A, B) != 0) with (A != B)
6192 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
6196 case Instruction::Or:
6197 // If bits are being or'd in that are not present in the constant we
6198 // are comparing against, then the comparison could never succeed!
6199 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
6200 Constant *NotCI = ConstantExpr::getNot(RHS);
6201 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
6202 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6207 case Instruction::And:
6208 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
6209 // If bits are being compared against that are and'd out, then the
6210 // comparison can never succeed!
6211 if ((RHSV & ~BOC->getValue()) != 0)
6212 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
6215 // If we have ((X & C) == C), turn it into ((X & C) != 0).
6216 if (RHS == BOC && RHSV.isPowerOf2())
6217 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
6218 ICmpInst::ICMP_NE, LHSI,
6219 Constant::getNullValue(RHS->getType()));
6221 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
6222 if (isSignBit(BOC)) {
6223 Value *X = BO->getOperand(0);
6224 Constant *Zero = Constant::getNullValue(X->getType());
6225 ICmpInst::Predicate pred = isICMP_NE ?
6226 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
6227 return new ICmpInst(pred, X, Zero);
6230 // ((X & ~7) == 0) --> X < 8
6231 if (RHSV == 0 && isHighOnes(BOC)) {
6232 Value *X = BO->getOperand(0);
6233 Constant *NegX = ConstantExpr::getNeg(BOC);
6234 ICmpInst::Predicate pred = isICMP_NE ?
6235 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
6236 return new ICmpInst(pred, X, NegX);
6241 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
6242 // Handle icmp {eq|ne} <intrinsic>, intcst.
6243 if (II->getIntrinsicID() == Intrinsic::bswap) {
6245 ICI.setOperand(0, II->getOperand(1));
6246 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
6250 } else { // Not a ICMP_EQ/ICMP_NE
6251 // If the LHS is a cast from an integral value of the same size,
6252 // then since we know the RHS is a constant, try to simlify.
6253 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
6254 Value *CastOp = Cast->getOperand(0);
6255 const Type *SrcTy = CastOp->getType();
6256 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
6257 if (SrcTy->isInteger() &&
6258 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
6259 // If this is an unsigned comparison, try to make the comparison use
6260 // smaller constant values.
6261 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
6262 // X u< 128 => X s> -1
6263 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
6264 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
6265 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
6266 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
6267 // X u> 127 => X s< 0
6268 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
6269 Constant::getNullValue(SrcTy));
6277 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
6278 /// We only handle extending casts so far.
6280 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
6281 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
6282 Value *LHSCIOp = LHSCI->getOperand(0);
6283 const Type *SrcTy = LHSCIOp->getType();
6284 const Type *DestTy = LHSCI->getType();
6287 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
6288 // integer type is the same size as the pointer type.
6289 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
6290 getTargetData().getPointerSizeInBits() ==
6291 cast<IntegerType>(DestTy)->getBitWidth()) {
6293 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
6294 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
6295 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
6296 RHSOp = RHSC->getOperand(0);
6297 // If the pointer types don't match, insert a bitcast.
6298 if (LHSCIOp->getType() != RHSOp->getType())
6299 RHSOp = InsertBitCastBefore(RHSOp, LHSCIOp->getType(), ICI);
6303 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
6306 // The code below only handles extension cast instructions, so far.
6308 if (LHSCI->getOpcode() != Instruction::ZExt &&
6309 LHSCI->getOpcode() != Instruction::SExt)
6312 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
6313 bool isSignedCmp = ICI.isSignedPredicate();
6315 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
6316 // Not an extension from the same type?
6317 RHSCIOp = CI->getOperand(0);
6318 if (RHSCIOp->getType() != LHSCIOp->getType())
6321 // If the signedness of the two casts doesn't agree (i.e. one is a sext
6322 // and the other is a zext), then we can't handle this.
6323 if (CI->getOpcode() != LHSCI->getOpcode())
6326 // Deal with equality cases early.
6327 if (ICI.isEquality())
6328 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6330 // A signed comparison of sign extended values simplifies into a
6331 // signed comparison.
6332 if (isSignedCmp && isSignedExt)
6333 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
6335 // The other three cases all fold into an unsigned comparison.
6336 return new ICmpInst(ICI.getUnsignedPredicate(), LHSCIOp, RHSCIOp);
6339 // If we aren't dealing with a constant on the RHS, exit early
6340 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
6344 // Compute the constant that would happen if we truncated to SrcTy then
6345 // reextended to DestTy.
6346 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
6347 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
6349 // If the re-extended constant didn't change...
6351 // Make sure that sign of the Cmp and the sign of the Cast are the same.
6352 // For example, we might have:
6353 // %A = sext short %X to uint
6354 // %B = icmp ugt uint %A, 1330
6355 // It is incorrect to transform this into
6356 // %B = icmp ugt short %X, 1330
6357 // because %A may have negative value.
6359 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
6360 // OR operation is EQ/NE.
6361 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
6362 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
6367 // The re-extended constant changed so the constant cannot be represented
6368 // in the shorter type. Consequently, we cannot emit a simple comparison.
6370 // First, handle some easy cases. We know the result cannot be equal at this
6371 // point so handle the ICI.isEquality() cases
6372 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
6373 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
6374 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
6375 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
6377 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
6378 // should have been folded away previously and not enter in here.
6381 // We're performing a signed comparison.
6382 if (cast<ConstantInt>(CI)->getValue().isNegative())
6383 Result = ConstantInt::getFalse(); // X < (small) --> false
6385 Result = ConstantInt::getTrue(); // X < (large) --> true
6387 // We're performing an unsigned comparison.
6389 // We're performing an unsigned comp with a sign extended value.
6390 // This is true if the input is >= 0. [aka >s -1]
6391 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
6392 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
6393 NegOne, ICI.getName()), ICI);
6395 // Unsigned extend & unsigned compare -> always true.
6396 Result = ConstantInt::getTrue();
6400 // Finally, return the value computed.
6401 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
6402 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
6403 return ReplaceInstUsesWith(ICI, Result);
6405 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
6406 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
6407 "ICmp should be folded!");
6408 if (Constant *CI = dyn_cast<Constant>(Result))
6409 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
6411 return BinaryOperator::createNot(Result);
6415 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
6416 return commonShiftTransforms(I);
6419 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
6420 return commonShiftTransforms(I);
6423 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
6424 if (Instruction *R = commonShiftTransforms(I))
6427 Value *Op0 = I.getOperand(0);
6429 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
6430 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
6431 if (CSI->isAllOnesValue())
6432 return ReplaceInstUsesWith(I, CSI);
6434 // See if we can turn a signed shr into an unsigned shr.
6435 if (MaskedValueIsZero(Op0,
6436 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits())))
6437 return BinaryOperator::createLShr(Op0, I.getOperand(1));
6442 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
6443 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
6444 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
6446 // shl X, 0 == X and shr X, 0 == X
6447 // shl 0, X == 0 and shr 0, X == 0
6448 if (Op1 == Constant::getNullValue(Op1->getType()) ||
6449 Op0 == Constant::getNullValue(Op0->getType()))
6450 return ReplaceInstUsesWith(I, Op0);
6452 if (isa<UndefValue>(Op0)) {
6453 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
6454 return ReplaceInstUsesWith(I, Op0);
6455 else // undef << X -> 0, undef >>u X -> 0
6456 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6458 if (isa<UndefValue>(Op1)) {
6459 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
6460 return ReplaceInstUsesWith(I, Op0);
6461 else // X << undef, X >>u undef -> 0
6462 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
6465 // Try to fold constant and into select arguments.
6466 if (isa<Constant>(Op0))
6467 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
6468 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6471 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
6472 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
6477 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
6478 BinaryOperator &I) {
6479 bool isLeftShift = I.getOpcode() == Instruction::Shl;
6481 // See if we can simplify any instructions used by the instruction whose sole
6482 // purpose is to compute bits we don't care about.
6483 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
6484 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
6485 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
6486 KnownZero, KnownOne))
6489 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
6490 // of a signed value.
6492 if (Op1->uge(TypeBits)) {
6493 if (I.getOpcode() != Instruction::AShr)
6494 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
6496 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
6501 // ((X*C1) << C2) == (X * (C1 << C2))
6502 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
6503 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
6504 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
6505 return BinaryOperator::createMul(BO->getOperand(0),
6506 ConstantExpr::getShl(BOOp, Op1));
6508 // Try to fold constant and into select arguments.
6509 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
6510 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
6512 if (isa<PHINode>(Op0))
6513 if (Instruction *NV = FoldOpIntoPhi(I))
6516 // Fold shift2(trunc(shift1(x,c1)), c2) -> trunc(shift2(shift1(x,c1),c2))
6517 if (TruncInst *TI = dyn_cast<TruncInst>(Op0)) {
6518 Instruction *TrOp = dyn_cast<Instruction>(TI->getOperand(0));
6519 // If 'shift2' is an ashr, we would have to get the sign bit into a funny
6520 // place. Don't try to do this transformation in this case. Also, we
6521 // require that the input operand is a shift-by-constant so that we have
6522 // confidence that the shifts will get folded together. We could do this
6523 // xform in more cases, but it is unlikely to be profitable.
6524 if (TrOp && I.isLogicalShift() && TrOp->isShift() &&
6525 isa<ConstantInt>(TrOp->getOperand(1))) {
6526 // Okay, we'll do this xform. Make the shift of shift.
6527 Constant *ShAmt = ConstantExpr::getZExt(Op1, TrOp->getType());
6528 Instruction *NSh = BinaryOperator::create(I.getOpcode(), TrOp, ShAmt,
6530 InsertNewInstBefore(NSh, I); // (shift2 (shift1 & 0x00FF), c2)
6532 // For logical shifts, the truncation has the effect of making the high
6533 // part of the register be zeros. Emulate this by inserting an AND to
6534 // clear the top bits as needed. This 'and' will usually be zapped by
6535 // other xforms later if dead.
6536 unsigned SrcSize = TrOp->getType()->getPrimitiveSizeInBits();
6537 unsigned DstSize = TI->getType()->getPrimitiveSizeInBits();
6538 APInt MaskV(APInt::getLowBitsSet(SrcSize, DstSize));
6540 // The mask we constructed says what the trunc would do if occurring
6541 // between the shifts. We want to know the effect *after* the second
6542 // shift. We know that it is a logical shift by a constant, so adjust the
6543 // mask as appropriate.
6544 if (I.getOpcode() == Instruction::Shl)
6545 MaskV <<= Op1->getZExtValue();
6547 assert(I.getOpcode() == Instruction::LShr && "Unknown logical shift");
6548 MaskV = MaskV.lshr(Op1->getZExtValue());
6551 Instruction *And = BinaryOperator::createAnd(NSh, ConstantInt::get(MaskV),
6553 InsertNewInstBefore(And, I); // shift1 & 0x00FF
6555 // Return the value truncated to the interesting size.
6556 return new TruncInst(And, I.getType());
6560 if (Op0->hasOneUse()) {
6561 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
6562 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6565 switch (Op0BO->getOpcode()) {
6567 case Instruction::Add:
6568 case Instruction::And:
6569 case Instruction::Or:
6570 case Instruction::Xor: {
6571 // These operators commute.
6572 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
6573 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
6574 match(Op0BO->getOperand(1),
6575 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6576 Instruction *YS = BinaryOperator::createShl(
6577 Op0BO->getOperand(0), Op1,
6579 InsertNewInstBefore(YS, I); // (Y << C)
6581 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
6582 Op0BO->getOperand(1)->getName());
6583 InsertNewInstBefore(X, I); // (X + (Y << C))
6584 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6585 return BinaryOperator::createAnd(X, ConstantInt::get(
6586 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6589 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
6590 Value *Op0BOOp1 = Op0BO->getOperand(1);
6591 if (isLeftShift && Op0BOOp1->hasOneUse() &&
6593 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
6594 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
6596 Instruction *YS = BinaryOperator::createShl(
6597 Op0BO->getOperand(0), Op1,
6599 InsertNewInstBefore(YS, I); // (Y << C)
6601 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6602 V1->getName()+".mask");
6603 InsertNewInstBefore(XM, I); // X & (CC << C)
6605 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
6610 case Instruction::Sub: {
6611 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
6612 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6613 match(Op0BO->getOperand(0),
6614 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
6615 Instruction *YS = BinaryOperator::createShl(
6616 Op0BO->getOperand(1), Op1,
6618 InsertNewInstBefore(YS, I); // (Y << C)
6620 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
6621 Op0BO->getOperand(0)->getName());
6622 InsertNewInstBefore(X, I); // (X + (Y << C))
6623 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
6624 return BinaryOperator::createAnd(X, ConstantInt::get(
6625 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
6628 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
6629 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
6630 match(Op0BO->getOperand(0),
6631 m_And(m_Shr(m_Value(V1), m_Value(V2)),
6632 m_ConstantInt(CC))) && V2 == Op1 &&
6633 cast<BinaryOperator>(Op0BO->getOperand(0))
6634 ->getOperand(0)->hasOneUse()) {
6635 Instruction *YS = BinaryOperator::createShl(
6636 Op0BO->getOperand(1), Op1,
6638 InsertNewInstBefore(YS, I); // (Y << C)
6640 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6641 V1->getName()+".mask");
6642 InsertNewInstBefore(XM, I); // X & (CC << C)
6644 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6652 // If the operand is an bitwise operator with a constant RHS, and the
6653 // shift is the only use, we can pull it out of the shift.
6654 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6655 bool isValid = true; // Valid only for And, Or, Xor
6656 bool highBitSet = false; // Transform if high bit of constant set?
6658 switch (Op0BO->getOpcode()) {
6659 default: isValid = false; break; // Do not perform transform!
6660 case Instruction::Add:
6661 isValid = isLeftShift;
6663 case Instruction::Or:
6664 case Instruction::Xor:
6667 case Instruction::And:
6672 // If this is a signed shift right, and the high bit is modified
6673 // by the logical operation, do not perform the transformation.
6674 // The highBitSet boolean indicates the value of the high bit of
6675 // the constant which would cause it to be modified for this
6678 if (isValid && I.getOpcode() == Instruction::AShr)
6679 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6682 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6684 Instruction *NewShift =
6685 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6686 InsertNewInstBefore(NewShift, I);
6687 NewShift->takeName(Op0BO);
6689 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6696 // Find out if this is a shift of a shift by a constant.
6697 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6698 if (ShiftOp && !ShiftOp->isShift())
6701 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6702 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6703 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6704 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6705 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6706 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6707 Value *X = ShiftOp->getOperand(0);
6709 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6710 if (AmtSum > TypeBits)
6713 const IntegerType *Ty = cast<IntegerType>(I.getType());
6715 // Check for (X << c1) << c2 and (X >> c1) >> c2
6716 if (I.getOpcode() == ShiftOp->getOpcode()) {
6717 return BinaryOperator::create(I.getOpcode(), X,
6718 ConstantInt::get(Ty, AmtSum));
6719 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6720 I.getOpcode() == Instruction::AShr) {
6721 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6722 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6723 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6724 I.getOpcode() == Instruction::LShr) {
6725 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6726 Instruction *Shift =
6727 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6728 InsertNewInstBefore(Shift, I);
6730 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6731 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6734 // Okay, if we get here, one shift must be left, and the other shift must be
6735 // right. See if the amounts are equal.
6736 if (ShiftAmt1 == ShiftAmt2) {
6737 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6738 if (I.getOpcode() == Instruction::Shl) {
6739 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6740 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6742 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6743 if (I.getOpcode() == Instruction::LShr) {
6744 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6745 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6747 // We can simplify ((X << C) >>s C) into a trunc + sext.
6748 // NOTE: we could do this for any C, but that would make 'unusual' integer
6749 // types. For now, just stick to ones well-supported by the code
6751 const Type *SExtType = 0;
6752 switch (Ty->getBitWidth() - ShiftAmt1) {
6759 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6764 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6765 InsertNewInstBefore(NewTrunc, I);
6766 return new SExtInst(NewTrunc, Ty);
6768 // Otherwise, we can't handle it yet.
6769 } else if (ShiftAmt1 < ShiftAmt2) {
6770 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6772 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6773 if (I.getOpcode() == Instruction::Shl) {
6774 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6775 ShiftOp->getOpcode() == Instruction::AShr);
6776 Instruction *Shift =
6777 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6778 InsertNewInstBefore(Shift, I);
6780 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6781 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6784 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6785 if (I.getOpcode() == Instruction::LShr) {
6786 assert(ShiftOp->getOpcode() == Instruction::Shl);
6787 Instruction *Shift =
6788 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6789 InsertNewInstBefore(Shift, I);
6791 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6792 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6795 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6797 assert(ShiftAmt2 < ShiftAmt1);
6798 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6800 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6801 if (I.getOpcode() == Instruction::Shl) {
6802 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6803 ShiftOp->getOpcode() == Instruction::AShr);
6804 Instruction *Shift =
6805 BinaryOperator::create(ShiftOp->getOpcode(), X,
6806 ConstantInt::get(Ty, ShiftDiff));
6807 InsertNewInstBefore(Shift, I);
6809 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6810 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6813 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6814 if (I.getOpcode() == Instruction::LShr) {
6815 assert(ShiftOp->getOpcode() == Instruction::Shl);
6816 Instruction *Shift =
6817 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6818 InsertNewInstBefore(Shift, I);
6820 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6821 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6824 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6831 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6832 /// expression. If so, decompose it, returning some value X, such that Val is
6835 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6837 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6838 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6839 Offset = CI->getZExtValue();
6841 return ConstantInt::get(Type::Int32Ty, 0);
6842 } else if (BinaryOperator *I = dyn_cast<BinaryOperator>(Val)) {
6843 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
6844 if (I->getOpcode() == Instruction::Shl) {
6845 // This is a value scaled by '1 << the shift amt'.
6846 Scale = 1U << RHS->getZExtValue();
6848 return I->getOperand(0);
6849 } else if (I->getOpcode() == Instruction::Mul) {
6850 // This value is scaled by 'RHS'.
6851 Scale = RHS->getZExtValue();
6853 return I->getOperand(0);
6854 } else if (I->getOpcode() == Instruction::Add) {
6855 // We have X+C. Check to see if we really have (X*C2)+C1,
6856 // where C1 is divisible by C2.
6859 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6860 Offset += RHS->getZExtValue();
6867 // Otherwise, we can't look past this.
6874 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6875 /// try to eliminate the cast by moving the type information into the alloc.
6876 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6877 AllocationInst &AI) {
6878 const PointerType *PTy = cast<PointerType>(CI.getType());
6880 // Remove any uses of AI that are dead.
6881 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6883 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6884 Instruction *User = cast<Instruction>(*UI++);
6885 if (isInstructionTriviallyDead(User)) {
6886 while (UI != E && *UI == User)
6887 ++UI; // If this instruction uses AI more than once, don't break UI.
6890 DOUT << "IC: DCE: " << *User;
6891 EraseInstFromFunction(*User);
6895 // Get the type really allocated and the type casted to.
6896 const Type *AllocElTy = AI.getAllocatedType();
6897 const Type *CastElTy = PTy->getElementType();
6898 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6900 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6901 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6902 if (CastElTyAlign < AllocElTyAlign) return 0;
6904 // If the allocation has multiple uses, only promote it if we are strictly
6905 // increasing the alignment of the resultant allocation. If we keep it the
6906 // same, we open the door to infinite loops of various kinds.
6907 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6909 uint64_t AllocElTySize = TD->getABITypeSize(AllocElTy);
6910 uint64_t CastElTySize = TD->getABITypeSize(CastElTy);
6911 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6913 // See if we can satisfy the modulus by pulling a scale out of the array
6915 unsigned ArraySizeScale;
6917 Value *NumElements = // See if the array size is a decomposable linear expr.
6918 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6920 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6922 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6923 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6925 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6930 // If the allocation size is constant, form a constant mul expression
6931 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6932 if (isa<ConstantInt>(NumElements))
6933 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6934 // otherwise multiply the amount and the number of elements
6935 else if (Scale != 1) {
6936 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6937 Amt = InsertNewInstBefore(Tmp, AI);
6941 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6942 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6943 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6944 Amt = InsertNewInstBefore(Tmp, AI);
6947 AllocationInst *New;
6948 if (isa<MallocInst>(AI))
6949 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6951 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6952 InsertNewInstBefore(New, AI);
6955 // If the allocation has multiple uses, insert a cast and change all things
6956 // that used it to use the new cast. This will also hack on CI, but it will
6958 if (!AI.hasOneUse()) {
6959 AddUsesToWorkList(AI);
6960 // New is the allocation instruction, pointer typed. AI is the original
6961 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6962 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6963 InsertNewInstBefore(NewCast, AI);
6964 AI.replaceAllUsesWith(NewCast);
6966 return ReplaceInstUsesWith(CI, New);
6969 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6970 /// and return it as type Ty without inserting any new casts and without
6971 /// changing the computed value. This is used by code that tries to decide
6972 /// whether promoting or shrinking integer operations to wider or smaller types
6973 /// will allow us to eliminate a truncate or extend.
6975 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6976 /// extension operation if Ty is larger.
6977 bool InstCombiner::CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6979 int &NumCastsRemoved) {
6980 // We can always evaluate constants in another type.
6981 if (isa<ConstantInt>(V))
6984 Instruction *I = dyn_cast<Instruction>(V);
6985 if (!I) return false;
6987 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6989 // If this is an extension or truncate, we can often eliminate it.
6990 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6991 // If this is a cast from the destination type, we can trivially eliminate
6992 // it, and this will remove a cast overall.
6993 if (I->getOperand(0)->getType() == Ty) {
6994 // If the first operand is itself a cast, and is eliminable, do not count
6995 // this as an eliminable cast. We would prefer to eliminate those two
6997 if (!isa<CastInst>(I->getOperand(0)))
7003 // We can't extend or shrink something that has multiple uses: doing so would
7004 // require duplicating the instruction in general, which isn't profitable.
7005 if (!I->hasOneUse()) return false;
7007 switch (I->getOpcode()) {
7008 case Instruction::Add:
7009 case Instruction::Sub:
7010 case Instruction::And:
7011 case Instruction::Or:
7012 case Instruction::Xor:
7013 // These operators can all arbitrarily be extended or truncated.
7014 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7016 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7019 case Instruction::Mul:
7020 // A multiply can be truncated by truncating its operands.
7021 return Ty->getBitWidth() < OrigTy->getBitWidth() &&
7022 CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7024 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
7027 case Instruction::Shl:
7028 // If we are truncating the result of this SHL, and if it's a shift of a
7029 // constant amount, we can always perform a SHL in a smaller type.
7030 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7031 uint32_t BitWidth = Ty->getBitWidth();
7032 if (BitWidth < OrigTy->getBitWidth() &&
7033 CI->getLimitedValue(BitWidth) < BitWidth)
7034 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7038 case Instruction::LShr:
7039 // If this is a truncate of a logical shr, we can truncate it to a smaller
7040 // lshr iff we know that the bits we would otherwise be shifting in are
7042 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
7043 uint32_t OrigBitWidth = OrigTy->getBitWidth();
7044 uint32_t BitWidth = Ty->getBitWidth();
7045 if (BitWidth < OrigBitWidth &&
7046 MaskedValueIsZero(I->getOperand(0),
7047 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
7048 CI->getLimitedValue(BitWidth) < BitWidth) {
7049 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
7054 case Instruction::ZExt:
7055 case Instruction::SExt:
7056 case Instruction::Trunc:
7057 // If this is the same kind of case as our original (e.g. zext+zext), we
7058 // can safely replace it. Note that replacing it does not reduce the number
7059 // of casts in the input.
7060 if (I->getOpcode() == CastOpc)
7065 // TODO: Can handle more cases here.
7072 /// EvaluateInDifferentType - Given an expression that
7073 /// CanEvaluateInDifferentType returns true for, actually insert the code to
7074 /// evaluate the expression.
7075 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
7077 if (Constant *C = dyn_cast<Constant>(V))
7078 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
7080 // Otherwise, it must be an instruction.
7081 Instruction *I = cast<Instruction>(V);
7082 Instruction *Res = 0;
7083 switch (I->getOpcode()) {
7084 case Instruction::Add:
7085 case Instruction::Sub:
7086 case Instruction::Mul:
7087 case Instruction::And:
7088 case Instruction::Or:
7089 case Instruction::Xor:
7090 case Instruction::AShr:
7091 case Instruction::LShr:
7092 case Instruction::Shl: {
7093 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
7094 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
7095 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
7096 LHS, RHS, I->getName());
7099 case Instruction::Trunc:
7100 case Instruction::ZExt:
7101 case Instruction::SExt:
7102 // If the source type of the cast is the type we're trying for then we can
7103 // just return the source. There's no need to insert it because it is not
7105 if (I->getOperand(0)->getType() == Ty)
7106 return I->getOperand(0);
7108 // Otherwise, must be the same type of case, so just reinsert a new one.
7109 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
7113 // TODO: Can handle more cases here.
7114 assert(0 && "Unreachable!");
7118 return InsertNewInstBefore(Res, *I);
7121 /// @brief Implement the transforms common to all CastInst visitors.
7122 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
7123 Value *Src = CI.getOperand(0);
7125 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
7126 // eliminate it now.
7127 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7128 if (Instruction::CastOps opc =
7129 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
7130 // The first cast (CSrc) is eliminable so we need to fix up or replace
7131 // the second cast (CI). CSrc will then have a good chance of being dead.
7132 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
7136 // If we are casting a select then fold the cast into the select
7137 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
7138 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
7141 // If we are casting a PHI then fold the cast into the PHI
7142 if (isa<PHINode>(Src))
7143 if (Instruction *NV = FoldOpIntoPhi(CI))
7149 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
7150 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
7151 Value *Src = CI.getOperand(0);
7153 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
7154 // If casting the result of a getelementptr instruction with no offset, turn
7155 // this into a cast of the original pointer!
7156 if (GEP->hasAllZeroIndices()) {
7157 // Changing the cast operand is usually not a good idea but it is safe
7158 // here because the pointer operand is being replaced with another
7159 // pointer operand so the opcode doesn't need to change.
7161 CI.setOperand(0, GEP->getOperand(0));
7165 // If the GEP has a single use, and the base pointer is a bitcast, and the
7166 // GEP computes a constant offset, see if we can convert these three
7167 // instructions into fewer. This typically happens with unions and other
7168 // non-type-safe code.
7169 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
7170 if (GEP->hasAllConstantIndices()) {
7171 // We are guaranteed to get a constant from EmitGEPOffset.
7172 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
7173 int64_t Offset = OffsetV->getSExtValue();
7175 // Get the base pointer input of the bitcast, and the type it points to.
7176 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
7177 const Type *GEPIdxTy =
7178 cast<PointerType>(OrigBase->getType())->getElementType();
7179 if (GEPIdxTy->isSized()) {
7180 SmallVector<Value*, 8> NewIndices;
7182 // Start with the index over the outer type. Note that the type size
7183 // might be zero (even if the offset isn't zero) if the indexed type
7184 // is something like [0 x {int, int}]
7185 const Type *IntPtrTy = TD->getIntPtrType();
7186 int64_t FirstIdx = 0;
7187 if (int64_t TySize = TD->getABITypeSize(GEPIdxTy)) {
7188 FirstIdx = Offset/TySize;
7191 // Handle silly modulus not returning values values [0..TySize).
7195 assert(Offset >= 0);
7197 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
7200 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
7202 // Index into the types. If we fail, set OrigBase to null.
7204 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
7205 const StructLayout *SL = TD->getStructLayout(STy);
7206 if (Offset < (int64_t)SL->getSizeInBytes()) {
7207 unsigned Elt = SL->getElementContainingOffset(Offset);
7208 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
7210 Offset -= SL->getElementOffset(Elt);
7211 GEPIdxTy = STy->getElementType(Elt);
7213 // Otherwise, we can't index into this, bail out.
7217 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
7218 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
7219 if (uint64_t EltSize = TD->getABITypeSize(STy->getElementType())){
7220 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
7223 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
7225 GEPIdxTy = STy->getElementType();
7227 // Otherwise, we can't index into this, bail out.
7233 // If we were able to index down into an element, create the GEP
7234 // and bitcast the result. This eliminates one bitcast, potentially
7236 Instruction *NGEP = GetElementPtrInst::Create(OrigBase,
7238 NewIndices.end(), "");
7239 InsertNewInstBefore(NGEP, CI);
7240 NGEP->takeName(GEP);
7242 if (isa<BitCastInst>(CI))
7243 return new BitCastInst(NGEP, CI.getType());
7244 assert(isa<PtrToIntInst>(CI));
7245 return new PtrToIntInst(NGEP, CI.getType());
7252 return commonCastTransforms(CI);
7257 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
7258 /// integer types. This function implements the common transforms for all those
7260 /// @brief Implement the transforms common to CastInst with integer operands
7261 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
7262 if (Instruction *Result = commonCastTransforms(CI))
7265 Value *Src = CI.getOperand(0);
7266 const Type *SrcTy = Src->getType();
7267 const Type *DestTy = CI.getType();
7268 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
7269 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
7271 // See if we can simplify any instructions used by the LHS whose sole
7272 // purpose is to compute bits we don't care about.
7273 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
7274 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
7275 KnownZero, KnownOne))
7278 // If the source isn't an instruction or has more than one use then we
7279 // can't do anything more.
7280 Instruction *SrcI = dyn_cast<Instruction>(Src);
7281 if (!SrcI || !Src->hasOneUse())
7284 // Attempt to propagate the cast into the instruction for int->int casts.
7285 int NumCastsRemoved = 0;
7286 if (!isa<BitCastInst>(CI) &&
7287 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
7288 CI.getOpcode(), NumCastsRemoved)) {
7289 // If this cast is a truncate, evaluting in a different type always
7290 // eliminates the cast, so it is always a win. If this is a zero-extension,
7291 // we need to do an AND to maintain the clear top-part of the computation,
7292 // so we require that the input have eliminated at least one cast. If this
7293 // is a sign extension, we insert two new casts (to do the extension) so we
7294 // require that two casts have been eliminated.
7296 switch (CI.getOpcode()) {
7298 // All the others use floating point so we shouldn't actually
7299 // get here because of the check above.
7300 assert(0 && "Unknown cast type");
7301 case Instruction::Trunc:
7304 case Instruction::ZExt:
7305 DoXForm = NumCastsRemoved >= 1;
7307 case Instruction::SExt:
7308 DoXForm = NumCastsRemoved >= 2;
7313 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
7314 CI.getOpcode() == Instruction::SExt);
7315 assert(Res->getType() == DestTy);
7316 switch (CI.getOpcode()) {
7317 default: assert(0 && "Unknown cast type!");
7318 case Instruction::Trunc:
7319 case Instruction::BitCast:
7320 // Just replace this cast with the result.
7321 return ReplaceInstUsesWith(CI, Res);
7322 case Instruction::ZExt: {
7323 // We need to emit an AND to clear the high bits.
7324 assert(SrcBitSize < DestBitSize && "Not a zext?");
7325 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
7327 return BinaryOperator::createAnd(Res, C);
7329 case Instruction::SExt:
7330 // We need to emit a cast to truncate, then a cast to sext.
7331 return CastInst::create(Instruction::SExt,
7332 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
7338 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
7339 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
7341 switch (SrcI->getOpcode()) {
7342 case Instruction::Add:
7343 case Instruction::Mul:
7344 case Instruction::And:
7345 case Instruction::Or:
7346 case Instruction::Xor:
7347 // If we are discarding information, rewrite.
7348 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
7349 // Don't insert two casts if they cannot be eliminated. We allow
7350 // two casts to be inserted if the sizes are the same. This could
7351 // only be converting signedness, which is a noop.
7352 if (DestBitSize == SrcBitSize ||
7353 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
7354 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7355 Instruction::CastOps opcode = CI.getOpcode();
7356 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7357 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7358 return BinaryOperator::create(
7359 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7363 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
7364 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
7365 SrcI->getOpcode() == Instruction::Xor &&
7366 Op1 == ConstantInt::getTrue() &&
7367 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
7368 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
7369 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
7372 case Instruction::SDiv:
7373 case Instruction::UDiv:
7374 case Instruction::SRem:
7375 case Instruction::URem:
7376 // If we are just changing the sign, rewrite.
7377 if (DestBitSize == SrcBitSize) {
7378 // Don't insert two casts if they cannot be eliminated. We allow
7379 // two casts to be inserted if the sizes are the same. This could
7380 // only be converting signedness, which is a noop.
7381 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
7382 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
7383 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
7385 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
7387 return BinaryOperator::create(
7388 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
7393 case Instruction::Shl:
7394 // Allow changing the sign of the source operand. Do not allow
7395 // changing the size of the shift, UNLESS the shift amount is a
7396 // constant. We must not change variable sized shifts to a smaller
7397 // size, because it is undefined to shift more bits out than exist
7399 if (DestBitSize == SrcBitSize ||
7400 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
7401 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
7402 Instruction::BitCast : Instruction::Trunc);
7403 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
7404 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
7405 return BinaryOperator::createShl(Op0c, Op1c);
7408 case Instruction::AShr:
7409 // If this is a signed shr, and if all bits shifted in are about to be
7410 // truncated off, turn it into an unsigned shr to allow greater
7412 if (DestBitSize < SrcBitSize &&
7413 isa<ConstantInt>(Op1)) {
7414 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
7415 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
7416 // Insert the new logical shift right.
7417 return BinaryOperator::createLShr(Op0, Op1);
7425 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
7426 if (Instruction *Result = commonIntCastTransforms(CI))
7429 Value *Src = CI.getOperand(0);
7430 const Type *Ty = CI.getType();
7431 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
7432 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
7434 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
7435 switch (SrcI->getOpcode()) {
7437 case Instruction::LShr:
7438 // We can shrink lshr to something smaller if we know the bits shifted in
7439 // are already zeros.
7440 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
7441 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
7443 // Get a mask for the bits shifting in.
7444 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
7445 Value* SrcIOp0 = SrcI->getOperand(0);
7446 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
7447 if (ShAmt >= DestBitWidth) // All zeros.
7448 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
7450 // Okay, we can shrink this. Truncate the input, then return a new
7452 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
7453 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
7455 return BinaryOperator::createLShr(V1, V2);
7457 } else { // This is a variable shr.
7459 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
7460 // more LLVM instructions, but allows '1 << Y' to be hoisted if
7461 // loop-invariant and CSE'd.
7462 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
7463 Value *One = ConstantInt::get(SrcI->getType(), 1);
7465 Value *V = InsertNewInstBefore(
7466 BinaryOperator::createShl(One, SrcI->getOperand(1),
7468 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
7469 SrcI->getOperand(0),
7471 Value *Zero = Constant::getNullValue(V->getType());
7472 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
7482 /// transformZExtICmp - Transform (zext icmp) to bitwise / integer operations
7483 /// in order to eliminate the icmp.
7484 Instruction *InstCombiner::transformZExtICmp(ICmpInst *ICI, Instruction &CI,
7486 // If we are just checking for a icmp eq of a single bit and zext'ing it
7487 // to an integer, then shift the bit to the appropriate place and then
7488 // cast to integer to avoid the comparison.
7489 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7490 const APInt &Op1CV = Op1C->getValue();
7492 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
7493 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
7494 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7495 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())) {
7496 if (!DoXform) return ICI;
7498 Value *In = ICI->getOperand(0);
7499 Value *Sh = ConstantInt::get(In->getType(),
7500 In->getType()->getPrimitiveSizeInBits()-1);
7501 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
7502 In->getName()+".lobit"),
7504 if (In->getType() != CI.getType())
7505 In = CastInst::createIntegerCast(In, CI.getType(),
7506 false/*ZExt*/, "tmp", &CI);
7508 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
7509 Constant *One = ConstantInt::get(In->getType(), 1);
7510 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
7511 In->getName()+".not"),
7515 return ReplaceInstUsesWith(CI, In);
7520 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
7521 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7522 // zext (X == 1) to i32 --> X iff X has only the low bit set.
7523 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
7524 // zext (X != 0) to i32 --> X iff X has only the low bit set.
7525 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
7526 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
7527 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
7528 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
7529 // This only works for EQ and NE
7530 ICI->isEquality()) {
7531 // If Op1C some other power of two, convert:
7532 uint32_t BitWidth = Op1C->getType()->getBitWidth();
7533 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
7534 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
7535 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
7537 APInt KnownZeroMask(~KnownZero);
7538 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
7539 if (!DoXform) return ICI;
7541 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
7542 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
7543 // (X&4) == 2 --> false
7544 // (X&4) != 2 --> true
7545 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
7546 Res = ConstantExpr::getZExt(Res, CI.getType());
7547 return ReplaceInstUsesWith(CI, Res);
7550 uint32_t ShiftAmt = KnownZeroMask.logBase2();
7551 Value *In = ICI->getOperand(0);
7553 // Perform a logical shr by shiftamt.
7554 // Insert the shift to put the result in the low bit.
7555 In = InsertNewInstBefore(BinaryOperator::createLShr(In,
7556 ConstantInt::get(In->getType(), ShiftAmt),
7557 In->getName()+".lobit"), CI);
7560 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
7561 Constant *One = ConstantInt::get(In->getType(), 1);
7562 In = BinaryOperator::createXor(In, One, "tmp");
7563 InsertNewInstBefore(cast<Instruction>(In), CI);
7566 if (CI.getType() == In->getType())
7567 return ReplaceInstUsesWith(CI, In);
7569 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
7577 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
7578 // If one of the common conversion will work ..
7579 if (Instruction *Result = commonIntCastTransforms(CI))
7582 Value *Src = CI.getOperand(0);
7584 // If this is a cast of a cast
7585 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
7586 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
7587 // types and if the sizes are just right we can convert this into a logical
7588 // 'and' which will be much cheaper than the pair of casts.
7589 if (isa<TruncInst>(CSrc)) {
7590 // Get the sizes of the types involved
7591 Value *A = CSrc->getOperand(0);
7592 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
7593 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
7594 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
7595 // If we're actually extending zero bits and the trunc is a no-op
7596 if (MidSize < DstSize && SrcSize == DstSize) {
7597 // Replace both of the casts with an And of the type mask.
7598 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
7599 Constant *AndConst = ConstantInt::get(AndValue);
7601 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
7602 // Unfortunately, if the type changed, we need to cast it back.
7603 if (And->getType() != CI.getType()) {
7604 And->setName(CSrc->getName()+".mask");
7605 InsertNewInstBefore(And, CI);
7606 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
7613 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src))
7614 return transformZExtICmp(ICI, CI);
7616 BinaryOperator *SrcI = dyn_cast<BinaryOperator>(Src);
7617 if (SrcI && SrcI->getOpcode() == Instruction::Or) {
7618 // zext (or icmp, icmp) --> or (zext icmp), (zext icmp) if at least one
7619 // of the (zext icmp) will be transformed.
7620 ICmpInst *LHS = dyn_cast<ICmpInst>(SrcI->getOperand(0));
7621 ICmpInst *RHS = dyn_cast<ICmpInst>(SrcI->getOperand(1));
7622 if (LHS && RHS && LHS->hasOneUse() && RHS->hasOneUse() &&
7623 (transformZExtICmp(LHS, CI, false) ||
7624 transformZExtICmp(RHS, CI, false))) {
7625 Value *LCast = InsertCastBefore(Instruction::ZExt, LHS, CI.getType(), CI);
7626 Value *RCast = InsertCastBefore(Instruction::ZExt, RHS, CI.getType(), CI);
7627 return BinaryOperator::create(Instruction::Or, LCast, RCast);
7634 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
7635 if (Instruction *I = commonIntCastTransforms(CI))
7638 Value *Src = CI.getOperand(0);
7640 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
7641 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
7642 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
7643 // If we are just checking for a icmp eq of a single bit and zext'ing it
7644 // to an integer, then shift the bit to the appropriate place and then
7645 // cast to integer to avoid the comparison.
7646 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
7647 const APInt &Op1CV = Op1C->getValue();
7649 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
7650 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
7651 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
7652 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
7653 Value *In = ICI->getOperand(0);
7654 Value *Sh = ConstantInt::get(In->getType(),
7655 In->getType()->getPrimitiveSizeInBits()-1);
7656 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
7657 In->getName()+".lobit"),
7659 if (In->getType() != CI.getType())
7660 In = CastInst::createIntegerCast(In, CI.getType(),
7661 true/*SExt*/, "tmp", &CI);
7663 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
7664 In = InsertNewInstBefore(BinaryOperator::createNot(In,
7665 In->getName()+".not"), CI);
7667 return ReplaceInstUsesWith(CI, In);
7675 /// FitsInFPType - Return a Constant* for the specified FP constant if it fits
7676 /// in the specified FP type without changing its value.
7677 static Constant *FitsInFPType(ConstantFP *CFP, const fltSemantics &Sem) {
7678 APFloat F = CFP->getValueAPF();
7679 if (F.convert(Sem, APFloat::rmNearestTiesToEven) == APFloat::opOK)
7680 return ConstantFP::get(F);
7684 /// LookThroughFPExtensions - If this is an fp extension instruction, look
7685 /// through it until we get the source value.
7686 static Value *LookThroughFPExtensions(Value *V) {
7687 if (Instruction *I = dyn_cast<Instruction>(V))
7688 if (I->getOpcode() == Instruction::FPExt)
7689 return LookThroughFPExtensions(I->getOperand(0));
7691 // If this value is a constant, return the constant in the smallest FP type
7692 // that can accurately represent it. This allows us to turn
7693 // (float)((double)X+2.0) into x+2.0f.
7694 if (ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
7695 if (CFP->getType() == Type::PPC_FP128Ty)
7696 return V; // No constant folding of this.
7697 // See if the value can be truncated to float and then reextended.
7698 if (Value *V = FitsInFPType(CFP, APFloat::IEEEsingle))
7700 if (CFP->getType() == Type::DoubleTy)
7701 return V; // Won't shrink.
7702 if (Value *V = FitsInFPType(CFP, APFloat::IEEEdouble))
7704 // Don't try to shrink to various long double types.
7710 Instruction *InstCombiner::visitFPTrunc(FPTruncInst &CI) {
7711 if (Instruction *I = commonCastTransforms(CI))
7714 // If we have fptrunc(add (fpextend x), (fpextend y)), where x and y are
7715 // smaller than the destination type, we can eliminate the truncate by doing
7716 // the add as the smaller type. This applies to add/sub/mul/div as well as
7717 // many builtins (sqrt, etc).
7718 BinaryOperator *OpI = dyn_cast<BinaryOperator>(CI.getOperand(0));
7719 if (OpI && OpI->hasOneUse()) {
7720 switch (OpI->getOpcode()) {
7722 case Instruction::Add:
7723 case Instruction::Sub:
7724 case Instruction::Mul:
7725 case Instruction::FDiv:
7726 case Instruction::FRem:
7727 const Type *SrcTy = OpI->getType();
7728 Value *LHSTrunc = LookThroughFPExtensions(OpI->getOperand(0));
7729 Value *RHSTrunc = LookThroughFPExtensions(OpI->getOperand(1));
7730 if (LHSTrunc->getType() != SrcTy &&
7731 RHSTrunc->getType() != SrcTy) {
7732 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
7733 // If the source types were both smaller than the destination type of
7734 // the cast, do this xform.
7735 if (LHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize &&
7736 RHSTrunc->getType()->getPrimitiveSizeInBits() <= DstSize) {
7737 LHSTrunc = InsertCastBefore(Instruction::FPExt, LHSTrunc,
7739 RHSTrunc = InsertCastBefore(Instruction::FPExt, RHSTrunc,
7741 return BinaryOperator::create(OpI->getOpcode(), LHSTrunc, RHSTrunc);
7750 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7751 return commonCastTransforms(CI);
7754 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7755 return commonCastTransforms(CI);
7758 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7759 return commonCastTransforms(CI);
7762 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7763 return commonCastTransforms(CI);
7766 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7767 return commonCastTransforms(CI);
7770 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7771 return commonPointerCastTransforms(CI);
7774 Instruction *InstCombiner::visitIntToPtr(IntToPtrInst &CI) {
7775 if (Instruction *I = commonCastTransforms(CI))
7778 const Type *DestPointee = cast<PointerType>(CI.getType())->getElementType();
7779 if (!DestPointee->isSized()) return 0;
7781 // If this is inttoptr(add (ptrtoint x), cst), try to turn this into a GEP.
7784 if (match(CI.getOperand(0), m_Add(m_Cast<PtrToIntInst>(m_Value(X)),
7785 m_ConstantInt(Cst)))) {
7786 // If the source and destination operands have the same type, see if this
7787 // is a single-index GEP.
7788 if (X->getType() == CI.getType()) {
7789 // Get the size of the pointee type.
7790 uint64_t Size = TD->getABITypeSize(DestPointee);
7792 // Convert the constant to intptr type.
7793 APInt Offset = Cst->getValue();
7794 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7796 // If Offset is evenly divisible by Size, we can do this xform.
7797 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7798 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7799 return GetElementPtrInst::Create(X, ConstantInt::get(Offset));
7802 // TODO: Could handle other cases, e.g. where add is indexing into field of
7804 } else if (CI.getOperand(0)->hasOneUse() &&
7805 match(CI.getOperand(0), m_Add(m_Value(X), m_ConstantInt(Cst)))) {
7806 // Otherwise, if this is inttoptr(add x, cst), try to turn this into an
7807 // "inttoptr+GEP" instead of "add+intptr".
7809 // Get the size of the pointee type.
7810 uint64_t Size = TD->getABITypeSize(DestPointee);
7812 // Convert the constant to intptr type.
7813 APInt Offset = Cst->getValue();
7814 Offset.sextOrTrunc(TD->getPointerSizeInBits());
7816 // If Offset is evenly divisible by Size, we can do this xform.
7817 if (Size && !APIntOps::srem(Offset, APInt(Offset.getBitWidth(), Size))){
7818 Offset = APIntOps::sdiv(Offset, APInt(Offset.getBitWidth(), Size));
7820 Instruction *P = InsertNewInstBefore(new IntToPtrInst(X, CI.getType(),
7822 return GetElementPtrInst::Create(P, ConstantInt::get(Offset), "tmp");
7828 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7829 // If the operands are integer typed then apply the integer transforms,
7830 // otherwise just apply the common ones.
7831 Value *Src = CI.getOperand(0);
7832 const Type *SrcTy = Src->getType();
7833 const Type *DestTy = CI.getType();
7835 if (SrcTy->isInteger() && DestTy->isInteger()) {
7836 if (Instruction *Result = commonIntCastTransforms(CI))
7838 } else if (isa<PointerType>(SrcTy)) {
7839 if (Instruction *I = commonPointerCastTransforms(CI))
7842 if (Instruction *Result = commonCastTransforms(CI))
7847 // Get rid of casts from one type to the same type. These are useless and can
7848 // be replaced by the operand.
7849 if (DestTy == Src->getType())
7850 return ReplaceInstUsesWith(CI, Src);
7852 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7853 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7854 const Type *DstElTy = DstPTy->getElementType();
7855 const Type *SrcElTy = SrcPTy->getElementType();
7857 // If the address spaces don't match, don't eliminate the bitcast, which is
7858 // required for changing types.
7859 if (SrcPTy->getAddressSpace() != DstPTy->getAddressSpace())
7862 // If we are casting a malloc or alloca to a pointer to a type of the same
7863 // size, rewrite the allocation instruction to allocate the "right" type.
7864 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7865 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7868 // If the source and destination are pointers, and this cast is equivalent
7869 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7870 // This can enhance SROA and other transforms that want type-safe pointers.
7871 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7872 unsigned NumZeros = 0;
7873 while (SrcElTy != DstElTy &&
7874 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7875 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7876 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7880 // If we found a path from the src to dest, create the getelementptr now.
7881 if (SrcElTy == DstElTy) {
7882 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7883 return GetElementPtrInst::Create(Src, Idxs.begin(), Idxs.end(), "",
7884 ((Instruction*) NULL));
7888 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7889 if (SVI->hasOneUse()) {
7890 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7891 // a bitconvert to a vector with the same # elts.
7892 if (isa<VectorType>(DestTy) &&
7893 cast<VectorType>(DestTy)->getNumElements() ==
7894 SVI->getType()->getNumElements()) {
7896 // If either of the operands is a cast from CI.getType(), then
7897 // evaluating the shuffle in the casted destination's type will allow
7898 // us to eliminate at least one cast.
7899 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7900 Tmp->getOperand(0)->getType() == DestTy) ||
7901 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7902 Tmp->getOperand(0)->getType() == DestTy)) {
7903 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7904 SVI->getOperand(0), DestTy, &CI);
7905 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7906 SVI->getOperand(1), DestTy, &CI);
7907 // Return a new shuffle vector. Use the same element ID's, as we
7908 // know the vector types match #elts.
7909 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7917 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7919 /// %D = select %cond, %C, %A
7921 /// %C = select %cond, %B, 0
7924 /// Assuming that the specified instruction is an operand to the select, return
7925 /// a bitmask indicating which operands of this instruction are foldable if they
7926 /// equal the other incoming value of the select.
7928 static unsigned GetSelectFoldableOperands(Instruction *I) {
7929 switch (I->getOpcode()) {
7930 case Instruction::Add:
7931 case Instruction::Mul:
7932 case Instruction::And:
7933 case Instruction::Or:
7934 case Instruction::Xor:
7935 return 3; // Can fold through either operand.
7936 case Instruction::Sub: // Can only fold on the amount subtracted.
7937 case Instruction::Shl: // Can only fold on the shift amount.
7938 case Instruction::LShr:
7939 case Instruction::AShr:
7942 return 0; // Cannot fold
7946 /// GetSelectFoldableConstant - For the same transformation as the previous
7947 /// function, return the identity constant that goes into the select.
7948 static Constant *GetSelectFoldableConstant(Instruction *I) {
7949 switch (I->getOpcode()) {
7950 default: assert(0 && "This cannot happen!"); abort();
7951 case Instruction::Add:
7952 case Instruction::Sub:
7953 case Instruction::Or:
7954 case Instruction::Xor:
7955 case Instruction::Shl:
7956 case Instruction::LShr:
7957 case Instruction::AShr:
7958 return Constant::getNullValue(I->getType());
7959 case Instruction::And:
7960 return Constant::getAllOnesValue(I->getType());
7961 case Instruction::Mul:
7962 return ConstantInt::get(I->getType(), 1);
7966 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7967 /// have the same opcode and only one use each. Try to simplify this.
7968 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7970 if (TI->getNumOperands() == 1) {
7971 // If this is a non-volatile load or a cast from the same type,
7974 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7977 return 0; // unknown unary op.
7980 // Fold this by inserting a select from the input values.
7981 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), TI->getOperand(0),
7982 FI->getOperand(0), SI.getName()+".v");
7983 InsertNewInstBefore(NewSI, SI);
7984 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7988 // Only handle binary operators here.
7989 if (!isa<BinaryOperator>(TI))
7992 // Figure out if the operations have any operands in common.
7993 Value *MatchOp, *OtherOpT, *OtherOpF;
7995 if (TI->getOperand(0) == FI->getOperand(0)) {
7996 MatchOp = TI->getOperand(0);
7997 OtherOpT = TI->getOperand(1);
7998 OtherOpF = FI->getOperand(1);
7999 MatchIsOpZero = true;
8000 } else if (TI->getOperand(1) == FI->getOperand(1)) {
8001 MatchOp = TI->getOperand(1);
8002 OtherOpT = TI->getOperand(0);
8003 OtherOpF = FI->getOperand(0);
8004 MatchIsOpZero = false;
8005 } else if (!TI->isCommutative()) {
8007 } else if (TI->getOperand(0) == FI->getOperand(1)) {
8008 MatchOp = TI->getOperand(0);
8009 OtherOpT = TI->getOperand(1);
8010 OtherOpF = FI->getOperand(0);
8011 MatchIsOpZero = true;
8012 } else if (TI->getOperand(1) == FI->getOperand(0)) {
8013 MatchOp = TI->getOperand(1);
8014 OtherOpT = TI->getOperand(0);
8015 OtherOpF = FI->getOperand(1);
8016 MatchIsOpZero = true;
8021 // If we reach here, they do have operations in common.
8022 SelectInst *NewSI = SelectInst::Create(SI.getCondition(), OtherOpT,
8023 OtherOpF, SI.getName()+".v");
8024 InsertNewInstBefore(NewSI, SI);
8026 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
8028 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
8030 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
8032 assert(0 && "Shouldn't get here");
8036 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
8037 Value *CondVal = SI.getCondition();
8038 Value *TrueVal = SI.getTrueValue();
8039 Value *FalseVal = SI.getFalseValue();
8041 // select true, X, Y -> X
8042 // select false, X, Y -> Y
8043 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
8044 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
8046 // select C, X, X -> X
8047 if (TrueVal == FalseVal)
8048 return ReplaceInstUsesWith(SI, TrueVal);
8050 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
8051 return ReplaceInstUsesWith(SI, FalseVal);
8052 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
8053 return ReplaceInstUsesWith(SI, TrueVal);
8054 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
8055 if (isa<Constant>(TrueVal))
8056 return ReplaceInstUsesWith(SI, TrueVal);
8058 return ReplaceInstUsesWith(SI, FalseVal);
8061 if (SI.getType() == Type::Int1Ty) {
8062 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
8063 if (C->getZExtValue()) {
8064 // Change: A = select B, true, C --> A = or B, C
8065 return BinaryOperator::createOr(CondVal, FalseVal);
8067 // Change: A = select B, false, C --> A = and !B, C
8069 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
8070 "not."+CondVal->getName()), SI);
8071 return BinaryOperator::createAnd(NotCond, FalseVal);
8073 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
8074 if (C->getZExtValue() == false) {
8075 // Change: A = select B, C, false --> A = and B, C
8076 return BinaryOperator::createAnd(CondVal, TrueVal);
8078 // Change: A = select B, C, true --> A = or !B, C
8080 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
8081 "not."+CondVal->getName()), SI);
8082 return BinaryOperator::createOr(NotCond, TrueVal);
8086 // select a, b, a -> a&b
8087 // select a, a, b -> a|b
8088 if (CondVal == TrueVal)
8089 return BinaryOperator::createOr(CondVal, FalseVal);
8090 else if (CondVal == FalseVal)
8091 return BinaryOperator::createAnd(CondVal, TrueVal);
8094 // Selecting between two integer constants?
8095 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
8096 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
8097 // select C, 1, 0 -> zext C to int
8098 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
8099 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
8100 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
8101 // select C, 0, 1 -> zext !C to int
8103 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
8104 "not."+CondVal->getName()), SI);
8105 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
8108 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
8110 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
8112 // (x <s 0) ? -1 : 0 -> ashr x, 31
8113 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
8114 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
8115 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
8116 // The comparison constant and the result are not neccessarily the
8117 // same width. Make an all-ones value by inserting a AShr.
8118 Value *X = IC->getOperand(0);
8119 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
8120 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
8121 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
8123 InsertNewInstBefore(SRA, SI);
8125 // Finally, convert to the type of the select RHS. We figure out
8126 // if this requires a SExt, Trunc or BitCast based on the sizes.
8127 Instruction::CastOps opc = Instruction::BitCast;
8128 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
8129 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
8130 if (SRASize < SISize)
8131 opc = Instruction::SExt;
8132 else if (SRASize > SISize)
8133 opc = Instruction::Trunc;
8134 return CastInst::create(opc, SRA, SI.getType());
8139 // If one of the constants is zero (we know they can't both be) and we
8140 // have an icmp instruction with zero, and we have an 'and' with the
8141 // non-constant value, eliminate this whole mess. This corresponds to
8142 // cases like this: ((X & 27) ? 27 : 0)
8143 if (TrueValC->isZero() || FalseValC->isZero())
8144 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
8145 cast<Constant>(IC->getOperand(1))->isNullValue())
8146 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
8147 if (ICA->getOpcode() == Instruction::And &&
8148 isa<ConstantInt>(ICA->getOperand(1)) &&
8149 (ICA->getOperand(1) == TrueValC ||
8150 ICA->getOperand(1) == FalseValC) &&
8151 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
8152 // Okay, now we know that everything is set up, we just don't
8153 // know whether we have a icmp_ne or icmp_eq and whether the
8154 // true or false val is the zero.
8155 bool ShouldNotVal = !TrueValC->isZero();
8156 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
8159 V = InsertNewInstBefore(BinaryOperator::create(
8160 Instruction::Xor, V, ICA->getOperand(1)), SI);
8161 return ReplaceInstUsesWith(SI, V);
8166 // See if we are selecting two values based on a comparison of the two values.
8167 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
8168 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
8169 // Transform (X == Y) ? X : Y -> Y
8170 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8171 // This is not safe in general for floating point:
8172 // consider X== -0, Y== +0.
8173 // It becomes safe if either operand is a nonzero constant.
8174 ConstantFP *CFPt, *CFPf;
8175 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8176 !CFPt->getValueAPF().isZero()) ||
8177 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8178 !CFPf->getValueAPF().isZero()))
8179 return ReplaceInstUsesWith(SI, FalseVal);
8181 // Transform (X != Y) ? X : Y -> X
8182 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8183 return ReplaceInstUsesWith(SI, TrueVal);
8184 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8186 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
8187 // Transform (X == Y) ? Y : X -> X
8188 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ) {
8189 // This is not safe in general for floating point:
8190 // consider X== -0, Y== +0.
8191 // It becomes safe if either operand is a nonzero constant.
8192 ConstantFP *CFPt, *CFPf;
8193 if (((CFPt = dyn_cast<ConstantFP>(TrueVal)) &&
8194 !CFPt->getValueAPF().isZero()) ||
8195 ((CFPf = dyn_cast<ConstantFP>(FalseVal)) &&
8196 !CFPf->getValueAPF().isZero()))
8197 return ReplaceInstUsesWith(SI, FalseVal);
8199 // Transform (X != Y) ? Y : X -> Y
8200 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
8201 return ReplaceInstUsesWith(SI, TrueVal);
8202 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8206 // See if we are selecting two values based on a comparison of the two values.
8207 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
8208 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
8209 // Transform (X == Y) ? X : Y -> Y
8210 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8211 return ReplaceInstUsesWith(SI, FalseVal);
8212 // Transform (X != Y) ? X : Y -> X
8213 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8214 return ReplaceInstUsesWith(SI, TrueVal);
8215 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8217 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
8218 // Transform (X == Y) ? Y : X -> X
8219 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
8220 return ReplaceInstUsesWith(SI, FalseVal);
8221 // Transform (X != Y) ? Y : X -> Y
8222 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
8223 return ReplaceInstUsesWith(SI, TrueVal);
8224 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
8228 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
8229 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
8230 if (TI->hasOneUse() && FI->hasOneUse()) {
8231 Instruction *AddOp = 0, *SubOp = 0;
8233 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
8234 if (TI->getOpcode() == FI->getOpcode())
8235 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
8238 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
8239 // even legal for FP.
8240 if (TI->getOpcode() == Instruction::Sub &&
8241 FI->getOpcode() == Instruction::Add) {
8242 AddOp = FI; SubOp = TI;
8243 } else if (FI->getOpcode() == Instruction::Sub &&
8244 TI->getOpcode() == Instruction::Add) {
8245 AddOp = TI; SubOp = FI;
8249 Value *OtherAddOp = 0;
8250 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
8251 OtherAddOp = AddOp->getOperand(1);
8252 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
8253 OtherAddOp = AddOp->getOperand(0);
8257 // So at this point we know we have (Y -> OtherAddOp):
8258 // select C, (add X, Y), (sub X, Z)
8259 Value *NegVal; // Compute -Z
8260 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
8261 NegVal = ConstantExpr::getNeg(C);
8263 NegVal = InsertNewInstBefore(
8264 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
8267 Value *NewTrueOp = OtherAddOp;
8268 Value *NewFalseOp = NegVal;
8270 std::swap(NewTrueOp, NewFalseOp);
8271 Instruction *NewSel =
8272 SelectInst::Create(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
8274 NewSel = InsertNewInstBefore(NewSel, SI);
8275 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
8280 // See if we can fold the select into one of our operands.
8281 if (SI.getType()->isInteger()) {
8282 // See the comment above GetSelectFoldableOperands for a description of the
8283 // transformation we are doing here.
8284 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
8285 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
8286 !isa<Constant>(FalseVal))
8287 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
8288 unsigned OpToFold = 0;
8289 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
8291 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
8296 Constant *C = GetSelectFoldableConstant(TVI);
8297 Instruction *NewSel =
8298 SelectInst::Create(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
8299 InsertNewInstBefore(NewSel, SI);
8300 NewSel->takeName(TVI);
8301 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
8302 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
8304 assert(0 && "Unknown instruction!!");
8309 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
8310 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
8311 !isa<Constant>(TrueVal))
8312 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
8313 unsigned OpToFold = 0;
8314 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
8316 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
8321 Constant *C = GetSelectFoldableConstant(FVI);
8322 Instruction *NewSel =
8323 SelectInst::Create(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
8324 InsertNewInstBefore(NewSel, SI);
8325 NewSel->takeName(FVI);
8326 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
8327 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
8329 assert(0 && "Unknown instruction!!");
8334 if (BinaryOperator::isNot(CondVal)) {
8335 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
8336 SI.setOperand(1, FalseVal);
8337 SI.setOperand(2, TrueVal);
8344 /// EnforceKnownAlignment - If the specified pointer points to an object that
8345 /// we control, modify the object's alignment to PrefAlign. This isn't
8346 /// often possible though. If alignment is important, a more reliable approach
8347 /// is to simply align all global variables and allocation instructions to
8348 /// their preferred alignment from the beginning.
8350 static unsigned EnforceKnownAlignment(Value *V,
8351 unsigned Align, unsigned PrefAlign) {
8353 User *U = dyn_cast<User>(V);
8354 if (!U) return Align;
8356 switch (getOpcode(U)) {
8358 case Instruction::BitCast:
8359 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8360 case Instruction::GetElementPtr: {
8361 // If all indexes are zero, it is just the alignment of the base pointer.
8362 bool AllZeroOperands = true;
8363 for (unsigned i = 1, e = U->getNumOperands(); i != e; ++i)
8364 if (!isa<Constant>(U->getOperand(i)) ||
8365 !cast<Constant>(U->getOperand(i))->isNullValue()) {
8366 AllZeroOperands = false;
8370 if (AllZeroOperands) {
8371 // Treat this like a bitcast.
8372 return EnforceKnownAlignment(U->getOperand(0), Align, PrefAlign);
8378 if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
8379 // If there is a large requested alignment and we can, bump up the alignment
8381 if (!GV->isDeclaration()) {
8382 GV->setAlignment(PrefAlign);
8385 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
8386 // If there is a requested alignment and if this is an alloca, round up. We
8387 // don't do this for malloc, because some systems can't respect the request.
8388 if (isa<AllocaInst>(AI)) {
8389 AI->setAlignment(PrefAlign);
8397 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
8398 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
8399 /// and it is more than the alignment of the ultimate object, see if we can
8400 /// increase the alignment of the ultimate object, making this check succeed.
8401 unsigned InstCombiner::GetOrEnforceKnownAlignment(Value *V,
8402 unsigned PrefAlign) {
8403 unsigned BitWidth = TD ? TD->getTypeSizeInBits(V->getType()) :
8404 sizeof(PrefAlign) * CHAR_BIT;
8405 APInt Mask = APInt::getAllOnesValue(BitWidth);
8406 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
8407 ComputeMaskedBits(V, Mask, KnownZero, KnownOne);
8408 unsigned TrailZ = KnownZero.countTrailingOnes();
8409 unsigned Align = 1u << std::min(BitWidth - 1, TrailZ);
8411 if (PrefAlign > Align)
8412 Align = EnforceKnownAlignment(V, Align, PrefAlign);
8414 // We don't need to make any adjustment.
8418 Instruction *InstCombiner::SimplifyMemTransfer(MemIntrinsic *MI) {
8419 unsigned DstAlign = GetOrEnforceKnownAlignment(MI->getOperand(1));
8420 unsigned SrcAlign = GetOrEnforceKnownAlignment(MI->getOperand(2));
8421 unsigned MinAlign = std::min(DstAlign, SrcAlign);
8422 unsigned CopyAlign = MI->getAlignment()->getZExtValue();
8424 if (CopyAlign < MinAlign) {
8425 MI->setAlignment(ConstantInt::get(Type::Int32Ty, MinAlign));
8429 // If MemCpyInst length is 1/2/4/8 bytes then replace memcpy with
8431 ConstantInt *MemOpLength = dyn_cast<ConstantInt>(MI->getOperand(3));
8432 if (MemOpLength == 0) return 0;
8434 // Source and destination pointer types are always "i8*" for intrinsic. See
8435 // if the size is something we can handle with a single primitive load/store.
8436 // A single load+store correctly handles overlapping memory in the memmove
8438 unsigned Size = MemOpLength->getZExtValue();
8439 if (Size == 0 || Size > 8 || (Size&(Size-1)))
8440 return 0; // If not 1/2/4/8 bytes, exit.
8442 // Use an integer load+store unless we can find something better.
8443 Type *NewPtrTy = PointerType::getUnqual(IntegerType::get(Size<<3));
8445 // Memcpy forces the use of i8* for the source and destination. That means
8446 // that if you're using memcpy to move one double around, you'll get a cast
8447 // from double* to i8*. We'd much rather use a double load+store rather than
8448 // an i64 load+store, here because this improves the odds that the source or
8449 // dest address will be promotable. See if we can find a better type than the
8450 // integer datatype.
8451 if (Value *Op = getBitCastOperand(MI->getOperand(1))) {
8452 const Type *SrcETy = cast<PointerType>(Op->getType())->getElementType();
8453 if (SrcETy->isSized() && TD->getTypeStoreSize(SrcETy) == Size) {
8454 // The SrcETy might be something like {{{double}}} or [1 x double]. Rip
8455 // down through these levels if so.
8456 while (!SrcETy->isFirstClassType()) {
8457 if (const StructType *STy = dyn_cast<StructType>(SrcETy)) {
8458 if (STy->getNumElements() == 1)
8459 SrcETy = STy->getElementType(0);
8462 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(SrcETy)) {
8463 if (ATy->getNumElements() == 1)
8464 SrcETy = ATy->getElementType();
8471 if (SrcETy->isFirstClassType())
8472 NewPtrTy = PointerType::getUnqual(SrcETy);
8477 // If the memcpy/memmove provides better alignment info than we can
8479 SrcAlign = std::max(SrcAlign, CopyAlign);
8480 DstAlign = std::max(DstAlign, CopyAlign);
8482 Value *Src = InsertBitCastBefore(MI->getOperand(2), NewPtrTy, *MI);
8483 Value *Dest = InsertBitCastBefore(MI->getOperand(1), NewPtrTy, *MI);
8484 Instruction *L = new LoadInst(Src, "tmp", false, SrcAlign);
8485 InsertNewInstBefore(L, *MI);
8486 InsertNewInstBefore(new StoreInst(L, Dest, false, DstAlign), *MI);
8488 // Set the size of the copy to 0, it will be deleted on the next iteration.
8489 MI->setOperand(3, Constant::getNullValue(MemOpLength->getType()));
8493 /// visitCallInst - CallInst simplification. This mostly only handles folding
8494 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
8495 /// the heavy lifting.
8497 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
8498 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
8499 if (!II) return visitCallSite(&CI);
8501 // Intrinsics cannot occur in an invoke, so handle them here instead of in
8503 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
8504 bool Changed = false;
8506 // memmove/cpy/set of zero bytes is a noop.
8507 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
8508 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
8510 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
8511 if (CI->getZExtValue() == 1) {
8512 // Replace the instruction with just byte operations. We would
8513 // transform other cases to loads/stores, but we don't know if
8514 // alignment is sufficient.
8518 // If we have a memmove and the source operation is a constant global,
8519 // then the source and dest pointers can't alias, so we can change this
8520 // into a call to memcpy.
8521 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(MI)) {
8522 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
8523 if (GVSrc->isConstant()) {
8524 Module *M = CI.getParent()->getParent()->getParent();
8525 Intrinsic::ID MemCpyID;
8526 if (CI.getOperand(3)->getType() == Type::Int32Ty)
8527 MemCpyID = Intrinsic::memcpy_i32;
8529 MemCpyID = Intrinsic::memcpy_i64;
8530 CI.setOperand(0, Intrinsic::getDeclaration(M, MemCpyID));
8535 // If we can determine a pointer alignment that is bigger than currently
8536 // set, update the alignment.
8537 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
8538 if (Instruction *I = SimplifyMemTransfer(MI))
8540 } else if (isa<MemSetInst>(MI)) {
8541 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest());
8542 if (MI->getAlignment()->getZExtValue() < Alignment) {
8543 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
8548 if (Changed) return II;
8550 switch (II->getIntrinsicID()) {
8552 case Intrinsic::ppc_altivec_lvx:
8553 case Intrinsic::ppc_altivec_lvxl:
8554 case Intrinsic::x86_sse_loadu_ps:
8555 case Intrinsic::x86_sse2_loadu_pd:
8556 case Intrinsic::x86_sse2_loadu_dq:
8557 // Turn PPC lvx -> load if the pointer is known aligned.
8558 // Turn X86 loadups -> load if the pointer is known aligned.
8559 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8560 Value *Ptr = InsertBitCastBefore(II->getOperand(1),
8561 PointerType::getUnqual(II->getType()),
8563 return new LoadInst(Ptr);
8566 case Intrinsic::ppc_altivec_stvx:
8567 case Intrinsic::ppc_altivec_stvxl:
8568 // Turn stvx -> store if the pointer is known aligned.
8569 if (GetOrEnforceKnownAlignment(II->getOperand(2), 16) >= 16) {
8570 const Type *OpPtrTy =
8571 PointerType::getUnqual(II->getOperand(1)->getType());
8572 Value *Ptr = InsertBitCastBefore(II->getOperand(2), OpPtrTy, CI);
8573 return new StoreInst(II->getOperand(1), Ptr);
8576 case Intrinsic::x86_sse_storeu_ps:
8577 case Intrinsic::x86_sse2_storeu_pd:
8578 case Intrinsic::x86_sse2_storeu_dq:
8579 case Intrinsic::x86_sse2_storel_dq:
8580 // Turn X86 storeu -> store if the pointer is known aligned.
8581 if (GetOrEnforceKnownAlignment(II->getOperand(1), 16) >= 16) {
8582 const Type *OpPtrTy =
8583 PointerType::getUnqual(II->getOperand(2)->getType());
8584 Value *Ptr = InsertBitCastBefore(II->getOperand(1), OpPtrTy, CI);
8585 return new StoreInst(II->getOperand(2), Ptr);
8589 case Intrinsic::x86_sse_cvttss2si: {
8590 // These intrinsics only demands the 0th element of its input vector. If
8591 // we can simplify the input based on that, do so now.
8593 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
8595 II->setOperand(1, V);
8601 case Intrinsic::ppc_altivec_vperm:
8602 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
8603 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
8604 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
8606 // Check that all of the elements are integer constants or undefs.
8607 bool AllEltsOk = true;
8608 for (unsigned i = 0; i != 16; ++i) {
8609 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
8610 !isa<UndefValue>(Mask->getOperand(i))) {
8617 // Cast the input vectors to byte vectors.
8618 Value *Op0 =InsertBitCastBefore(II->getOperand(1),Mask->getType(),CI);
8619 Value *Op1 =InsertBitCastBefore(II->getOperand(2),Mask->getType(),CI);
8620 Value *Result = UndefValue::get(Op0->getType());
8622 // Only extract each element once.
8623 Value *ExtractedElts[32];
8624 memset(ExtractedElts, 0, sizeof(ExtractedElts));
8626 for (unsigned i = 0; i != 16; ++i) {
8627 if (isa<UndefValue>(Mask->getOperand(i)))
8629 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
8630 Idx &= 31; // Match the hardware behavior.
8632 if (ExtractedElts[Idx] == 0) {
8634 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
8635 InsertNewInstBefore(Elt, CI);
8636 ExtractedElts[Idx] = Elt;
8639 // Insert this value into the result vector.
8640 Result = InsertElementInst::Create(Result, ExtractedElts[Idx], i, "tmp");
8641 InsertNewInstBefore(cast<Instruction>(Result), CI);
8643 return CastInst::create(Instruction::BitCast, Result, CI.getType());
8648 case Intrinsic::stackrestore: {
8649 // If the save is right next to the restore, remove the restore. This can
8650 // happen when variable allocas are DCE'd.
8651 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
8652 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
8653 BasicBlock::iterator BI = SS;
8655 return EraseInstFromFunction(CI);
8659 // Scan down this block to see if there is another stack restore in the
8660 // same block without an intervening call/alloca.
8661 BasicBlock::iterator BI = II;
8662 TerminatorInst *TI = II->getParent()->getTerminator();
8663 bool CannotRemove = false;
8664 for (++BI; &*BI != TI; ++BI) {
8665 if (isa<AllocaInst>(BI)) {
8666 CannotRemove = true;
8669 if (isa<CallInst>(BI)) {
8670 if (!isa<IntrinsicInst>(BI)) {
8671 CannotRemove = true;
8674 // If there is a stackrestore below this one, remove this one.
8675 return EraseInstFromFunction(CI);
8679 // If the stack restore is in a return/unwind block and if there are no
8680 // allocas or calls between the restore and the return, nuke the restore.
8681 if (!CannotRemove && (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)))
8682 return EraseInstFromFunction(CI);
8688 return visitCallSite(II);
8691 // InvokeInst simplification
8693 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
8694 return visitCallSite(&II);
8697 // visitCallSite - Improvements for call and invoke instructions.
8699 Instruction *InstCombiner::visitCallSite(CallSite CS) {
8700 bool Changed = false;
8702 // If the callee is a constexpr cast of a function, attempt to move the cast
8703 // to the arguments of the call/invoke.
8704 if (transformConstExprCastCall(CS)) return 0;
8706 Value *Callee = CS.getCalledValue();
8708 if (Function *CalleeF = dyn_cast<Function>(Callee))
8709 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
8710 Instruction *OldCall = CS.getInstruction();
8711 // If the call and callee calling conventions don't match, this call must
8712 // be unreachable, as the call is undefined.
8713 new StoreInst(ConstantInt::getTrue(),
8714 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8716 if (!OldCall->use_empty())
8717 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
8718 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
8719 return EraseInstFromFunction(*OldCall);
8723 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
8724 // This instruction is not reachable, just remove it. We insert a store to
8725 // undef so that we know that this code is not reachable, despite the fact
8726 // that we can't modify the CFG here.
8727 new StoreInst(ConstantInt::getTrue(),
8728 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)),
8729 CS.getInstruction());
8731 if (!CS.getInstruction()->use_empty())
8732 CS.getInstruction()->
8733 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
8735 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
8736 // Don't break the CFG, insert a dummy cond branch.
8737 BranchInst::Create(II->getNormalDest(), II->getUnwindDest(),
8738 ConstantInt::getTrue(), II);
8740 return EraseInstFromFunction(*CS.getInstruction());
8743 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee))
8744 if (IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0)))
8745 if (In->getIntrinsicID() == Intrinsic::init_trampoline)
8746 return transformCallThroughTrampoline(CS);
8748 const PointerType *PTy = cast<PointerType>(Callee->getType());
8749 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
8750 if (FTy->isVarArg()) {
8751 // See if we can optimize any arguments passed through the varargs area of
8753 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
8754 E = CS.arg_end(); I != E; ++I)
8755 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
8756 // If this cast does not effect the value passed through the varargs
8757 // area, we can eliminate the use of the cast.
8758 Value *Op = CI->getOperand(0);
8759 if (CI->isLosslessCast()) {
8766 if (isa<InlineAsm>(Callee) && !CS.doesNotThrow()) {
8767 // Inline asm calls cannot throw - mark them 'nounwind'.
8768 CS.setDoesNotThrow();
8772 return Changed ? CS.getInstruction() : 0;
8775 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
8776 // attempt to move the cast to the arguments of the call/invoke.
8778 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
8779 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
8780 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
8781 if (CE->getOpcode() != Instruction::BitCast ||
8782 !isa<Function>(CE->getOperand(0)))
8784 Function *Callee = cast<Function>(CE->getOperand(0));
8785 Instruction *Caller = CS.getInstruction();
8786 const PAListPtr &CallerPAL = CS.getParamAttrs();
8788 // Okay, this is a cast from a function to a different type. Unless doing so
8789 // would cause a type conversion of one of our arguments, change this call to
8790 // be a direct call with arguments casted to the appropriate types.
8792 const FunctionType *FT = Callee->getFunctionType();
8793 const Type *OldRetTy = Caller->getType();
8795 if (isa<StructType>(FT->getReturnType()))
8796 return false; // TODO: Handle multiple return values.
8798 // Check to see if we are changing the return type...
8799 if (OldRetTy != FT->getReturnType()) {
8800 if (Callee->isDeclaration() && !Caller->use_empty() &&
8801 // Conversion is ok if changing from pointer to int of same size.
8802 !(isa<PointerType>(FT->getReturnType()) &&
8803 TD->getIntPtrType() == OldRetTy))
8804 return false; // Cannot transform this return value.
8806 if (!Caller->use_empty() &&
8807 // void -> non-void is handled specially
8808 FT->getReturnType() != Type::VoidTy &&
8809 !CastInst::isCastable(FT->getReturnType(), OldRetTy))
8810 return false; // Cannot transform this return value.
8812 if (!CallerPAL.isEmpty() && !Caller->use_empty()) {
8813 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8814 if (RAttrs & ParamAttr::typeIncompatible(FT->getReturnType()))
8815 return false; // Attribute not compatible with transformed value.
8818 // If the callsite is an invoke instruction, and the return value is used by
8819 // a PHI node in a successor, we cannot change the return type of the call
8820 // because there is no place to put the cast instruction (without breaking
8821 // the critical edge). Bail out in this case.
8822 if (!Caller->use_empty())
8823 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8824 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8826 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8827 if (PN->getParent() == II->getNormalDest() ||
8828 PN->getParent() == II->getUnwindDest())
8832 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8833 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8835 CallSite::arg_iterator AI = CS.arg_begin();
8836 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8837 const Type *ParamTy = FT->getParamType(i);
8838 const Type *ActTy = (*AI)->getType();
8840 if (!CastInst::isCastable(ActTy, ParamTy))
8841 return false; // Cannot transform this parameter value.
8843 if (CallerPAL.getParamAttrs(i + 1) & ParamAttr::typeIncompatible(ParamTy))
8844 return false; // Attribute not compatible with transformed value.
8846 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8847 // Some conversions are safe even if we do not have a body.
8848 // Either we can cast directly, or we can upconvert the argument
8849 bool isConvertible = ActTy == ParamTy ||
8850 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8851 (ParamTy->isInteger() && ActTy->isInteger() &&
8852 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8853 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8854 && c->getValue().isStrictlyPositive());
8855 if (Callee->isDeclaration() && !isConvertible) return false;
8858 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8859 Callee->isDeclaration())
8860 return false; // Do not delete arguments unless we have a function body.
8862 if (FT->getNumParams() < NumActualArgs && FT->isVarArg() &&
8863 !CallerPAL.isEmpty())
8864 // In this case we have more arguments than the new function type, but we
8865 // won't be dropping them. Check that these extra arguments have attributes
8866 // that are compatible with being a vararg call argument.
8867 for (unsigned i = CallerPAL.getNumSlots(); i; --i) {
8868 if (CallerPAL.getSlot(i - 1).Index <= FT->getNumParams())
8870 ParameterAttributes PAttrs = CallerPAL.getSlot(i - 1).Attrs;
8871 if (PAttrs & ParamAttr::VarArgsIncompatible)
8875 // Okay, we decided that this is a safe thing to do: go ahead and start
8876 // inserting cast instructions as necessary...
8877 std::vector<Value*> Args;
8878 Args.reserve(NumActualArgs);
8879 SmallVector<ParamAttrsWithIndex, 8> attrVec;
8880 attrVec.reserve(NumCommonArgs);
8882 // Get any return attributes.
8883 ParameterAttributes RAttrs = CallerPAL.getParamAttrs(0);
8885 // If the return value is not being used, the type may not be compatible
8886 // with the existing attributes. Wipe out any problematic attributes.
8887 RAttrs &= ~ParamAttr::typeIncompatible(FT->getReturnType());
8889 // Add the new return attributes.
8891 attrVec.push_back(ParamAttrsWithIndex::get(0, RAttrs));
8893 AI = CS.arg_begin();
8894 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8895 const Type *ParamTy = FT->getParamType(i);
8896 if ((*AI)->getType() == ParamTy) {
8897 Args.push_back(*AI);
8899 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8900 false, ParamTy, false);
8901 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8902 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8905 // Add any parameter attributes.
8906 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8907 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8910 // If the function takes more arguments than the call was taking, add them
8912 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8913 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8915 // If we are removing arguments to the function, emit an obnoxious warning...
8916 if (FT->getNumParams() < NumActualArgs) {
8917 if (!FT->isVarArg()) {
8918 cerr << "WARNING: While resolving call to function '"
8919 << Callee->getName() << "' arguments were dropped!\n";
8921 // Add all of the arguments in their promoted form to the arg list...
8922 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8923 const Type *PTy = getPromotedType((*AI)->getType());
8924 if (PTy != (*AI)->getType()) {
8925 // Must promote to pass through va_arg area!
8926 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8928 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8929 InsertNewInstBefore(Cast, *Caller);
8930 Args.push_back(Cast);
8932 Args.push_back(*AI);
8935 // Add any parameter attributes.
8936 if (ParameterAttributes PAttrs = CallerPAL.getParamAttrs(i + 1))
8937 attrVec.push_back(ParamAttrsWithIndex::get(i + 1, PAttrs));
8942 if (FT->getReturnType() == Type::VoidTy)
8943 Caller->setName(""); // Void type should not have a name.
8945 const PAListPtr &NewCallerPAL = PAListPtr::get(attrVec.begin(),attrVec.end());
8948 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8949 NC = InvokeInst::Create(Callee, II->getNormalDest(), II->getUnwindDest(),
8950 Args.begin(), Args.end(), Caller->getName(), Caller);
8951 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8952 cast<InvokeInst>(NC)->setParamAttrs(NewCallerPAL);
8954 NC = CallInst::Create(Callee, Args.begin(), Args.end(),
8955 Caller->getName(), Caller);
8956 CallInst *CI = cast<CallInst>(Caller);
8957 if (CI->isTailCall())
8958 cast<CallInst>(NC)->setTailCall();
8959 cast<CallInst>(NC)->setCallingConv(CI->getCallingConv());
8960 cast<CallInst>(NC)->setParamAttrs(NewCallerPAL);
8963 // Insert a cast of the return type as necessary.
8965 if (OldRetTy != NV->getType() && !Caller->use_empty()) {
8966 if (NV->getType() != Type::VoidTy) {
8967 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8969 NV = NC = CastInst::create(opcode, NC, OldRetTy, "tmp");
8971 // If this is an invoke instruction, we should insert it after the first
8972 // non-phi, instruction in the normal successor block.
8973 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8974 BasicBlock::iterator I = II->getNormalDest()->begin();
8975 while (isa<PHINode>(I)) ++I;
8976 InsertNewInstBefore(NC, *I);
8978 // Otherwise, it's a call, just insert cast right after the call instr
8979 InsertNewInstBefore(NC, *Caller);
8981 AddUsersToWorkList(*Caller);
8983 NV = UndefValue::get(Caller->getType());
8987 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8988 Caller->replaceAllUsesWith(NV);
8989 Caller->eraseFromParent();
8990 RemoveFromWorkList(Caller);
8994 // transformCallThroughTrampoline - Turn a call to a function created by the
8995 // init_trampoline intrinsic into a direct call to the underlying function.
8997 Instruction *InstCombiner::transformCallThroughTrampoline(CallSite CS) {
8998 Value *Callee = CS.getCalledValue();
8999 const PointerType *PTy = cast<PointerType>(Callee->getType());
9000 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
9001 const PAListPtr &Attrs = CS.getParamAttrs();
9003 // If the call already has the 'nest' attribute somewhere then give up -
9004 // otherwise 'nest' would occur twice after splicing in the chain.
9005 if (Attrs.hasAttrSomewhere(ParamAttr::Nest))
9008 IntrinsicInst *Tramp =
9009 cast<IntrinsicInst>(cast<BitCastInst>(Callee)->getOperand(0));
9012 cast<Function>(IntrinsicInst::StripPointerCasts(Tramp->getOperand(2)));
9013 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
9014 const FunctionType *NestFTy = cast<FunctionType>(NestFPTy->getElementType());
9016 const PAListPtr &NestAttrs = NestF->getParamAttrs();
9017 if (!NestAttrs.isEmpty()) {
9018 unsigned NestIdx = 1;
9019 const Type *NestTy = 0;
9020 ParameterAttributes NestAttr = ParamAttr::None;
9022 // Look for a parameter marked with the 'nest' attribute.
9023 for (FunctionType::param_iterator I = NestFTy->param_begin(),
9024 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
9025 if (NestAttrs.paramHasAttr(NestIdx, ParamAttr::Nest)) {
9026 // Record the parameter type and any other attributes.
9028 NestAttr = NestAttrs.getParamAttrs(NestIdx);
9033 Instruction *Caller = CS.getInstruction();
9034 std::vector<Value*> NewArgs;
9035 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
9037 SmallVector<ParamAttrsWithIndex, 8> NewAttrs;
9038 NewAttrs.reserve(Attrs.getNumSlots() + 1);
9040 // Insert the nest argument into the call argument list, which may
9041 // mean appending it. Likewise for attributes.
9043 // Add any function result attributes.
9044 if (ParameterAttributes Attr = Attrs.getParamAttrs(0))
9045 NewAttrs.push_back(ParamAttrsWithIndex::get(0, Attr));
9049 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
9051 if (Idx == NestIdx) {
9052 // Add the chain argument and attributes.
9053 Value *NestVal = Tramp->getOperand(3);
9054 if (NestVal->getType() != NestTy)
9055 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
9056 NewArgs.push_back(NestVal);
9057 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
9063 // Add the original argument and attributes.
9064 NewArgs.push_back(*I);
9065 if (ParameterAttributes Attr = Attrs.getParamAttrs(Idx))
9067 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
9073 // The trampoline may have been bitcast to a bogus type (FTy).
9074 // Handle this by synthesizing a new function type, equal to FTy
9075 // with the chain parameter inserted.
9077 std::vector<const Type*> NewTypes;
9078 NewTypes.reserve(FTy->getNumParams()+1);
9080 // Insert the chain's type into the list of parameter types, which may
9081 // mean appending it.
9084 FunctionType::param_iterator I = FTy->param_begin(),
9085 E = FTy->param_end();
9089 // Add the chain's type.
9090 NewTypes.push_back(NestTy);
9095 // Add the original type.
9096 NewTypes.push_back(*I);
9102 // Replace the trampoline call with a direct call. Let the generic
9103 // code sort out any function type mismatches.
9104 FunctionType *NewFTy =
9105 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg());
9106 Constant *NewCallee = NestF->getType() == PointerType::getUnqual(NewFTy) ?
9107 NestF : ConstantExpr::getBitCast(NestF, PointerType::getUnqual(NewFTy));
9108 const PAListPtr &NewPAL = PAListPtr::get(NewAttrs.begin(),NewAttrs.end());
9110 Instruction *NewCaller;
9111 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
9112 NewCaller = InvokeInst::Create(NewCallee,
9113 II->getNormalDest(), II->getUnwindDest(),
9114 NewArgs.begin(), NewArgs.end(),
9115 Caller->getName(), Caller);
9116 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
9117 cast<InvokeInst>(NewCaller)->setParamAttrs(NewPAL);
9119 NewCaller = CallInst::Create(NewCallee, NewArgs.begin(), NewArgs.end(),
9120 Caller->getName(), Caller);
9121 if (cast<CallInst>(Caller)->isTailCall())
9122 cast<CallInst>(NewCaller)->setTailCall();
9123 cast<CallInst>(NewCaller)->
9124 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
9125 cast<CallInst>(NewCaller)->setParamAttrs(NewPAL);
9127 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
9128 Caller->replaceAllUsesWith(NewCaller);
9129 Caller->eraseFromParent();
9130 RemoveFromWorkList(Caller);
9135 // Replace the trampoline call with a direct call. Since there is no 'nest'
9136 // parameter, there is no need to adjust the argument list. Let the generic
9137 // code sort out any function type mismatches.
9138 Constant *NewCallee =
9139 NestF->getType() == PTy ? NestF : ConstantExpr::getBitCast(NestF, PTy);
9140 CS.setCalledFunction(NewCallee);
9141 return CS.getInstruction();
9144 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
9145 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
9146 /// and a single binop.
9147 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
9148 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9149 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
9150 isa<CmpInst>(FirstInst));
9151 unsigned Opc = FirstInst->getOpcode();
9152 Value *LHSVal = FirstInst->getOperand(0);
9153 Value *RHSVal = FirstInst->getOperand(1);
9155 const Type *LHSType = LHSVal->getType();
9156 const Type *RHSType = RHSVal->getType();
9158 // Scan to see if all operands are the same opcode, all have one use, and all
9159 // kill their operands (i.e. the operands have one use).
9160 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
9161 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
9162 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
9163 // Verify type of the LHS matches so we don't fold cmp's of different
9164 // types or GEP's with different index types.
9165 I->getOperand(0)->getType() != LHSType ||
9166 I->getOperand(1)->getType() != RHSType)
9169 // If they are CmpInst instructions, check their predicates
9170 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
9171 if (cast<CmpInst>(I)->getPredicate() !=
9172 cast<CmpInst>(FirstInst)->getPredicate())
9175 // Keep track of which operand needs a phi node.
9176 if (I->getOperand(0) != LHSVal) LHSVal = 0;
9177 if (I->getOperand(1) != RHSVal) RHSVal = 0;
9180 // Otherwise, this is safe to transform, determine if it is profitable.
9182 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
9183 // Indexes are often folded into load/store instructions, so we don't want to
9184 // hide them behind a phi.
9185 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
9188 Value *InLHS = FirstInst->getOperand(0);
9189 Value *InRHS = FirstInst->getOperand(1);
9190 PHINode *NewLHS = 0, *NewRHS = 0;
9192 NewLHS = PHINode::Create(LHSType, FirstInst->getOperand(0)->getName()+".pn");
9193 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
9194 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
9195 InsertNewInstBefore(NewLHS, PN);
9200 NewRHS = PHINode::Create(RHSType, FirstInst->getOperand(1)->getName()+".pn");
9201 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
9202 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
9203 InsertNewInstBefore(NewRHS, PN);
9207 // Add all operands to the new PHIs.
9208 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9210 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9211 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
9214 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
9215 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
9219 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9220 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
9221 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9222 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
9225 assert(isa<GetElementPtrInst>(FirstInst));
9226 return GetElementPtrInst::Create(LHSVal, RHSVal);
9230 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
9231 /// of the block that defines it. This means that it must be obvious the value
9232 /// of the load is not changed from the point of the load to the end of the
9235 /// Finally, it is safe, but not profitable, to sink a load targetting a
9236 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
9238 static bool isSafeToSinkLoad(LoadInst *L) {
9239 BasicBlock::iterator BBI = L, E = L->getParent()->end();
9241 for (++BBI; BBI != E; ++BBI)
9242 if (BBI->mayWriteToMemory())
9245 // Check for non-address taken alloca. If not address-taken already, it isn't
9246 // profitable to do this xform.
9247 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
9248 bool isAddressTaken = false;
9249 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
9251 if (isa<LoadInst>(UI)) continue;
9252 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
9253 // If storing TO the alloca, then the address isn't taken.
9254 if (SI->getOperand(1) == AI) continue;
9256 isAddressTaken = true;
9260 if (!isAddressTaken)
9268 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
9269 // operator and they all are only used by the PHI, PHI together their
9270 // inputs, and do the operation once, to the result of the PHI.
9271 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
9272 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
9274 // Scan the instruction, looking for input operations that can be folded away.
9275 // If all input operands to the phi are the same instruction (e.g. a cast from
9276 // the same type or "+42") we can pull the operation through the PHI, reducing
9277 // code size and simplifying code.
9278 Constant *ConstantOp = 0;
9279 const Type *CastSrcTy = 0;
9280 bool isVolatile = false;
9281 if (isa<CastInst>(FirstInst)) {
9282 CastSrcTy = FirstInst->getOperand(0)->getType();
9283 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
9284 // Can fold binop, compare or shift here if the RHS is a constant,
9285 // otherwise call FoldPHIArgBinOpIntoPHI.
9286 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
9287 if (ConstantOp == 0)
9288 return FoldPHIArgBinOpIntoPHI(PN);
9289 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
9290 isVolatile = LI->isVolatile();
9291 // We can't sink the load if the loaded value could be modified between the
9292 // load and the PHI.
9293 if (LI->getParent() != PN.getIncomingBlock(0) ||
9294 !isSafeToSinkLoad(LI))
9296 } else if (isa<GetElementPtrInst>(FirstInst)) {
9297 if (FirstInst->getNumOperands() == 2)
9298 return FoldPHIArgBinOpIntoPHI(PN);
9299 // Can't handle general GEPs yet.
9302 return 0; // Cannot fold this operation.
9305 // Check to see if all arguments are the same operation.
9306 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9307 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
9308 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
9309 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
9312 if (I->getOperand(0)->getType() != CastSrcTy)
9313 return 0; // Cast operation must match.
9314 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9315 // We can't sink the load if the loaded value could be modified between
9316 // the load and the PHI.
9317 if (LI->isVolatile() != isVolatile ||
9318 LI->getParent() != PN.getIncomingBlock(i) ||
9319 !isSafeToSinkLoad(LI))
9321 } else if (I->getOperand(1) != ConstantOp) {
9326 // Okay, they are all the same operation. Create a new PHI node of the
9327 // correct type, and PHI together all of the LHS's of the instructions.
9328 PHINode *NewPN = PHINode::Create(FirstInst->getOperand(0)->getType(),
9329 PN.getName()+".in");
9330 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
9332 Value *InVal = FirstInst->getOperand(0);
9333 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
9335 // Add all operands to the new PHI.
9336 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
9337 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
9338 if (NewInVal != InVal)
9340 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
9345 // The new PHI unions all of the same values together. This is really
9346 // common, so we handle it intelligently here for compile-time speed.
9350 InsertNewInstBefore(NewPN, PN);
9354 // Insert and return the new operation.
9355 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
9356 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
9357 else if (isa<LoadInst>(FirstInst))
9358 return new LoadInst(PhiVal, "", isVolatile);
9359 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
9360 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
9361 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
9362 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
9363 PhiVal, ConstantOp);
9365 assert(0 && "Unknown operation");
9369 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
9371 static bool DeadPHICycle(PHINode *PN,
9372 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
9373 if (PN->use_empty()) return true;
9374 if (!PN->hasOneUse()) return false;
9376 // Remember this node, and if we find the cycle, return.
9377 if (!PotentiallyDeadPHIs.insert(PN))
9380 // Don't scan crazily complex things.
9381 if (PotentiallyDeadPHIs.size() == 16)
9384 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
9385 return DeadPHICycle(PU, PotentiallyDeadPHIs);
9390 /// PHIsEqualValue - Return true if this phi node is always equal to
9391 /// NonPhiInVal. This happens with mutually cyclic phi nodes like:
9392 /// z = some value; x = phi (y, z); y = phi (x, z)
9393 static bool PHIsEqualValue(PHINode *PN, Value *NonPhiInVal,
9394 SmallPtrSet<PHINode*, 16> &ValueEqualPHIs) {
9395 // See if we already saw this PHI node.
9396 if (!ValueEqualPHIs.insert(PN))
9399 // Don't scan crazily complex things.
9400 if (ValueEqualPHIs.size() == 16)
9403 // Scan the operands to see if they are either phi nodes or are equal to
9405 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
9406 Value *Op = PN->getIncomingValue(i);
9407 if (PHINode *OpPN = dyn_cast<PHINode>(Op)) {
9408 if (!PHIsEqualValue(OpPN, NonPhiInVal, ValueEqualPHIs))
9410 } else if (Op != NonPhiInVal)
9418 // PHINode simplification
9420 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
9421 // If LCSSA is around, don't mess with Phi nodes
9422 if (MustPreserveLCSSA) return 0;
9424 if (Value *V = PN.hasConstantValue())
9425 return ReplaceInstUsesWith(PN, V);
9427 // If all PHI operands are the same operation, pull them through the PHI,
9428 // reducing code size.
9429 if (isa<Instruction>(PN.getIncomingValue(0)) &&
9430 PN.getIncomingValue(0)->hasOneUse())
9431 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
9434 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
9435 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
9436 // PHI)... break the cycle.
9437 if (PN.hasOneUse()) {
9438 Instruction *PHIUser = cast<Instruction>(PN.use_back());
9439 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
9440 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
9441 PotentiallyDeadPHIs.insert(&PN);
9442 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
9443 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9446 // If this phi has a single use, and if that use just computes a value for
9447 // the next iteration of a loop, delete the phi. This occurs with unused
9448 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
9449 // common case here is good because the only other things that catch this
9450 // are induction variable analysis (sometimes) and ADCE, which is only run
9452 if (PHIUser->hasOneUse() &&
9453 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
9454 PHIUser->use_back() == &PN) {
9455 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
9459 // We sometimes end up with phi cycles that non-obviously end up being the
9460 // same value, for example:
9461 // z = some value; x = phi (y, z); y = phi (x, z)
9462 // where the phi nodes don't necessarily need to be in the same block. Do a
9463 // quick check to see if the PHI node only contains a single non-phi value, if
9464 // so, scan to see if the phi cycle is actually equal to that value.
9466 unsigned InValNo = 0, NumOperandVals = PN.getNumIncomingValues();
9467 // Scan for the first non-phi operand.
9468 while (InValNo != NumOperandVals &&
9469 isa<PHINode>(PN.getIncomingValue(InValNo)))
9472 if (InValNo != NumOperandVals) {
9473 Value *NonPhiInVal = PN.getOperand(InValNo);
9475 // Scan the rest of the operands to see if there are any conflicts, if so
9476 // there is no need to recursively scan other phis.
9477 for (++InValNo; InValNo != NumOperandVals; ++InValNo) {
9478 Value *OpVal = PN.getIncomingValue(InValNo);
9479 if (OpVal != NonPhiInVal && !isa<PHINode>(OpVal))
9483 // If we scanned over all operands, then we have one unique value plus
9484 // phi values. Scan PHI nodes to see if they all merge in each other or
9486 if (InValNo == NumOperandVals) {
9487 SmallPtrSet<PHINode*, 16> ValueEqualPHIs;
9488 if (PHIsEqualValue(&PN, NonPhiInVal, ValueEqualPHIs))
9489 return ReplaceInstUsesWith(PN, NonPhiInVal);
9496 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
9497 Instruction *InsertPoint,
9499 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
9500 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
9501 // We must cast correctly to the pointer type. Ensure that we
9502 // sign extend the integer value if it is smaller as this is
9503 // used for address computation.
9504 Instruction::CastOps opcode =
9505 (VTySize < PtrSize ? Instruction::SExt :
9506 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
9507 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
9511 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
9512 Value *PtrOp = GEP.getOperand(0);
9513 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
9514 // If so, eliminate the noop.
9515 if (GEP.getNumOperands() == 1)
9516 return ReplaceInstUsesWith(GEP, PtrOp);
9518 if (isa<UndefValue>(GEP.getOperand(0)))
9519 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
9521 bool HasZeroPointerIndex = false;
9522 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
9523 HasZeroPointerIndex = C->isNullValue();
9525 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
9526 return ReplaceInstUsesWith(GEP, PtrOp);
9528 // Eliminate unneeded casts for indices.
9529 bool MadeChange = false;
9531 gep_type_iterator GTI = gep_type_begin(GEP);
9532 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
9533 if (isa<SequentialType>(*GTI)) {
9534 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
9535 if (CI->getOpcode() == Instruction::ZExt ||
9536 CI->getOpcode() == Instruction::SExt) {
9537 const Type *SrcTy = CI->getOperand(0)->getType();
9538 // We can eliminate a cast from i32 to i64 iff the target
9539 // is a 32-bit pointer target.
9540 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
9542 GEP.setOperand(i, CI->getOperand(0));
9546 // If we are using a wider index than needed for this platform, shrink it
9547 // to what we need. If the incoming value needs a cast instruction,
9548 // insert it. This explicit cast can make subsequent optimizations more
9550 Value *Op = GEP.getOperand(i);
9551 if (TD->getTypeSizeInBits(Op->getType()) > TD->getPointerSizeInBits()) {
9552 if (Constant *C = dyn_cast<Constant>(Op)) {
9553 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
9556 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
9558 GEP.setOperand(i, Op);
9564 if (MadeChange) return &GEP;
9566 // If this GEP instruction doesn't move the pointer, and if the input operand
9567 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
9568 // real input to the dest type.
9569 if (GEP.hasAllZeroIndices()) {
9570 if (BitCastInst *BCI = dyn_cast<BitCastInst>(GEP.getOperand(0))) {
9571 // If the bitcast is of an allocation, and the allocation will be
9572 // converted to match the type of the cast, don't touch this.
9573 if (isa<AllocationInst>(BCI->getOperand(0))) {
9574 // See if the bitcast simplifies, if so, don't nuke this GEP yet.
9575 if (Instruction *I = visitBitCast(*BCI)) {
9578 BCI->getParent()->getInstList().insert(BCI, I);
9579 ReplaceInstUsesWith(*BCI, I);
9584 return new BitCastInst(BCI->getOperand(0), GEP.getType());
9588 // Combine Indices - If the source pointer to this getelementptr instruction
9589 // is a getelementptr instruction, combine the indices of the two
9590 // getelementptr instructions into a single instruction.
9592 SmallVector<Value*, 8> SrcGEPOperands;
9593 if (User *Src = dyn_castGetElementPtr(PtrOp))
9594 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
9596 if (!SrcGEPOperands.empty()) {
9597 // Note that if our source is a gep chain itself that we wait for that
9598 // chain to be resolved before we perform this transformation. This
9599 // avoids us creating a TON of code in some cases.
9601 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
9602 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
9603 return 0; // Wait until our source is folded to completion.
9605 SmallVector<Value*, 8> Indices;
9607 // Find out whether the last index in the source GEP is a sequential idx.
9608 bool EndsWithSequential = false;
9609 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
9610 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
9611 EndsWithSequential = !isa<StructType>(*I);
9613 // Can we combine the two pointer arithmetics offsets?
9614 if (EndsWithSequential) {
9615 // Replace: gep (gep %P, long B), long A, ...
9616 // With: T = long A+B; gep %P, T, ...
9618 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
9619 if (SO1 == Constant::getNullValue(SO1->getType())) {
9621 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
9624 // If they aren't the same type, convert both to an integer of the
9625 // target's pointer size.
9626 if (SO1->getType() != GO1->getType()) {
9627 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
9628 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
9629 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
9630 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
9632 unsigned PS = TD->getPointerSizeInBits();
9633 if (TD->getTypeSizeInBits(SO1->getType()) == PS) {
9634 // Convert GO1 to SO1's type.
9635 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
9637 } else if (TD->getTypeSizeInBits(GO1->getType()) == PS) {
9638 // Convert SO1 to GO1's type.
9639 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
9641 const Type *PT = TD->getIntPtrType();
9642 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
9643 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
9647 if (isa<Constant>(SO1) && isa<Constant>(GO1))
9648 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
9650 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
9651 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
9655 // Recycle the GEP we already have if possible.
9656 if (SrcGEPOperands.size() == 2) {
9657 GEP.setOperand(0, SrcGEPOperands[0]);
9658 GEP.setOperand(1, Sum);
9661 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9662 SrcGEPOperands.end()-1);
9663 Indices.push_back(Sum);
9664 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
9666 } else if (isa<Constant>(*GEP.idx_begin()) &&
9667 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
9668 SrcGEPOperands.size() != 1) {
9669 // Otherwise we can do the fold if the first index of the GEP is a zero
9670 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
9671 SrcGEPOperands.end());
9672 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
9675 if (!Indices.empty())
9676 return GetElementPtrInst::Create(SrcGEPOperands[0], Indices.begin(),
9677 Indices.end(), GEP.getName());
9679 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
9680 // GEP of global variable. If all of the indices for this GEP are
9681 // constants, we can promote this to a constexpr instead of an instruction.
9683 // Scan for nonconstants...
9684 SmallVector<Constant*, 8> Indices;
9685 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
9686 for (; I != E && isa<Constant>(*I); ++I)
9687 Indices.push_back(cast<Constant>(*I));
9689 if (I == E) { // If they are all constants...
9690 Constant *CE = ConstantExpr::getGetElementPtr(GV,
9691 &Indices[0],Indices.size());
9693 // Replace all uses of the GEP with the new constexpr...
9694 return ReplaceInstUsesWith(GEP, CE);
9696 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
9697 if (!isa<PointerType>(X->getType())) {
9698 // Not interesting. Source pointer must be a cast from pointer.
9699 } else if (HasZeroPointerIndex) {
9700 // transform: GEP (bitcast [10 x i8]* X to [0 x i8]*), i32 0, ...
9701 // into : GEP [10 x i8]* X, i32 0, ...
9703 // This occurs when the program declares an array extern like "int X[];"
9705 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
9706 const PointerType *XTy = cast<PointerType>(X->getType());
9707 if (const ArrayType *XATy =
9708 dyn_cast<ArrayType>(XTy->getElementType()))
9709 if (const ArrayType *CATy =
9710 dyn_cast<ArrayType>(CPTy->getElementType()))
9711 if (CATy->getElementType() == XATy->getElementType()) {
9712 // At this point, we know that the cast source type is a pointer
9713 // to an array of the same type as the destination pointer
9714 // array. Because the array type is never stepped over (there
9715 // is a leading zero) we can fold the cast into this GEP.
9716 GEP.setOperand(0, X);
9719 } else if (GEP.getNumOperands() == 2) {
9720 // Transform things like:
9721 // %t = getelementptr i32* bitcast ([2 x i32]* %str to i32*), i32 %V
9722 // into: %t1 = getelementptr [2 x i32]* %str, i32 0, i32 %V; bitcast
9723 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
9724 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
9725 if (isa<ArrayType>(SrcElTy) &&
9726 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
9727 TD->getABITypeSize(ResElTy)) {
9729 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9730 Idx[1] = GEP.getOperand(1);
9731 Value *V = InsertNewInstBefore(
9732 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName()), GEP);
9733 // V and GEP are both pointer types --> BitCast
9734 return new BitCastInst(V, GEP.getType());
9737 // Transform things like:
9738 // getelementptr i8* bitcast ([100 x double]* X to i8*), i32 %tmp
9739 // (where tmp = 8*tmp2) into:
9740 // getelementptr [100 x double]* %arr, i32 0, i32 %tmp2; bitcast
9742 if (isa<ArrayType>(SrcElTy) && ResElTy == Type::Int8Ty) {
9743 uint64_t ArrayEltSize =
9744 TD->getABITypeSize(cast<ArrayType>(SrcElTy)->getElementType());
9746 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
9747 // allow either a mul, shift, or constant here.
9749 ConstantInt *Scale = 0;
9750 if (ArrayEltSize == 1) {
9751 NewIdx = GEP.getOperand(1);
9752 Scale = ConstantInt::get(NewIdx->getType(), 1);
9753 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
9754 NewIdx = ConstantInt::get(CI->getType(), 1);
9756 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
9757 if (Inst->getOpcode() == Instruction::Shl &&
9758 isa<ConstantInt>(Inst->getOperand(1))) {
9759 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
9760 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
9761 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
9762 NewIdx = Inst->getOperand(0);
9763 } else if (Inst->getOpcode() == Instruction::Mul &&
9764 isa<ConstantInt>(Inst->getOperand(1))) {
9765 Scale = cast<ConstantInt>(Inst->getOperand(1));
9766 NewIdx = Inst->getOperand(0);
9770 // If the index will be to exactly the right offset with the scale taken
9771 // out, perform the transformation. Note, we don't know whether Scale is
9772 // signed or not. We'll use unsigned version of division/modulo
9773 // operation after making sure Scale doesn't have the sign bit set.
9774 if (Scale && Scale->getSExtValue() >= 0LL &&
9775 Scale->getZExtValue() % ArrayEltSize == 0) {
9776 Scale = ConstantInt::get(Scale->getType(),
9777 Scale->getZExtValue() / ArrayEltSize);
9778 if (Scale->getZExtValue() != 1) {
9779 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
9781 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
9782 NewIdx = InsertNewInstBefore(Sc, GEP);
9785 // Insert the new GEP instruction.
9787 Idx[0] = Constant::getNullValue(Type::Int32Ty);
9789 Instruction *NewGEP =
9790 GetElementPtrInst::Create(X, Idx, Idx + 2, GEP.getName());
9791 NewGEP = InsertNewInstBefore(NewGEP, GEP);
9792 // The NewGEP must be pointer typed, so must the old one -> BitCast
9793 return new BitCastInst(NewGEP, GEP.getType());
9802 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
9803 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
9804 if (AI.isArrayAllocation()) { // Check C != 1
9805 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
9807 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
9808 AllocationInst *New = 0;
9810 // Create and insert the replacement instruction...
9811 if (isa<MallocInst>(AI))
9812 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
9814 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
9815 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
9818 InsertNewInstBefore(New, AI);
9820 // Scan to the end of the allocation instructions, to skip over a block of
9821 // allocas if possible...
9823 BasicBlock::iterator It = New;
9824 while (isa<AllocationInst>(*It)) ++It;
9826 // Now that I is pointing to the first non-allocation-inst in the block,
9827 // insert our getelementptr instruction...
9829 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
9833 Value *V = GetElementPtrInst::Create(New, Idx, Idx + 2,
9834 New->getName()+".sub", It);
9836 // Now make everything use the getelementptr instead of the original
9838 return ReplaceInstUsesWith(AI, V);
9839 } else if (isa<UndefValue>(AI.getArraySize())) {
9840 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9844 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
9845 // Note that we only do this for alloca's, because malloc should allocate and
9846 // return a unique pointer, even for a zero byte allocation.
9847 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
9848 TD->getABITypeSize(AI.getAllocatedType()) == 0)
9849 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
9854 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
9855 Value *Op = FI.getOperand(0);
9857 // free undef -> unreachable.
9858 if (isa<UndefValue>(Op)) {
9859 // Insert a new store to null because we cannot modify the CFG here.
9860 new StoreInst(ConstantInt::getTrue(),
9861 UndefValue::get(PointerType::getUnqual(Type::Int1Ty)), &FI);
9862 return EraseInstFromFunction(FI);
9865 // If we have 'free null' delete the instruction. This can happen in stl code
9866 // when lots of inlining happens.
9867 if (isa<ConstantPointerNull>(Op))
9868 return EraseInstFromFunction(FI);
9870 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
9871 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
9872 FI.setOperand(0, CI->getOperand(0));
9876 // Change free (gep X, 0,0,0,0) into free(X)
9877 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
9878 if (GEPI->hasAllZeroIndices()) {
9879 AddToWorkList(GEPI);
9880 FI.setOperand(0, GEPI->getOperand(0));
9885 // Change free(malloc) into nothing, if the malloc has a single use.
9886 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
9887 if (MI->hasOneUse()) {
9888 EraseInstFromFunction(FI);
9889 return EraseInstFromFunction(*MI);
9896 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
9897 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI,
9898 const TargetData *TD) {
9899 User *CI = cast<User>(LI.getOperand(0));
9900 Value *CastOp = CI->getOperand(0);
9902 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(CI)) {
9903 // Instead of loading constant c string, use corresponding integer value
9904 // directly if string length is small enough.
9905 const std::string &Str = CE->getOperand(0)->getStringValue();
9907 unsigned len = Str.length();
9908 const Type *Ty = cast<PointerType>(CE->getType())->getElementType();
9909 unsigned numBits = Ty->getPrimitiveSizeInBits();
9910 // Replace LI with immediate integer store.
9911 if ((numBits >> 3) == len + 1) {
9912 APInt StrVal(numBits, 0);
9913 APInt SingleChar(numBits, 0);
9914 if (TD->isLittleEndian()) {
9915 for (signed i = len-1; i >= 0; i--) {
9916 SingleChar = (uint64_t) Str[i];
9917 StrVal = (StrVal << 8) | SingleChar;
9920 for (unsigned i = 0; i < len; i++) {
9921 SingleChar = (uint64_t) Str[i];
9922 StrVal = (StrVal << 8) | SingleChar;
9924 // Append NULL at the end.
9926 StrVal = (StrVal << 8) | SingleChar;
9928 Value *NL = ConstantInt::get(StrVal);
9929 return IC.ReplaceInstUsesWith(LI, NL);
9934 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9935 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9936 const Type *SrcPTy = SrcTy->getElementType();
9938 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
9939 isa<VectorType>(DestPTy)) {
9940 // If the source is an array, the code below will not succeed. Check to
9941 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9943 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9944 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9945 if (ASrcTy->getNumElements() != 0) {
9947 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9948 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9949 SrcTy = cast<PointerType>(CastOp->getType());
9950 SrcPTy = SrcTy->getElementType();
9953 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
9954 isa<VectorType>(SrcPTy)) &&
9955 // Do not allow turning this into a load of an integer, which is then
9956 // casted to a pointer, this pessimizes pointer analysis a lot.
9957 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
9958 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9959 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9961 // Okay, we are casting from one integer or pointer type to another of
9962 // the same size. Instead of casting the pointer before the load, cast
9963 // the result of the loaded value.
9964 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
9966 LI.isVolatile()),LI);
9967 // Now cast the result of the load.
9968 return new BitCastInst(NewLoad, LI.getType());
9975 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
9976 /// from this value cannot trap. If it is not obviously safe to load from the
9977 /// specified pointer, we do a quick local scan of the basic block containing
9978 /// ScanFrom, to determine if the address is already accessed.
9979 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
9980 // If it is an alloca it is always safe to load from.
9981 if (isa<AllocaInst>(V)) return true;
9983 // If it is a global variable it is mostly safe to load from.
9984 if (const GlobalValue *GV = dyn_cast<GlobalVariable>(V))
9985 // Don't try to evaluate aliases. External weak GV can be null.
9986 return !isa<GlobalAlias>(GV) && !GV->hasExternalWeakLinkage();
9988 // Otherwise, be a little bit agressive by scanning the local block where we
9989 // want to check to see if the pointer is already being loaded or stored
9990 // from/to. If so, the previous load or store would have already trapped,
9991 // so there is no harm doing an extra load (also, CSE will later eliminate
9992 // the load entirely).
9993 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
9998 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9999 if (LI->getOperand(0) == V) return true;
10000 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10001 if (SI->getOperand(1) == V) return true;
10007 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
10008 /// until we find the underlying object a pointer is referring to or something
10009 /// we don't understand. Note that the returned pointer may be offset from the
10010 /// input, because we ignore GEP indices.
10011 static Value *GetUnderlyingObject(Value *Ptr) {
10013 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
10014 if (CE->getOpcode() == Instruction::BitCast ||
10015 CE->getOpcode() == Instruction::GetElementPtr)
10016 Ptr = CE->getOperand(0);
10019 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
10020 Ptr = BCI->getOperand(0);
10021 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
10022 Ptr = GEP->getOperand(0);
10029 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
10030 Value *Op = LI.getOperand(0);
10032 // Attempt to improve the alignment.
10033 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op);
10035 (LI.getAlignment() == 0 ? TD->getABITypeAlignment(LI.getType()) :
10036 LI.getAlignment()))
10037 LI.setAlignment(KnownAlign);
10039 // load (cast X) --> cast (load X) iff safe
10040 if (isa<CastInst>(Op))
10041 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10044 // None of the following transforms are legal for volatile loads.
10045 if (LI.isVolatile()) return 0;
10047 if (&LI.getParent()->front() != &LI) {
10048 BasicBlock::iterator BBI = &LI; --BBI;
10049 // If the instruction immediately before this is a store to the same
10050 // address, do a simple form of store->load forwarding.
10051 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
10052 if (SI->getOperand(1) == LI.getOperand(0))
10053 return ReplaceInstUsesWith(LI, SI->getOperand(0));
10054 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
10055 if (LIB->getOperand(0) == LI.getOperand(0))
10056 return ReplaceInstUsesWith(LI, LIB);
10059 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
10060 const Value *GEPI0 = GEPI->getOperand(0);
10061 // TODO: Consider a target hook for valid address spaces for this xform.
10062 if (isa<ConstantPointerNull>(GEPI0) &&
10063 cast<PointerType>(GEPI0->getType())->getAddressSpace() == 0) {
10064 // Insert a new store to null instruction before the load to indicate
10065 // that this code is not reachable. We do this instead of inserting
10066 // an unreachable instruction directly because we cannot modify the
10068 new StoreInst(UndefValue::get(LI.getType()),
10069 Constant::getNullValue(Op->getType()), &LI);
10070 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10074 if (Constant *C = dyn_cast<Constant>(Op)) {
10075 // load null/undef -> undef
10076 // TODO: Consider a target hook for valid address spaces for this xform.
10077 if (isa<UndefValue>(C) || (C->isNullValue() &&
10078 cast<PointerType>(Op->getType())->getAddressSpace() == 0)) {
10079 // Insert a new store to null instruction before the load to indicate that
10080 // this code is not reachable. We do this instead of inserting an
10081 // unreachable instruction directly because we cannot modify the CFG.
10082 new StoreInst(UndefValue::get(LI.getType()),
10083 Constant::getNullValue(Op->getType()), &LI);
10084 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10087 // Instcombine load (constant global) into the value loaded.
10088 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
10089 if (GV->isConstant() && !GV->isDeclaration())
10090 return ReplaceInstUsesWith(LI, GV->getInitializer());
10092 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
10093 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op)) {
10094 if (CE->getOpcode() == Instruction::GetElementPtr) {
10095 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
10096 if (GV->isConstant() && !GV->isDeclaration())
10098 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
10099 return ReplaceInstUsesWith(LI, V);
10100 if (CE->getOperand(0)->isNullValue()) {
10101 // Insert a new store to null instruction before the load to indicate
10102 // that this code is not reachable. We do this instead of inserting
10103 // an unreachable instruction directly because we cannot modify the
10105 new StoreInst(UndefValue::get(LI.getType()),
10106 Constant::getNullValue(Op->getType()), &LI);
10107 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10110 } else if (CE->isCast()) {
10111 if (Instruction *Res = InstCombineLoadCast(*this, LI, TD))
10117 // If this load comes from anywhere in a constant global, and if the global
10118 // is all undef or zero, we know what it loads.
10119 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
10120 if (GV->isConstant() && GV->hasInitializer()) {
10121 if (GV->getInitializer()->isNullValue())
10122 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
10123 else if (isa<UndefValue>(GV->getInitializer()))
10124 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
10128 if (Op->hasOneUse()) {
10129 // Change select and PHI nodes to select values instead of addresses: this
10130 // helps alias analysis out a lot, allows many others simplifications, and
10131 // exposes redundancy in the code.
10133 // Note that we cannot do the transformation unless we know that the
10134 // introduced loads cannot trap! Something like this is valid as long as
10135 // the condition is always false: load (select bool %C, int* null, int* %G),
10136 // but it would not be valid if we transformed it to load from null
10137 // unconditionally.
10139 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
10140 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
10141 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
10142 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
10143 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
10144 SI->getOperand(1)->getName()+".val"), LI);
10145 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
10146 SI->getOperand(2)->getName()+".val"), LI);
10147 return SelectInst::Create(SI->getCondition(), V1, V2);
10150 // load (select (cond, null, P)) -> load P
10151 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
10152 if (C->isNullValue()) {
10153 LI.setOperand(0, SI->getOperand(2));
10157 // load (select (cond, P, null)) -> load P
10158 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
10159 if (C->isNullValue()) {
10160 LI.setOperand(0, SI->getOperand(1));
10168 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
10170 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
10171 User *CI = cast<User>(SI.getOperand(1));
10172 Value *CastOp = CI->getOperand(0);
10174 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
10175 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
10176 const Type *SrcPTy = SrcTy->getElementType();
10178 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
10179 // If the source is an array, the code below will not succeed. Check to
10180 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
10182 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
10183 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
10184 if (ASrcTy->getNumElements() != 0) {
10186 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
10187 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
10188 SrcTy = cast<PointerType>(CastOp->getType());
10189 SrcPTy = SrcTy->getElementType();
10192 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
10193 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
10194 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
10196 // Okay, we are casting from one integer or pointer type to another of
10197 // the same size. Instead of casting the pointer before
10198 // the store, cast the value to be stored.
10200 Value *SIOp0 = SI.getOperand(0);
10201 Instruction::CastOps opcode = Instruction::BitCast;
10202 const Type* CastSrcTy = SIOp0->getType();
10203 const Type* CastDstTy = SrcPTy;
10204 if (isa<PointerType>(CastDstTy)) {
10205 if (CastSrcTy->isInteger())
10206 opcode = Instruction::IntToPtr;
10207 } else if (isa<IntegerType>(CastDstTy)) {
10208 if (isa<PointerType>(SIOp0->getType()))
10209 opcode = Instruction::PtrToInt;
10211 if (Constant *C = dyn_cast<Constant>(SIOp0))
10212 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
10214 NewCast = IC.InsertNewInstBefore(
10215 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
10217 return new StoreInst(NewCast, CastOp);
10224 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
10225 Value *Val = SI.getOperand(0);
10226 Value *Ptr = SI.getOperand(1);
10228 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
10229 EraseInstFromFunction(SI);
10234 // If the RHS is an alloca with a single use, zapify the store, making the
10236 if (Ptr->hasOneUse()) {
10237 if (isa<AllocaInst>(Ptr)) {
10238 EraseInstFromFunction(SI);
10243 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
10244 if (isa<AllocaInst>(GEP->getOperand(0)) &&
10245 GEP->getOperand(0)->hasOneUse()) {
10246 EraseInstFromFunction(SI);
10252 // Attempt to improve the alignment.
10253 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr);
10255 (SI.getAlignment() == 0 ? TD->getABITypeAlignment(Val->getType()) :
10256 SI.getAlignment()))
10257 SI.setAlignment(KnownAlign);
10259 // Do really simple DSE, to catch cases where there are several consequtive
10260 // stores to the same location, separated by a few arithmetic operations. This
10261 // situation often occurs with bitfield accesses.
10262 BasicBlock::iterator BBI = &SI;
10263 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
10267 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
10268 // Prev store isn't volatile, and stores to the same location?
10269 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
10272 EraseInstFromFunction(*PrevSI);
10278 // If this is a load, we have to stop. However, if the loaded value is from
10279 // the pointer we're loading and is producing the pointer we're storing,
10280 // then *this* store is dead (X = load P; store X -> P).
10281 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
10282 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
10283 EraseInstFromFunction(SI);
10287 // Otherwise, this is a load from some other location. Stores before it
10288 // may not be dead.
10292 // Don't skip over loads or things that can modify memory.
10293 if (BBI->mayWriteToMemory())
10298 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
10300 // store X, null -> turns into 'unreachable' in SimplifyCFG
10301 if (isa<ConstantPointerNull>(Ptr)) {
10302 if (!isa<UndefValue>(Val)) {
10303 SI.setOperand(0, UndefValue::get(Val->getType()));
10304 if (Instruction *U = dyn_cast<Instruction>(Val))
10305 AddToWorkList(U); // Dropped a use.
10308 return 0; // Do not modify these!
10311 // store undef, Ptr -> noop
10312 if (isa<UndefValue>(Val)) {
10313 EraseInstFromFunction(SI);
10318 // If the pointer destination is a cast, see if we can fold the cast into the
10320 if (isa<CastInst>(Ptr))
10321 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10323 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
10325 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
10329 // If this store is the last instruction in the basic block, and if the block
10330 // ends with an unconditional branch, try to move it to the successor block.
10332 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
10333 if (BI->isUnconditional())
10334 if (SimplifyStoreAtEndOfBlock(SI))
10335 return 0; // xform done!
10340 /// SimplifyStoreAtEndOfBlock - Turn things like:
10341 /// if () { *P = v1; } else { *P = v2 }
10342 /// into a phi node with a store in the successor.
10344 /// Simplify things like:
10345 /// *P = v1; if () { *P = v2; }
10346 /// into a phi node with a store in the successor.
10348 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
10349 BasicBlock *StoreBB = SI.getParent();
10351 // Check to see if the successor block has exactly two incoming edges. If
10352 // so, see if the other predecessor contains a store to the same location.
10353 // if so, insert a PHI node (if needed) and move the stores down.
10354 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
10356 // Determine whether Dest has exactly two predecessors and, if so, compute
10357 // the other predecessor.
10358 pred_iterator PI = pred_begin(DestBB);
10359 BasicBlock *OtherBB = 0;
10360 if (*PI != StoreBB)
10363 if (PI == pred_end(DestBB))
10366 if (*PI != StoreBB) {
10371 if (++PI != pred_end(DestBB))
10375 // Verify that the other block ends in a branch and is not otherwise empty.
10376 BasicBlock::iterator BBI = OtherBB->getTerminator();
10377 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
10378 if (!OtherBr || BBI == OtherBB->begin())
10381 // If the other block ends in an unconditional branch, check for the 'if then
10382 // else' case. there is an instruction before the branch.
10383 StoreInst *OtherStore = 0;
10384 if (OtherBr->isUnconditional()) {
10385 // If this isn't a store, or isn't a store to the same location, bail out.
10387 OtherStore = dyn_cast<StoreInst>(BBI);
10388 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
10391 // Otherwise, the other block ended with a conditional branch. If one of the
10392 // destinations is StoreBB, then we have the if/then case.
10393 if (OtherBr->getSuccessor(0) != StoreBB &&
10394 OtherBr->getSuccessor(1) != StoreBB)
10397 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
10398 // if/then triangle. See if there is a store to the same ptr as SI that
10399 // lives in OtherBB.
10401 // Check to see if we find the matching store.
10402 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
10403 if (OtherStore->getOperand(1) != SI.getOperand(1))
10407 // If we find something that may be using the stored value, or if we run
10408 // out of instructions, we can't do the xform.
10409 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
10410 BBI == OtherBB->begin())
10414 // In order to eliminate the store in OtherBr, we have to
10415 // make sure nothing reads the stored value in StoreBB.
10416 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
10417 // FIXME: This should really be AA driven.
10418 if (isa<LoadInst>(I) || I->mayWriteToMemory())
10423 // Insert a PHI node now if we need it.
10424 Value *MergedVal = OtherStore->getOperand(0);
10425 if (MergedVal != SI.getOperand(0)) {
10426 PHINode *PN = PHINode::Create(MergedVal->getType(), "storemerge");
10427 PN->reserveOperandSpace(2);
10428 PN->addIncoming(SI.getOperand(0), SI.getParent());
10429 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
10430 MergedVal = InsertNewInstBefore(PN, DestBB->front());
10433 // Advance to a place where it is safe to insert the new store and
10435 BBI = DestBB->begin();
10436 while (isa<PHINode>(BBI)) ++BBI;
10437 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
10438 OtherStore->isVolatile()), *BBI);
10440 // Nuke the old stores.
10441 EraseInstFromFunction(SI);
10442 EraseInstFromFunction(*OtherStore);
10448 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
10449 // Change br (not X), label True, label False to: br X, label False, True
10451 BasicBlock *TrueDest;
10452 BasicBlock *FalseDest;
10453 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
10454 !isa<Constant>(X)) {
10455 // Swap Destinations and condition...
10456 BI.setCondition(X);
10457 BI.setSuccessor(0, FalseDest);
10458 BI.setSuccessor(1, TrueDest);
10462 // Cannonicalize fcmp_one -> fcmp_oeq
10463 FCmpInst::Predicate FPred; Value *Y;
10464 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
10465 TrueDest, FalseDest)))
10466 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
10467 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
10468 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
10469 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
10470 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
10471 NewSCC->takeName(I);
10472 // Swap Destinations and condition...
10473 BI.setCondition(NewSCC);
10474 BI.setSuccessor(0, FalseDest);
10475 BI.setSuccessor(1, TrueDest);
10476 RemoveFromWorkList(I);
10477 I->eraseFromParent();
10478 AddToWorkList(NewSCC);
10482 // Cannonicalize icmp_ne -> icmp_eq
10483 ICmpInst::Predicate IPred;
10484 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
10485 TrueDest, FalseDest)))
10486 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
10487 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
10488 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
10489 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
10490 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
10491 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
10492 NewSCC->takeName(I);
10493 // Swap Destinations and condition...
10494 BI.setCondition(NewSCC);
10495 BI.setSuccessor(0, FalseDest);
10496 BI.setSuccessor(1, TrueDest);
10497 RemoveFromWorkList(I);
10498 I->eraseFromParent();;
10499 AddToWorkList(NewSCC);
10506 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
10507 Value *Cond = SI.getCondition();
10508 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
10509 if (I->getOpcode() == Instruction::Add)
10510 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
10511 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
10512 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
10513 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
10515 SI.setOperand(0, I->getOperand(0));
10523 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
10524 /// is to leave as a vector operation.
10525 static bool CheapToScalarize(Value *V, bool isConstant) {
10526 if (isa<ConstantAggregateZero>(V))
10528 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
10529 if (isConstant) return true;
10530 // If all elts are the same, we can extract.
10531 Constant *Op0 = C->getOperand(0);
10532 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10533 if (C->getOperand(i) != Op0)
10537 Instruction *I = dyn_cast<Instruction>(V);
10538 if (!I) return false;
10540 // Insert element gets simplified to the inserted element or is deleted if
10541 // this is constant idx extract element and its a constant idx insertelt.
10542 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
10543 isa<ConstantInt>(I->getOperand(2)))
10545 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
10547 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
10548 if (BO->hasOneUse() &&
10549 (CheapToScalarize(BO->getOperand(0), isConstant) ||
10550 CheapToScalarize(BO->getOperand(1), isConstant)))
10552 if (CmpInst *CI = dyn_cast<CmpInst>(I))
10553 if (CI->hasOneUse() &&
10554 (CheapToScalarize(CI->getOperand(0), isConstant) ||
10555 CheapToScalarize(CI->getOperand(1), isConstant)))
10561 /// Read and decode a shufflevector mask.
10563 /// It turns undef elements into values that are larger than the number of
10564 /// elements in the input.
10565 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
10566 unsigned NElts = SVI->getType()->getNumElements();
10567 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
10568 return std::vector<unsigned>(NElts, 0);
10569 if (isa<UndefValue>(SVI->getOperand(2)))
10570 return std::vector<unsigned>(NElts, 2*NElts);
10572 std::vector<unsigned> Result;
10573 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
10574 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
10575 if (isa<UndefValue>(CP->getOperand(i)))
10576 Result.push_back(NElts*2); // undef -> 8
10578 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
10582 /// FindScalarElement - Given a vector and an element number, see if the scalar
10583 /// value is already around as a register, for example if it were inserted then
10584 /// extracted from the vector.
10585 static Value *FindScalarElement(Value *V, unsigned EltNo) {
10586 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
10587 const VectorType *PTy = cast<VectorType>(V->getType());
10588 unsigned Width = PTy->getNumElements();
10589 if (EltNo >= Width) // Out of range access.
10590 return UndefValue::get(PTy->getElementType());
10592 if (isa<UndefValue>(V))
10593 return UndefValue::get(PTy->getElementType());
10594 else if (isa<ConstantAggregateZero>(V))
10595 return Constant::getNullValue(PTy->getElementType());
10596 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
10597 return CP->getOperand(EltNo);
10598 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
10599 // If this is an insert to a variable element, we don't know what it is.
10600 if (!isa<ConstantInt>(III->getOperand(2)))
10602 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
10604 // If this is an insert to the element we are looking for, return the
10606 if (EltNo == IIElt)
10607 return III->getOperand(1);
10609 // Otherwise, the insertelement doesn't modify the value, recurse on its
10611 return FindScalarElement(III->getOperand(0), EltNo);
10612 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
10613 unsigned InEl = getShuffleMask(SVI)[EltNo];
10615 return FindScalarElement(SVI->getOperand(0), InEl);
10616 else if (InEl < Width*2)
10617 return FindScalarElement(SVI->getOperand(1), InEl - Width);
10619 return UndefValue::get(PTy->getElementType());
10622 // Otherwise, we don't know.
10626 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
10628 // If vector val is undef, replace extract with scalar undef.
10629 if (isa<UndefValue>(EI.getOperand(0)))
10630 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10632 // If vector val is constant 0, replace extract with scalar 0.
10633 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
10634 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
10636 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
10637 // If vector val is constant with uniform operands, replace EI
10638 // with that operand
10639 Constant *op0 = C->getOperand(0);
10640 for (unsigned i = 1; i < C->getNumOperands(); ++i)
10641 if (C->getOperand(i) != op0) {
10646 return ReplaceInstUsesWith(EI, op0);
10649 // If extracting a specified index from the vector, see if we can recursively
10650 // find a previously computed scalar that was inserted into the vector.
10651 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10652 unsigned IndexVal = IdxC->getZExtValue();
10653 unsigned VectorWidth =
10654 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
10656 // If this is extracting an invalid index, turn this into undef, to avoid
10657 // crashing the code below.
10658 if (IndexVal >= VectorWidth)
10659 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10661 // This instruction only demands the single element from the input vector.
10662 // If the input vector has a single use, simplify it based on this use
10664 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
10665 uint64_t UndefElts;
10666 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
10669 EI.setOperand(0, V);
10674 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
10675 return ReplaceInstUsesWith(EI, Elt);
10677 // If the this extractelement is directly using a bitcast from a vector of
10678 // the same number of elements, see if we can find the source element from
10679 // it. In this case, we will end up needing to bitcast the scalars.
10680 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
10681 if (const VectorType *VT =
10682 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
10683 if (VT->getNumElements() == VectorWidth)
10684 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
10685 return new BitCastInst(Elt, EI.getType());
10689 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
10690 if (I->hasOneUse()) {
10691 // Push extractelement into predecessor operation if legal and
10692 // profitable to do so
10693 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
10694 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
10695 if (CheapToScalarize(BO, isConstantElt)) {
10696 ExtractElementInst *newEI0 =
10697 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
10698 EI.getName()+".lhs");
10699 ExtractElementInst *newEI1 =
10700 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
10701 EI.getName()+".rhs");
10702 InsertNewInstBefore(newEI0, EI);
10703 InsertNewInstBefore(newEI1, EI);
10704 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
10706 } else if (isa<LoadInst>(I)) {
10708 cast<PointerType>(I->getOperand(0)->getType())->getAddressSpace();
10709 Value *Ptr = InsertBitCastBefore(I->getOperand(0),
10710 PointerType::get(EI.getType(), AS),EI);
10711 GetElementPtrInst *GEP =
10712 GetElementPtrInst::Create(Ptr, EI.getOperand(1), I->getName() + ".gep");
10713 InsertNewInstBefore(GEP, EI);
10714 return new LoadInst(GEP);
10717 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
10718 // Extracting the inserted element?
10719 if (IE->getOperand(2) == EI.getOperand(1))
10720 return ReplaceInstUsesWith(EI, IE->getOperand(1));
10721 // If the inserted and extracted elements are constants, they must not
10722 // be the same value, extract from the pre-inserted value instead.
10723 if (isa<Constant>(IE->getOperand(2)) &&
10724 isa<Constant>(EI.getOperand(1))) {
10725 AddUsesToWorkList(EI);
10726 EI.setOperand(0, IE->getOperand(0));
10729 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
10730 // If this is extracting an element from a shufflevector, figure out where
10731 // it came from and extract from the appropriate input element instead.
10732 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
10733 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
10735 if (SrcIdx < SVI->getType()->getNumElements())
10736 Src = SVI->getOperand(0);
10737 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
10738 SrcIdx -= SVI->getType()->getNumElements();
10739 Src = SVI->getOperand(1);
10741 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
10743 return new ExtractElementInst(Src, SrcIdx);
10750 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
10751 /// elements from either LHS or RHS, return the shuffle mask and true.
10752 /// Otherwise, return false.
10753 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
10754 std::vector<Constant*> &Mask) {
10755 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
10756 "Invalid CollectSingleShuffleElements");
10757 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10759 if (isa<UndefValue>(V)) {
10760 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10762 } else if (V == LHS) {
10763 for (unsigned i = 0; i != NumElts; ++i)
10764 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10766 } else if (V == RHS) {
10767 for (unsigned i = 0; i != NumElts; ++i)
10768 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
10770 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10771 // If this is an insert of an extract from some other vector, include it.
10772 Value *VecOp = IEI->getOperand(0);
10773 Value *ScalarOp = IEI->getOperand(1);
10774 Value *IdxOp = IEI->getOperand(2);
10776 if (!isa<ConstantInt>(IdxOp))
10778 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10780 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
10781 // Okay, we can handle this if the vector we are insertinting into is
10782 // transitively ok.
10783 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10784 // If so, update the mask to reflect the inserted undef.
10785 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
10788 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
10789 if (isa<ConstantInt>(EI->getOperand(1)) &&
10790 EI->getOperand(0)->getType() == V->getType()) {
10791 unsigned ExtractedIdx =
10792 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10794 // This must be extracting from either LHS or RHS.
10795 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
10796 // Okay, we can handle this if the vector we are insertinting into is
10797 // transitively ok.
10798 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
10799 // If so, update the mask to reflect the inserted value.
10800 if (EI->getOperand(0) == LHS) {
10801 Mask[InsertedIdx & (NumElts-1)] =
10802 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10804 assert(EI->getOperand(0) == RHS);
10805 Mask[InsertedIdx & (NumElts-1)] =
10806 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
10815 // TODO: Handle shufflevector here!
10820 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
10821 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
10822 /// that computes V and the LHS value of the shuffle.
10823 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
10825 assert(isa<VectorType>(V->getType()) &&
10826 (RHS == 0 || V->getType() == RHS->getType()) &&
10827 "Invalid shuffle!");
10828 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
10830 if (isa<UndefValue>(V)) {
10831 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
10833 } else if (isa<ConstantAggregateZero>(V)) {
10834 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
10836 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
10837 // If this is an insert of an extract from some other vector, include it.
10838 Value *VecOp = IEI->getOperand(0);
10839 Value *ScalarOp = IEI->getOperand(1);
10840 Value *IdxOp = IEI->getOperand(2);
10842 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10843 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10844 EI->getOperand(0)->getType() == V->getType()) {
10845 unsigned ExtractedIdx =
10846 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10847 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10849 // Either the extracted from or inserted into vector must be RHSVec,
10850 // otherwise we'd end up with a shuffle of three inputs.
10851 if (EI->getOperand(0) == RHS || RHS == 0) {
10852 RHS = EI->getOperand(0);
10853 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
10854 Mask[InsertedIdx & (NumElts-1)] =
10855 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
10859 if (VecOp == RHS) {
10860 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
10861 // Everything but the extracted element is replaced with the RHS.
10862 for (unsigned i = 0; i != NumElts; ++i) {
10863 if (i != InsertedIdx)
10864 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
10869 // If this insertelement is a chain that comes from exactly these two
10870 // vectors, return the vector and the effective shuffle.
10871 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
10872 return EI->getOperand(0);
10877 // TODO: Handle shufflevector here!
10879 // Otherwise, can't do anything fancy. Return an identity vector.
10880 for (unsigned i = 0; i != NumElts; ++i)
10881 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
10885 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
10886 Value *VecOp = IE.getOperand(0);
10887 Value *ScalarOp = IE.getOperand(1);
10888 Value *IdxOp = IE.getOperand(2);
10890 // Inserting an undef or into an undefined place, remove this.
10891 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
10892 ReplaceInstUsesWith(IE, VecOp);
10894 // If the inserted element was extracted from some other vector, and if the
10895 // indexes are constant, try to turn this into a shufflevector operation.
10896 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
10897 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
10898 EI->getOperand(0)->getType() == IE.getType()) {
10899 unsigned NumVectorElts = IE.getType()->getNumElements();
10900 unsigned ExtractedIdx =
10901 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
10902 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
10904 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
10905 return ReplaceInstUsesWith(IE, VecOp);
10907 if (InsertedIdx >= NumVectorElts) // Out of range insert.
10908 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
10910 // If we are extracting a value from a vector, then inserting it right
10911 // back into the same place, just use the input vector.
10912 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
10913 return ReplaceInstUsesWith(IE, VecOp);
10915 // We could theoretically do this for ANY input. However, doing so could
10916 // turn chains of insertelement instructions into a chain of shufflevector
10917 // instructions, and right now we do not merge shufflevectors. As such,
10918 // only do this in a situation where it is clear that there is benefit.
10919 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
10920 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
10921 // the values of VecOp, except then one read from EIOp0.
10922 // Build a new shuffle mask.
10923 std::vector<Constant*> Mask;
10924 if (isa<UndefValue>(VecOp))
10925 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
10927 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
10928 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
10931 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
10932 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
10933 ConstantVector::get(Mask));
10936 // If this insertelement isn't used by some other insertelement, turn it
10937 // (and any insertelements it points to), into one big shuffle.
10938 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
10939 std::vector<Constant*> Mask;
10941 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
10942 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
10943 // We now have a shuffle of LHS, RHS, Mask.
10944 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
10953 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
10954 Value *LHS = SVI.getOperand(0);
10955 Value *RHS = SVI.getOperand(1);
10956 std::vector<unsigned> Mask = getShuffleMask(&SVI);
10958 bool MadeChange = false;
10960 // Undefined shuffle mask -> undefined value.
10961 if (isa<UndefValue>(SVI.getOperand(2)))
10962 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
10964 // If we have shuffle(x, undef, mask) and any elements of mask refer to
10965 // the undef, change them to undefs.
10966 if (isa<UndefValue>(SVI.getOperand(1))) {
10967 // Scan to see if there are any references to the RHS. If so, replace them
10968 // with undef element refs and set MadeChange to true.
10969 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10970 if (Mask[i] >= e && Mask[i] != 2*e) {
10977 // Remap any references to RHS to use LHS.
10978 std::vector<Constant*> Elts;
10979 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
10980 if (Mask[i] == 2*e)
10981 Elts.push_back(UndefValue::get(Type::Int32Ty));
10983 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
10985 SVI.setOperand(2, ConstantVector::get(Elts));
10989 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
10990 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
10991 if (LHS == RHS || isa<UndefValue>(LHS)) {
10992 if (isa<UndefValue>(LHS) && LHS == RHS) {
10993 // shuffle(undef,undef,mask) -> undef.
10994 return ReplaceInstUsesWith(SVI, LHS);
10997 // Remap any references to RHS to use LHS.
10998 std::vector<Constant*> Elts;
10999 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11000 if (Mask[i] >= 2*e)
11001 Elts.push_back(UndefValue::get(Type::Int32Ty));
11003 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
11004 (Mask[i] < e && isa<UndefValue>(LHS)))
11005 Mask[i] = 2*e; // Turn into undef.
11007 Mask[i] &= (e-1); // Force to LHS.
11008 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
11011 SVI.setOperand(0, SVI.getOperand(1));
11012 SVI.setOperand(1, UndefValue::get(RHS->getType()));
11013 SVI.setOperand(2, ConstantVector::get(Elts));
11014 LHS = SVI.getOperand(0);
11015 RHS = SVI.getOperand(1);
11019 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
11020 bool isLHSID = true, isRHSID = true;
11022 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
11023 if (Mask[i] >= e*2) continue; // Ignore undef values.
11024 // Is this an identity shuffle of the LHS value?
11025 isLHSID &= (Mask[i] == i);
11027 // Is this an identity shuffle of the RHS value?
11028 isRHSID &= (Mask[i]-e == i);
11031 // Eliminate identity shuffles.
11032 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
11033 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
11035 // If the LHS is a shufflevector itself, see if we can combine it with this
11036 // one without producing an unusual shuffle. Here we are really conservative:
11037 // we are absolutely afraid of producing a shuffle mask not in the input
11038 // program, because the code gen may not be smart enough to turn a merged
11039 // shuffle into two specific shuffles: it may produce worse code. As such,
11040 // we only merge two shuffles if the result is one of the two input shuffle
11041 // masks. In this case, merging the shuffles just removes one instruction,
11042 // which we know is safe. This is good for things like turning:
11043 // (splat(splat)) -> splat.
11044 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
11045 if (isa<UndefValue>(RHS)) {
11046 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
11048 std::vector<unsigned> NewMask;
11049 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
11050 if (Mask[i] >= 2*e)
11051 NewMask.push_back(2*e);
11053 NewMask.push_back(LHSMask[Mask[i]]);
11055 // If the result mask is equal to the src shuffle or this shuffle mask, do
11056 // the replacement.
11057 if (NewMask == LHSMask || NewMask == Mask) {
11058 std::vector<Constant*> Elts;
11059 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
11060 if (NewMask[i] >= e*2) {
11061 Elts.push_back(UndefValue::get(Type::Int32Ty));
11063 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
11066 return new ShuffleVectorInst(LHSSVI->getOperand(0),
11067 LHSSVI->getOperand(1),
11068 ConstantVector::get(Elts));
11073 return MadeChange ? &SVI : 0;
11079 /// TryToSinkInstruction - Try to move the specified instruction from its
11080 /// current block into the beginning of DestBlock, which can only happen if it's
11081 /// safe to move the instruction past all of the instructions between it and the
11082 /// end of its block.
11083 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
11084 assert(I->hasOneUse() && "Invariants didn't hold!");
11086 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
11087 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
11089 // Do not sink alloca instructions out of the entry block.
11090 if (isa<AllocaInst>(I) && I->getParent() ==
11091 &DestBlock->getParent()->getEntryBlock())
11094 // We can only sink load instructions if there is nothing between the load and
11095 // the end of block that could change the value.
11096 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
11097 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
11099 if (Scan->mayWriteToMemory())
11103 BasicBlock::iterator InsertPos = DestBlock->begin();
11104 while (isa<PHINode>(InsertPos)) ++InsertPos;
11106 I->moveBefore(InsertPos);
11112 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
11113 /// all reachable code to the worklist.
11115 /// This has a couple of tricks to make the code faster and more powerful. In
11116 /// particular, we constant fold and DCE instructions as we go, to avoid adding
11117 /// them to the worklist (this significantly speeds up instcombine on code where
11118 /// many instructions are dead or constant). Additionally, if we find a branch
11119 /// whose condition is a known constant, we only visit the reachable successors.
11121 static void AddReachableCodeToWorklist(BasicBlock *BB,
11122 SmallPtrSet<BasicBlock*, 64> &Visited,
11124 const TargetData *TD) {
11125 std::vector<BasicBlock*> Worklist;
11126 Worklist.push_back(BB);
11128 while (!Worklist.empty()) {
11129 BB = Worklist.back();
11130 Worklist.pop_back();
11132 // We have now visited this block! If we've already been here, ignore it.
11133 if (!Visited.insert(BB)) continue;
11135 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
11136 Instruction *Inst = BBI++;
11138 // DCE instruction if trivially dead.
11139 if (isInstructionTriviallyDead(Inst)) {
11141 DOUT << "IC: DCE: " << *Inst;
11142 Inst->eraseFromParent();
11146 // ConstantProp instruction if trivially constant.
11147 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
11148 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
11149 Inst->replaceAllUsesWith(C);
11151 Inst->eraseFromParent();
11155 IC.AddToWorkList(Inst);
11158 // Recursively visit successors. If this is a branch or switch on a
11159 // constant, only visit the reachable successor.
11160 if (BB->getUnwindDest())
11161 Worklist.push_back(BB->getUnwindDest());
11162 TerminatorInst *TI = BB->getTerminator();
11163 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
11164 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
11165 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
11166 BasicBlock *ReachableBB = BI->getSuccessor(!CondVal);
11167 if (ReachableBB != BB->getUnwindDest())
11168 Worklist.push_back(ReachableBB);
11171 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
11172 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
11173 // See if this is an explicit destination.
11174 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
11175 if (SI->getCaseValue(i) == Cond) {
11176 BasicBlock *ReachableBB = SI->getSuccessor(i);
11177 if (ReachableBB != BB->getUnwindDest())
11178 Worklist.push_back(ReachableBB);
11182 // Otherwise it is the default destination.
11183 Worklist.push_back(SI->getSuccessor(0));
11188 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
11189 Worklist.push_back(TI->getSuccessor(i));
11193 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
11194 bool Changed = false;
11195 TD = &getAnalysis<TargetData>();
11197 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
11198 << F.getNameStr() << "\n");
11201 // Do a depth-first traversal of the function, populate the worklist with
11202 // the reachable instructions. Ignore blocks that are not reachable. Keep
11203 // track of which blocks we visit.
11204 SmallPtrSet<BasicBlock*, 64> Visited;
11205 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
11207 // Do a quick scan over the function. If we find any blocks that are
11208 // unreachable, remove any instructions inside of them. This prevents
11209 // the instcombine code from having to deal with some bad special cases.
11210 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
11211 if (!Visited.count(BB)) {
11212 Instruction *Term = BB->getTerminator();
11213 while (Term != BB->begin()) { // Remove instrs bottom-up
11214 BasicBlock::iterator I = Term; --I;
11216 DOUT << "IC: DCE: " << *I;
11219 if (!I->use_empty())
11220 I->replaceAllUsesWith(UndefValue::get(I->getType()));
11221 I->eraseFromParent();
11226 while (!Worklist.empty()) {
11227 Instruction *I = RemoveOneFromWorkList();
11228 if (I == 0) continue; // skip null values.
11230 // Check to see if we can DCE the instruction.
11231 if (isInstructionTriviallyDead(I)) {
11232 // Add operands to the worklist.
11233 if (I->getNumOperands() < 4)
11234 AddUsesToWorkList(*I);
11237 DOUT << "IC: DCE: " << *I;
11239 I->eraseFromParent();
11240 RemoveFromWorkList(I);
11244 // Instruction isn't dead, see if we can constant propagate it.
11245 if (Constant *C = ConstantFoldInstruction(I, TD)) {
11246 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
11248 // Add operands to the worklist.
11249 AddUsesToWorkList(*I);
11250 ReplaceInstUsesWith(*I, C);
11253 I->eraseFromParent();
11254 RemoveFromWorkList(I);
11258 // See if we can trivially sink this instruction to a successor basic block.
11259 if (I->hasOneUse()) {
11260 BasicBlock *BB = I->getParent();
11261 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
11262 if (UserParent != BB) {
11263 bool UserIsSuccessor = false;
11264 // See if the user is one of our successors.
11265 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
11266 if (*SI == UserParent) {
11267 UserIsSuccessor = true;
11271 // If the user is one of our immediate successors, and if that successor
11272 // only has us as a predecessors (we'd have to split the critical edge
11273 // otherwise), we can keep going.
11274 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
11275 next(pred_begin(UserParent)) == pred_end(UserParent))
11276 // Okay, the CFG is simple enough, try to sink this instruction.
11277 Changed |= TryToSinkInstruction(I, UserParent);
11281 // Now that we have an instruction, try combining it to simplify it...
11285 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
11286 if (Instruction *Result = visit(*I)) {
11288 // Should we replace the old instruction with a new one?
11290 DOUT << "IC: Old = " << *I
11291 << " New = " << *Result;
11293 // Everything uses the new instruction now.
11294 I->replaceAllUsesWith(Result);
11296 // Push the new instruction and any users onto the worklist.
11297 AddToWorkList(Result);
11298 AddUsersToWorkList(*Result);
11300 // Move the name to the new instruction first.
11301 Result->takeName(I);
11303 // Insert the new instruction into the basic block...
11304 BasicBlock *InstParent = I->getParent();
11305 BasicBlock::iterator InsertPos = I;
11307 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
11308 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
11311 InstParent->getInstList().insert(InsertPos, Result);
11313 // Make sure that we reprocess all operands now that we reduced their
11315 AddUsesToWorkList(*I);
11317 // Instructions can end up on the worklist more than once. Make sure
11318 // we do not process an instruction that has been deleted.
11319 RemoveFromWorkList(I);
11321 // Erase the old instruction.
11322 InstParent->getInstList().erase(I);
11325 DOUT << "IC: Mod = " << OrigI
11326 << " New = " << *I;
11329 // If the instruction was modified, it's possible that it is now dead.
11330 // if so, remove it.
11331 if (isInstructionTriviallyDead(I)) {
11332 // Make sure we process all operands now that we are reducing their
11334 AddUsesToWorkList(*I);
11336 // Instructions may end up in the worklist more than once. Erase all
11337 // occurrences of this instruction.
11338 RemoveFromWorkList(I);
11339 I->eraseFromParent();
11342 AddUsersToWorkList(*I);
11349 assert(WorklistMap.empty() && "Worklist empty, but map not?");
11351 // Do an explicit clear, this shrinks the map if needed.
11352 WorklistMap.clear();
11357 bool InstCombiner::runOnFunction(Function &F) {
11358 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
11360 bool EverMadeChange = false;
11362 // Iterate while there is work to do.
11363 unsigned Iteration = 0;
11364 while (DoOneIteration(F, Iteration++))
11365 EverMadeChange = true;
11366 return EverMadeChange;
11369 FunctionPass *llvm::createInstructionCombiningPass() {
11370 return new InstCombiner();