1 //===- InstructionCombining.cpp - Combine multiple instructions -----------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source 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/Debug.h"
48 #include "llvm/Support/GetElementPtrTypeIterator.h"
49 #include "llvm/Support/InstVisitor.h"
50 #include "llvm/Support/MathExtras.h"
51 #include "llvm/Support/PatternMatch.h"
52 #include "llvm/Support/Compiler.h"
53 #include "llvm/ADT/DenseMap.h"
54 #include "llvm/ADT/SmallVector.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/ADT/STLExtras.h"
61 using namespace llvm::PatternMatch;
63 STATISTIC(NumCombined , "Number of insts combined");
64 STATISTIC(NumConstProp, "Number of constant folds");
65 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
66 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
67 STATISTIC(NumSunkInst , "Number of instructions sunk");
70 class VISIBILITY_HIDDEN InstCombiner
71 : public FunctionPass,
72 public InstVisitor<InstCombiner, Instruction*> {
73 // Worklist of all of the instructions that need to be simplified.
74 std::vector<Instruction*> Worklist;
75 DenseMap<Instruction*, unsigned> WorklistMap;
77 bool MustPreserveLCSSA;
79 static char ID; // Pass identification, replacement for typeid
80 InstCombiner() : FunctionPass((intptr_t)&ID) {}
82 /// AddToWorkList - Add the specified instruction to the worklist if it
83 /// isn't already in it.
84 void AddToWorkList(Instruction *I) {
85 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
86 Worklist.push_back(I);
89 // RemoveFromWorkList - remove I from the worklist if it exists.
90 void RemoveFromWorkList(Instruction *I) {
91 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
92 if (It == WorklistMap.end()) return; // Not in worklist.
94 // Don't bother moving everything down, just null out the slot.
95 Worklist[It->second] = 0;
97 WorklistMap.erase(It);
100 Instruction *RemoveOneFromWorkList() {
101 Instruction *I = Worklist.back();
103 WorklistMap.erase(I);
108 /// AddUsersToWorkList - When an instruction is simplified, add all users of
109 /// the instruction to the work lists because they might get more simplified
112 void AddUsersToWorkList(Value &I) {
113 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
115 AddToWorkList(cast<Instruction>(*UI));
118 /// AddUsesToWorkList - When an instruction is simplified, add operands to
119 /// the work lists because they might get more simplified now.
121 void AddUsesToWorkList(Instruction &I) {
122 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
123 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
127 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
128 /// dead. Add all of its operands to the worklist, turning them into
129 /// undef's to reduce the number of uses of those instructions.
131 /// Return the specified operand before it is turned into an undef.
133 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
134 Value *R = I.getOperand(op);
136 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
137 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
139 // Set the operand to undef to drop the use.
140 I.setOperand(i, UndefValue::get(Op->getType()));
147 virtual bool runOnFunction(Function &F);
149 bool DoOneIteration(Function &F, unsigned ItNum);
151 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
152 AU.addRequired<TargetData>();
153 AU.addPreservedID(LCSSAID);
154 AU.setPreservesCFG();
157 TargetData &getTargetData() const { return *TD; }
159 // Visitation implementation - Implement instruction combining for different
160 // instruction types. The semantics are as follows:
162 // null - No change was made
163 // I - Change was made, I is still valid, I may be dead though
164 // otherwise - Change was made, replace I with returned instruction
166 Instruction *visitAdd(BinaryOperator &I);
167 Instruction *visitSub(BinaryOperator &I);
168 Instruction *visitMul(BinaryOperator &I);
169 Instruction *visitURem(BinaryOperator &I);
170 Instruction *visitSRem(BinaryOperator &I);
171 Instruction *visitFRem(BinaryOperator &I);
172 Instruction *commonRemTransforms(BinaryOperator &I);
173 Instruction *commonIRemTransforms(BinaryOperator &I);
174 Instruction *commonDivTransforms(BinaryOperator &I);
175 Instruction *commonIDivTransforms(BinaryOperator &I);
176 Instruction *visitUDiv(BinaryOperator &I);
177 Instruction *visitSDiv(BinaryOperator &I);
178 Instruction *visitFDiv(BinaryOperator &I);
179 Instruction *visitAnd(BinaryOperator &I);
180 Instruction *visitOr (BinaryOperator &I);
181 Instruction *visitXor(BinaryOperator &I);
182 Instruction *visitShl(BinaryOperator &I);
183 Instruction *visitAShr(BinaryOperator &I);
184 Instruction *visitLShr(BinaryOperator &I);
185 Instruction *commonShiftTransforms(BinaryOperator &I);
186 Instruction *visitFCmpInst(FCmpInst &I);
187 Instruction *visitICmpInst(ICmpInst &I);
188 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
189 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
192 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
193 ConstantInt *DivRHS);
195 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
196 ICmpInst::Predicate Cond, Instruction &I);
197 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
199 Instruction *commonCastTransforms(CastInst &CI);
200 Instruction *commonIntCastTransforms(CastInst &CI);
201 Instruction *commonPointerCastTransforms(CastInst &CI);
202 Instruction *visitTrunc(TruncInst &CI);
203 Instruction *visitZExt(ZExtInst &CI);
204 Instruction *visitSExt(SExtInst &CI);
205 Instruction *visitFPTrunc(CastInst &CI);
206 Instruction *visitFPExt(CastInst &CI);
207 Instruction *visitFPToUI(CastInst &CI);
208 Instruction *visitFPToSI(CastInst &CI);
209 Instruction *visitUIToFP(CastInst &CI);
210 Instruction *visitSIToFP(CastInst &CI);
211 Instruction *visitPtrToInt(CastInst &CI);
212 Instruction *visitIntToPtr(CastInst &CI);
213 Instruction *visitBitCast(BitCastInst &CI);
214 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
216 Instruction *visitSelectInst(SelectInst &CI);
217 Instruction *visitCallInst(CallInst &CI);
218 Instruction *visitInvokeInst(InvokeInst &II);
219 Instruction *visitPHINode(PHINode &PN);
220 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
221 Instruction *visitAllocationInst(AllocationInst &AI);
222 Instruction *visitFreeInst(FreeInst &FI);
223 Instruction *visitLoadInst(LoadInst &LI);
224 Instruction *visitStoreInst(StoreInst &SI);
225 Instruction *visitBranchInst(BranchInst &BI);
226 Instruction *visitSwitchInst(SwitchInst &SI);
227 Instruction *visitInsertElementInst(InsertElementInst &IE);
228 Instruction *visitExtractElementInst(ExtractElementInst &EI);
229 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
231 // visitInstruction - Specify what to return for unhandled instructions...
232 Instruction *visitInstruction(Instruction &I) { return 0; }
235 Instruction *visitCallSite(CallSite CS);
236 bool transformConstExprCastCall(CallSite CS);
239 // InsertNewInstBefore - insert an instruction New before instruction Old
240 // in the program. Add the new instruction to the worklist.
242 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
243 assert(New && New->getParent() == 0 &&
244 "New instruction already inserted into a basic block!");
245 BasicBlock *BB = Old.getParent();
246 BB->getInstList().insert(&Old, New); // Insert inst
251 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
252 /// This also adds the cast to the worklist. Finally, this returns the
254 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
256 if (V->getType() == Ty) return V;
258 if (Constant *CV = dyn_cast<Constant>(V))
259 return ConstantExpr::getCast(opc, CV, Ty);
261 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
266 // ReplaceInstUsesWith - This method is to be used when an instruction is
267 // found to be dead, replacable with another preexisting expression. Here
268 // we add all uses of I to the worklist, replace all uses of I with the new
269 // value, then return I, so that the inst combiner will know that I was
272 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
273 AddUsersToWorkList(I); // Add all modified instrs to worklist
275 I.replaceAllUsesWith(V);
278 // If we are replacing the instruction with itself, this must be in a
279 // segment of unreachable code, so just clobber the instruction.
280 I.replaceAllUsesWith(UndefValue::get(I.getType()));
285 // UpdateValueUsesWith - This method is to be used when an value is
286 // found to be replacable with another preexisting expression or was
287 // updated. Here we add all uses of I to the worklist, replace all uses of
288 // I with the new value (unless the instruction was just updated), then
289 // return true, so that the inst combiner will know that I was modified.
291 bool UpdateValueUsesWith(Value *Old, Value *New) {
292 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
294 Old->replaceAllUsesWith(New);
295 if (Instruction *I = dyn_cast<Instruction>(Old))
297 if (Instruction *I = dyn_cast<Instruction>(New))
302 // EraseInstFromFunction - When dealing with an instruction that has side
303 // effects or produces a void value, we can't rely on DCE to delete the
304 // instruction. Instead, visit methods should return the value returned by
306 Instruction *EraseInstFromFunction(Instruction &I) {
307 assert(I.use_empty() && "Cannot erase instruction that is used!");
308 AddUsesToWorkList(I);
309 RemoveFromWorkList(&I);
311 return 0; // Don't do anything with FI
315 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
316 /// InsertBefore instruction. This is specialized a bit to avoid inserting
317 /// casts that are known to not do anything...
319 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
320 Value *V, const Type *DestTy,
321 Instruction *InsertBefore);
323 /// SimplifyCommutative - This performs a few simplifications for
324 /// commutative operators.
325 bool SimplifyCommutative(BinaryOperator &I);
327 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
328 /// most-complex to least-complex order.
329 bool SimplifyCompare(CmpInst &I);
331 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
332 /// on the demanded bits.
333 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
334 APInt& KnownZero, APInt& KnownOne,
337 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
338 uint64_t &UndefElts, unsigned Depth = 0);
340 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
341 // PHI node as operand #0, see if we can fold the instruction into the PHI
342 // (which is only possible if all operands to the PHI are constants).
343 Instruction *FoldOpIntoPhi(Instruction &I);
345 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
346 // operator and they all are only used by the PHI, PHI together their
347 // inputs, and do the operation once, to the result of the PHI.
348 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
349 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
352 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
353 ConstantInt *AndRHS, BinaryOperator &TheAnd);
355 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
356 bool isSub, Instruction &I);
357 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
358 bool isSigned, bool Inside, Instruction &IB);
359 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
360 Instruction *MatchBSwap(BinaryOperator &I);
361 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
363 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
366 char InstCombiner::ID = 0;
367 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
370 // getComplexity: Assign a complexity or rank value to LLVM Values...
371 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
372 static unsigned getComplexity(Value *V) {
373 if (isa<Instruction>(V)) {
374 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
378 if (isa<Argument>(V)) return 3;
379 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
382 // isOnlyUse - Return true if this instruction will be deleted if we stop using
384 static bool isOnlyUse(Value *V) {
385 return V->hasOneUse() || isa<Constant>(V);
388 // getPromotedType - Return the specified type promoted as it would be to pass
389 // though a va_arg area...
390 static const Type *getPromotedType(const Type *Ty) {
391 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
392 if (ITy->getBitWidth() < 32)
393 return Type::Int32Ty;
398 /// getBitCastOperand - If the specified operand is a CastInst or a constant
399 /// expression bitcast, return the operand value, otherwise return null.
400 static Value *getBitCastOperand(Value *V) {
401 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
402 return I->getOperand(0);
403 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
404 if (CE->getOpcode() == Instruction::BitCast)
405 return CE->getOperand(0);
409 /// This function is a wrapper around CastInst::isEliminableCastPair. It
410 /// simply extracts arguments and returns what that function returns.
411 static Instruction::CastOps
412 isEliminableCastPair(
413 const CastInst *CI, ///< The first cast instruction
414 unsigned opcode, ///< The opcode of the second cast instruction
415 const Type *DstTy, ///< The target type for the second cast instruction
416 TargetData *TD ///< The target data for pointer size
419 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
420 const Type *MidTy = CI->getType(); // B from above
422 // Get the opcodes of the two Cast instructions
423 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
424 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
426 return Instruction::CastOps(
427 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
428 DstTy, TD->getIntPtrType()));
431 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
432 /// in any code being generated. It does not require codegen if V is simple
433 /// enough or if the cast can be folded into other casts.
434 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
435 const Type *Ty, TargetData *TD) {
436 if (V->getType() == Ty || isa<Constant>(V)) return false;
438 // If this is another cast that can be eliminated, it isn't codegen either.
439 if (const CastInst *CI = dyn_cast<CastInst>(V))
440 if (isEliminableCastPair(CI, opcode, Ty, TD))
445 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
446 /// InsertBefore instruction. This is specialized a bit to avoid inserting
447 /// casts that are known to not do anything...
449 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
450 Value *V, const Type *DestTy,
451 Instruction *InsertBefore) {
452 if (V->getType() == DestTy) return V;
453 if (Constant *C = dyn_cast<Constant>(V))
454 return ConstantExpr::getCast(opcode, C, DestTy);
456 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
459 // SimplifyCommutative - This performs a few simplifications for commutative
462 // 1. Order operands such that they are listed from right (least complex) to
463 // left (most complex). This puts constants before unary operators before
466 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
467 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
469 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
470 bool Changed = false;
471 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
472 Changed = !I.swapOperands();
474 if (!I.isAssociative()) return Changed;
475 Instruction::BinaryOps Opcode = I.getOpcode();
476 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
477 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
478 if (isa<Constant>(I.getOperand(1))) {
479 Constant *Folded = ConstantExpr::get(I.getOpcode(),
480 cast<Constant>(I.getOperand(1)),
481 cast<Constant>(Op->getOperand(1)));
482 I.setOperand(0, Op->getOperand(0));
483 I.setOperand(1, Folded);
485 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
486 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
487 isOnlyUse(Op) && isOnlyUse(Op1)) {
488 Constant *C1 = cast<Constant>(Op->getOperand(1));
489 Constant *C2 = cast<Constant>(Op1->getOperand(1));
491 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
492 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
493 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
497 I.setOperand(0, New);
498 I.setOperand(1, Folded);
505 /// SimplifyCompare - For a CmpInst this function just orders the operands
506 /// so that theyare listed from right (least complex) to left (most complex).
507 /// This puts constants before unary operators before binary operators.
508 bool InstCombiner::SimplifyCompare(CmpInst &I) {
509 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
512 // Compare instructions are not associative so there's nothing else we can do.
516 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
517 // if the LHS is a constant zero (which is the 'negate' form).
519 static inline Value *dyn_castNegVal(Value *V) {
520 if (BinaryOperator::isNeg(V))
521 return BinaryOperator::getNegArgument(V);
523 // Constants can be considered to be negated values if they can be folded.
524 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
525 return ConstantExpr::getNeg(C);
529 static inline Value *dyn_castNotVal(Value *V) {
530 if (BinaryOperator::isNot(V))
531 return BinaryOperator::getNotArgument(V);
533 // Constants can be considered to be not'ed values...
534 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
535 return ConstantInt::get(~C->getValue());
539 // dyn_castFoldableMul - If this value is a multiply that can be folded into
540 // other computations (because it has a constant operand), return the
541 // non-constant operand of the multiply, and set CST to point to the multiplier.
542 // Otherwise, return null.
544 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
545 if (V->hasOneUse() && V->getType()->isInteger())
546 if (Instruction *I = dyn_cast<Instruction>(V)) {
547 if (I->getOpcode() == Instruction::Mul)
548 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
549 return I->getOperand(0);
550 if (I->getOpcode() == Instruction::Shl)
551 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
552 // The multiplier is really 1 << CST.
553 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
554 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
555 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
556 return I->getOperand(0);
562 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
563 /// expression, return it.
564 static User *dyn_castGetElementPtr(Value *V) {
565 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
566 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
567 if (CE->getOpcode() == Instruction::GetElementPtr)
568 return cast<User>(V);
572 /// AddOne - Add one to a ConstantInt
573 static ConstantInt *AddOne(ConstantInt *C) {
574 APInt Val(C->getValue());
575 return ConstantInt::get(++Val);
577 /// SubOne - Subtract one from a ConstantInt
578 static ConstantInt *SubOne(ConstantInt *C) {
579 APInt Val(C->getValue());
580 return ConstantInt::get(--Val);
582 /// Add - Add two ConstantInts together
583 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
584 return ConstantInt::get(C1->getValue() + C2->getValue());
586 /// And - Bitwise AND two ConstantInts together
587 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
588 return ConstantInt::get(C1->getValue() & C2->getValue());
590 /// Subtract - Subtract one ConstantInt from another
591 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
592 return ConstantInt::get(C1->getValue() - C2->getValue());
594 /// Multiply - Multiply two ConstantInts together
595 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
596 return ConstantInt::get(C1->getValue() * C2->getValue());
599 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
600 /// known to be either zero or one and return them in the KnownZero/KnownOne
601 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
603 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
604 /// we cannot optimize based on the assumption that it is zero without changing
605 /// it to be an explicit zero. If we don't change it to zero, other code could
606 /// optimized based on the contradictory assumption that it is non-zero.
607 /// Because instcombine aggressively folds operations with undef args anyway,
608 /// this won't lose us code quality.
609 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
610 APInt& KnownOne, unsigned Depth = 0) {
611 assert(V && "No Value?");
612 assert(Depth <= 6 && "Limit Search Depth");
613 uint32_t BitWidth = Mask.getBitWidth();
614 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
615 KnownZero.getBitWidth() == BitWidth &&
616 KnownOne.getBitWidth() == BitWidth &&
617 "V, Mask, KnownOne and KnownZero should have same BitWidth");
618 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
619 // We know all of the bits for a constant!
620 KnownOne = CI->getValue() & Mask;
621 KnownZero = ~KnownOne & Mask;
625 if (Depth == 6 || Mask == 0)
626 return; // Limit search depth.
628 Instruction *I = dyn_cast<Instruction>(V);
631 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
632 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
634 switch (I->getOpcode()) {
635 case Instruction::And: {
636 // If either the LHS or the RHS are Zero, the result is zero.
637 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
638 APInt Mask2(Mask & ~KnownZero);
639 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
640 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
641 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
643 // Output known-1 bits are only known if set in both the LHS & RHS.
644 KnownOne &= KnownOne2;
645 // Output known-0 are known to be clear if zero in either the LHS | RHS.
646 KnownZero |= KnownZero2;
649 case Instruction::Or: {
650 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
651 APInt Mask2(Mask & ~KnownOne);
652 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
653 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
654 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
656 // Output known-0 bits are only known if clear in both the LHS & RHS.
657 KnownZero &= KnownZero2;
658 // Output known-1 are known to be set if set in either the LHS | RHS.
659 KnownOne |= KnownOne2;
662 case Instruction::Xor: {
663 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
664 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
665 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
666 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
668 // Output known-0 bits are known if clear or set in both the LHS & RHS.
669 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
670 // Output known-1 are known to be set if set in only one of the LHS, RHS.
671 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
672 KnownZero = KnownZeroOut;
675 case Instruction::Select:
676 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
677 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
678 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
679 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
681 // Only known if known in both the LHS and RHS.
682 KnownOne &= KnownOne2;
683 KnownZero &= KnownZero2;
685 case Instruction::FPTrunc:
686 case Instruction::FPExt:
687 case Instruction::FPToUI:
688 case Instruction::FPToSI:
689 case Instruction::SIToFP:
690 case Instruction::PtrToInt:
691 case Instruction::UIToFP:
692 case Instruction::IntToPtr:
693 return; // Can't work with floating point or pointers
694 case Instruction::Trunc: {
695 // All these have integer operands
696 uint32_t SrcBitWidth =
697 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
699 MaskIn.zext(SrcBitWidth);
700 KnownZero.zext(SrcBitWidth);
701 KnownOne.zext(SrcBitWidth);
702 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
703 KnownZero.trunc(BitWidth);
704 KnownOne.trunc(BitWidth);
707 case Instruction::BitCast: {
708 const Type *SrcTy = I->getOperand(0)->getType();
709 if (SrcTy->isInteger()) {
710 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
715 case Instruction::ZExt: {
716 // Compute the bits in the result that are not present in the input.
717 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
718 uint32_t SrcBitWidth = SrcTy->getBitWidth();
721 MaskIn.trunc(SrcBitWidth);
722 KnownZero.trunc(SrcBitWidth);
723 KnownOne.trunc(SrcBitWidth);
724 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
725 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
726 // The top bits are known to be zero.
727 KnownZero.zext(BitWidth);
728 KnownOne.zext(BitWidth);
729 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
732 case Instruction::SExt: {
733 // Compute the bits in the result that are not present in the input.
734 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
735 uint32_t SrcBitWidth = SrcTy->getBitWidth();
738 MaskIn.trunc(SrcBitWidth);
739 KnownZero.trunc(SrcBitWidth);
740 KnownOne.trunc(SrcBitWidth);
741 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
742 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
743 KnownZero.zext(BitWidth);
744 KnownOne.zext(BitWidth);
746 // If the sign bit of the input is known set or clear, then we know the
747 // top bits of the result.
748 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
749 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
750 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
751 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
754 case Instruction::Shl:
755 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
756 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
757 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
758 APInt Mask2(Mask.lshr(ShiftAmt));
759 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
760 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
761 KnownZero <<= ShiftAmt;
762 KnownOne <<= ShiftAmt;
763 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
767 case Instruction::LShr:
768 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
769 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
770 // Compute the new bits that are at the top now.
771 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
773 // Unsigned shift right.
774 APInt Mask2(Mask.shl(ShiftAmt));
775 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
776 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
777 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
778 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
779 // high bits known zero.
780 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
784 case Instruction::AShr:
785 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
786 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
787 // Compute the new bits that are at the top now.
788 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
790 // Signed shift right.
791 APInt Mask2(Mask.shl(ShiftAmt));
792 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
793 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
794 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
795 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
797 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
798 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
799 KnownZero |= HighBits;
800 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
801 KnownOne |= HighBits;
808 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
809 /// this predicate to simplify operations downstream. Mask is known to be zero
810 /// for bits that V cannot have.
811 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
812 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
813 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
814 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
815 return (KnownZero & Mask) == Mask;
818 /// ShrinkDemandedConstant - Check to see if the specified operand of the
819 /// specified instruction is a constant integer. If so, check to see if there
820 /// are any bits set in the constant that are not demanded. If so, shrink the
821 /// constant and return true.
822 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
824 assert(I && "No instruction?");
825 assert(OpNo < I->getNumOperands() && "Operand index too large");
827 // If the operand is not a constant integer, nothing to do.
828 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
829 if (!OpC) return false;
831 // If there are no bits set that aren't demanded, nothing to do.
832 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
833 if ((~Demanded & OpC->getValue()) == 0)
836 // This instruction is producing bits that are not demanded. Shrink the RHS.
837 Demanded &= OpC->getValue();
838 I->setOperand(OpNo, ConstantInt::get(Demanded));
842 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
843 // set of known zero and one bits, compute the maximum and minimum values that
844 // could have the specified known zero and known one bits, returning them in
846 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
847 const APInt& KnownZero,
848 const APInt& KnownOne,
849 APInt& Min, APInt& Max) {
850 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
851 assert(KnownZero.getBitWidth() == BitWidth &&
852 KnownOne.getBitWidth() == BitWidth &&
853 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
854 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
855 APInt UnknownBits = ~(KnownZero|KnownOne);
857 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
858 // bit if it is unknown.
860 Max = KnownOne|UnknownBits;
862 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
864 Max.clear(BitWidth-1);
868 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
869 // a set of known zero and one bits, compute the maximum and minimum values that
870 // could have the specified known zero and known one bits, returning them in
872 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
873 const APInt& KnownZero,
874 const APInt& KnownOne,
877 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
878 assert(KnownZero.getBitWidth() == BitWidth &&
879 KnownOne.getBitWidth() == BitWidth &&
880 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
881 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
882 APInt UnknownBits = ~(KnownZero|KnownOne);
884 // The minimum value is when the unknown bits are all zeros.
886 // The maximum value is when the unknown bits are all ones.
887 Max = KnownOne|UnknownBits;
890 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
891 /// value based on the demanded bits. When this function is called, it is known
892 /// that only the bits set in DemandedMask of the result of V are ever used
893 /// downstream. Consequently, depending on the mask and V, it may be possible
894 /// to replace V with a constant or one of its operands. In such cases, this
895 /// function does the replacement and returns true. In all other cases, it
896 /// returns false after analyzing the expression and setting KnownOne and known
897 /// to be one in the expression. KnownZero contains all the bits that are known
898 /// to be zero in the expression. These are provided to potentially allow the
899 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
900 /// the expression. KnownOne and KnownZero always follow the invariant that
901 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
902 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
903 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
904 /// and KnownOne must all be the same.
905 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
906 APInt& KnownZero, APInt& KnownOne,
908 assert(V != 0 && "Null pointer of Value???");
909 assert(Depth <= 6 && "Limit Search Depth");
910 uint32_t BitWidth = DemandedMask.getBitWidth();
911 const IntegerType *VTy = cast<IntegerType>(V->getType());
912 assert(VTy->getBitWidth() == BitWidth &&
913 KnownZero.getBitWidth() == BitWidth &&
914 KnownOne.getBitWidth() == BitWidth &&
915 "Value *V, DemandedMask, KnownZero and KnownOne \
916 must have same BitWidth");
917 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
918 // We know all of the bits for a constant!
919 KnownOne = CI->getValue() & DemandedMask;
920 KnownZero = ~KnownOne & DemandedMask;
926 if (!V->hasOneUse()) { // Other users may use these bits.
927 if (Depth != 0) { // Not at the root.
928 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
929 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
932 // If this is the root being simplified, allow it to have multiple uses,
933 // just set the DemandedMask to all bits.
934 DemandedMask = APInt::getAllOnesValue(BitWidth);
935 } else if (DemandedMask == 0) { // Not demanding any bits from V.
936 if (V != UndefValue::get(VTy))
937 return UpdateValueUsesWith(V, UndefValue::get(VTy));
939 } else if (Depth == 6) { // Limit search depth.
943 Instruction *I = dyn_cast<Instruction>(V);
944 if (!I) return false; // Only analyze instructions.
946 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
947 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
948 switch (I->getOpcode()) {
950 case Instruction::And:
951 // If either the LHS or the RHS are Zero, the result is zero.
952 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
953 RHSKnownZero, RHSKnownOne, Depth+1))
955 assert((RHSKnownZero & RHSKnownOne) == 0 &&
956 "Bits known to be one AND zero?");
958 // If something is known zero on the RHS, the bits aren't demanded on the
960 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
961 LHSKnownZero, LHSKnownOne, Depth+1))
963 assert((LHSKnownZero & LHSKnownOne) == 0 &&
964 "Bits known to be one AND zero?");
966 // If all of the demanded bits are known 1 on one side, return the other.
967 // These bits cannot contribute to the result of the 'and'.
968 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
969 (DemandedMask & ~LHSKnownZero))
970 return UpdateValueUsesWith(I, I->getOperand(0));
971 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
972 (DemandedMask & ~RHSKnownZero))
973 return UpdateValueUsesWith(I, I->getOperand(1));
975 // If all of the demanded bits in the inputs are known zeros, return zero.
976 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
977 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
979 // If the RHS is a constant, see if we can simplify it.
980 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
981 return UpdateValueUsesWith(I, I);
983 // Output known-1 bits are only known if set in both the LHS & RHS.
984 RHSKnownOne &= LHSKnownOne;
985 // Output known-0 are known to be clear if zero in either the LHS | RHS.
986 RHSKnownZero |= LHSKnownZero;
988 case Instruction::Or:
989 // If either the LHS or the RHS are One, the result is One.
990 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
991 RHSKnownZero, RHSKnownOne, Depth+1))
993 assert((RHSKnownZero & RHSKnownOne) == 0 &&
994 "Bits known to be one AND zero?");
995 // If something is known one on the RHS, the bits aren't demanded on the
997 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
998 LHSKnownZero, LHSKnownOne, Depth+1))
1000 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1001 "Bits known to be one AND zero?");
1003 // If all of the demanded bits are known zero on one side, return the other.
1004 // These bits cannot contribute to the result of the 'or'.
1005 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1006 (DemandedMask & ~LHSKnownOne))
1007 return UpdateValueUsesWith(I, I->getOperand(0));
1008 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1009 (DemandedMask & ~RHSKnownOne))
1010 return UpdateValueUsesWith(I, I->getOperand(1));
1012 // If all of the potentially set bits on one side are known to be set on
1013 // the other side, just use the 'other' side.
1014 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1015 (DemandedMask & (~RHSKnownZero)))
1016 return UpdateValueUsesWith(I, I->getOperand(0));
1017 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1018 (DemandedMask & (~LHSKnownZero)))
1019 return UpdateValueUsesWith(I, I->getOperand(1));
1021 // If the RHS is a constant, see if we can simplify it.
1022 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1023 return UpdateValueUsesWith(I, I);
1025 // Output known-0 bits are only known if clear in both the LHS & RHS.
1026 RHSKnownZero &= LHSKnownZero;
1027 // Output known-1 are known to be set if set in either the LHS | RHS.
1028 RHSKnownOne |= LHSKnownOne;
1030 case Instruction::Xor: {
1031 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1032 RHSKnownZero, RHSKnownOne, Depth+1))
1034 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1035 "Bits known to be one AND zero?");
1036 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1037 LHSKnownZero, LHSKnownOne, Depth+1))
1039 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1040 "Bits known to be one AND zero?");
1042 // If all of the demanded bits are known zero on one side, return the other.
1043 // These bits cannot contribute to the result of the 'xor'.
1044 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1045 return UpdateValueUsesWith(I, I->getOperand(0));
1046 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1047 return UpdateValueUsesWith(I, I->getOperand(1));
1049 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1050 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1051 (RHSKnownOne & LHSKnownOne);
1052 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1053 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1054 (RHSKnownOne & LHSKnownZero);
1056 // If all of the demanded bits are known to be zero on one side or the
1057 // other, turn this into an *inclusive* or.
1058 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1059 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1061 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1063 InsertNewInstBefore(Or, *I);
1064 return UpdateValueUsesWith(I, Or);
1067 // If all of the demanded bits on one side are known, and all of the set
1068 // bits on that side are also known to be set on the other side, turn this
1069 // into an AND, as we know the bits will be cleared.
1070 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1071 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1073 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1074 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1076 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1077 InsertNewInstBefore(And, *I);
1078 return UpdateValueUsesWith(I, And);
1082 // If the RHS is a constant, see if we can simplify it.
1083 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1084 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1085 return UpdateValueUsesWith(I, I);
1087 RHSKnownZero = KnownZeroOut;
1088 RHSKnownOne = KnownOneOut;
1091 case Instruction::Select:
1092 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1093 RHSKnownZero, RHSKnownOne, Depth+1))
1095 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1096 LHSKnownZero, LHSKnownOne, Depth+1))
1098 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1099 "Bits known to be one AND zero?");
1100 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1101 "Bits known to be one AND zero?");
1103 // If the operands are constants, see if we can simplify them.
1104 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1105 return UpdateValueUsesWith(I, I);
1106 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1107 return UpdateValueUsesWith(I, I);
1109 // Only known if known in both the LHS and RHS.
1110 RHSKnownOne &= LHSKnownOne;
1111 RHSKnownZero &= LHSKnownZero;
1113 case Instruction::Trunc: {
1115 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1116 DemandedMask.zext(truncBf);
1117 RHSKnownZero.zext(truncBf);
1118 RHSKnownOne.zext(truncBf);
1119 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1120 RHSKnownZero, RHSKnownOne, Depth+1))
1122 DemandedMask.trunc(BitWidth);
1123 RHSKnownZero.trunc(BitWidth);
1124 RHSKnownOne.trunc(BitWidth);
1125 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1126 "Bits known to be one AND zero?");
1129 case Instruction::BitCast:
1130 if (!I->getOperand(0)->getType()->isInteger())
1133 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1134 RHSKnownZero, RHSKnownOne, Depth+1))
1136 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1137 "Bits known to be one AND zero?");
1139 case Instruction::ZExt: {
1140 // Compute the bits in the result that are not present in the input.
1141 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1142 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1144 DemandedMask.trunc(SrcBitWidth);
1145 RHSKnownZero.trunc(SrcBitWidth);
1146 RHSKnownOne.trunc(SrcBitWidth);
1147 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1148 RHSKnownZero, RHSKnownOne, Depth+1))
1150 DemandedMask.zext(BitWidth);
1151 RHSKnownZero.zext(BitWidth);
1152 RHSKnownOne.zext(BitWidth);
1153 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1154 "Bits known to be one AND zero?");
1155 // The top bits are known to be zero.
1156 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1159 case Instruction::SExt: {
1160 // Compute the bits in the result that are not present in the input.
1161 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1162 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1164 APInt InputDemandedBits = DemandedMask &
1165 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1167 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1168 // If any of the sign extended bits are demanded, we know that the sign
1170 if ((NewBits & DemandedMask) != 0)
1171 InputDemandedBits.set(SrcBitWidth-1);
1173 InputDemandedBits.trunc(SrcBitWidth);
1174 RHSKnownZero.trunc(SrcBitWidth);
1175 RHSKnownOne.trunc(SrcBitWidth);
1176 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1177 RHSKnownZero, RHSKnownOne, Depth+1))
1179 InputDemandedBits.zext(BitWidth);
1180 RHSKnownZero.zext(BitWidth);
1181 RHSKnownOne.zext(BitWidth);
1182 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1183 "Bits known to be one AND zero?");
1185 // If the sign bit of the input is known set or clear, then we know the
1186 // top bits of the result.
1188 // If the input sign bit is known zero, or if the NewBits are not demanded
1189 // convert this into a zero extension.
1190 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1192 // Convert to ZExt cast
1193 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1194 return UpdateValueUsesWith(I, NewCast);
1195 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1196 RHSKnownOne |= NewBits;
1200 case Instruction::Add: {
1201 // Figure out what the input bits are. If the top bits of the and result
1202 // are not demanded, then the add doesn't demand them from its input
1204 uint32_t NLZ = DemandedMask.countLeadingZeros();
1206 // If there is a constant on the RHS, there are a variety of xformations
1208 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1209 // If null, this should be simplified elsewhere. Some of the xforms here
1210 // won't work if the RHS is zero.
1214 // If the top bit of the output is demanded, demand everything from the
1215 // input. Otherwise, we demand all the input bits except NLZ top bits.
1216 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1218 // Find information about known zero/one bits in the input.
1219 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1220 LHSKnownZero, LHSKnownOne, Depth+1))
1223 // If the RHS of the add has bits set that can't affect the input, reduce
1225 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1226 return UpdateValueUsesWith(I, I);
1228 // Avoid excess work.
1229 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1232 // Turn it into OR if input bits are zero.
1233 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1235 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1237 InsertNewInstBefore(Or, *I);
1238 return UpdateValueUsesWith(I, Or);
1241 // We can say something about the output known-zero and known-one bits,
1242 // depending on potential carries from the input constant and the
1243 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1244 // bits set and the RHS constant is 0x01001, then we know we have a known
1245 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1247 // To compute this, we first compute the potential carry bits. These are
1248 // the bits which may be modified. I'm not aware of a better way to do
1250 const APInt& RHSVal = RHS->getValue();
1251 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1253 // Now that we know which bits have carries, compute the known-1/0 sets.
1255 // Bits are known one if they are known zero in one operand and one in the
1256 // other, and there is no input carry.
1257 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1258 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1260 // Bits are known zero if they are known zero in both operands and there
1261 // is no input carry.
1262 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1264 // If the high-bits of this ADD are not demanded, then it does not demand
1265 // the high bits of its LHS or RHS.
1266 if (DemandedMask[BitWidth-1] == 0) {
1267 // Right fill the mask of bits for this ADD to demand the most
1268 // significant bit and all those below it.
1269 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1270 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1271 LHSKnownZero, LHSKnownOne, Depth+1))
1273 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1274 LHSKnownZero, LHSKnownOne, Depth+1))
1280 case Instruction::Sub:
1281 // If the high-bits of this SUB are not demanded, then it does not demand
1282 // the high bits of its LHS or RHS.
1283 if (DemandedMask[BitWidth-1] == 0) {
1284 // Right fill the mask of bits for this SUB to demand the most
1285 // significant bit and all those below it.
1286 uint32_t NLZ = DemandedMask.countLeadingZeros();
1287 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1288 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1289 LHSKnownZero, LHSKnownOne, Depth+1))
1291 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1292 LHSKnownZero, LHSKnownOne, Depth+1))
1296 case Instruction::Shl:
1297 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1298 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1299 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1300 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1301 RHSKnownZero, RHSKnownOne, Depth+1))
1303 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1304 "Bits known to be one AND zero?");
1305 RHSKnownZero <<= ShiftAmt;
1306 RHSKnownOne <<= ShiftAmt;
1307 // low bits known zero.
1309 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1312 case Instruction::LShr:
1313 // For a logical shift right
1314 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1315 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1317 // Unsigned shift right.
1318 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1319 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1320 RHSKnownZero, RHSKnownOne, Depth+1))
1322 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1323 "Bits known to be one AND zero?");
1324 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1325 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1327 // Compute the new bits that are at the top now.
1328 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1329 RHSKnownZero |= HighBits; // high bits known zero.
1333 case Instruction::AShr:
1334 // If this is an arithmetic shift right and only the low-bit is set, we can
1335 // always convert this into a logical shr, even if the shift amount is
1336 // variable. The low bit of the shift cannot be an input sign bit unless
1337 // the shift amount is >= the size of the datatype, which is undefined.
1338 if (DemandedMask == 1) {
1339 // Perform the logical shift right.
1340 Value *NewVal = BinaryOperator::createLShr(
1341 I->getOperand(0), I->getOperand(1), I->getName());
1342 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1343 return UpdateValueUsesWith(I, NewVal);
1346 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1347 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1349 // Signed shift right.
1350 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1351 // If any of the "high bits" are demanded, we should set the sign bit as
1353 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1354 DemandedMaskIn.set(BitWidth-1);
1355 if (SimplifyDemandedBits(I->getOperand(0),
1357 RHSKnownZero, RHSKnownOne, Depth+1))
1359 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1360 "Bits known to be one AND zero?");
1361 // Compute the new bits that are at the top now.
1362 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1363 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1364 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1366 // Handle the sign bits.
1367 APInt SignBit(APInt::getSignBit(BitWidth));
1368 // Adjust to where it is now in the mask.
1369 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1371 // If the input sign bit is known to be zero, or if none of the top bits
1372 // are demanded, turn this into an unsigned shift right.
1373 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1374 (HighBits & ~DemandedMask) == HighBits) {
1375 // Perform the logical shift right.
1376 Value *NewVal = BinaryOperator::createLShr(
1377 I->getOperand(0), SA, I->getName());
1378 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1379 return UpdateValueUsesWith(I, NewVal);
1380 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1381 RHSKnownOne |= HighBits;
1387 // If the client is only demanding bits that we know, return the known
1389 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1390 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1395 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1396 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1397 /// actually used by the caller. This method analyzes which elements of the
1398 /// operand are undef and returns that information in UndefElts.
1400 /// If the information about demanded elements can be used to simplify the
1401 /// operation, the operation is simplified, then the resultant value is
1402 /// returned. This returns null if no change was made.
1403 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1404 uint64_t &UndefElts,
1406 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1407 assert(VWidth <= 64 && "Vector too wide to analyze!");
1408 uint64_t EltMask = ~0ULL >> (64-VWidth);
1409 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1410 "Invalid DemandedElts!");
1412 if (isa<UndefValue>(V)) {
1413 // If the entire vector is undefined, just return this info.
1414 UndefElts = EltMask;
1416 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1417 UndefElts = EltMask;
1418 return UndefValue::get(V->getType());
1422 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1423 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1424 Constant *Undef = UndefValue::get(EltTy);
1426 std::vector<Constant*> Elts;
1427 for (unsigned i = 0; i != VWidth; ++i)
1428 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1429 Elts.push_back(Undef);
1430 UndefElts |= (1ULL << i);
1431 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1432 Elts.push_back(Undef);
1433 UndefElts |= (1ULL << i);
1434 } else { // Otherwise, defined.
1435 Elts.push_back(CP->getOperand(i));
1438 // If we changed the constant, return it.
1439 Constant *NewCP = ConstantVector::get(Elts);
1440 return NewCP != CP ? NewCP : 0;
1441 } else if (isa<ConstantAggregateZero>(V)) {
1442 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1444 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1445 Constant *Zero = Constant::getNullValue(EltTy);
1446 Constant *Undef = UndefValue::get(EltTy);
1447 std::vector<Constant*> Elts;
1448 for (unsigned i = 0; i != VWidth; ++i)
1449 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1450 UndefElts = DemandedElts ^ EltMask;
1451 return ConstantVector::get(Elts);
1454 if (!V->hasOneUse()) { // Other users may use these bits.
1455 if (Depth != 0) { // Not at the root.
1456 // TODO: Just compute the UndefElts information recursively.
1460 } else if (Depth == 10) { // Limit search depth.
1464 Instruction *I = dyn_cast<Instruction>(V);
1465 if (!I) return false; // Only analyze instructions.
1467 bool MadeChange = false;
1468 uint64_t UndefElts2;
1470 switch (I->getOpcode()) {
1473 case Instruction::InsertElement: {
1474 // If this is a variable index, we don't know which element it overwrites.
1475 // demand exactly the same input as we produce.
1476 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1478 // Note that we can't propagate undef elt info, because we don't know
1479 // which elt is getting updated.
1480 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1481 UndefElts2, Depth+1);
1482 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1486 // If this is inserting an element that isn't demanded, remove this
1488 unsigned IdxNo = Idx->getZExtValue();
1489 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1490 return AddSoonDeadInstToWorklist(*I, 0);
1492 // Otherwise, the element inserted overwrites whatever was there, so the
1493 // input demanded set is simpler than the output set.
1494 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1495 DemandedElts & ~(1ULL << IdxNo),
1496 UndefElts, Depth+1);
1497 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1499 // The inserted element is defined.
1500 UndefElts |= 1ULL << IdxNo;
1503 case Instruction::BitCast: {
1504 // Packed->packed casts only.
1505 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1507 unsigned InVWidth = VTy->getNumElements();
1508 uint64_t InputDemandedElts = 0;
1511 if (VWidth == InVWidth) {
1512 // If we are converting from <4x i32> -> <4 x f32>, we demand the same
1513 // elements as are demanded of us.
1515 InputDemandedElts = DemandedElts;
1516 } else if (VWidth > InVWidth) {
1520 // If there are more elements in the result than there are in the source,
1521 // then an input element is live if any of the corresponding output
1522 // elements are live.
1523 Ratio = VWidth/InVWidth;
1524 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1525 if (DemandedElts & (1ULL << OutIdx))
1526 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1532 // If there are more elements in the source than there are in the result,
1533 // then an input element is live if the corresponding output element is
1535 Ratio = InVWidth/VWidth;
1536 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1537 if (DemandedElts & (1ULL << InIdx/Ratio))
1538 InputDemandedElts |= 1ULL << InIdx;
1541 // div/rem demand all inputs, because they don't want divide by zero.
1542 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1543 UndefElts2, Depth+1);
1545 I->setOperand(0, TmpV);
1549 UndefElts = UndefElts2;
1550 if (VWidth > InVWidth) {
1551 assert(0 && "Unimp");
1552 // If there are more elements in the result than there are in the source,
1553 // then an output element is undef if the corresponding input element is
1555 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1556 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1557 UndefElts |= 1ULL << OutIdx;
1558 } else if (VWidth < InVWidth) {
1559 assert(0 && "Unimp");
1560 // If there are more elements in the source than there are in the result,
1561 // then a result element is undef if all of the corresponding input
1562 // elements are undef.
1563 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1564 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1565 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1566 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1570 case Instruction::And:
1571 case Instruction::Or:
1572 case Instruction::Xor:
1573 case Instruction::Add:
1574 case Instruction::Sub:
1575 case Instruction::Mul:
1576 // div/rem demand all inputs, because they don't want divide by zero.
1577 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1578 UndefElts, Depth+1);
1579 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1580 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1581 UndefElts2, Depth+1);
1582 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1584 // Output elements are undefined if both are undefined. Consider things
1585 // like undef&0. The result is known zero, not undef.
1586 UndefElts &= UndefElts2;
1589 case Instruction::Call: {
1590 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1592 switch (II->getIntrinsicID()) {
1595 // Binary vector operations that work column-wise. A dest element is a
1596 // function of the corresponding input elements from the two inputs.
1597 case Intrinsic::x86_sse_sub_ss:
1598 case Intrinsic::x86_sse_mul_ss:
1599 case Intrinsic::x86_sse_min_ss:
1600 case Intrinsic::x86_sse_max_ss:
1601 case Intrinsic::x86_sse2_sub_sd:
1602 case Intrinsic::x86_sse2_mul_sd:
1603 case Intrinsic::x86_sse2_min_sd:
1604 case Intrinsic::x86_sse2_max_sd:
1605 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1606 UndefElts, Depth+1);
1607 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1608 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1609 UndefElts2, Depth+1);
1610 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1612 // If only the low elt is demanded and this is a scalarizable intrinsic,
1613 // scalarize it now.
1614 if (DemandedElts == 1) {
1615 switch (II->getIntrinsicID()) {
1617 case Intrinsic::x86_sse_sub_ss:
1618 case Intrinsic::x86_sse_mul_ss:
1619 case Intrinsic::x86_sse2_sub_sd:
1620 case Intrinsic::x86_sse2_mul_sd:
1621 // TODO: Lower MIN/MAX/ABS/etc
1622 Value *LHS = II->getOperand(1);
1623 Value *RHS = II->getOperand(2);
1624 // Extract the element as scalars.
1625 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1626 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1628 switch (II->getIntrinsicID()) {
1629 default: assert(0 && "Case stmts out of sync!");
1630 case Intrinsic::x86_sse_sub_ss:
1631 case Intrinsic::x86_sse2_sub_sd:
1632 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1633 II->getName()), *II);
1635 case Intrinsic::x86_sse_mul_ss:
1636 case Intrinsic::x86_sse2_mul_sd:
1637 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1638 II->getName()), *II);
1643 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1645 InsertNewInstBefore(New, *II);
1646 AddSoonDeadInstToWorklist(*II, 0);
1651 // Output elements are undefined if both are undefined. Consider things
1652 // like undef&0. The result is known zero, not undef.
1653 UndefElts &= UndefElts2;
1659 return MadeChange ? I : 0;
1662 /// @returns true if the specified compare instruction is
1663 /// true when both operands are equal...
1664 /// @brief Determine if the ICmpInst returns true if both operands are equal
1665 static bool isTrueWhenEqual(ICmpInst &ICI) {
1666 ICmpInst::Predicate pred = ICI.getPredicate();
1667 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1668 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1669 pred == ICmpInst::ICMP_SLE;
1672 /// AssociativeOpt - Perform an optimization on an associative operator. This
1673 /// function is designed to check a chain of associative operators for a
1674 /// potential to apply a certain optimization. Since the optimization may be
1675 /// applicable if the expression was reassociated, this checks the chain, then
1676 /// reassociates the expression as necessary to expose the optimization
1677 /// opportunity. This makes use of a special Functor, which must define
1678 /// 'shouldApply' and 'apply' methods.
1680 template<typename Functor>
1681 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1682 unsigned Opcode = Root.getOpcode();
1683 Value *LHS = Root.getOperand(0);
1685 // Quick check, see if the immediate LHS matches...
1686 if (F.shouldApply(LHS))
1687 return F.apply(Root);
1689 // Otherwise, if the LHS is not of the same opcode as the root, return.
1690 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1691 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1692 // Should we apply this transform to the RHS?
1693 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1695 // If not to the RHS, check to see if we should apply to the LHS...
1696 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1697 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1701 // If the functor wants to apply the optimization to the RHS of LHSI,
1702 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1704 BasicBlock *BB = Root.getParent();
1706 // Now all of the instructions are in the current basic block, go ahead
1707 // and perform the reassociation.
1708 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1710 // First move the selected RHS to the LHS of the root...
1711 Root.setOperand(0, LHSI->getOperand(1));
1713 // Make what used to be the LHS of the root be the user of the root...
1714 Value *ExtraOperand = TmpLHSI->getOperand(1);
1715 if (&Root == TmpLHSI) {
1716 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1719 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1720 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1721 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1722 BasicBlock::iterator ARI = &Root; ++ARI;
1723 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1726 // Now propagate the ExtraOperand down the chain of instructions until we
1728 while (TmpLHSI != LHSI) {
1729 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1730 // Move the instruction to immediately before the chain we are
1731 // constructing to avoid breaking dominance properties.
1732 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1733 BB->getInstList().insert(ARI, NextLHSI);
1736 Value *NextOp = NextLHSI->getOperand(1);
1737 NextLHSI->setOperand(1, ExtraOperand);
1739 ExtraOperand = NextOp;
1742 // Now that the instructions are reassociated, have the functor perform
1743 // the transformation...
1744 return F.apply(Root);
1747 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1753 // AddRHS - Implements: X + X --> X << 1
1756 AddRHS(Value *rhs) : RHS(rhs) {}
1757 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1758 Instruction *apply(BinaryOperator &Add) const {
1759 return BinaryOperator::createShl(Add.getOperand(0),
1760 ConstantInt::get(Add.getType(), 1));
1764 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1766 struct AddMaskingAnd {
1768 AddMaskingAnd(Constant *c) : C2(c) {}
1769 bool shouldApply(Value *LHS) const {
1771 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1772 ConstantExpr::getAnd(C1, C2)->isNullValue();
1774 Instruction *apply(BinaryOperator &Add) const {
1775 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1779 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1781 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1782 if (Constant *SOC = dyn_cast<Constant>(SO))
1783 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1785 return IC->InsertNewInstBefore(CastInst::create(
1786 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1789 // Figure out if the constant is the left or the right argument.
1790 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1791 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1793 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1795 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1796 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1799 Value *Op0 = SO, *Op1 = ConstOperand;
1801 std::swap(Op0, Op1);
1803 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1804 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1805 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1806 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1807 SO->getName()+".cmp");
1809 assert(0 && "Unknown binary instruction type!");
1812 return IC->InsertNewInstBefore(New, I);
1815 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1816 // constant as the other operand, try to fold the binary operator into the
1817 // select arguments. This also works for Cast instructions, which obviously do
1818 // not have a second operand.
1819 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1821 // Don't modify shared select instructions
1822 if (!SI->hasOneUse()) return 0;
1823 Value *TV = SI->getOperand(1);
1824 Value *FV = SI->getOperand(2);
1826 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1827 // Bool selects with constant operands can be folded to logical ops.
1828 if (SI->getType() == Type::Int1Ty) return 0;
1830 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1831 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1833 return new SelectInst(SI->getCondition(), SelectTrueVal,
1840 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1841 /// node as operand #0, see if we can fold the instruction into the PHI (which
1842 /// is only possible if all operands to the PHI are constants).
1843 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1844 PHINode *PN = cast<PHINode>(I.getOperand(0));
1845 unsigned NumPHIValues = PN->getNumIncomingValues();
1846 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1848 // Check to see if all of the operands of the PHI are constants. If there is
1849 // one non-constant value, remember the BB it is. If there is more than one
1850 // or if *it* is a PHI, bail out.
1851 BasicBlock *NonConstBB = 0;
1852 for (unsigned i = 0; i != NumPHIValues; ++i)
1853 if (!isa<Constant>(PN->getIncomingValue(i))) {
1854 if (NonConstBB) return 0; // More than one non-const value.
1855 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1856 NonConstBB = PN->getIncomingBlock(i);
1858 // If the incoming non-constant value is in I's block, we have an infinite
1860 if (NonConstBB == I.getParent())
1864 // If there is exactly one non-constant value, we can insert a copy of the
1865 // operation in that block. However, if this is a critical edge, we would be
1866 // inserting the computation one some other paths (e.g. inside a loop). Only
1867 // do this if the pred block is unconditionally branching into the phi block.
1869 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1870 if (!BI || !BI->isUnconditional()) return 0;
1873 // Okay, we can do the transformation: create the new PHI node.
1874 PHINode *NewPN = new PHINode(I.getType(), "");
1875 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1876 InsertNewInstBefore(NewPN, *PN);
1877 NewPN->takeName(PN);
1879 // Next, add all of the operands to the PHI.
1880 if (I.getNumOperands() == 2) {
1881 Constant *C = cast<Constant>(I.getOperand(1));
1882 for (unsigned i = 0; i != NumPHIValues; ++i) {
1884 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1885 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1886 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1888 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1890 assert(PN->getIncomingBlock(i) == NonConstBB);
1891 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1892 InV = BinaryOperator::create(BO->getOpcode(),
1893 PN->getIncomingValue(i), C, "phitmp",
1894 NonConstBB->getTerminator());
1895 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1896 InV = CmpInst::create(CI->getOpcode(),
1898 PN->getIncomingValue(i), C, "phitmp",
1899 NonConstBB->getTerminator());
1901 assert(0 && "Unknown binop!");
1903 AddToWorkList(cast<Instruction>(InV));
1905 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1908 CastInst *CI = cast<CastInst>(&I);
1909 const Type *RetTy = CI->getType();
1910 for (unsigned i = 0; i != NumPHIValues; ++i) {
1912 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1913 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1915 assert(PN->getIncomingBlock(i) == NonConstBB);
1916 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1917 I.getType(), "phitmp",
1918 NonConstBB->getTerminator());
1919 AddToWorkList(cast<Instruction>(InV));
1921 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1924 return ReplaceInstUsesWith(I, NewPN);
1927 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1928 bool Changed = SimplifyCommutative(I);
1929 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1931 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1932 // X + undef -> undef
1933 if (isa<UndefValue>(RHS))
1934 return ReplaceInstUsesWith(I, RHS);
1937 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1938 if (RHSC->isNullValue())
1939 return ReplaceInstUsesWith(I, LHS);
1940 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1941 if (CFP->isExactlyValue(-0.0))
1942 return ReplaceInstUsesWith(I, LHS);
1945 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1946 // X + (signbit) --> X ^ signbit
1947 const APInt& Val = CI->getValue();
1948 uint32_t BitWidth = Val.getBitWidth();
1949 if (Val == APInt::getSignBit(BitWidth))
1950 return BinaryOperator::createXor(LHS, RHS);
1952 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1953 // (X & 254)+1 -> (X&254)|1
1954 if (!isa<VectorType>(I.getType())) {
1955 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1956 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1957 KnownZero, KnownOne))
1962 if (isa<PHINode>(LHS))
1963 if (Instruction *NV = FoldOpIntoPhi(I))
1966 ConstantInt *XorRHS = 0;
1968 if (isa<ConstantInt>(RHSC) &&
1969 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1970 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1971 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1973 uint32_t Size = TySizeBits / 2;
1974 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1975 APInt CFF80Val(-C0080Val);
1977 if (TySizeBits > Size) {
1978 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1979 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1980 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1981 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1982 // This is a sign extend if the top bits are known zero.
1983 if (!MaskedValueIsZero(XorLHS,
1984 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1985 Size = 0; // Not a sign ext, but can't be any others either.
1990 C0080Val = APIntOps::lshr(C0080Val, Size);
1991 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1992 } while (Size >= 1);
1994 // FIXME: This shouldn't be necessary. When the backends can handle types
1995 // with funny bit widths then this whole cascade of if statements should
1996 // be removed. It is just here to get the size of the "middle" type back
1997 // up to something that the back ends can handle.
1998 const Type *MiddleType = 0;
2001 case 32: MiddleType = Type::Int32Ty; break;
2002 case 16: MiddleType = Type::Int16Ty; break;
2003 case 8: MiddleType = Type::Int8Ty; break;
2006 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2007 InsertNewInstBefore(NewTrunc, I);
2008 return new SExtInst(NewTrunc, I.getType(), I.getName());
2014 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2015 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2017 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2018 if (RHSI->getOpcode() == Instruction::Sub)
2019 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2020 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2022 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2023 if (LHSI->getOpcode() == Instruction::Sub)
2024 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2025 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2030 if (Value *V = dyn_castNegVal(LHS))
2031 return BinaryOperator::createSub(RHS, V);
2034 if (!isa<Constant>(RHS))
2035 if (Value *V = dyn_castNegVal(RHS))
2036 return BinaryOperator::createSub(LHS, V);
2040 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2041 if (X == RHS) // X*C + X --> X * (C+1)
2042 return BinaryOperator::createMul(RHS, AddOne(C2));
2044 // X*C1 + X*C2 --> X * (C1+C2)
2046 if (X == dyn_castFoldableMul(RHS, C1))
2047 return BinaryOperator::createMul(X, Add(C1, C2));
2050 // X + X*C --> X * (C+1)
2051 if (dyn_castFoldableMul(RHS, C2) == LHS)
2052 return BinaryOperator::createMul(LHS, AddOne(C2));
2054 // X + ~X --> -1 since ~X = -X-1
2055 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2056 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2059 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2060 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2061 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2064 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2066 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2067 return BinaryOperator::createSub(SubOne(CRHS), X);
2069 // (X & FF00) + xx00 -> (X+xx00) & FF00
2070 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2071 Constant *Anded = And(CRHS, C2);
2072 if (Anded == CRHS) {
2073 // See if all bits from the first bit set in the Add RHS up are included
2074 // in the mask. First, get the rightmost bit.
2075 const APInt& AddRHSV = CRHS->getValue();
2077 // Form a mask of all bits from the lowest bit added through the top.
2078 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2080 // See if the and mask includes all of these bits.
2081 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2083 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2084 // Okay, the xform is safe. Insert the new add pronto.
2085 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2086 LHS->getName()), I);
2087 return BinaryOperator::createAnd(NewAdd, C2);
2092 // Try to fold constant add into select arguments.
2093 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2094 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2098 // add (cast *A to intptrtype) B ->
2099 // cast (GEP (cast *A to sbyte*) B) ->
2102 CastInst *CI = dyn_cast<CastInst>(LHS);
2105 CI = dyn_cast<CastInst>(RHS);
2108 if (CI && CI->getType()->isSized() &&
2109 (CI->getType()->getPrimitiveSizeInBits() ==
2110 TD->getIntPtrType()->getPrimitiveSizeInBits())
2111 && isa<PointerType>(CI->getOperand(0)->getType())) {
2112 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2113 PointerType::get(Type::Int8Ty), I);
2114 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2115 return new PtrToIntInst(I2, CI->getType());
2119 return Changed ? &I : 0;
2122 // isSignBit - Return true if the value represented by the constant only has the
2123 // highest order bit set.
2124 static bool isSignBit(ConstantInt *CI) {
2125 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2126 return CI->getValue() == APInt::getSignBit(NumBits);
2129 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2130 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2132 if (Op0 == Op1) // sub X, X -> 0
2133 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2135 // If this is a 'B = x-(-A)', change to B = x+A...
2136 if (Value *V = dyn_castNegVal(Op1))
2137 return BinaryOperator::createAdd(Op0, V);
2139 if (isa<UndefValue>(Op0))
2140 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2141 if (isa<UndefValue>(Op1))
2142 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2144 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2145 // Replace (-1 - A) with (~A)...
2146 if (C->isAllOnesValue())
2147 return BinaryOperator::createNot(Op1);
2149 // C - ~X == X + (1+C)
2151 if (match(Op1, m_Not(m_Value(X))))
2152 return BinaryOperator::createAdd(X, AddOne(C));
2154 // -(X >>u 31) -> (X >>s 31)
2155 // -(X >>s 31) -> (X >>u 31)
2157 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2158 if (SI->getOpcode() == Instruction::LShr) {
2159 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2160 // Check to see if we are shifting out everything but the sign bit.
2161 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2162 SI->getType()->getPrimitiveSizeInBits()-1) {
2163 // Ok, the transformation is safe. Insert AShr.
2164 return BinaryOperator::create(Instruction::AShr,
2165 SI->getOperand(0), CU, SI->getName());
2169 else if (SI->getOpcode() == Instruction::AShr) {
2170 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2171 // Check to see if we are shifting out everything but the sign bit.
2172 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2173 SI->getType()->getPrimitiveSizeInBits()-1) {
2174 // Ok, the transformation is safe. Insert LShr.
2175 return BinaryOperator::createLShr(
2176 SI->getOperand(0), CU, SI->getName());
2182 // Try to fold constant sub into select arguments.
2183 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2184 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2187 if (isa<PHINode>(Op0))
2188 if (Instruction *NV = FoldOpIntoPhi(I))
2192 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2193 if (Op1I->getOpcode() == Instruction::Add &&
2194 !Op0->getType()->isFPOrFPVector()) {
2195 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2196 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2197 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2198 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2199 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2200 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2201 // C1-(X+C2) --> (C1-C2)-X
2202 return BinaryOperator::createSub(Subtract(CI1, CI2),
2203 Op1I->getOperand(0));
2207 if (Op1I->hasOneUse()) {
2208 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2209 // is not used by anyone else...
2211 if (Op1I->getOpcode() == Instruction::Sub &&
2212 !Op1I->getType()->isFPOrFPVector()) {
2213 // Swap the two operands of the subexpr...
2214 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2215 Op1I->setOperand(0, IIOp1);
2216 Op1I->setOperand(1, IIOp0);
2218 // Create the new top level add instruction...
2219 return BinaryOperator::createAdd(Op0, Op1);
2222 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2224 if (Op1I->getOpcode() == Instruction::And &&
2225 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2226 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2229 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2230 return BinaryOperator::createAnd(Op0, NewNot);
2233 // 0 - (X sdiv C) -> (X sdiv -C)
2234 if (Op1I->getOpcode() == Instruction::SDiv)
2235 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2237 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2238 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2239 ConstantExpr::getNeg(DivRHS));
2241 // X - X*C --> X * (1-C)
2242 ConstantInt *C2 = 0;
2243 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2244 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2245 return BinaryOperator::createMul(Op0, CP1);
2250 if (!Op0->getType()->isFPOrFPVector())
2251 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2252 if (Op0I->getOpcode() == Instruction::Add) {
2253 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2254 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2255 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2256 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2257 } else if (Op0I->getOpcode() == Instruction::Sub) {
2258 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2259 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2263 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2264 if (X == Op1) // X*C - X --> X * (C-1)
2265 return BinaryOperator::createMul(Op1, SubOne(C1));
2267 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2268 if (X == dyn_castFoldableMul(Op1, C2))
2269 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2274 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2275 /// really just returns true if the most significant (sign) bit is set.
2276 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2278 case ICmpInst::ICMP_SLT:
2279 // True if LHS s< RHS and RHS == 0
2280 return RHS->isZero();
2281 case ICmpInst::ICMP_SLE:
2282 // True if LHS s<= RHS and RHS == -1
2283 return RHS->isAllOnesValue();
2284 case ICmpInst::ICMP_UGE:
2285 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2286 return RHS->getValue() ==
2287 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2288 case ICmpInst::ICMP_UGT:
2289 // True if LHS u> RHS and RHS == high-bit-mask - 1
2290 return RHS->getValue() ==
2291 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2297 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2298 bool Changed = SimplifyCommutative(I);
2299 Value *Op0 = I.getOperand(0);
2301 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2302 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2304 // Simplify mul instructions with a constant RHS...
2305 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2306 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2308 // ((X << C1)*C2) == (X * (C2 << C1))
2309 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2310 if (SI->getOpcode() == Instruction::Shl)
2311 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2312 return BinaryOperator::createMul(SI->getOperand(0),
2313 ConstantExpr::getShl(CI, ShOp));
2316 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2317 if (CI->equalsInt(1)) // X * 1 == X
2318 return ReplaceInstUsesWith(I, Op0);
2319 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2320 return BinaryOperator::createNeg(Op0, I.getName());
2322 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2323 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2324 return BinaryOperator::createShl(Op0,
2325 ConstantInt::get(Op0->getType(), Val.logBase2()));
2327 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2328 if (Op1F->isNullValue())
2329 return ReplaceInstUsesWith(I, Op1);
2331 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2332 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2333 if (Op1F->getValue() == 1.0)
2334 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2337 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2338 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2339 isa<ConstantInt>(Op0I->getOperand(1))) {
2340 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2341 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2343 InsertNewInstBefore(Add, I);
2344 Value *C1C2 = ConstantExpr::getMul(Op1,
2345 cast<Constant>(Op0I->getOperand(1)));
2346 return BinaryOperator::createAdd(Add, C1C2);
2350 // Try to fold constant mul into select arguments.
2351 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2352 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2355 if (isa<PHINode>(Op0))
2356 if (Instruction *NV = FoldOpIntoPhi(I))
2360 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2361 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2362 return BinaryOperator::createMul(Op0v, Op1v);
2364 // If one of the operands of the multiply is a cast from a boolean value, then
2365 // we know the bool is either zero or one, so this is a 'masking' multiply.
2366 // See if we can simplify things based on how the boolean was originally
2368 CastInst *BoolCast = 0;
2369 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2370 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2373 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2374 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2377 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2378 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2379 const Type *SCOpTy = SCIOp0->getType();
2381 // If the icmp is true iff the sign bit of X is set, then convert this
2382 // multiply into a shift/and combination.
2383 if (isa<ConstantInt>(SCIOp1) &&
2384 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2385 // Shift the X value right to turn it into "all signbits".
2386 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2387 SCOpTy->getPrimitiveSizeInBits()-1);
2389 InsertNewInstBefore(
2390 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2391 BoolCast->getOperand(0)->getName()+
2394 // If the multiply type is not the same as the source type, sign extend
2395 // or truncate to the multiply type.
2396 if (I.getType() != V->getType()) {
2397 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2398 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2399 Instruction::CastOps opcode =
2400 (SrcBits == DstBits ? Instruction::BitCast :
2401 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2402 V = InsertCastBefore(opcode, V, I.getType(), I);
2405 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2406 return BinaryOperator::createAnd(V, OtherOp);
2411 return Changed ? &I : 0;
2414 /// This function implements the transforms on div instructions that work
2415 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2416 /// used by the visitors to those instructions.
2417 /// @brief Transforms common to all three div instructions
2418 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2419 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2422 if (isa<UndefValue>(Op0))
2423 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2425 // X / undef -> undef
2426 if (isa<UndefValue>(Op1))
2427 return ReplaceInstUsesWith(I, Op1);
2429 // Handle cases involving: div X, (select Cond, Y, Z)
2430 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2431 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2432 // same basic block, then we replace the select with Y, and the condition
2433 // of the select with false (if the cond value is in the same BB). If the
2434 // select has uses other than the div, this allows them to be simplified
2435 // also. Note that div X, Y is just as good as div X, 0 (undef)
2436 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2437 if (ST->isNullValue()) {
2438 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2439 if (CondI && CondI->getParent() == I.getParent())
2440 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2441 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2442 I.setOperand(1, SI->getOperand(2));
2444 UpdateValueUsesWith(SI, SI->getOperand(2));
2448 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2449 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2450 if (ST->isNullValue()) {
2451 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2452 if (CondI && CondI->getParent() == I.getParent())
2453 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2454 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2455 I.setOperand(1, SI->getOperand(1));
2457 UpdateValueUsesWith(SI, SI->getOperand(1));
2465 /// This function implements the transforms common to both integer division
2466 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2467 /// division instructions.
2468 /// @brief Common integer divide transforms
2469 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2470 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2472 if (Instruction *Common = commonDivTransforms(I))
2475 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2477 if (RHS->equalsInt(1))
2478 return ReplaceInstUsesWith(I, Op0);
2480 // (X / C1) / C2 -> X / (C1*C2)
2481 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2482 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2483 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2484 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2485 Multiply(RHS, LHSRHS));
2488 if (!RHS->isZero()) { // avoid X udiv 0
2489 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2490 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2492 if (isa<PHINode>(Op0))
2493 if (Instruction *NV = FoldOpIntoPhi(I))
2498 // 0 / X == 0, we don't need to preserve faults!
2499 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2500 if (LHS->equalsInt(0))
2501 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2506 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2507 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2509 // Handle the integer div common cases
2510 if (Instruction *Common = commonIDivTransforms(I))
2513 // X udiv C^2 -> X >> C
2514 // Check to see if this is an unsigned division with an exact power of 2,
2515 // if so, convert to a right shift.
2516 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2517 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2518 return BinaryOperator::createLShr(Op0,
2519 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2522 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2523 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2524 if (RHSI->getOpcode() == Instruction::Shl &&
2525 isa<ConstantInt>(RHSI->getOperand(0))) {
2526 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2527 if (C1.isPowerOf2()) {
2528 Value *N = RHSI->getOperand(1);
2529 const Type *NTy = N->getType();
2530 if (uint32_t C2 = C1.logBase2()) {
2531 Constant *C2V = ConstantInt::get(NTy, C2);
2532 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2534 return BinaryOperator::createLShr(Op0, N);
2539 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2540 // where C1&C2 are powers of two.
2541 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2542 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2543 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2544 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2545 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2546 // Compute the shift amounts
2547 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2548 // Construct the "on true" case of the select
2549 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2550 Instruction *TSI = BinaryOperator::createLShr(
2551 Op0, TC, SI->getName()+".t");
2552 TSI = InsertNewInstBefore(TSI, I);
2554 // Construct the "on false" case of the select
2555 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2556 Instruction *FSI = BinaryOperator::createLShr(
2557 Op0, FC, SI->getName()+".f");
2558 FSI = InsertNewInstBefore(FSI, I);
2560 // construct the select instruction and return it.
2561 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2567 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2568 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2570 // Handle the integer div common cases
2571 if (Instruction *Common = commonIDivTransforms(I))
2574 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2576 if (RHS->isAllOnesValue())
2577 return BinaryOperator::createNeg(Op0);
2580 if (Value *LHSNeg = dyn_castNegVal(Op0))
2581 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2584 // If the sign bits of both operands are zero (i.e. we can prove they are
2585 // unsigned inputs), turn this into a udiv.
2586 if (I.getType()->isInteger()) {
2587 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2588 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2589 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2596 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2597 return commonDivTransforms(I);
2600 /// GetFactor - If we can prove that the specified value is at least a multiple
2601 /// of some factor, return that factor.
2602 static Constant *GetFactor(Value *V) {
2603 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2606 // Unless we can be tricky, we know this is a multiple of 1.
2607 Constant *Result = ConstantInt::get(V->getType(), 1);
2609 Instruction *I = dyn_cast<Instruction>(V);
2610 if (!I) return Result;
2612 if (I->getOpcode() == Instruction::Mul) {
2613 // Handle multiplies by a constant, etc.
2614 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2615 GetFactor(I->getOperand(1)));
2616 } else if (I->getOpcode() == Instruction::Shl) {
2617 // (X<<C) -> X * (1 << C)
2618 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2619 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2620 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2622 } else if (I->getOpcode() == Instruction::And) {
2623 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2624 // X & 0xFFF0 is known to be a multiple of 16.
2625 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2626 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2627 return ConstantExpr::getShl(Result,
2628 ConstantInt::get(Result->getType(), Zeros));
2630 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2631 // Only handle int->int casts.
2632 if (!CI->isIntegerCast())
2634 Value *Op = CI->getOperand(0);
2635 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2640 /// This function implements the transforms on rem instructions that work
2641 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2642 /// is used by the visitors to those instructions.
2643 /// @brief Transforms common to all three rem instructions
2644 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2645 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2647 // 0 % X == 0, we don't need to preserve faults!
2648 if (Constant *LHS = dyn_cast<Constant>(Op0))
2649 if (LHS->isNullValue())
2650 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2652 if (isa<UndefValue>(Op0)) // undef % X -> 0
2653 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2654 if (isa<UndefValue>(Op1))
2655 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2657 // Handle cases involving: rem X, (select Cond, Y, Z)
2658 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2659 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2660 // the same basic block, then we replace the select with Y, and the
2661 // condition of the select with false (if the cond value is in the same
2662 // BB). If the select has uses other than the div, this allows them to be
2664 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2665 if (ST->isNullValue()) {
2666 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2667 if (CondI && CondI->getParent() == I.getParent())
2668 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2669 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2670 I.setOperand(1, SI->getOperand(2));
2672 UpdateValueUsesWith(SI, SI->getOperand(2));
2675 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2676 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2677 if (ST->isNullValue()) {
2678 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2679 if (CondI && CondI->getParent() == I.getParent())
2680 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2681 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2682 I.setOperand(1, SI->getOperand(1));
2684 UpdateValueUsesWith(SI, SI->getOperand(1));
2692 /// This function implements the transforms common to both integer remainder
2693 /// instructions (urem and srem). It is called by the visitors to those integer
2694 /// remainder instructions.
2695 /// @brief Common integer remainder transforms
2696 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2697 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2699 if (Instruction *common = commonRemTransforms(I))
2702 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2703 // X % 0 == undef, we don't need to preserve faults!
2704 if (RHS->equalsInt(0))
2705 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2707 if (RHS->equalsInt(1)) // X % 1 == 0
2708 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2710 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2711 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2712 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2714 } else if (isa<PHINode>(Op0I)) {
2715 if (Instruction *NV = FoldOpIntoPhi(I))
2718 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2719 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2720 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2727 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2728 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2730 if (Instruction *common = commonIRemTransforms(I))
2733 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2734 // X urem C^2 -> X and C
2735 // Check to see if this is an unsigned remainder with an exact power of 2,
2736 // if so, convert to a bitwise and.
2737 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2738 if (C->getValue().isPowerOf2())
2739 return BinaryOperator::createAnd(Op0, SubOne(C));
2742 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2743 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2744 if (RHSI->getOpcode() == Instruction::Shl &&
2745 isa<ConstantInt>(RHSI->getOperand(0))) {
2746 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2747 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2748 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2750 return BinaryOperator::createAnd(Op0, Add);
2755 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2756 // where C1&C2 are powers of two.
2757 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2758 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2759 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2760 // STO == 0 and SFO == 0 handled above.
2761 if ((STO->getValue().isPowerOf2()) &&
2762 (SFO->getValue().isPowerOf2())) {
2763 Value *TrueAnd = InsertNewInstBefore(
2764 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2765 Value *FalseAnd = InsertNewInstBefore(
2766 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2767 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2775 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2776 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2778 if (Instruction *common = commonIRemTransforms(I))
2781 if (Value *RHSNeg = dyn_castNegVal(Op1))
2782 if (!isa<ConstantInt>(RHSNeg) ||
2783 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2785 AddUsesToWorkList(I);
2786 I.setOperand(1, RHSNeg);
2790 // If the top bits of both operands are zero (i.e. we can prove they are
2791 // unsigned inputs), turn this into a urem.
2792 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2793 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2794 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2795 return BinaryOperator::createURem(Op0, Op1, I.getName());
2801 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2802 return commonRemTransforms(I);
2805 // isMaxValueMinusOne - return true if this is Max-1
2806 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2807 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2809 // Calculate 0111111111..11111
2810 APInt Val(APInt::getSignedMaxValue(TypeBits));
2811 return C->getValue() == Val-1;
2813 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2816 // isMinValuePlusOne - return true if this is Min+1
2817 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2819 // Calculate 1111111111000000000000
2820 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2821 APInt Val(APInt::getSignedMinValue(TypeBits));
2822 return C->getValue() == Val+1;
2824 return C->getValue() == 1; // unsigned
2827 // isOneBitSet - Return true if there is exactly one bit set in the specified
2829 static bool isOneBitSet(const ConstantInt *CI) {
2830 return CI->getValue().isPowerOf2();
2833 // isHighOnes - Return true if the constant is of the form 1+0+.
2834 // This is the same as lowones(~X).
2835 static bool isHighOnes(const ConstantInt *CI) {
2836 return (~CI->getValue() + 1).isPowerOf2();
2839 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2840 /// are carefully arranged to allow folding of expressions such as:
2842 /// (A < B) | (A > B) --> (A != B)
2844 /// Note that this is only valid if the first and second predicates have the
2845 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2847 /// Three bits are used to represent the condition, as follows:
2852 /// <=> Value Definition
2853 /// 000 0 Always false
2860 /// 111 7 Always true
2862 static unsigned getICmpCode(const ICmpInst *ICI) {
2863 switch (ICI->getPredicate()) {
2865 case ICmpInst::ICMP_UGT: return 1; // 001
2866 case ICmpInst::ICMP_SGT: return 1; // 001
2867 case ICmpInst::ICMP_EQ: return 2; // 010
2868 case ICmpInst::ICMP_UGE: return 3; // 011
2869 case ICmpInst::ICMP_SGE: return 3; // 011
2870 case ICmpInst::ICMP_ULT: return 4; // 100
2871 case ICmpInst::ICMP_SLT: return 4; // 100
2872 case ICmpInst::ICMP_NE: return 5; // 101
2873 case ICmpInst::ICMP_ULE: return 6; // 110
2874 case ICmpInst::ICMP_SLE: return 6; // 110
2877 assert(0 && "Invalid ICmp predicate!");
2882 /// getICmpValue - This is the complement of getICmpCode, which turns an
2883 /// opcode and two operands into either a constant true or false, or a brand
2884 /// new /// ICmp instruction. The sign is passed in to determine which kind
2885 /// of predicate to use in new icmp instructions.
2886 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2888 default: assert(0 && "Illegal ICmp code!");
2889 case 0: return ConstantInt::getFalse();
2892 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2894 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2895 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2898 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2900 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2903 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2905 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2906 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2909 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2911 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2912 case 7: return ConstantInt::getTrue();
2916 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2917 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2918 (ICmpInst::isSignedPredicate(p1) &&
2919 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2920 (ICmpInst::isSignedPredicate(p2) &&
2921 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2925 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2926 struct FoldICmpLogical {
2929 ICmpInst::Predicate pred;
2930 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2931 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2932 pred(ICI->getPredicate()) {}
2933 bool shouldApply(Value *V) const {
2934 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2935 if (PredicatesFoldable(pred, ICI->getPredicate()))
2936 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2937 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2940 Instruction *apply(Instruction &Log) const {
2941 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2942 if (ICI->getOperand(0) != LHS) {
2943 assert(ICI->getOperand(1) == LHS);
2944 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2947 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2948 unsigned LHSCode = getICmpCode(ICI);
2949 unsigned RHSCode = getICmpCode(RHSICI);
2951 switch (Log.getOpcode()) {
2952 case Instruction::And: Code = LHSCode & RHSCode; break;
2953 case Instruction::Or: Code = LHSCode | RHSCode; break;
2954 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2955 default: assert(0 && "Illegal logical opcode!"); return 0;
2958 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2959 ICmpInst::isSignedPredicate(ICI->getPredicate());
2961 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2962 if (Instruction *I = dyn_cast<Instruction>(RV))
2964 // Otherwise, it's a constant boolean value...
2965 return IC.ReplaceInstUsesWith(Log, RV);
2968 } // end anonymous namespace
2970 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2971 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2972 // guaranteed to be a binary operator.
2973 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2975 ConstantInt *AndRHS,
2976 BinaryOperator &TheAnd) {
2977 Value *X = Op->getOperand(0);
2978 Constant *Together = 0;
2980 Together = And(AndRHS, OpRHS);
2982 switch (Op->getOpcode()) {
2983 case Instruction::Xor:
2984 if (Op->hasOneUse()) {
2985 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2986 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2987 InsertNewInstBefore(And, TheAnd);
2989 return BinaryOperator::createXor(And, Together);
2992 case Instruction::Or:
2993 if (Together == AndRHS) // (X | C) & C --> C
2994 return ReplaceInstUsesWith(TheAnd, AndRHS);
2996 if (Op->hasOneUse() && Together != OpRHS) {
2997 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2998 Instruction *Or = BinaryOperator::createOr(X, Together);
2999 InsertNewInstBefore(Or, TheAnd);
3001 return BinaryOperator::createAnd(Or, AndRHS);
3004 case Instruction::Add:
3005 if (Op->hasOneUse()) {
3006 // Adding a one to a single bit bit-field should be turned into an XOR
3007 // of the bit. First thing to check is to see if this AND is with a
3008 // single bit constant.
3009 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3011 // If there is only one bit set...
3012 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3013 // Ok, at this point, we know that we are masking the result of the
3014 // ADD down to exactly one bit. If the constant we are adding has
3015 // no bits set below this bit, then we can eliminate the ADD.
3016 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3018 // Check to see if any bits below the one bit set in AndRHSV are set.
3019 if ((AddRHS & (AndRHSV-1)) == 0) {
3020 // If not, the only thing that can effect the output of the AND is
3021 // the bit specified by AndRHSV. If that bit is set, the effect of
3022 // the XOR is to toggle the bit. If it is clear, then the ADD has
3024 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3025 TheAnd.setOperand(0, X);
3028 // Pull the XOR out of the AND.
3029 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3030 InsertNewInstBefore(NewAnd, TheAnd);
3031 NewAnd->takeName(Op);
3032 return BinaryOperator::createXor(NewAnd, AndRHS);
3039 case Instruction::Shl: {
3040 // We know that the AND will not produce any of the bits shifted in, so if
3041 // the anded constant includes them, clear them now!
3043 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3044 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3045 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3046 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3048 if (CI->getValue() == ShlMask) {
3049 // Masking out bits that the shift already masks
3050 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3051 } else if (CI != AndRHS) { // Reducing bits set in and.
3052 TheAnd.setOperand(1, CI);
3057 case Instruction::LShr:
3059 // We know that the AND will not produce any of the bits shifted in, so if
3060 // the anded constant includes them, clear them now! This only applies to
3061 // unsigned shifts, because a signed shr may bring in set bits!
3063 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3064 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3065 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3066 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3068 if (CI->getValue() == ShrMask) {
3069 // Masking out bits that the shift already masks.
3070 return ReplaceInstUsesWith(TheAnd, Op);
3071 } else if (CI != AndRHS) {
3072 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3077 case Instruction::AShr:
3079 // See if this is shifting in some sign extension, then masking it out
3081 if (Op->hasOneUse()) {
3082 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3083 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3084 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3085 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3086 if (C == AndRHS) { // Masking out bits shifted in.
3087 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3088 // Make the argument unsigned.
3089 Value *ShVal = Op->getOperand(0);
3090 ShVal = InsertNewInstBefore(
3091 BinaryOperator::createLShr(ShVal, OpRHS,
3092 Op->getName()), TheAnd);
3093 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3102 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3103 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3104 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3105 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3106 /// insert new instructions.
3107 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3108 bool isSigned, bool Inside,
3110 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3111 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3112 "Lo is not <= Hi in range emission code!");
3115 if (Lo == Hi) // Trivially false.
3116 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3118 // V >= Min && V < Hi --> V < Hi
3119 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3120 ICmpInst::Predicate pred = (isSigned ?
3121 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3122 return new ICmpInst(pred, V, Hi);
3125 // Emit V-Lo <u Hi-Lo
3126 Constant *NegLo = ConstantExpr::getNeg(Lo);
3127 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3128 InsertNewInstBefore(Add, IB);
3129 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3130 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3133 if (Lo == Hi) // Trivially true.
3134 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3136 // V < Min || V >= Hi -> V > Hi-1
3137 Hi = SubOne(cast<ConstantInt>(Hi));
3138 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3139 ICmpInst::Predicate pred = (isSigned ?
3140 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3141 return new ICmpInst(pred, V, Hi);
3144 // Emit V-Lo >u Hi-1-Lo
3145 // Note that Hi has already had one subtracted from it, above.
3146 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3147 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3148 InsertNewInstBefore(Add, IB);
3149 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3150 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3153 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3154 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3155 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3156 // not, since all 1s are not contiguous.
3157 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3158 const APInt& V = Val->getValue();
3159 uint32_t BitWidth = Val->getType()->getBitWidth();
3160 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3162 // look for the first zero bit after the run of ones
3163 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3164 // look for the first non-zero bit
3165 ME = V.getActiveBits();
3169 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3170 /// where isSub determines whether the operator is a sub. If we can fold one of
3171 /// the following xforms:
3173 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3174 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3175 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3177 /// return (A +/- B).
3179 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3180 ConstantInt *Mask, bool isSub,
3182 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3183 if (!LHSI || LHSI->getNumOperands() != 2 ||
3184 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3186 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3188 switch (LHSI->getOpcode()) {
3190 case Instruction::And:
3191 if (And(N, Mask) == Mask) {
3192 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3193 if ((Mask->getValue().countLeadingZeros() +
3194 Mask->getValue().countPopulation()) ==
3195 Mask->getValue().getBitWidth())
3198 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3199 // part, we don't need any explicit masks to take them out of A. If that
3200 // is all N is, ignore it.
3201 uint32_t MB = 0, ME = 0;
3202 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3203 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3204 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3205 if (MaskedValueIsZero(RHS, Mask))
3210 case Instruction::Or:
3211 case Instruction::Xor:
3212 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3213 if ((Mask->getValue().countLeadingZeros() +
3214 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3215 && And(N, Mask)->isZero())
3222 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3224 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3225 return InsertNewInstBefore(New, I);
3228 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3229 bool Changed = SimplifyCommutative(I);
3230 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3232 if (isa<UndefValue>(Op1)) // X & undef -> 0
3233 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3237 return ReplaceInstUsesWith(I, Op1);
3239 // See if we can simplify any instructions used by the instruction whose sole
3240 // purpose is to compute bits we don't care about.
3241 if (!isa<VectorType>(I.getType())) {
3242 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3243 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3244 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3245 KnownZero, KnownOne))
3248 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3249 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3250 return ReplaceInstUsesWith(I, I.getOperand(0));
3251 } else if (isa<ConstantAggregateZero>(Op1)) {
3252 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3256 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3257 const APInt& AndRHSMask = AndRHS->getValue();
3258 APInt NotAndRHS(~AndRHSMask);
3260 // Optimize a variety of ((val OP C1) & C2) combinations...
3261 if (isa<BinaryOperator>(Op0)) {
3262 Instruction *Op0I = cast<Instruction>(Op0);
3263 Value *Op0LHS = Op0I->getOperand(0);
3264 Value *Op0RHS = Op0I->getOperand(1);
3265 switch (Op0I->getOpcode()) {
3266 case Instruction::Xor:
3267 case Instruction::Or:
3268 // If the mask is only needed on one incoming arm, push it up.
3269 if (Op0I->hasOneUse()) {
3270 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3271 // Not masking anything out for the LHS, move to RHS.
3272 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3273 Op0RHS->getName()+".masked");
3274 InsertNewInstBefore(NewRHS, I);
3275 return BinaryOperator::create(
3276 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3278 if (!isa<Constant>(Op0RHS) &&
3279 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3280 // Not masking anything out for the RHS, move to LHS.
3281 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3282 Op0LHS->getName()+".masked");
3283 InsertNewInstBefore(NewLHS, I);
3284 return BinaryOperator::create(
3285 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3290 case Instruction::Add:
3291 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3292 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3293 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3294 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3295 return BinaryOperator::createAnd(V, AndRHS);
3296 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3297 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3300 case Instruction::Sub:
3301 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3302 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3303 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3304 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3305 return BinaryOperator::createAnd(V, AndRHS);
3309 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3310 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3312 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3313 // If this is an integer truncation or change from signed-to-unsigned, and
3314 // if the source is an and/or with immediate, transform it. This
3315 // frequently occurs for bitfield accesses.
3316 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3317 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3318 CastOp->getNumOperands() == 2)
3319 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3320 if (CastOp->getOpcode() == Instruction::And) {
3321 // Change: and (cast (and X, C1) to T), C2
3322 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3323 // This will fold the two constants together, which may allow
3324 // other simplifications.
3325 Instruction *NewCast = CastInst::createTruncOrBitCast(
3326 CastOp->getOperand(0), I.getType(),
3327 CastOp->getName()+".shrunk");
3328 NewCast = InsertNewInstBefore(NewCast, I);
3329 // trunc_or_bitcast(C1)&C2
3330 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3331 C3 = ConstantExpr::getAnd(C3, AndRHS);
3332 return BinaryOperator::createAnd(NewCast, C3);
3333 } else if (CastOp->getOpcode() == Instruction::Or) {
3334 // Change: and (cast (or X, C1) to T), C2
3335 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3336 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3337 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3338 return ReplaceInstUsesWith(I, AndRHS);
3343 // Try to fold constant and into select arguments.
3344 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3345 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3347 if (isa<PHINode>(Op0))
3348 if (Instruction *NV = FoldOpIntoPhi(I))
3352 Value *Op0NotVal = dyn_castNotVal(Op0);
3353 Value *Op1NotVal = dyn_castNotVal(Op1);
3355 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3356 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3358 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3359 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3360 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3361 I.getName()+".demorgan");
3362 InsertNewInstBefore(Or, I);
3363 return BinaryOperator::createNot(Or);
3367 Value *A = 0, *B = 0, *C = 0, *D = 0;
3368 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3369 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3370 return ReplaceInstUsesWith(I, Op1);
3372 // (A|B) & ~(A&B) -> A^B
3373 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3374 if ((A == C && B == D) || (A == D && B == C))
3375 return BinaryOperator::createXor(A, B);
3379 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3380 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3381 return ReplaceInstUsesWith(I, Op0);
3383 // ~(A&B) & (A|B) -> A^B
3384 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3385 if ((A == C && B == D) || (A == D && B == C))
3386 return BinaryOperator::createXor(A, B);
3390 if (Op0->hasOneUse() &&
3391 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3392 if (A == Op1) { // (A^B)&A -> A&(A^B)
3393 I.swapOperands(); // Simplify below
3394 std::swap(Op0, Op1);
3395 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3396 cast<BinaryOperator>(Op0)->swapOperands();
3397 I.swapOperands(); // Simplify below
3398 std::swap(Op0, Op1);
3401 if (Op1->hasOneUse() &&
3402 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3403 if (B == Op0) { // B&(A^B) -> B&(B^A)
3404 cast<BinaryOperator>(Op1)->swapOperands();
3407 if (A == Op0) { // A&(A^B) -> A & ~B
3408 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3409 InsertNewInstBefore(NotB, I);
3410 return BinaryOperator::createAnd(A, NotB);
3415 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3416 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3417 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3420 Value *LHSVal, *RHSVal;
3421 ConstantInt *LHSCst, *RHSCst;
3422 ICmpInst::Predicate LHSCC, RHSCC;
3423 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3424 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3425 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3426 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3427 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3428 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3429 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3430 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3431 // Ensure that the larger constant is on the RHS.
3432 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3433 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3434 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3435 ICmpInst *LHS = cast<ICmpInst>(Op0);
3436 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3437 std::swap(LHS, RHS);
3438 std::swap(LHSCst, RHSCst);
3439 std::swap(LHSCC, RHSCC);
3442 // At this point, we know we have have two icmp instructions
3443 // comparing a value against two constants and and'ing the result
3444 // together. Because of the above check, we know that we only have
3445 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3446 // (from the FoldICmpLogical check above), that the two constants
3447 // are not equal and that the larger constant is on the RHS
3448 assert(LHSCst != RHSCst && "Compares not folded above?");
3451 default: assert(0 && "Unknown integer condition code!");
3452 case ICmpInst::ICMP_EQ:
3454 default: assert(0 && "Unknown integer condition code!");
3455 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3456 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3457 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3458 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3459 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3460 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3461 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3462 return ReplaceInstUsesWith(I, LHS);
3464 case ICmpInst::ICMP_NE:
3466 default: assert(0 && "Unknown integer condition code!");
3467 case ICmpInst::ICMP_ULT:
3468 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3469 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3470 break; // (X != 13 & X u< 15) -> no change
3471 case ICmpInst::ICMP_SLT:
3472 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3473 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3474 break; // (X != 13 & X s< 15) -> no change
3475 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3476 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3477 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3478 return ReplaceInstUsesWith(I, RHS);
3479 case ICmpInst::ICMP_NE:
3480 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3481 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3482 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3483 LHSVal->getName()+".off");
3484 InsertNewInstBefore(Add, I);
3485 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3486 ConstantInt::get(Add->getType(), 1));
3488 break; // (X != 13 & X != 15) -> no change
3491 case ICmpInst::ICMP_ULT:
3493 default: assert(0 && "Unknown integer condition code!");
3494 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3495 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3496 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3497 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3499 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3500 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3501 return ReplaceInstUsesWith(I, LHS);
3502 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3506 case ICmpInst::ICMP_SLT:
3508 default: assert(0 && "Unknown integer condition code!");
3509 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3510 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3511 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3512 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3514 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3515 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3516 return ReplaceInstUsesWith(I, LHS);
3517 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3521 case ICmpInst::ICMP_UGT:
3523 default: assert(0 && "Unknown integer condition code!");
3524 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3525 return ReplaceInstUsesWith(I, LHS);
3526 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3527 return ReplaceInstUsesWith(I, RHS);
3528 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3530 case ICmpInst::ICMP_NE:
3531 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3532 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3533 break; // (X u> 13 & X != 15) -> no change
3534 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3535 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3537 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3541 case ICmpInst::ICMP_SGT:
3543 default: assert(0 && "Unknown integer condition code!");
3544 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3545 return ReplaceInstUsesWith(I, LHS);
3546 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3547 return ReplaceInstUsesWith(I, RHS);
3548 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3550 case ICmpInst::ICMP_NE:
3551 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3552 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3553 break; // (X s> 13 & X != 15) -> no change
3554 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3555 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3557 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3565 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3566 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3567 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3568 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3569 const Type *SrcTy = Op0C->getOperand(0)->getType();
3570 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3571 // Only do this if the casts both really cause code to be generated.
3572 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3574 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3576 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3577 Op1C->getOperand(0),
3579 InsertNewInstBefore(NewOp, I);
3580 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3584 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3585 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3586 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3587 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3588 SI0->getOperand(1) == SI1->getOperand(1) &&
3589 (SI0->hasOneUse() || SI1->hasOneUse())) {
3590 Instruction *NewOp =
3591 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3593 SI0->getName()), I);
3594 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3595 SI1->getOperand(1));
3599 return Changed ? &I : 0;
3602 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3603 /// in the result. If it does, and if the specified byte hasn't been filled in
3604 /// yet, fill it in and return false.
3605 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3606 Instruction *I = dyn_cast<Instruction>(V);
3607 if (I == 0) return true;
3609 // If this is an or instruction, it is an inner node of the bswap.
3610 if (I->getOpcode() == Instruction::Or)
3611 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3612 CollectBSwapParts(I->getOperand(1), ByteValues);
3614 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3615 // If this is a shift by a constant int, and it is "24", then its operand
3616 // defines a byte. We only handle unsigned types here.
3617 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3618 // Not shifting the entire input by N-1 bytes?
3619 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3620 8*(ByteValues.size()-1))
3624 if (I->getOpcode() == Instruction::Shl) {
3625 // X << 24 defines the top byte with the lowest of the input bytes.
3626 DestNo = ByteValues.size()-1;
3628 // X >>u 24 defines the low byte with the highest of the input bytes.
3632 // If the destination byte value is already defined, the values are or'd
3633 // together, which isn't a bswap (unless it's an or of the same bits).
3634 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3636 ByteValues[DestNo] = I->getOperand(0);
3640 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3642 Value *Shift = 0, *ShiftLHS = 0;
3643 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3644 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3645 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3647 Instruction *SI = cast<Instruction>(Shift);
3649 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3650 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3651 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3654 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3656 if (AndAmt->getValue().getActiveBits() > 64)
3658 uint64_t AndAmtVal = AndAmt->getZExtValue();
3659 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3660 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3662 // Unknown mask for bswap.
3663 if (DestByte == ByteValues.size()) return true;
3665 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3667 if (SI->getOpcode() == Instruction::Shl)
3668 SrcByte = DestByte - ShiftBytes;
3670 SrcByte = DestByte + ShiftBytes;
3672 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3673 if (SrcByte != ByteValues.size()-DestByte-1)
3676 // If the destination byte value is already defined, the values are or'd
3677 // together, which isn't a bswap (unless it's an or of the same bits).
3678 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3680 ByteValues[DestByte] = SI->getOperand(0);
3684 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3685 /// If so, insert the new bswap intrinsic and return it.
3686 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3687 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3688 if (!ITy || ITy->getBitWidth() % 16)
3689 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3691 /// ByteValues - For each byte of the result, we keep track of which value
3692 /// defines each byte.
3693 SmallVector<Value*, 8> ByteValues;
3694 ByteValues.resize(ITy->getBitWidth()/8);
3696 // Try to find all the pieces corresponding to the bswap.
3697 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3698 CollectBSwapParts(I.getOperand(1), ByteValues))
3701 // Check to see if all of the bytes come from the same value.
3702 Value *V = ByteValues[0];
3703 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3705 // Check to make sure that all of the bytes come from the same value.
3706 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3707 if (ByteValues[i] != V)
3709 const Type *Tys[] = { ITy, ITy };
3710 Module *M = I.getParent()->getParent()->getParent();
3711 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 2);
3712 return new CallInst(F, V);
3716 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3717 bool Changed = SimplifyCommutative(I);
3718 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3720 if (isa<UndefValue>(Op1)) // X | undef -> -1
3721 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3725 return ReplaceInstUsesWith(I, Op0);
3727 // See if we can simplify any instructions used by the instruction whose sole
3728 // purpose is to compute bits we don't care about.
3729 if (!isa<VectorType>(I.getType())) {
3730 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3731 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3732 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3733 KnownZero, KnownOne))
3735 } else if (isa<ConstantAggregateZero>(Op1)) {
3736 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3737 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3738 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3739 return ReplaceInstUsesWith(I, I.getOperand(1));
3745 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3746 ConstantInt *C1 = 0; Value *X = 0;
3747 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3748 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3749 Instruction *Or = BinaryOperator::createOr(X, RHS);
3750 InsertNewInstBefore(Or, I);
3752 return BinaryOperator::createAnd(Or,
3753 ConstantInt::get(RHS->getValue() | C1->getValue()));
3756 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3757 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3758 Instruction *Or = BinaryOperator::createOr(X, RHS);
3759 InsertNewInstBefore(Or, I);
3761 return BinaryOperator::createXor(Or,
3762 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3765 // Try to fold constant and into select arguments.
3766 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3767 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3769 if (isa<PHINode>(Op0))
3770 if (Instruction *NV = FoldOpIntoPhi(I))
3774 Value *A = 0, *B = 0;
3775 ConstantInt *C1 = 0, *C2 = 0;
3777 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3778 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3779 return ReplaceInstUsesWith(I, Op1);
3780 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3781 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3782 return ReplaceInstUsesWith(I, Op0);
3784 // (A | B) | C and A | (B | C) -> bswap if possible.
3785 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3786 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3787 match(Op1, m_Or(m_Value(), m_Value())) ||
3788 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3789 match(Op1, m_Shift(m_Value(), m_Value())))) {
3790 if (Instruction *BSwap = MatchBSwap(I))
3794 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3795 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3796 MaskedValueIsZero(Op1, C1->getValue())) {
3797 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3798 InsertNewInstBefore(NOr, I);
3800 return BinaryOperator::createXor(NOr, C1);
3803 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3804 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3805 MaskedValueIsZero(Op0, C1->getValue())) {
3806 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3807 InsertNewInstBefore(NOr, I);
3809 return BinaryOperator::createXor(NOr, C1);
3813 Value *C = 0, *D = 0;
3814 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3815 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3816 Value *V1 = 0, *V2 = 0, *V3 = 0;
3817 C1 = dyn_cast<ConstantInt>(C);
3818 C2 = dyn_cast<ConstantInt>(D);
3819 if (C1 && C2) { // (A & C1)|(B & C2)
3820 // If we have: ((V + N) & C1) | (V & C2)
3821 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3822 // replace with V+N.
3823 if (C1->getValue() == ~C2->getValue()) {
3824 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3825 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3826 // Add commutes, try both ways.
3827 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3828 return ReplaceInstUsesWith(I, A);
3829 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3830 return ReplaceInstUsesWith(I, A);
3832 // Or commutes, try both ways.
3833 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3834 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3835 // Add commutes, try both ways.
3836 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3837 return ReplaceInstUsesWith(I, B);
3838 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3839 return ReplaceInstUsesWith(I, B);
3842 V1 = 0; V2 = 0; V3 = 0;
3845 // Check to see if we have any common things being and'ed. If so, find the
3846 // terms for V1 & (V2|V3).
3847 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3848 if (A == B) // (A & C)|(A & D) == A & (C|D)
3849 V1 = A, V2 = C, V3 = D;
3850 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3851 V1 = A, V2 = B, V3 = C;
3852 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3853 V1 = C, V2 = A, V3 = D;
3854 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3855 V1 = C, V2 = A, V3 = B;
3859 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3860 return BinaryOperator::createAnd(V1, Or);
3863 // (V1 & V3)|(V2 & ~V3) -> ((V1 ^ V2) & V3) ^ V2
3864 if (isOnlyUse(Op0) && isOnlyUse(Op1)) {
3865 // Try all combination of terms to find V3 and ~V3.
3866 if (A->hasOneUse() && match(A, m_Not(m_Value(V3)))) {
3872 if (B->hasOneUse() && match(B, m_Not(m_Value(V3)))) {
3878 if (C->hasOneUse() && match(C, m_Not(m_Value(V3)))) {
3884 if (D->hasOneUse() && match(D, m_Not(m_Value(V3)))) {
3891 A = InsertNewInstBefore(BinaryOperator::createXor(V1, V2, "tmp"), I);
3892 A = InsertNewInstBefore(BinaryOperator::createAnd(A, V3, "tmp"), I);
3893 return BinaryOperator::createXor(A, V2);
3899 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3900 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3901 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3902 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3903 SI0->getOperand(1) == SI1->getOperand(1) &&
3904 (SI0->hasOneUse() || SI1->hasOneUse())) {
3905 Instruction *NewOp =
3906 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3908 SI0->getName()), I);
3909 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3910 SI1->getOperand(1));
3914 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3915 if (A == Op1) // ~A | A == -1
3916 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3920 // Note, A is still live here!
3921 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3923 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3925 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3926 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3927 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3928 I.getName()+".demorgan"), I);
3929 return BinaryOperator::createNot(And);
3933 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3934 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3935 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3938 Value *LHSVal, *RHSVal;
3939 ConstantInt *LHSCst, *RHSCst;
3940 ICmpInst::Predicate LHSCC, RHSCC;
3941 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3942 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3943 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3944 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3945 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3946 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3947 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3948 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3949 // We can't fold (ugt x, C) | (sgt x, C2).
3950 PredicatesFoldable(LHSCC, RHSCC)) {
3951 // Ensure that the larger constant is on the RHS.
3952 ICmpInst *LHS = cast<ICmpInst>(Op0);
3954 if (ICmpInst::isSignedPredicate(LHSCC))
3955 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3957 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3960 std::swap(LHS, RHS);
3961 std::swap(LHSCst, RHSCst);
3962 std::swap(LHSCC, RHSCC);
3965 // At this point, we know we have have two icmp instructions
3966 // comparing a value against two constants and or'ing the result
3967 // together. Because of the above check, we know that we only have
3968 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3969 // FoldICmpLogical check above), that the two constants are not
3971 assert(LHSCst != RHSCst && "Compares not folded above?");
3974 default: assert(0 && "Unknown integer condition code!");
3975 case ICmpInst::ICMP_EQ:
3977 default: assert(0 && "Unknown integer condition code!");
3978 case ICmpInst::ICMP_EQ:
3979 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3980 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3981 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3982 LHSVal->getName()+".off");
3983 InsertNewInstBefore(Add, I);
3984 AddCST = Subtract(AddOne(RHSCst), LHSCst);
3985 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3987 break; // (X == 13 | X == 15) -> no change
3988 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3989 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3991 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3992 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3993 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3994 return ReplaceInstUsesWith(I, RHS);
3997 case ICmpInst::ICMP_NE:
3999 default: assert(0 && "Unknown integer condition code!");
4000 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4001 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4002 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4003 return ReplaceInstUsesWith(I, LHS);
4004 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4005 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4006 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4007 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4010 case ICmpInst::ICMP_ULT:
4012 default: assert(0 && "Unknown integer condition code!");
4013 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4015 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4016 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4018 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4020 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4021 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4022 return ReplaceInstUsesWith(I, RHS);
4023 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4027 case ICmpInst::ICMP_SLT:
4029 default: assert(0 && "Unknown integer condition code!");
4030 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4032 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4033 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4035 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4037 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4038 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4039 return ReplaceInstUsesWith(I, RHS);
4040 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4044 case ICmpInst::ICMP_UGT:
4046 default: assert(0 && "Unknown integer condition code!");
4047 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4048 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4049 return ReplaceInstUsesWith(I, LHS);
4050 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4052 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4053 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4054 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4055 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4059 case ICmpInst::ICMP_SGT:
4061 default: assert(0 && "Unknown integer condition code!");
4062 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4063 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4064 return ReplaceInstUsesWith(I, LHS);
4065 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4067 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4068 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4069 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4070 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4078 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4079 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4080 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4081 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4082 const Type *SrcTy = Op0C->getOperand(0)->getType();
4083 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4084 // Only do this if the casts both really cause code to be generated.
4085 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4087 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4089 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4090 Op1C->getOperand(0),
4092 InsertNewInstBefore(NewOp, I);
4093 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4098 return Changed ? &I : 0;
4101 // XorSelf - Implements: X ^ X --> 0
4104 XorSelf(Value *rhs) : RHS(rhs) {}
4105 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4106 Instruction *apply(BinaryOperator &Xor) const {
4112 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4113 bool Changed = SimplifyCommutative(I);
4114 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4116 if (isa<UndefValue>(Op1))
4117 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4119 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4120 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4121 assert(Result == &I && "AssociativeOpt didn't work?");
4122 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4125 // See if we can simplify any instructions used by the instruction whose sole
4126 // purpose is to compute bits we don't care about.
4127 if (!isa<VectorType>(I.getType())) {
4128 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4129 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4130 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4131 KnownZero, KnownOne))
4133 } else if (isa<ConstantAggregateZero>(Op1)) {
4134 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4137 // Is this a ~ operation?
4138 if (Value *NotOp = dyn_castNotVal(&I)) {
4139 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4140 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4141 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4142 if (Op0I->getOpcode() == Instruction::And ||
4143 Op0I->getOpcode() == Instruction::Or) {
4144 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4145 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4147 BinaryOperator::createNot(Op0I->getOperand(1),
4148 Op0I->getOperand(1)->getName()+".not");
4149 InsertNewInstBefore(NotY, I);
4150 if (Op0I->getOpcode() == Instruction::And)
4151 return BinaryOperator::createOr(Op0NotVal, NotY);
4153 return BinaryOperator::createAnd(Op0NotVal, NotY);
4160 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4161 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
4162 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4163 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
4164 return new ICmpInst(ICI->getInversePredicate(),
4165 ICI->getOperand(0), ICI->getOperand(1));
4167 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4168 // ~(c-X) == X-c-1 == X+(-c-1)
4169 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4170 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4171 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4172 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4173 ConstantInt::get(I.getType(), 1));
4174 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4177 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4178 if (Op0I->getOpcode() == Instruction::Add) {
4179 // ~(X-c) --> (-c-1)-X
4180 if (RHS->isAllOnesValue()) {
4181 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4182 return BinaryOperator::createSub(
4183 ConstantExpr::getSub(NegOp0CI,
4184 ConstantInt::get(I.getType(), 1)),
4185 Op0I->getOperand(0));
4186 } else if (RHS->getValue().isSignBit()) {
4187 // (X + C) ^ signbit -> (X + C + signbit)
4188 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4189 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4192 } else if (Op0I->getOpcode() == Instruction::Or) {
4193 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4194 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4195 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4196 // Anything in both C1 and C2 is known to be zero, remove it from
4198 Constant *CommonBits = And(Op0CI, RHS);
4199 NewRHS = ConstantExpr::getAnd(NewRHS,
4200 ConstantExpr::getNot(CommonBits));
4201 AddToWorkList(Op0I);
4202 I.setOperand(0, Op0I->getOperand(0));
4203 I.setOperand(1, NewRHS);
4209 // Try to fold constant and into select arguments.
4210 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4211 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4213 if (isa<PHINode>(Op0))
4214 if (Instruction *NV = FoldOpIntoPhi(I))
4218 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4220 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4222 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4224 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4227 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4230 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4231 if (A == Op0) { // B^(B|A) == (A|B)^B
4232 Op1I->swapOperands();
4234 std::swap(Op0, Op1);
4235 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4236 I.swapOperands(); // Simplified below.
4237 std::swap(Op0, Op1);
4239 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4240 if (Op0 == A) // A^(A^B) == B
4241 return ReplaceInstUsesWith(I, B);
4242 else if (Op0 == B) // A^(B^A) == B
4243 return ReplaceInstUsesWith(I, A);
4244 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4245 if (A == Op0) { // A^(A&B) -> A^(B&A)
4246 Op1I->swapOperands();
4249 if (B == Op0) { // A^(B&A) -> (B&A)^A
4250 I.swapOperands(); // Simplified below.
4251 std::swap(Op0, Op1);
4256 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4259 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4260 if (A == Op1) // (B|A)^B == (A|B)^B
4262 if (B == Op1) { // (A|B)^B == A & ~B
4264 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4265 return BinaryOperator::createAnd(A, NotB);
4267 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4268 if (Op1 == A) // (A^B)^A == B
4269 return ReplaceInstUsesWith(I, B);
4270 else if (Op1 == B) // (B^A)^A == B
4271 return ReplaceInstUsesWith(I, A);
4272 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4273 if (A == Op1) // (A&B)^A -> (B&A)^A
4275 if (B == Op1 && // (B&A)^A == ~B & A
4276 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4278 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4279 return BinaryOperator::createAnd(N, Op1);
4284 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4285 if (Op0I && Op1I && Op0I->isShift() &&
4286 Op0I->getOpcode() == Op1I->getOpcode() &&
4287 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4288 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4289 Instruction *NewOp =
4290 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4291 Op1I->getOperand(0),
4292 Op0I->getName()), I);
4293 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4294 Op1I->getOperand(1));
4298 Value *A, *B, *C, *D;
4299 // (A & B)^(A | B) -> A ^ B
4300 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4301 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4302 if ((A == C && B == D) || (A == D && B == C))
4303 return BinaryOperator::createXor(A, B);
4305 // (A | B)^(A & B) -> A ^ B
4306 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4307 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4308 if ((A == C && B == D) || (A == D && B == C))
4309 return BinaryOperator::createXor(A, B);
4313 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4314 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4315 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4316 // (X & Y)^(X & Y) -> (Y^Z) & X
4317 Value *X = 0, *Y = 0, *Z = 0;
4319 X = A, Y = B, Z = D;
4321 X = A, Y = B, Z = C;
4323 X = B, Y = A, Z = D;
4325 X = B, Y = A, Z = C;
4328 Instruction *NewOp =
4329 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4330 return BinaryOperator::createAnd(NewOp, X);
4335 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4336 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4337 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4340 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4341 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4342 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4343 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4344 const Type *SrcTy = Op0C->getOperand(0)->getType();
4345 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4346 // Only do this if the casts both really cause code to be generated.
4347 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4349 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4351 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4352 Op1C->getOperand(0),
4354 InsertNewInstBefore(NewOp, I);
4355 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4359 return Changed ? &I : 0;
4362 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4363 /// overflowed for this type.
4364 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4365 ConstantInt *In2, bool IsSigned = false) {
4366 Result = cast<ConstantInt>(Add(In1, In2));
4369 if (In2->getValue().isNegative())
4370 return Result->getValue().sgt(In1->getValue());
4372 return Result->getValue().slt(In1->getValue());
4374 return Result->getValue().ult(In1->getValue());
4377 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4378 /// code necessary to compute the offset from the base pointer (without adding
4379 /// in the base pointer). Return the result as a signed integer of intptr size.
4380 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4381 TargetData &TD = IC.getTargetData();
4382 gep_type_iterator GTI = gep_type_begin(GEP);
4383 const Type *IntPtrTy = TD.getIntPtrType();
4384 Value *Result = Constant::getNullValue(IntPtrTy);
4386 // Build a mask for high order bits.
4387 unsigned IntPtrWidth = TD.getPointerSize()*8;
4388 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4390 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4391 Value *Op = GEP->getOperand(i);
4392 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4393 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4394 if (OpC->isZero()) continue;
4396 // Handle a struct index, which adds its field offset to the pointer.
4397 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4398 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4400 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4401 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4403 Result = IC.InsertNewInstBefore(
4404 BinaryOperator::createAdd(Result,
4405 ConstantInt::get(IntPtrTy, Size),
4406 GEP->getName()+".offs"), I);
4410 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4411 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4412 Scale = ConstantExpr::getMul(OC, Scale);
4413 if (Constant *RC = dyn_cast<Constant>(Result))
4414 Result = ConstantExpr::getAdd(RC, Scale);
4416 // Emit an add instruction.
4417 Result = IC.InsertNewInstBefore(
4418 BinaryOperator::createAdd(Result, Scale,
4419 GEP->getName()+".offs"), I);
4423 // Convert to correct type.
4424 if (Op->getType() != IntPtrTy) {
4425 if (Constant *OpC = dyn_cast<Constant>(Op))
4426 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4428 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4429 Op->getName()+".c"), I);
4432 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4433 if (Constant *OpC = dyn_cast<Constant>(Op))
4434 Op = ConstantExpr::getMul(OpC, Scale);
4435 else // We'll let instcombine(mul) convert this to a shl if possible.
4436 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4437 GEP->getName()+".idx"), I);
4440 // Emit an add instruction.
4441 if (isa<Constant>(Op) && isa<Constant>(Result))
4442 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4443 cast<Constant>(Result));
4445 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4446 GEP->getName()+".offs"), I);
4451 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4452 /// else. At this point we know that the GEP is on the LHS of the comparison.
4453 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4454 ICmpInst::Predicate Cond,
4456 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4458 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4459 if (isa<PointerType>(CI->getOperand(0)->getType()))
4460 RHS = CI->getOperand(0);
4462 Value *PtrBase = GEPLHS->getOperand(0);
4463 if (PtrBase == RHS) {
4464 // As an optimization, we don't actually have to compute the actual value of
4465 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4466 // each index is zero or not.
4467 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4468 Instruction *InVal = 0;
4469 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4470 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4472 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4473 if (isa<UndefValue>(C)) // undef index -> undef.
4474 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4475 if (C->isNullValue())
4477 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4478 EmitIt = false; // This is indexing into a zero sized array?
4479 } else if (isa<ConstantInt>(C))
4480 return ReplaceInstUsesWith(I, // No comparison is needed here.
4481 ConstantInt::get(Type::Int1Ty,
4482 Cond == ICmpInst::ICMP_NE));
4487 new ICmpInst(Cond, GEPLHS->getOperand(i),
4488 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4492 InVal = InsertNewInstBefore(InVal, I);
4493 InsertNewInstBefore(Comp, I);
4494 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4495 InVal = BinaryOperator::createOr(InVal, Comp);
4496 else // True if all are equal
4497 InVal = BinaryOperator::createAnd(InVal, Comp);
4505 // No comparison is needed here, all indexes = 0
4506 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4507 Cond == ICmpInst::ICMP_EQ));
4510 // Only lower this if the icmp is the only user of the GEP or if we expect
4511 // the result to fold to a constant!
4512 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4513 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4514 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4515 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4516 Constant::getNullValue(Offset->getType()));
4518 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4519 // If the base pointers are different, but the indices are the same, just
4520 // compare the base pointer.
4521 if (PtrBase != GEPRHS->getOperand(0)) {
4522 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4523 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4524 GEPRHS->getOperand(0)->getType();
4526 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4527 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4528 IndicesTheSame = false;
4532 // If all indices are the same, just compare the base pointers.
4534 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4535 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4537 // Otherwise, the base pointers are different and the indices are
4538 // different, bail out.
4542 // If one of the GEPs has all zero indices, recurse.
4543 bool AllZeros = true;
4544 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4545 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4546 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4551 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4552 ICmpInst::getSwappedPredicate(Cond), I);
4554 // If the other GEP has all zero indices, recurse.
4556 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4557 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4558 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4563 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4565 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4566 // If the GEPs only differ by one index, compare it.
4567 unsigned NumDifferences = 0; // Keep track of # differences.
4568 unsigned DiffOperand = 0; // The operand that differs.
4569 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4570 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4571 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4572 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4573 // Irreconcilable differences.
4577 if (NumDifferences++) break;
4582 if (NumDifferences == 0) // SAME GEP?
4583 return ReplaceInstUsesWith(I, // No comparison is needed here.
4584 ConstantInt::get(Type::Int1Ty,
4585 Cond == ICmpInst::ICMP_EQ));
4586 else if (NumDifferences == 1) {
4587 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4588 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4589 // Make sure we do a signed comparison here.
4590 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4594 // Only lower this if the icmp is the only user of the GEP or if we expect
4595 // the result to fold to a constant!
4596 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4597 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4598 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4599 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4600 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4601 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4607 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4608 bool Changed = SimplifyCompare(I);
4609 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4611 // Fold trivial predicates.
4612 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4613 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4614 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4615 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4617 // Simplify 'fcmp pred X, X'
4619 switch (I.getPredicate()) {
4620 default: assert(0 && "Unknown predicate!");
4621 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4622 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4623 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4624 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4625 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4626 case FCmpInst::FCMP_OLT: // True if ordered and less than
4627 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4628 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4630 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4631 case FCmpInst::FCMP_ULT: // True if unordered or less than
4632 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4633 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4634 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4635 I.setPredicate(FCmpInst::FCMP_UNO);
4636 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4639 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4640 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4641 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4642 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4643 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4644 I.setPredicate(FCmpInst::FCMP_ORD);
4645 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4650 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4651 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4653 // Handle fcmp with constant RHS
4654 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4655 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4656 switch (LHSI->getOpcode()) {
4657 case Instruction::PHI:
4658 if (Instruction *NV = FoldOpIntoPhi(I))
4661 case Instruction::Select:
4662 // If either operand of the select is a constant, we can fold the
4663 // comparison into the select arms, which will cause one to be
4664 // constant folded and the select turned into a bitwise or.
4665 Value *Op1 = 0, *Op2 = 0;
4666 if (LHSI->hasOneUse()) {
4667 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4668 // Fold the known value into the constant operand.
4669 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4670 // Insert a new FCmp of the other select operand.
4671 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4672 LHSI->getOperand(2), RHSC,
4674 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4675 // Fold the known value into the constant operand.
4676 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4677 // Insert a new FCmp of the other select operand.
4678 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4679 LHSI->getOperand(1), RHSC,
4685 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4690 return Changed ? &I : 0;
4693 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4694 bool Changed = SimplifyCompare(I);
4695 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4696 const Type *Ty = Op0->getType();
4700 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4701 isTrueWhenEqual(I)));
4703 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4704 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4706 // icmp of GlobalValues can never equal each other as long as they aren't
4707 // external weak linkage type.
4708 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4709 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4710 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4711 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4712 !isTrueWhenEqual(I)));
4714 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4715 // addresses never equal each other! We already know that Op0 != Op1.
4716 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4717 isa<ConstantPointerNull>(Op0)) &&
4718 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4719 isa<ConstantPointerNull>(Op1)))
4720 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4721 !isTrueWhenEqual(I)));
4723 // icmp's with boolean values can always be turned into bitwise operations
4724 if (Ty == Type::Int1Ty) {
4725 switch (I.getPredicate()) {
4726 default: assert(0 && "Invalid icmp instruction!");
4727 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4728 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4729 InsertNewInstBefore(Xor, I);
4730 return BinaryOperator::createNot(Xor);
4732 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4733 return BinaryOperator::createXor(Op0, Op1);
4735 case ICmpInst::ICMP_UGT:
4736 case ICmpInst::ICMP_SGT:
4737 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4739 case ICmpInst::ICMP_ULT:
4740 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4741 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4742 InsertNewInstBefore(Not, I);
4743 return BinaryOperator::createAnd(Not, Op1);
4745 case ICmpInst::ICMP_UGE:
4746 case ICmpInst::ICMP_SGE:
4747 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4749 case ICmpInst::ICMP_ULE:
4750 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4751 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4752 InsertNewInstBefore(Not, I);
4753 return BinaryOperator::createOr(Not, Op1);
4758 // See if we are doing a comparison between a constant and an instruction that
4759 // can be folded into the comparison.
4760 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4761 switch (I.getPredicate()) {
4763 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4764 if (CI->isMinValue(false))
4765 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4766 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4767 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4768 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4769 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4770 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4771 if (CI->isMinValue(true))
4772 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4773 ConstantInt::getAllOnesValue(Op0->getType()));
4777 case ICmpInst::ICMP_SLT:
4778 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4779 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4780 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4781 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4782 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4783 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4786 case ICmpInst::ICMP_UGT:
4787 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4788 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4789 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4790 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4791 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4792 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4794 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4795 if (CI->isMaxValue(true))
4796 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4797 ConstantInt::getNullValue(Op0->getType()));
4800 case ICmpInst::ICMP_SGT:
4801 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4802 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4803 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4804 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4805 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4806 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4809 case ICmpInst::ICMP_ULE:
4810 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4811 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4812 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4813 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4814 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4815 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4818 case ICmpInst::ICMP_SLE:
4819 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4820 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4821 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4822 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4823 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4824 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4827 case ICmpInst::ICMP_UGE:
4828 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4829 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4830 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4831 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4832 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4833 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4836 case ICmpInst::ICMP_SGE:
4837 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4838 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4839 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4840 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4841 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4842 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4846 // If we still have a icmp le or icmp ge instruction, turn it into the
4847 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4848 // already been handled above, this requires little checking.
4850 switch (I.getPredicate()) {
4852 case ICmpInst::ICMP_ULE:
4853 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4854 case ICmpInst::ICMP_SLE:
4855 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4856 case ICmpInst::ICMP_UGE:
4857 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4858 case ICmpInst::ICMP_SGE:
4859 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4862 // See if we can fold the comparison based on bits known to be zero or one
4864 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4865 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4866 if (SimplifyDemandedBits(Op0, APInt::getAllOnesValue(BitWidth),
4867 KnownZero, KnownOne, 0))
4870 // Given the known and unknown bits, compute a range that the LHS could be
4872 if ((KnownOne | KnownZero) != 0) {
4873 // Compute the Min, Max and RHS values based on the known bits. For the
4874 // EQ and NE we use unsigned values.
4875 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4876 const APInt& RHSVal = CI->getValue();
4877 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4878 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4881 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4884 switch (I.getPredicate()) { // LE/GE have been folded already.
4885 default: assert(0 && "Unknown icmp opcode!");
4886 case ICmpInst::ICMP_EQ:
4887 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4888 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4890 case ICmpInst::ICMP_NE:
4891 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4892 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4894 case ICmpInst::ICMP_ULT:
4895 if (Max.ult(RHSVal))
4896 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4897 if (Min.uge(RHSVal))
4898 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4900 case ICmpInst::ICMP_UGT:
4901 if (Min.ugt(RHSVal))
4902 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4903 if (Max.ule(RHSVal))
4904 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4906 case ICmpInst::ICMP_SLT:
4907 if (Max.slt(RHSVal))
4908 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4909 if (Min.sgt(RHSVal))
4910 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4912 case ICmpInst::ICMP_SGT:
4913 if (Min.sgt(RHSVal))
4914 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4915 if (Max.sle(RHSVal))
4916 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4921 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4922 // instruction, see if that instruction also has constants so that the
4923 // instruction can be folded into the icmp
4924 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4925 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4929 // Handle icmp with constant (but not simple integer constant) RHS
4930 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4931 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4932 switch (LHSI->getOpcode()) {
4933 case Instruction::GetElementPtr:
4934 if (RHSC->isNullValue()) {
4935 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4936 bool isAllZeros = true;
4937 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4938 if (!isa<Constant>(LHSI->getOperand(i)) ||
4939 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4944 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
4945 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4949 case Instruction::PHI:
4950 if (Instruction *NV = FoldOpIntoPhi(I))
4953 case Instruction::Select: {
4954 // If either operand of the select is a constant, we can fold the
4955 // comparison into the select arms, which will cause one to be
4956 // constant folded and the select turned into a bitwise or.
4957 Value *Op1 = 0, *Op2 = 0;
4958 if (LHSI->hasOneUse()) {
4959 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4960 // Fold the known value into the constant operand.
4961 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4962 // Insert a new ICmp of the other select operand.
4963 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4964 LHSI->getOperand(2), RHSC,
4966 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4967 // Fold the known value into the constant operand.
4968 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4969 // Insert a new ICmp of the other select operand.
4970 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4971 LHSI->getOperand(1), RHSC,
4977 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4980 case Instruction::Malloc:
4981 // If we have (malloc != null), and if the malloc has a single use, we
4982 // can assume it is successful and remove the malloc.
4983 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
4984 AddToWorkList(LHSI);
4985 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4986 !isTrueWhenEqual(I)));
4992 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4993 if (User *GEP = dyn_castGetElementPtr(Op0))
4994 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
4996 if (User *GEP = dyn_castGetElementPtr(Op1))
4997 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
4998 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5001 // Test to see if the operands of the icmp are casted versions of other
5002 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5004 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5005 if (isa<PointerType>(Op0->getType()) &&
5006 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5007 // We keep moving the cast from the left operand over to the right
5008 // operand, where it can often be eliminated completely.
5009 Op0 = CI->getOperand(0);
5011 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5012 // so eliminate it as well.
5013 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5014 Op1 = CI2->getOperand(0);
5016 // If Op1 is a constant, we can fold the cast into the constant.
5017 if (Op0->getType() != Op1->getType())
5018 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5019 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5021 // Otherwise, cast the RHS right before the icmp
5022 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5024 return new ICmpInst(I.getPredicate(), Op0, Op1);
5028 if (isa<CastInst>(Op0)) {
5029 // Handle the special case of: icmp (cast bool to X), <cst>
5030 // This comes up when you have code like
5033 // For generality, we handle any zero-extension of any operand comparison
5034 // with a constant or another cast from the same type.
5035 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5036 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5040 if (I.isEquality()) {
5041 Value *A, *B, *C, *D;
5042 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5043 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5044 Value *OtherVal = A == Op1 ? B : A;
5045 return new ICmpInst(I.getPredicate(), OtherVal,
5046 Constant::getNullValue(A->getType()));
5049 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5050 // A^c1 == C^c2 --> A == C^(c1^c2)
5051 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5052 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5053 if (Op1->hasOneUse()) {
5054 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5055 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5056 return new ICmpInst(I.getPredicate(), A,
5057 InsertNewInstBefore(Xor, I));
5060 // A^B == A^D -> B == D
5061 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5062 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5063 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5064 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5068 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5069 (A == Op0 || B == Op0)) {
5070 // A == (A^B) -> B == 0
5071 Value *OtherVal = A == Op0 ? B : A;
5072 return new ICmpInst(I.getPredicate(), OtherVal,
5073 Constant::getNullValue(A->getType()));
5075 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5076 // (A-B) == A -> B == 0
5077 return new ICmpInst(I.getPredicate(), B,
5078 Constant::getNullValue(B->getType()));
5080 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5081 // A == (A-B) -> B == 0
5082 return new ICmpInst(I.getPredicate(), B,
5083 Constant::getNullValue(B->getType()));
5086 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5087 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5088 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5089 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5090 Value *X = 0, *Y = 0, *Z = 0;
5093 X = B; Y = D; Z = A;
5094 } else if (A == D) {
5095 X = B; Y = C; Z = A;
5096 } else if (B == C) {
5097 X = A; Y = D; Z = B;
5098 } else if (B == D) {
5099 X = A; Y = C; Z = B;
5102 if (X) { // Build (X^Y) & Z
5103 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5104 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5105 I.setOperand(0, Op1);
5106 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5111 return Changed ? &I : 0;
5115 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5116 /// and CmpRHS are both known to be integer constants.
5117 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5118 ConstantInt *DivRHS) {
5119 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5120 const APInt &CmpRHSV = CmpRHS->getValue();
5122 // FIXME: If the operand types don't match the type of the divide
5123 // then don't attempt this transform. The code below doesn't have the
5124 // logic to deal with a signed divide and an unsigned compare (and
5125 // vice versa). This is because (x /s C1) <s C2 produces different
5126 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5127 // (x /u C1) <u C2. Simply casting the operands and result won't
5128 // work. :( The if statement below tests that condition and bails
5130 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5131 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5133 if (DivRHS->isZero())
5134 return 0; // The ProdOV computation fails on divide by zero.
5136 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5137 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5138 // C2 (CI). By solving for X we can turn this into a range check
5139 // instead of computing a divide.
5140 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5142 // Determine if the product overflows by seeing if the product is
5143 // not equal to the divide. Make sure we do the same kind of divide
5144 // as in the LHS instruction that we're folding.
5145 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5146 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5148 // Get the ICmp opcode
5149 ICmpInst::Predicate Pred = ICI.getPredicate();
5151 // Figure out the interval that is being checked. For example, a comparison
5152 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5153 // Compute this interval based on the constants involved and the signedness of
5154 // the compare/divide. This computes a half-open interval, keeping track of
5155 // whether either value in the interval overflows. After analysis each
5156 // overflow variable is set to 0 if it's corresponding bound variable is valid
5157 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5158 int LoOverflow = 0, HiOverflow = 0;
5159 ConstantInt *LoBound = 0, *HiBound = 0;
5162 if (!DivIsSigned) { // udiv
5163 // e.g. X/5 op 3 --> [15, 20)
5165 HiOverflow = LoOverflow = ProdOV;
5167 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5168 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5169 if (CmpRHSV == 0) { // (X / pos) op 0
5170 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5171 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5173 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5174 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5175 HiOverflow = LoOverflow = ProdOV;
5177 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5178 } else { // (X / pos) op neg
5179 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5180 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5181 LoOverflow = AddWithOverflow(LoBound, Prod,
5182 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5183 HiBound = AddOne(Prod);
5184 HiOverflow = ProdOV ? -1 : 0;
5186 } else { // Divisor is < 0.
5187 if (CmpRHSV == 0) { // (X / neg) op 0
5188 // e.g. X/-5 op 0 --> [-4, 5)
5189 LoBound = AddOne(DivRHS);
5190 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5191 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5192 HiOverflow = 1; // [INTMIN+1, overflow)
5193 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5195 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5196 // e.g. X/-5 op 3 --> [-19, -14)
5197 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5199 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5200 HiBound = AddOne(Prod);
5201 } else { // (X / neg) op neg
5202 // e.g. X/-5 op -3 --> [15, 20)
5204 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5205 HiBound = Subtract(Prod, DivRHS);
5208 // Dividing by a negative swaps the condition. LT <-> GT
5209 Pred = ICmpInst::getSwappedPredicate(Pred);
5212 Value *X = DivI->getOperand(0);
5214 default: assert(0 && "Unhandled icmp opcode!");
5215 case ICmpInst::ICMP_EQ:
5216 if (LoOverflow && HiOverflow)
5217 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5218 else if (HiOverflow)
5219 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5220 ICmpInst::ICMP_UGE, X, LoBound);
5221 else if (LoOverflow)
5222 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5223 ICmpInst::ICMP_ULT, X, HiBound);
5225 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5226 case ICmpInst::ICMP_NE:
5227 if (LoOverflow && HiOverflow)
5228 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5229 else if (HiOverflow)
5230 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5231 ICmpInst::ICMP_ULT, X, LoBound);
5232 else if (LoOverflow)
5233 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5234 ICmpInst::ICMP_UGE, X, HiBound);
5236 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5237 case ICmpInst::ICMP_ULT:
5238 case ICmpInst::ICMP_SLT:
5239 if (LoOverflow == +1) // Low bound is greater than input range.
5240 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5241 if (LoOverflow == -1) // Low bound is less than input range.
5242 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5243 return new ICmpInst(Pred, X, LoBound);
5244 case ICmpInst::ICMP_UGT:
5245 case ICmpInst::ICMP_SGT:
5246 if (HiOverflow == +1) // High bound greater than input range.
5247 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5248 else if (HiOverflow == -1) // High bound less than input range.
5249 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5250 if (Pred == ICmpInst::ICMP_UGT)
5251 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5253 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5258 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5260 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5263 const APInt &RHSV = RHS->getValue();
5265 switch (LHSI->getOpcode()) {
5266 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5267 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5268 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5270 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5271 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5272 Value *CompareVal = LHSI->getOperand(0);
5274 // If the sign bit of the XorCST is not set, there is no change to
5275 // the operation, just stop using the Xor.
5276 if (!XorCST->getValue().isNegative()) {
5277 ICI.setOperand(0, CompareVal);
5278 AddToWorkList(LHSI);
5282 // Was the old condition true if the operand is positive?
5283 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5285 // If so, the new one isn't.
5286 isTrueIfPositive ^= true;
5288 if (isTrueIfPositive)
5289 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5291 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5295 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5296 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5297 LHSI->getOperand(0)->hasOneUse()) {
5298 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5300 // If the LHS is an AND of a truncating cast, we can widen the
5301 // and/compare to be the input width without changing the value
5302 // produced, eliminating a cast.
5303 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5304 // We can do this transformation if either the AND constant does not
5305 // have its sign bit set or if it is an equality comparison.
5306 // Extending a relational comparison when we're checking the sign
5307 // bit would not work.
5308 if (Cast->hasOneUse() &&
5309 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5310 RHSV.isPositive())) {
5312 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5313 APInt NewCST = AndCST->getValue();
5314 NewCST.zext(BitWidth);
5316 NewCI.zext(BitWidth);
5317 Instruction *NewAnd =
5318 BinaryOperator::createAnd(Cast->getOperand(0),
5319 ConstantInt::get(NewCST),LHSI->getName());
5320 InsertNewInstBefore(NewAnd, ICI);
5321 return new ICmpInst(ICI.getPredicate(), NewAnd,
5322 ConstantInt::get(NewCI));
5326 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5327 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5328 // happens a LOT in code produced by the C front-end, for bitfield
5330 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5331 if (Shift && !Shift->isShift())
5335 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5336 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5337 const Type *AndTy = AndCST->getType(); // Type of the and.
5339 // We can fold this as long as we can't shift unknown bits
5340 // into the mask. This can only happen with signed shift
5341 // rights, as they sign-extend.
5343 bool CanFold = Shift->isLogicalShift();
5345 // To test for the bad case of the signed shr, see if any
5346 // of the bits shifted in could be tested after the mask.
5347 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5348 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5350 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5351 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5352 AndCST->getValue()) == 0)
5358 if (Shift->getOpcode() == Instruction::Shl)
5359 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5361 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5363 // Check to see if we are shifting out any of the bits being
5365 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5366 // If we shifted bits out, the fold is not going to work out.
5367 // As a special case, check to see if this means that the
5368 // result is always true or false now.
5369 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5370 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5371 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5372 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5374 ICI.setOperand(1, NewCst);
5375 Constant *NewAndCST;
5376 if (Shift->getOpcode() == Instruction::Shl)
5377 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5379 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5380 LHSI->setOperand(1, NewAndCST);
5381 LHSI->setOperand(0, Shift->getOperand(0));
5382 AddToWorkList(Shift); // Shift is dead.
5383 AddUsesToWorkList(ICI);
5389 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5390 // preferable because it allows the C<<Y expression to be hoisted out
5391 // of a loop if Y is invariant and X is not.
5392 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5393 ICI.isEquality() && !Shift->isArithmeticShift() &&
5394 isa<Instruction>(Shift->getOperand(0))) {
5397 if (Shift->getOpcode() == Instruction::LShr) {
5398 NS = BinaryOperator::createShl(AndCST,
5399 Shift->getOperand(1), "tmp");
5401 // Insert a logical shift.
5402 NS = BinaryOperator::createLShr(AndCST,
5403 Shift->getOperand(1), "tmp");
5405 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5407 // Compute X & (C << Y).
5408 Instruction *NewAnd =
5409 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5410 InsertNewInstBefore(NewAnd, ICI);
5412 ICI.setOperand(0, NewAnd);
5418 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
5419 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5420 if (ICI.isEquality()) {
5421 uint32_t TypeBits = RHSV.getBitWidth();
5423 // Check that the shift amount is in range. If not, don't perform
5424 // undefined shifts. When the shift is visited it will be
5426 if (ShAmt->uge(TypeBits))
5429 // If we are comparing against bits always shifted out, the
5430 // comparison cannot succeed.
5432 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5433 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5434 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5435 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5436 return ReplaceInstUsesWith(ICI, Cst);
5439 if (LHSI->hasOneUse()) {
5440 // Otherwise strength reduce the shift into an and.
5441 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5443 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5446 BinaryOperator::createAnd(LHSI->getOperand(0),
5447 Mask, LHSI->getName()+".mask");
5448 Value *And = InsertNewInstBefore(AndI, ICI);
5449 return new ICmpInst(ICI.getPredicate(), And,
5450 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5456 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5457 case Instruction::AShr:
5458 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5459 if (ICI.isEquality()) {
5460 // Check that the shift amount is in range. If not, don't perform
5461 // undefined shifts. When the shift is visited it will be
5463 uint32_t TypeBits = RHSV.getBitWidth();
5464 if (ShAmt->uge(TypeBits))
5466 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5468 // If we are comparing against bits always shifted out, the
5469 // comparison cannot succeed.
5470 APInt Comp = RHSV << ShAmtVal;
5471 if (LHSI->getOpcode() == Instruction::LShr)
5472 Comp = Comp.lshr(ShAmtVal);
5474 Comp = Comp.ashr(ShAmtVal);
5476 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5477 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5478 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5479 return ReplaceInstUsesWith(ICI, Cst);
5482 if (LHSI->hasOneUse() || RHSV == 0) {
5483 // Otherwise strength reduce the shift into an and.
5484 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5485 Constant *Mask = ConstantInt::get(Val);
5488 BinaryOperator::createAnd(LHSI->getOperand(0),
5489 Mask, LHSI->getName()+".mask");
5490 Value *And = InsertNewInstBefore(AndI, ICI);
5491 return new ICmpInst(ICI.getPredicate(), And,
5492 ConstantExpr::getShl(RHS, ShAmt));
5498 case Instruction::SDiv:
5499 case Instruction::UDiv:
5500 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5501 // Fold this div into the comparison, producing a range check.
5502 // Determine, based on the divide type, what the range is being
5503 // checked. If there is an overflow on the low or high side, remember
5504 // it, otherwise compute the range [low, hi) bounding the new value.
5505 // See: InsertRangeTest above for the kinds of replacements possible.
5506 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5507 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5513 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5514 if (ICI.isEquality()) {
5515 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5517 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5518 // the second operand is a constant, simplify a bit.
5519 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5520 switch (BO->getOpcode()) {
5521 case Instruction::SRem:
5522 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5523 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5524 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5525 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5526 Instruction *NewRem =
5527 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5529 InsertNewInstBefore(NewRem, ICI);
5530 return new ICmpInst(ICI.getPredicate(), NewRem,
5531 Constant::getNullValue(BO->getType()));
5535 case Instruction::Add:
5536 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5537 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5538 if (BO->hasOneUse())
5539 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5540 Subtract(RHS, BOp1C));
5541 } else if (RHSV == 0) {
5542 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5543 // efficiently invertible, or if the add has just this one use.
5544 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5546 if (Value *NegVal = dyn_castNegVal(BOp1))
5547 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5548 else if (Value *NegVal = dyn_castNegVal(BOp0))
5549 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5550 else if (BO->hasOneUse()) {
5551 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5552 InsertNewInstBefore(Neg, ICI);
5554 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5558 case Instruction::Xor:
5559 // For the xor case, we can xor two constants together, eliminating
5560 // the explicit xor.
5561 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5562 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5563 ConstantExpr::getXor(RHS, BOC));
5566 case Instruction::Sub:
5567 // Replace (([sub|xor] A, B) != 0) with (A != B)
5569 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5573 case Instruction::Or:
5574 // If bits are being or'd in that are not present in the constant we
5575 // are comparing against, then the comparison could never succeed!
5576 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5577 Constant *NotCI = ConstantExpr::getNot(RHS);
5578 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5579 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5584 case Instruction::And:
5585 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5586 // If bits are being compared against that are and'd out, then the
5587 // comparison can never succeed!
5588 if ((RHSV & ~BOC->getValue()) != 0)
5589 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5592 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5593 if (RHS == BOC && RHSV.isPowerOf2())
5594 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5595 ICmpInst::ICMP_NE, LHSI,
5596 Constant::getNullValue(RHS->getType()));
5598 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5599 if (isSignBit(BOC)) {
5600 Value *X = BO->getOperand(0);
5601 Constant *Zero = Constant::getNullValue(X->getType());
5602 ICmpInst::Predicate pred = isICMP_NE ?
5603 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5604 return new ICmpInst(pred, X, Zero);
5607 // ((X & ~7) == 0) --> X < 8
5608 if (RHSV == 0 && isHighOnes(BOC)) {
5609 Value *X = BO->getOperand(0);
5610 Constant *NegX = ConstantExpr::getNeg(BOC);
5611 ICmpInst::Predicate pred = isICMP_NE ?
5612 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5613 return new ICmpInst(pred, X, NegX);
5618 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5619 // Handle icmp {eq|ne} <intrinsic>, intcst.
5620 if (II->getIntrinsicID() == Intrinsic::bswap) {
5622 ICI.setOperand(0, II->getOperand(1));
5623 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5627 } else { // Not a ICMP_EQ/ICMP_NE
5628 // If the LHS is a cast from an integral value of the same size,
5629 // then since we know the RHS is a constant, try to simlify.
5630 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5631 Value *CastOp = Cast->getOperand(0);
5632 const Type *SrcTy = CastOp->getType();
5633 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5634 if (SrcTy->isInteger() &&
5635 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5636 // If this is an unsigned comparison, try to make the comparison use
5637 // smaller constant values.
5638 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5639 // X u< 128 => X s> -1
5640 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5641 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5642 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5643 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5644 // X u> 127 => X s< 0
5645 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5646 Constant::getNullValue(SrcTy));
5654 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5655 /// We only handle extending casts so far.
5657 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5658 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5659 Value *LHSCIOp = LHSCI->getOperand(0);
5660 const Type *SrcTy = LHSCIOp->getType();
5661 const Type *DestTy = LHSCI->getType();
5664 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5665 // integer type is the same size as the pointer type.
5666 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5667 getTargetData().getPointerSizeInBits() ==
5668 cast<IntegerType>(DestTy)->getBitWidth()) {
5670 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5671 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5672 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5673 RHSOp = RHSC->getOperand(0);
5674 // If the pointer types don't match, insert a bitcast.
5675 if (LHSCIOp->getType() != RHSOp->getType())
5676 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5677 LHSCIOp->getType(), ICI);
5681 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5684 // The code below only handles extension cast instructions, so far.
5686 if (LHSCI->getOpcode() != Instruction::ZExt &&
5687 LHSCI->getOpcode() != Instruction::SExt)
5690 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5691 bool isSignedCmp = ICI.isSignedPredicate();
5693 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5694 // Not an extension from the same type?
5695 RHSCIOp = CI->getOperand(0);
5696 if (RHSCIOp->getType() != LHSCIOp->getType())
5699 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5700 // and the other is a zext), then we can't handle this.
5701 if (CI->getOpcode() != LHSCI->getOpcode())
5704 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5705 // then we can't handle this.
5706 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5709 // Okay, just insert a compare of the reduced operands now!
5710 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5713 // If we aren't dealing with a constant on the RHS, exit early
5714 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5718 // Compute the constant that would happen if we truncated to SrcTy then
5719 // reextended to DestTy.
5720 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5721 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5723 // If the re-extended constant didn't change...
5725 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5726 // For example, we might have:
5727 // %A = sext short %X to uint
5728 // %B = icmp ugt uint %A, 1330
5729 // It is incorrect to transform this into
5730 // %B = icmp ugt short %X, 1330
5731 // because %A may have negative value.
5733 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5734 // OR operation is EQ/NE.
5735 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5736 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5741 // The re-extended constant changed so the constant cannot be represented
5742 // in the shorter type. Consequently, we cannot emit a simple comparison.
5744 // First, handle some easy cases. We know the result cannot be equal at this
5745 // point so handle the ICI.isEquality() cases
5746 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5747 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5748 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5749 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5751 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5752 // should have been folded away previously and not enter in here.
5755 // We're performing a signed comparison.
5756 if (cast<ConstantInt>(CI)->getValue().isNegative())
5757 Result = ConstantInt::getFalse(); // X < (small) --> false
5759 Result = ConstantInt::getTrue(); // X < (large) --> true
5761 // We're performing an unsigned comparison.
5763 // We're performing an unsigned comp with a sign extended value.
5764 // This is true if the input is >= 0. [aka >s -1]
5765 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5766 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5767 NegOne, ICI.getName()), ICI);
5769 // Unsigned extend & unsigned compare -> always true.
5770 Result = ConstantInt::getTrue();
5774 // Finally, return the value computed.
5775 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5776 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5777 return ReplaceInstUsesWith(ICI, Result);
5779 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5780 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5781 "ICmp should be folded!");
5782 if (Constant *CI = dyn_cast<Constant>(Result))
5783 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5785 return BinaryOperator::createNot(Result);
5789 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5790 return commonShiftTransforms(I);
5793 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5794 return commonShiftTransforms(I);
5797 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5798 return commonShiftTransforms(I);
5801 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5802 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5803 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5805 // shl X, 0 == X and shr X, 0 == X
5806 // shl 0, X == 0 and shr 0, X == 0
5807 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5808 Op0 == Constant::getNullValue(Op0->getType()))
5809 return ReplaceInstUsesWith(I, Op0);
5811 if (isa<UndefValue>(Op0)) {
5812 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5813 return ReplaceInstUsesWith(I, Op0);
5814 else // undef << X -> 0, undef >>u X -> 0
5815 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5817 if (isa<UndefValue>(Op1)) {
5818 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5819 return ReplaceInstUsesWith(I, Op0);
5820 else // X << undef, X >>u undef -> 0
5821 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5824 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5825 if (I.getOpcode() == Instruction::AShr)
5826 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5827 if (CSI->isAllOnesValue())
5828 return ReplaceInstUsesWith(I, CSI);
5830 // Try to fold constant and into select arguments.
5831 if (isa<Constant>(Op0))
5832 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5833 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5836 // See if we can turn a signed shr into an unsigned shr.
5837 if (I.isArithmeticShift()) {
5838 if (MaskedValueIsZero(Op0,
5839 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5840 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5844 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5845 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5850 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5851 BinaryOperator &I) {
5852 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5854 // See if we can simplify any instructions used by the instruction whose sole
5855 // purpose is to compute bits we don't care about.
5856 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5857 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5858 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5859 KnownZero, KnownOne))
5862 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5863 // of a signed value.
5865 if (Op1->uge(TypeBits)) {
5866 if (I.getOpcode() != Instruction::AShr)
5867 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5869 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5874 // ((X*C1) << C2) == (X * (C1 << C2))
5875 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5876 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5877 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5878 return BinaryOperator::createMul(BO->getOperand(0),
5879 ConstantExpr::getShl(BOOp, Op1));
5881 // Try to fold constant and into select arguments.
5882 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5883 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5885 if (isa<PHINode>(Op0))
5886 if (Instruction *NV = FoldOpIntoPhi(I))
5889 if (Op0->hasOneUse()) {
5890 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5891 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5894 switch (Op0BO->getOpcode()) {
5896 case Instruction::Add:
5897 case Instruction::And:
5898 case Instruction::Or:
5899 case Instruction::Xor: {
5900 // These operators commute.
5901 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5902 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5903 match(Op0BO->getOperand(1),
5904 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5905 Instruction *YS = BinaryOperator::createShl(
5906 Op0BO->getOperand(0), Op1,
5908 InsertNewInstBefore(YS, I); // (Y << C)
5910 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5911 Op0BO->getOperand(1)->getName());
5912 InsertNewInstBefore(X, I); // (X + (Y << C))
5913 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5914 return BinaryOperator::createAnd(X, ConstantInt::get(
5915 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5918 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5919 Value *Op0BOOp1 = Op0BO->getOperand(1);
5920 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5922 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5923 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5925 Instruction *YS = BinaryOperator::createShl(
5926 Op0BO->getOperand(0), Op1,
5928 InsertNewInstBefore(YS, I); // (Y << C)
5930 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5931 V1->getName()+".mask");
5932 InsertNewInstBefore(XM, I); // X & (CC << C)
5934 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5939 case Instruction::Sub: {
5940 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5941 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5942 match(Op0BO->getOperand(0),
5943 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5944 Instruction *YS = BinaryOperator::createShl(
5945 Op0BO->getOperand(1), Op1,
5947 InsertNewInstBefore(YS, I); // (Y << C)
5949 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5950 Op0BO->getOperand(0)->getName());
5951 InsertNewInstBefore(X, I); // (X + (Y << C))
5952 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5953 return BinaryOperator::createAnd(X, ConstantInt::get(
5954 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5957 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5958 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5959 match(Op0BO->getOperand(0),
5960 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5961 m_ConstantInt(CC))) && V2 == Op1 &&
5962 cast<BinaryOperator>(Op0BO->getOperand(0))
5963 ->getOperand(0)->hasOneUse()) {
5964 Instruction *YS = BinaryOperator::createShl(
5965 Op0BO->getOperand(1), Op1,
5967 InsertNewInstBefore(YS, I); // (Y << C)
5969 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5970 V1->getName()+".mask");
5971 InsertNewInstBefore(XM, I); // X & (CC << C)
5973 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5981 // If the operand is an bitwise operator with a constant RHS, and the
5982 // shift is the only use, we can pull it out of the shift.
5983 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5984 bool isValid = true; // Valid only for And, Or, Xor
5985 bool highBitSet = false; // Transform if high bit of constant set?
5987 switch (Op0BO->getOpcode()) {
5988 default: isValid = false; break; // Do not perform transform!
5989 case Instruction::Add:
5990 isValid = isLeftShift;
5992 case Instruction::Or:
5993 case Instruction::Xor:
5996 case Instruction::And:
6001 // If this is a signed shift right, and the high bit is modified
6002 // by the logical operation, do not perform the transformation.
6003 // The highBitSet boolean indicates the value of the high bit of
6004 // the constant which would cause it to be modified for this
6007 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
6008 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6012 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6014 Instruction *NewShift =
6015 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6016 InsertNewInstBefore(NewShift, I);
6017 NewShift->takeName(Op0BO);
6019 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6026 // Find out if this is a shift of a shift by a constant.
6027 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6028 if (ShiftOp && !ShiftOp->isShift())
6031 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6032 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6033 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6034 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6035 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6036 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6037 Value *X = ShiftOp->getOperand(0);
6039 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6040 if (AmtSum > TypeBits)
6043 const IntegerType *Ty = cast<IntegerType>(I.getType());
6045 // Check for (X << c1) << c2 and (X >> c1) >> c2
6046 if (I.getOpcode() == ShiftOp->getOpcode()) {
6047 return BinaryOperator::create(I.getOpcode(), X,
6048 ConstantInt::get(Ty, AmtSum));
6049 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6050 I.getOpcode() == Instruction::AShr) {
6051 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6052 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6053 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6054 I.getOpcode() == Instruction::LShr) {
6055 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6056 Instruction *Shift =
6057 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6058 InsertNewInstBefore(Shift, I);
6060 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6061 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6064 // Okay, if we get here, one shift must be left, and the other shift must be
6065 // right. See if the amounts are equal.
6066 if (ShiftAmt1 == ShiftAmt2) {
6067 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6068 if (I.getOpcode() == Instruction::Shl) {
6069 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6070 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6072 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6073 if (I.getOpcode() == Instruction::LShr) {
6074 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6075 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6077 // We can simplify ((X << C) >>s C) into a trunc + sext.
6078 // NOTE: we could do this for any C, but that would make 'unusual' integer
6079 // types. For now, just stick to ones well-supported by the code
6081 const Type *SExtType = 0;
6082 switch (Ty->getBitWidth() - ShiftAmt1) {
6089 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6094 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6095 InsertNewInstBefore(NewTrunc, I);
6096 return new SExtInst(NewTrunc, Ty);
6098 // Otherwise, we can't handle it yet.
6099 } else if (ShiftAmt1 < ShiftAmt2) {
6100 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6102 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6103 if (I.getOpcode() == Instruction::Shl) {
6104 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6105 ShiftOp->getOpcode() == Instruction::AShr);
6106 Instruction *Shift =
6107 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6108 InsertNewInstBefore(Shift, I);
6110 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6111 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6114 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6115 if (I.getOpcode() == Instruction::LShr) {
6116 assert(ShiftOp->getOpcode() == Instruction::Shl);
6117 Instruction *Shift =
6118 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6119 InsertNewInstBefore(Shift, I);
6121 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6122 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6125 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6127 assert(ShiftAmt2 < ShiftAmt1);
6128 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6130 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6131 if (I.getOpcode() == Instruction::Shl) {
6132 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6133 ShiftOp->getOpcode() == Instruction::AShr);
6134 Instruction *Shift =
6135 BinaryOperator::create(ShiftOp->getOpcode(), X,
6136 ConstantInt::get(Ty, ShiftDiff));
6137 InsertNewInstBefore(Shift, I);
6139 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6140 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6143 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6144 if (I.getOpcode() == Instruction::LShr) {
6145 assert(ShiftOp->getOpcode() == Instruction::Shl);
6146 Instruction *Shift =
6147 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6148 InsertNewInstBefore(Shift, I);
6150 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6151 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6154 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6161 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6162 /// expression. If so, decompose it, returning some value X, such that Val is
6165 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6167 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6168 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6169 Offset = CI->getZExtValue();
6171 return ConstantInt::get(Type::Int32Ty, 0);
6172 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
6173 if (I->getNumOperands() == 2) {
6174 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6175 if (I->getOpcode() == Instruction::Shl) {
6176 // This is a value scaled by '1 << the shift amt'.
6177 Scale = 1U << CUI->getZExtValue();
6179 return I->getOperand(0);
6180 } else if (I->getOpcode() == Instruction::Mul) {
6181 // This value is scaled by 'CUI'.
6182 Scale = CUI->getZExtValue();
6184 return I->getOperand(0);
6185 } else if (I->getOpcode() == Instruction::Add) {
6186 // We have X+C. Check to see if we really have (X*C2)+C1,
6187 // where C1 is divisible by C2.
6190 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6191 Offset += CUI->getZExtValue();
6192 if (SubScale > 1 && (Offset % SubScale == 0)) {
6201 // Otherwise, we can't look past this.
6208 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6209 /// try to eliminate the cast by moving the type information into the alloc.
6210 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6211 AllocationInst &AI) {
6212 const PointerType *PTy = cast<PointerType>(CI.getType());
6214 // Remove any uses of AI that are dead.
6215 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6217 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6218 Instruction *User = cast<Instruction>(*UI++);
6219 if (isInstructionTriviallyDead(User)) {
6220 while (UI != E && *UI == User)
6221 ++UI; // If this instruction uses AI more than once, don't break UI.
6224 DOUT << "IC: DCE: " << *User;
6225 EraseInstFromFunction(*User);
6229 // Get the type really allocated and the type casted to.
6230 const Type *AllocElTy = AI.getAllocatedType();
6231 const Type *CastElTy = PTy->getElementType();
6232 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6234 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6235 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6236 if (CastElTyAlign < AllocElTyAlign) return 0;
6238 // If the allocation has multiple uses, only promote it if we are strictly
6239 // increasing the alignment of the resultant allocation. If we keep it the
6240 // same, we open the door to infinite loops of various kinds.
6241 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6243 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
6244 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
6245 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6247 // See if we can satisfy the modulus by pulling a scale out of the array
6249 unsigned ArraySizeScale;
6251 Value *NumElements = // See if the array size is a decomposable linear expr.
6252 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6254 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6256 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6257 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6259 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6264 // If the allocation size is constant, form a constant mul expression
6265 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6266 if (isa<ConstantInt>(NumElements))
6267 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6268 // otherwise multiply the amount and the number of elements
6269 else if (Scale != 1) {
6270 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6271 Amt = InsertNewInstBefore(Tmp, AI);
6275 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6276 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6277 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6278 Amt = InsertNewInstBefore(Tmp, AI);
6281 AllocationInst *New;
6282 if (isa<MallocInst>(AI))
6283 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6285 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6286 InsertNewInstBefore(New, AI);
6289 // If the allocation has multiple uses, insert a cast and change all things
6290 // that used it to use the new cast. This will also hack on CI, but it will
6292 if (!AI.hasOneUse()) {
6293 AddUsesToWorkList(AI);
6294 // New is the allocation instruction, pointer typed. AI is the original
6295 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6296 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6297 InsertNewInstBefore(NewCast, AI);
6298 AI.replaceAllUsesWith(NewCast);
6300 return ReplaceInstUsesWith(CI, New);
6303 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6304 /// and return it as type Ty without inserting any new casts and without
6305 /// changing the computed value. This is used by code that tries to decide
6306 /// whether promoting or shrinking integer operations to wider or smaller types
6307 /// will allow us to eliminate a truncate or extend.
6309 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6310 /// extension operation if Ty is larger.
6311 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6312 int &NumCastsRemoved) {
6313 // We can always evaluate constants in another type.
6314 if (isa<ConstantInt>(V))
6317 Instruction *I = dyn_cast<Instruction>(V);
6318 if (!I) return false;
6320 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6322 switch (I->getOpcode()) {
6323 case Instruction::Add:
6324 case Instruction::Sub:
6325 case Instruction::And:
6326 case Instruction::Or:
6327 case Instruction::Xor:
6328 if (!I->hasOneUse()) return false;
6329 // These operators can all arbitrarily be extended or truncated.
6330 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
6331 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
6333 case Instruction::Shl:
6334 if (!I->hasOneUse()) return false;
6335 // If we are truncating the result of this SHL, and if it's a shift of a
6336 // constant amount, we can always perform a SHL in a smaller type.
6337 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6338 uint32_t BitWidth = Ty->getBitWidth();
6339 if (BitWidth < OrigTy->getBitWidth() &&
6340 CI->getLimitedValue(BitWidth) < BitWidth)
6341 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6344 case Instruction::LShr:
6345 if (!I->hasOneUse()) return false;
6346 // If this is a truncate of a logical shr, we can truncate it to a smaller
6347 // lshr iff we know that the bits we would otherwise be shifting in are
6349 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6350 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6351 uint32_t BitWidth = Ty->getBitWidth();
6352 if (BitWidth < OrigBitWidth &&
6353 MaskedValueIsZero(I->getOperand(0),
6354 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6355 CI->getLimitedValue(BitWidth) < BitWidth) {
6356 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6360 case Instruction::Trunc:
6361 case Instruction::ZExt:
6362 case Instruction::SExt:
6363 // If this is a cast from the destination type, we can trivially eliminate
6364 // it, and this will remove a cast overall.
6365 if (I->getOperand(0)->getType() == Ty) {
6366 // If the first operand is itself a cast, and is eliminable, do not count
6367 // this as an eliminable cast. We would prefer to eliminate those two
6369 if (isa<CastInst>(I->getOperand(0)))
6377 // TODO: Can handle more cases here.
6384 /// EvaluateInDifferentType - Given an expression that
6385 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6386 /// evaluate the expression.
6387 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6389 if (Constant *C = dyn_cast<Constant>(V))
6390 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6392 // Otherwise, it must be an instruction.
6393 Instruction *I = cast<Instruction>(V);
6394 Instruction *Res = 0;
6395 switch (I->getOpcode()) {
6396 case Instruction::Add:
6397 case Instruction::Sub:
6398 case Instruction::And:
6399 case Instruction::Or:
6400 case Instruction::Xor:
6401 case Instruction::AShr:
6402 case Instruction::LShr:
6403 case Instruction::Shl: {
6404 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6405 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6406 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6407 LHS, RHS, I->getName());
6410 case Instruction::Trunc:
6411 case Instruction::ZExt:
6412 case Instruction::SExt:
6413 case Instruction::BitCast:
6414 // If the source type of the cast is the type we're trying for then we can
6415 // just return the source. There's no need to insert it because its not new.
6416 if (I->getOperand(0)->getType() == Ty)
6417 return I->getOperand(0);
6419 // Some other kind of cast, which shouldn't happen, so just ..
6422 // TODO: Can handle more cases here.
6423 assert(0 && "Unreachable!");
6427 return InsertNewInstBefore(Res, *I);
6430 /// @brief Implement the transforms common to all CastInst visitors.
6431 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6432 Value *Src = CI.getOperand(0);
6434 // Casting undef to anything results in undef so might as just replace it and
6435 // get rid of the cast.
6436 if (isa<UndefValue>(Src)) // cast undef -> undef
6437 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6439 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6440 // eliminate it now.
6441 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6442 if (Instruction::CastOps opc =
6443 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6444 // The first cast (CSrc) is eliminable so we need to fix up or replace
6445 // the second cast (CI). CSrc will then have a good chance of being dead.
6446 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6450 // If we are casting a select then fold the cast into the select
6451 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6452 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6455 // If we are casting a PHI then fold the cast into the PHI
6456 if (isa<PHINode>(Src))
6457 if (Instruction *NV = FoldOpIntoPhi(CI))
6463 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6464 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6465 Value *Src = CI.getOperand(0);
6467 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6468 // If casting the result of a getelementptr instruction with no offset, turn
6469 // this into a cast of the original pointer!
6470 if (GEP->hasAllZeroIndices()) {
6471 // Changing the cast operand is usually not a good idea but it is safe
6472 // here because the pointer operand is being replaced with another
6473 // pointer operand so the opcode doesn't need to change.
6475 CI.setOperand(0, GEP->getOperand(0));
6479 // If the GEP has a single use, and the base pointer is a bitcast, and the
6480 // GEP computes a constant offset, see if we can convert these three
6481 // instructions into fewer. This typically happens with unions and other
6482 // non-type-safe code.
6483 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6484 if (GEP->hasAllConstantIndices()) {
6485 // We are guaranteed to get a constant from EmitGEPOffset.
6486 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6487 int64_t Offset = OffsetV->getSExtValue();
6489 // Get the base pointer input of the bitcast, and the type it points to.
6490 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6491 const Type *GEPIdxTy =
6492 cast<PointerType>(OrigBase->getType())->getElementType();
6493 if (GEPIdxTy->isSized()) {
6494 SmallVector<Value*, 8> NewIndices;
6496 // Start with the index over the outer type. Note that the type size
6497 // might be zero (even if the offset isn't zero) if the indexed type
6498 // is something like [0 x {int, int}]
6499 const Type *IntPtrTy = TD->getIntPtrType();
6500 int64_t FirstIdx = 0;
6501 if (int64_t TySize = TD->getTypeSize(GEPIdxTy)) {
6502 FirstIdx = Offset/TySize;
6505 // Handle silly modulus not returning values values [0..TySize).
6509 assert(Offset >= 0);
6511 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6514 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6516 // Index into the types. If we fail, set OrigBase to null.
6518 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6519 const StructLayout *SL = TD->getStructLayout(STy);
6520 if (Offset < (int64_t)SL->getSizeInBytes()) {
6521 unsigned Elt = SL->getElementContainingOffset(Offset);
6522 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6524 Offset -= SL->getElementOffset(Elt);
6525 GEPIdxTy = STy->getElementType(Elt);
6527 // Otherwise, we can't index into this, bail out.
6531 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6532 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6533 if (uint64_t EltSize = TD->getTypeSize(STy->getElementType())) {
6534 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6537 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6539 GEPIdxTy = STy->getElementType();
6541 // Otherwise, we can't index into this, bail out.
6547 // If we were able to index down into an element, create the GEP
6548 // and bitcast the result. This eliminates one bitcast, potentially
6550 Instruction *NGEP = new GetElementPtrInst(OrigBase, &NewIndices[0],
6551 NewIndices.size(), "");
6552 InsertNewInstBefore(NGEP, CI);
6553 NGEP->takeName(GEP);
6555 if (isa<BitCastInst>(CI))
6556 return new BitCastInst(NGEP, CI.getType());
6557 assert(isa<PtrToIntInst>(CI));
6558 return new PtrToIntInst(NGEP, CI.getType());
6565 return commonCastTransforms(CI);
6570 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6571 /// integer types. This function implements the common transforms for all those
6573 /// @brief Implement the transforms common to CastInst with integer operands
6574 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6575 if (Instruction *Result = commonCastTransforms(CI))
6578 Value *Src = CI.getOperand(0);
6579 const Type *SrcTy = Src->getType();
6580 const Type *DestTy = CI.getType();
6581 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6582 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6584 // See if we can simplify any instructions used by the LHS whose sole
6585 // purpose is to compute bits we don't care about.
6586 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6587 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6588 KnownZero, KnownOne))
6591 // If the source isn't an instruction or has more than one use then we
6592 // can't do anything more.
6593 Instruction *SrcI = dyn_cast<Instruction>(Src);
6594 if (!SrcI || !Src->hasOneUse())
6597 // Attempt to propagate the cast into the instruction for int->int casts.
6598 int NumCastsRemoved = 0;
6599 if (!isa<BitCastInst>(CI) &&
6600 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6602 // If this cast is a truncate, evaluting in a different type always
6603 // eliminates the cast, so it is always a win. If this is a noop-cast
6604 // this just removes a noop cast which isn't pointful, but simplifies
6605 // the code. If this is a zero-extension, we need to do an AND to
6606 // maintain the clear top-part of the computation, so we require that
6607 // the input have eliminated at least one cast. If this is a sign
6608 // extension, we insert two new casts (to do the extension) so we
6609 // require that two casts have been eliminated.
6611 switch (CI.getOpcode()) {
6613 // All the others use floating point so we shouldn't actually
6614 // get here because of the check above.
6615 assert(0 && "Unknown cast type");
6616 case Instruction::Trunc:
6619 case Instruction::ZExt:
6620 DoXForm = NumCastsRemoved >= 1;
6622 case Instruction::SExt:
6623 DoXForm = NumCastsRemoved >= 2;
6625 case Instruction::BitCast:
6631 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6632 CI.getOpcode() == Instruction::SExt);
6633 assert(Res->getType() == DestTy);
6634 switch (CI.getOpcode()) {
6635 default: assert(0 && "Unknown cast type!");
6636 case Instruction::Trunc:
6637 case Instruction::BitCast:
6638 // Just replace this cast with the result.
6639 return ReplaceInstUsesWith(CI, Res);
6640 case Instruction::ZExt: {
6641 // We need to emit an AND to clear the high bits.
6642 assert(SrcBitSize < DestBitSize && "Not a zext?");
6643 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6645 return BinaryOperator::createAnd(Res, C);
6647 case Instruction::SExt:
6648 // We need to emit a cast to truncate, then a cast to sext.
6649 return CastInst::create(Instruction::SExt,
6650 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6656 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6657 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6659 switch (SrcI->getOpcode()) {
6660 case Instruction::Add:
6661 case Instruction::Mul:
6662 case Instruction::And:
6663 case Instruction::Or:
6664 case Instruction::Xor:
6665 // If we are discarding information, rewrite.
6666 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6667 // Don't insert two casts if they cannot be eliminated. We allow
6668 // two casts to be inserted if the sizes are the same. This could
6669 // only be converting signedness, which is a noop.
6670 if (DestBitSize == SrcBitSize ||
6671 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6672 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6673 Instruction::CastOps opcode = CI.getOpcode();
6674 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6675 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6676 return BinaryOperator::create(
6677 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6681 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6682 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6683 SrcI->getOpcode() == Instruction::Xor &&
6684 Op1 == ConstantInt::getTrue() &&
6685 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6686 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6687 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6690 case Instruction::SDiv:
6691 case Instruction::UDiv:
6692 case Instruction::SRem:
6693 case Instruction::URem:
6694 // If we are just changing the sign, rewrite.
6695 if (DestBitSize == SrcBitSize) {
6696 // Don't insert two casts if they cannot be eliminated. We allow
6697 // two casts to be inserted if the sizes are the same. This could
6698 // only be converting signedness, which is a noop.
6699 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6700 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6701 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6703 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6705 return BinaryOperator::create(
6706 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6711 case Instruction::Shl:
6712 // Allow changing the sign of the source operand. Do not allow
6713 // changing the size of the shift, UNLESS the shift amount is a
6714 // constant. We must not change variable sized shifts to a smaller
6715 // size, because it is undefined to shift more bits out than exist
6717 if (DestBitSize == SrcBitSize ||
6718 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6719 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6720 Instruction::BitCast : Instruction::Trunc);
6721 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6722 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6723 return BinaryOperator::createShl(Op0c, Op1c);
6726 case Instruction::AShr:
6727 // If this is a signed shr, and if all bits shifted in are about to be
6728 // truncated off, turn it into an unsigned shr to allow greater
6730 if (DestBitSize < SrcBitSize &&
6731 isa<ConstantInt>(Op1)) {
6732 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6733 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6734 // Insert the new logical shift right.
6735 return BinaryOperator::createLShr(Op0, Op1);
6743 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6744 if (Instruction *Result = commonIntCastTransforms(CI))
6747 Value *Src = CI.getOperand(0);
6748 const Type *Ty = CI.getType();
6749 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6750 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6752 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6753 switch (SrcI->getOpcode()) {
6755 case Instruction::LShr:
6756 // We can shrink lshr to something smaller if we know the bits shifted in
6757 // are already zeros.
6758 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6759 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6761 // Get a mask for the bits shifting in.
6762 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6763 Value* SrcIOp0 = SrcI->getOperand(0);
6764 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6765 if (ShAmt >= DestBitWidth) // All zeros.
6766 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6768 // Okay, we can shrink this. Truncate the input, then return a new
6770 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6771 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6773 return BinaryOperator::createLShr(V1, V2);
6775 } else { // This is a variable shr.
6777 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6778 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6779 // loop-invariant and CSE'd.
6780 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6781 Value *One = ConstantInt::get(SrcI->getType(), 1);
6783 Value *V = InsertNewInstBefore(
6784 BinaryOperator::createShl(One, SrcI->getOperand(1),
6786 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6787 SrcI->getOperand(0),
6789 Value *Zero = Constant::getNullValue(V->getType());
6790 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6800 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6801 // If one of the common conversion will work ..
6802 if (Instruction *Result = commonIntCastTransforms(CI))
6805 Value *Src = CI.getOperand(0);
6807 // If this is a cast of a cast
6808 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6809 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6810 // types and if the sizes are just right we can convert this into a logical
6811 // 'and' which will be much cheaper than the pair of casts.
6812 if (isa<TruncInst>(CSrc)) {
6813 // Get the sizes of the types involved
6814 Value *A = CSrc->getOperand(0);
6815 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6816 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6817 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6818 // If we're actually extending zero bits and the trunc is a no-op
6819 if (MidSize < DstSize && SrcSize == DstSize) {
6820 // Replace both of the casts with an And of the type mask.
6821 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6822 Constant *AndConst = ConstantInt::get(AndValue);
6824 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6825 // Unfortunately, if the type changed, we need to cast it back.
6826 if (And->getType() != CI.getType()) {
6827 And->setName(CSrc->getName()+".mask");
6828 InsertNewInstBefore(And, CI);
6829 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6836 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6837 // If we are just checking for a icmp eq of a single bit and zext'ing it
6838 // to an integer, then shift the bit to the appropriate place and then
6839 // cast to integer to avoid the comparison.
6840 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6841 const APInt &Op1CV = Op1C->getValue();
6843 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
6844 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
6845 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6846 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6847 Value *In = ICI->getOperand(0);
6848 Value *Sh = ConstantInt::get(In->getType(),
6849 In->getType()->getPrimitiveSizeInBits()-1);
6850 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
6851 In->getName()+".lobit"),
6853 if (In->getType() != CI.getType())
6854 In = CastInst::createIntegerCast(In, CI.getType(),
6855 false/*ZExt*/, "tmp", &CI);
6857 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
6858 Constant *One = ConstantInt::get(In->getType(), 1);
6859 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
6860 In->getName()+".not"),
6864 return ReplaceInstUsesWith(CI, In);
6869 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
6870 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6871 // zext (X == 1) to i32 --> X iff X has only the low bit set.
6872 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
6873 // zext (X != 0) to i32 --> X iff X has only the low bit set.
6874 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
6875 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
6876 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6877 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
6878 // This only works for EQ and NE
6879 ICI->isEquality()) {
6880 // If Op1C some other power of two, convert:
6881 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6882 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6883 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6884 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
6886 APInt KnownZeroMask(~KnownZero);
6887 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6888 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
6889 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6890 // (X&4) == 2 --> false
6891 // (X&4) != 2 --> true
6892 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6893 Res = ConstantExpr::getZExt(Res, CI.getType());
6894 return ReplaceInstUsesWith(CI, Res);
6897 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6898 Value *In = ICI->getOperand(0);
6900 // Perform a logical shr by shiftamt.
6901 // Insert the shift to put the result in the low bit.
6902 In = InsertNewInstBefore(
6903 BinaryOperator::createLShr(In,
6904 ConstantInt::get(In->getType(), ShiftAmt),
6905 In->getName()+".lobit"), CI);
6908 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6909 Constant *One = ConstantInt::get(In->getType(), 1);
6910 In = BinaryOperator::createXor(In, One, "tmp");
6911 InsertNewInstBefore(cast<Instruction>(In), CI);
6914 if (CI.getType() == In->getType())
6915 return ReplaceInstUsesWith(CI, In);
6917 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6925 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
6926 if (Instruction *I = commonIntCastTransforms(CI))
6929 Value *Src = CI.getOperand(0);
6931 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
6932 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
6933 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6934 // If we are just checking for a icmp eq of a single bit and zext'ing it
6935 // to an integer, then shift the bit to the appropriate place and then
6936 // cast to integer to avoid the comparison.
6937 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6938 const APInt &Op1CV = Op1C->getValue();
6940 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
6941 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
6942 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6943 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6944 Value *In = ICI->getOperand(0);
6945 Value *Sh = ConstantInt::get(In->getType(),
6946 In->getType()->getPrimitiveSizeInBits()-1);
6947 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
6948 In->getName()+".lobit"),
6950 if (In->getType() != CI.getType())
6951 In = CastInst::createIntegerCast(In, CI.getType(),
6952 true/*SExt*/, "tmp", &CI);
6954 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
6955 In = InsertNewInstBefore(BinaryOperator::createNot(In,
6956 In->getName()+".not"), CI);
6958 return ReplaceInstUsesWith(CI, In);
6966 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6967 return commonCastTransforms(CI);
6970 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6971 return commonCastTransforms(CI);
6974 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6975 return commonCastTransforms(CI);
6978 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6979 return commonCastTransforms(CI);
6982 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6983 return commonCastTransforms(CI);
6986 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6987 return commonCastTransforms(CI);
6990 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6991 return commonPointerCastTransforms(CI);
6994 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6995 return commonCastTransforms(CI);
6998 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
6999 // If the operands are integer typed then apply the integer transforms,
7000 // otherwise just apply the common ones.
7001 Value *Src = CI.getOperand(0);
7002 const Type *SrcTy = Src->getType();
7003 const Type *DestTy = CI.getType();
7005 if (SrcTy->isInteger() && DestTy->isInteger()) {
7006 if (Instruction *Result = commonIntCastTransforms(CI))
7008 } else if (isa<PointerType>(SrcTy)) {
7009 if (Instruction *I = commonPointerCastTransforms(CI))
7012 if (Instruction *Result = commonCastTransforms(CI))
7017 // Get rid of casts from one type to the same type. These are useless and can
7018 // be replaced by the operand.
7019 if (DestTy == Src->getType())
7020 return ReplaceInstUsesWith(CI, Src);
7022 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7023 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7024 const Type *DstElTy = DstPTy->getElementType();
7025 const Type *SrcElTy = SrcPTy->getElementType();
7027 // If we are casting a malloc or alloca to a pointer to a type of the same
7028 // size, rewrite the allocation instruction to allocate the "right" type.
7029 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7030 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7033 // If the source and destination are pointers, and this cast is equivalent
7034 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7035 // This can enhance SROA and other transforms that want type-safe pointers.
7036 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7037 unsigned NumZeros = 0;
7038 while (SrcElTy != DstElTy &&
7039 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7040 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7041 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7045 // If we found a path from the src to dest, create the getelementptr now.
7046 if (SrcElTy == DstElTy) {
7047 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7048 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
7052 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7053 if (SVI->hasOneUse()) {
7054 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7055 // a bitconvert to a vector with the same # elts.
7056 if (isa<VectorType>(DestTy) &&
7057 cast<VectorType>(DestTy)->getNumElements() ==
7058 SVI->getType()->getNumElements()) {
7060 // If either of the operands is a cast from CI.getType(), then
7061 // evaluating the shuffle in the casted destination's type will allow
7062 // us to eliminate at least one cast.
7063 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7064 Tmp->getOperand(0)->getType() == DestTy) ||
7065 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7066 Tmp->getOperand(0)->getType() == DestTy)) {
7067 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7068 SVI->getOperand(0), DestTy, &CI);
7069 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7070 SVI->getOperand(1), DestTy, &CI);
7071 // Return a new shuffle vector. Use the same element ID's, as we
7072 // know the vector types match #elts.
7073 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7081 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7083 /// %D = select %cond, %C, %A
7085 /// %C = select %cond, %B, 0
7088 /// Assuming that the specified instruction is an operand to the select, return
7089 /// a bitmask indicating which operands of this instruction are foldable if they
7090 /// equal the other incoming value of the select.
7092 static unsigned GetSelectFoldableOperands(Instruction *I) {
7093 switch (I->getOpcode()) {
7094 case Instruction::Add:
7095 case Instruction::Mul:
7096 case Instruction::And:
7097 case Instruction::Or:
7098 case Instruction::Xor:
7099 return 3; // Can fold through either operand.
7100 case Instruction::Sub: // Can only fold on the amount subtracted.
7101 case Instruction::Shl: // Can only fold on the shift amount.
7102 case Instruction::LShr:
7103 case Instruction::AShr:
7106 return 0; // Cannot fold
7110 /// GetSelectFoldableConstant - For the same transformation as the previous
7111 /// function, return the identity constant that goes into the select.
7112 static Constant *GetSelectFoldableConstant(Instruction *I) {
7113 switch (I->getOpcode()) {
7114 default: assert(0 && "This cannot happen!"); abort();
7115 case Instruction::Add:
7116 case Instruction::Sub:
7117 case Instruction::Or:
7118 case Instruction::Xor:
7119 case Instruction::Shl:
7120 case Instruction::LShr:
7121 case Instruction::AShr:
7122 return Constant::getNullValue(I->getType());
7123 case Instruction::And:
7124 return Constant::getAllOnesValue(I->getType());
7125 case Instruction::Mul:
7126 return ConstantInt::get(I->getType(), 1);
7130 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7131 /// have the same opcode and only one use each. Try to simplify this.
7132 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7134 if (TI->getNumOperands() == 1) {
7135 // If this is a non-volatile load or a cast from the same type,
7138 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7141 return 0; // unknown unary op.
7144 // Fold this by inserting a select from the input values.
7145 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7146 FI->getOperand(0), SI.getName()+".v");
7147 InsertNewInstBefore(NewSI, SI);
7148 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7152 // Only handle binary operators here.
7153 if (!isa<BinaryOperator>(TI))
7156 // Figure out if the operations have any operands in common.
7157 Value *MatchOp, *OtherOpT, *OtherOpF;
7159 if (TI->getOperand(0) == FI->getOperand(0)) {
7160 MatchOp = TI->getOperand(0);
7161 OtherOpT = TI->getOperand(1);
7162 OtherOpF = FI->getOperand(1);
7163 MatchIsOpZero = true;
7164 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7165 MatchOp = TI->getOperand(1);
7166 OtherOpT = TI->getOperand(0);
7167 OtherOpF = FI->getOperand(0);
7168 MatchIsOpZero = false;
7169 } else if (!TI->isCommutative()) {
7171 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7172 MatchOp = TI->getOperand(0);
7173 OtherOpT = TI->getOperand(1);
7174 OtherOpF = FI->getOperand(0);
7175 MatchIsOpZero = true;
7176 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7177 MatchOp = TI->getOperand(1);
7178 OtherOpT = TI->getOperand(0);
7179 OtherOpF = FI->getOperand(1);
7180 MatchIsOpZero = true;
7185 // If we reach here, they do have operations in common.
7186 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7187 OtherOpF, SI.getName()+".v");
7188 InsertNewInstBefore(NewSI, SI);
7190 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7192 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7194 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7196 assert(0 && "Shouldn't get here");
7200 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7201 Value *CondVal = SI.getCondition();
7202 Value *TrueVal = SI.getTrueValue();
7203 Value *FalseVal = SI.getFalseValue();
7205 // select true, X, Y -> X
7206 // select false, X, Y -> Y
7207 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7208 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7210 // select C, X, X -> X
7211 if (TrueVal == FalseVal)
7212 return ReplaceInstUsesWith(SI, TrueVal);
7214 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7215 return ReplaceInstUsesWith(SI, FalseVal);
7216 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7217 return ReplaceInstUsesWith(SI, TrueVal);
7218 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7219 if (isa<Constant>(TrueVal))
7220 return ReplaceInstUsesWith(SI, TrueVal);
7222 return ReplaceInstUsesWith(SI, FalseVal);
7225 if (SI.getType() == Type::Int1Ty) {
7226 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7227 if (C->getZExtValue()) {
7228 // Change: A = select B, true, C --> A = or B, C
7229 return BinaryOperator::createOr(CondVal, FalseVal);
7231 // Change: A = select B, false, C --> A = and !B, C
7233 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7234 "not."+CondVal->getName()), SI);
7235 return BinaryOperator::createAnd(NotCond, FalseVal);
7237 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7238 if (C->getZExtValue() == false) {
7239 // Change: A = select B, C, false --> A = and B, C
7240 return BinaryOperator::createAnd(CondVal, TrueVal);
7242 // Change: A = select B, C, true --> A = or !B, C
7244 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7245 "not."+CondVal->getName()), SI);
7246 return BinaryOperator::createOr(NotCond, TrueVal);
7251 // Selecting between two integer constants?
7252 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7253 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7254 // select C, 1, 0 -> zext C to int
7255 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7256 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7257 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7258 // select C, 0, 1 -> zext !C to int
7260 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7261 "not."+CondVal->getName()), SI);
7262 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7265 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7267 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7269 // (x <s 0) ? -1 : 0 -> ashr x, 31
7270 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7271 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7272 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7273 // The comparison constant and the result are not neccessarily the
7274 // same width. Make an all-ones value by inserting a AShr.
7275 Value *X = IC->getOperand(0);
7276 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7277 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7278 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7280 InsertNewInstBefore(SRA, SI);
7282 // Finally, convert to the type of the select RHS. We figure out
7283 // if this requires a SExt, Trunc or BitCast based on the sizes.
7284 Instruction::CastOps opc = Instruction::BitCast;
7285 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7286 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7287 if (SRASize < SISize)
7288 opc = Instruction::SExt;
7289 else if (SRASize > SISize)
7290 opc = Instruction::Trunc;
7291 return CastInst::create(opc, SRA, SI.getType());
7296 // If one of the constants is zero (we know they can't both be) and we
7297 // have an icmp instruction with zero, and we have an 'and' with the
7298 // non-constant value, eliminate this whole mess. This corresponds to
7299 // cases like this: ((X & 27) ? 27 : 0)
7300 if (TrueValC->isZero() || FalseValC->isZero())
7301 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7302 cast<Constant>(IC->getOperand(1))->isNullValue())
7303 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7304 if (ICA->getOpcode() == Instruction::And &&
7305 isa<ConstantInt>(ICA->getOperand(1)) &&
7306 (ICA->getOperand(1) == TrueValC ||
7307 ICA->getOperand(1) == FalseValC) &&
7308 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7309 // Okay, now we know that everything is set up, we just don't
7310 // know whether we have a icmp_ne or icmp_eq and whether the
7311 // true or false val is the zero.
7312 bool ShouldNotVal = !TrueValC->isZero();
7313 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7316 V = InsertNewInstBefore(BinaryOperator::create(
7317 Instruction::Xor, V, ICA->getOperand(1)), SI);
7318 return ReplaceInstUsesWith(SI, V);
7323 // See if we are selecting two values based on a comparison of the two values.
7324 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7325 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7326 // Transform (X == Y) ? X : Y -> Y
7327 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7328 return ReplaceInstUsesWith(SI, FalseVal);
7329 // Transform (X != Y) ? X : Y -> X
7330 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7331 return ReplaceInstUsesWith(SI, TrueVal);
7332 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7334 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7335 // Transform (X == Y) ? Y : X -> X
7336 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7337 return ReplaceInstUsesWith(SI, FalseVal);
7338 // Transform (X != Y) ? Y : X -> Y
7339 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7340 return ReplaceInstUsesWith(SI, TrueVal);
7341 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7345 // See if we are selecting two values based on a comparison of the two values.
7346 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7347 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7348 // Transform (X == Y) ? X : Y -> Y
7349 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7350 return ReplaceInstUsesWith(SI, FalseVal);
7351 // Transform (X != Y) ? X : Y -> X
7352 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7353 return ReplaceInstUsesWith(SI, TrueVal);
7354 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7356 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7357 // Transform (X == Y) ? Y : X -> X
7358 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7359 return ReplaceInstUsesWith(SI, FalseVal);
7360 // Transform (X != Y) ? Y : X -> Y
7361 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7362 return ReplaceInstUsesWith(SI, TrueVal);
7363 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7367 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7368 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7369 if (TI->hasOneUse() && FI->hasOneUse()) {
7370 Instruction *AddOp = 0, *SubOp = 0;
7372 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7373 if (TI->getOpcode() == FI->getOpcode())
7374 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7377 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7378 // even legal for FP.
7379 if (TI->getOpcode() == Instruction::Sub &&
7380 FI->getOpcode() == Instruction::Add) {
7381 AddOp = FI; SubOp = TI;
7382 } else if (FI->getOpcode() == Instruction::Sub &&
7383 TI->getOpcode() == Instruction::Add) {
7384 AddOp = TI; SubOp = FI;
7388 Value *OtherAddOp = 0;
7389 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7390 OtherAddOp = AddOp->getOperand(1);
7391 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7392 OtherAddOp = AddOp->getOperand(0);
7396 // So at this point we know we have (Y -> OtherAddOp):
7397 // select C, (add X, Y), (sub X, Z)
7398 Value *NegVal; // Compute -Z
7399 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7400 NegVal = ConstantExpr::getNeg(C);
7402 NegVal = InsertNewInstBefore(
7403 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7406 Value *NewTrueOp = OtherAddOp;
7407 Value *NewFalseOp = NegVal;
7409 std::swap(NewTrueOp, NewFalseOp);
7410 Instruction *NewSel =
7411 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7413 NewSel = InsertNewInstBefore(NewSel, SI);
7414 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7419 // See if we can fold the select into one of our operands.
7420 if (SI.getType()->isInteger()) {
7421 // See the comment above GetSelectFoldableOperands for a description of the
7422 // transformation we are doing here.
7423 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7424 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7425 !isa<Constant>(FalseVal))
7426 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7427 unsigned OpToFold = 0;
7428 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7430 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7435 Constant *C = GetSelectFoldableConstant(TVI);
7436 Instruction *NewSel =
7437 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7438 InsertNewInstBefore(NewSel, SI);
7439 NewSel->takeName(TVI);
7440 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7441 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7443 assert(0 && "Unknown instruction!!");
7448 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7449 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7450 !isa<Constant>(TrueVal))
7451 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7452 unsigned OpToFold = 0;
7453 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7455 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7460 Constant *C = GetSelectFoldableConstant(FVI);
7461 Instruction *NewSel =
7462 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7463 InsertNewInstBefore(NewSel, SI);
7464 NewSel->takeName(FVI);
7465 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7466 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7468 assert(0 && "Unknown instruction!!");
7473 if (BinaryOperator::isNot(CondVal)) {
7474 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7475 SI.setOperand(1, FalseVal);
7476 SI.setOperand(2, TrueVal);
7483 /// GetKnownAlignment - If the specified pointer has an alignment that we can
7484 /// determine, return it, otherwise return 0.
7485 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
7486 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7487 unsigned Align = GV->getAlignment();
7488 if (Align == 0 && TD)
7489 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7491 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7492 unsigned Align = AI->getAlignment();
7493 if (Align == 0 && TD) {
7494 if (isa<AllocaInst>(AI))
7495 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7496 else if (isa<MallocInst>(AI)) {
7497 // Malloc returns maximally aligned memory.
7498 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7501 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7504 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7508 } else if (isa<BitCastInst>(V) ||
7509 (isa<ConstantExpr>(V) &&
7510 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7511 User *CI = cast<User>(V);
7512 if (isa<PointerType>(CI->getOperand(0)->getType()))
7513 return GetKnownAlignment(CI->getOperand(0), TD);
7515 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7516 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7517 if (BaseAlignment == 0) return 0;
7519 // If all indexes are zero, it is just the alignment of the base pointer.
7520 bool AllZeroOperands = true;
7521 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7522 if (!isa<Constant>(GEPI->getOperand(i)) ||
7523 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7524 AllZeroOperands = false;
7527 if (AllZeroOperands)
7528 return BaseAlignment;
7530 // Otherwise, if the base alignment is >= the alignment we expect for the
7531 // base pointer type, then we know that the resultant pointer is aligned at
7532 // least as much as its type requires.
7535 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7536 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7537 if (TD->getABITypeAlignment(PtrTy->getElementType())
7539 const Type *GEPTy = GEPI->getType();
7540 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7541 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7549 /// visitCallInst - CallInst simplification. This mostly only handles folding
7550 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7551 /// the heavy lifting.
7553 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7554 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7555 if (!II) return visitCallSite(&CI);
7557 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7559 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7560 bool Changed = false;
7562 // memmove/cpy/set of zero bytes is a noop.
7563 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7564 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7566 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7567 if (CI->getZExtValue() == 1) {
7568 // Replace the instruction with just byte operations. We would
7569 // transform other cases to loads/stores, but we don't know if
7570 // alignment is sufficient.
7574 // If we have a memmove and the source operation is a constant global,
7575 // then the source and dest pointers can't alias, so we can change this
7576 // into a call to memcpy.
7577 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7578 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7579 if (GVSrc->isConstant()) {
7580 Module *M = CI.getParent()->getParent()->getParent();
7582 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7584 Name = "llvm.memcpy.i32";
7586 Name = "llvm.memcpy.i64";
7587 Constant *MemCpy = M->getOrInsertFunction(Name,
7588 CI.getCalledFunction()->getFunctionType());
7589 CI.setOperand(0, MemCpy);
7594 // If we can determine a pointer alignment that is bigger than currently
7595 // set, update the alignment.
7596 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7597 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7598 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7599 unsigned Align = std::min(Alignment1, Alignment2);
7600 if (MI->getAlignment()->getZExtValue() < Align) {
7601 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7604 } else if (isa<MemSetInst>(MI)) {
7605 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7606 if (MI->getAlignment()->getZExtValue() < Alignment) {
7607 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7612 if (Changed) return II;
7614 switch (II->getIntrinsicID()) {
7616 case Intrinsic::ppc_altivec_lvx:
7617 case Intrinsic::ppc_altivec_lvxl:
7618 case Intrinsic::x86_sse_loadu_ps:
7619 case Intrinsic::x86_sse2_loadu_pd:
7620 case Intrinsic::x86_sse2_loadu_dq:
7621 // Turn PPC lvx -> load if the pointer is known aligned.
7622 // Turn X86 loadups -> load if the pointer is known aligned.
7623 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7624 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7625 PointerType::get(II->getType()), CI);
7626 return new LoadInst(Ptr);
7629 case Intrinsic::ppc_altivec_stvx:
7630 case Intrinsic::ppc_altivec_stvxl:
7631 // Turn stvx -> store if the pointer is known aligned.
7632 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7633 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7634 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7636 return new StoreInst(II->getOperand(1), Ptr);
7639 case Intrinsic::x86_sse_storeu_ps:
7640 case Intrinsic::x86_sse2_storeu_pd:
7641 case Intrinsic::x86_sse2_storeu_dq:
7642 case Intrinsic::x86_sse2_storel_dq:
7643 // Turn X86 storeu -> store if the pointer is known aligned.
7644 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7645 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7646 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7648 return new StoreInst(II->getOperand(2), Ptr);
7652 case Intrinsic::x86_sse_cvttss2si: {
7653 // These intrinsics only demands the 0th element of its input vector. If
7654 // we can simplify the input based on that, do so now.
7656 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7658 II->setOperand(1, V);
7664 case Intrinsic::ppc_altivec_vperm:
7665 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7666 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7667 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7669 // Check that all of the elements are integer constants or undefs.
7670 bool AllEltsOk = true;
7671 for (unsigned i = 0; i != 16; ++i) {
7672 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7673 !isa<UndefValue>(Mask->getOperand(i))) {
7680 // Cast the input vectors to byte vectors.
7681 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7682 II->getOperand(1), Mask->getType(), CI);
7683 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7684 II->getOperand(2), Mask->getType(), CI);
7685 Value *Result = UndefValue::get(Op0->getType());
7687 // Only extract each element once.
7688 Value *ExtractedElts[32];
7689 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7691 for (unsigned i = 0; i != 16; ++i) {
7692 if (isa<UndefValue>(Mask->getOperand(i)))
7694 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7695 Idx &= 31; // Match the hardware behavior.
7697 if (ExtractedElts[Idx] == 0) {
7699 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7700 InsertNewInstBefore(Elt, CI);
7701 ExtractedElts[Idx] = Elt;
7704 // Insert this value into the result vector.
7705 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7706 InsertNewInstBefore(cast<Instruction>(Result), CI);
7708 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7713 case Intrinsic::stackrestore: {
7714 // If the save is right next to the restore, remove the restore. This can
7715 // happen when variable allocas are DCE'd.
7716 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7717 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7718 BasicBlock::iterator BI = SS;
7720 return EraseInstFromFunction(CI);
7724 // If the stack restore is in a return/unwind block and if there are no
7725 // allocas or calls between the restore and the return, nuke the restore.
7726 TerminatorInst *TI = II->getParent()->getTerminator();
7727 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7728 BasicBlock::iterator BI = II;
7729 bool CannotRemove = false;
7730 for (++BI; &*BI != TI; ++BI) {
7731 if (isa<AllocaInst>(BI) ||
7732 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7733 CannotRemove = true;
7738 return EraseInstFromFunction(CI);
7745 return visitCallSite(II);
7748 // InvokeInst simplification
7750 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7751 return visitCallSite(&II);
7754 // visitCallSite - Improvements for call and invoke instructions.
7756 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7757 bool Changed = false;
7759 // If the callee is a constexpr cast of a function, attempt to move the cast
7760 // to the arguments of the call/invoke.
7761 if (transformConstExprCastCall(CS)) return 0;
7763 Value *Callee = CS.getCalledValue();
7765 if (Function *CalleeF = dyn_cast<Function>(Callee))
7766 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7767 Instruction *OldCall = CS.getInstruction();
7768 // If the call and callee calling conventions don't match, this call must
7769 // be unreachable, as the call is undefined.
7770 new StoreInst(ConstantInt::getTrue(),
7771 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7772 if (!OldCall->use_empty())
7773 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7774 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7775 return EraseInstFromFunction(*OldCall);
7779 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7780 // This instruction is not reachable, just remove it. We insert a store to
7781 // undef so that we know that this code is not reachable, despite the fact
7782 // that we can't modify the CFG here.
7783 new StoreInst(ConstantInt::getTrue(),
7784 UndefValue::get(PointerType::get(Type::Int1Ty)),
7785 CS.getInstruction());
7787 if (!CS.getInstruction()->use_empty())
7788 CS.getInstruction()->
7789 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7791 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7792 // Don't break the CFG, insert a dummy cond branch.
7793 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7794 ConstantInt::getTrue(), II);
7796 return EraseInstFromFunction(*CS.getInstruction());
7799 const PointerType *PTy = cast<PointerType>(Callee->getType());
7800 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7801 if (FTy->isVarArg()) {
7802 // See if we can optimize any arguments passed through the varargs area of
7804 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7805 E = CS.arg_end(); I != E; ++I)
7806 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7807 // If this cast does not effect the value passed through the varargs
7808 // area, we can eliminate the use of the cast.
7809 Value *Op = CI->getOperand(0);
7810 if (CI->isLosslessCast()) {
7817 return Changed ? CS.getInstruction() : 0;
7820 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7821 // attempt to move the cast to the arguments of the call/invoke.
7823 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7824 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7825 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7826 if (CE->getOpcode() != Instruction::BitCast ||
7827 !isa<Function>(CE->getOperand(0)))
7829 Function *Callee = cast<Function>(CE->getOperand(0));
7830 Instruction *Caller = CS.getInstruction();
7832 // Okay, this is a cast from a function to a different type. Unless doing so
7833 // would cause a type conversion of one of our arguments, change this call to
7834 // be a direct call with arguments casted to the appropriate types.
7836 const FunctionType *FT = Callee->getFunctionType();
7837 const Type *OldRetTy = Caller->getType();
7839 const FunctionType *ActualFT =
7840 cast<FunctionType>(cast<PointerType>(CE->getType())->getElementType());
7842 // If the parameter attributes don't match up, don't do the xform. We don't
7843 // want to lose an sret attribute or something.
7844 if (FT->getParamAttrs() != ActualFT->getParamAttrs())
7847 // Check to see if we are changing the return type...
7848 if (OldRetTy != FT->getReturnType()) {
7849 if (Callee->isDeclaration() && !Caller->use_empty() &&
7850 // Conversion is ok if changing from pointer to int of same size.
7851 !(isa<PointerType>(FT->getReturnType()) &&
7852 TD->getIntPtrType() == OldRetTy))
7853 return false; // Cannot transform this return value.
7855 // If the callsite is an invoke instruction, and the return value is used by
7856 // a PHI node in a successor, we cannot change the return type of the call
7857 // because there is no place to put the cast instruction (without breaking
7858 // the critical edge). Bail out in this case.
7859 if (!Caller->use_empty())
7860 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7861 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7863 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7864 if (PN->getParent() == II->getNormalDest() ||
7865 PN->getParent() == II->getUnwindDest())
7869 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7870 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7872 CallSite::arg_iterator AI = CS.arg_begin();
7873 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7874 const Type *ParamTy = FT->getParamType(i);
7875 const Type *ActTy = (*AI)->getType();
7876 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7877 //Some conversions are safe even if we do not have a body.
7878 //Either we can cast directly, or we can upconvert the argument
7879 bool isConvertible = ActTy == ParamTy ||
7880 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7881 (ParamTy->isInteger() && ActTy->isInteger() &&
7882 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7883 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7884 && c->getValue().isStrictlyPositive());
7885 if (Callee->isDeclaration() && !isConvertible) return false;
7887 // Most other conversions can be done if we have a body, even if these
7888 // lose information, e.g. int->short.
7889 // Some conversions cannot be done at all, e.g. float to pointer.
7890 // Logic here parallels CastInst::getCastOpcode (the design there
7891 // requires legality checks like this be done before calling it).
7892 if (ParamTy->isInteger()) {
7893 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7894 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7897 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
7898 !isa<PointerType>(ActTy))
7900 } else if (ParamTy->isFloatingPoint()) {
7901 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7902 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7905 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
7907 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
7908 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7909 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
7912 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
7914 } else if (isa<PointerType>(ParamTy)) {
7915 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
7922 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7923 Callee->isDeclaration())
7924 return false; // Do not delete arguments unless we have a function body...
7926 // Okay, we decided that this is a safe thing to do: go ahead and start
7927 // inserting cast instructions as necessary...
7928 std::vector<Value*> Args;
7929 Args.reserve(NumActualArgs);
7931 AI = CS.arg_begin();
7932 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7933 const Type *ParamTy = FT->getParamType(i);
7934 if ((*AI)->getType() == ParamTy) {
7935 Args.push_back(*AI);
7937 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7938 false, ParamTy, false);
7939 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7940 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7944 // If the function takes more arguments than the call was taking, add them
7946 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7947 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7949 // If we are removing arguments to the function, emit an obnoxious warning...
7950 if (FT->getNumParams() < NumActualArgs)
7951 if (!FT->isVarArg()) {
7952 cerr << "WARNING: While resolving call to function '"
7953 << Callee->getName() << "' arguments were dropped!\n";
7955 // Add all of the arguments in their promoted form to the arg list...
7956 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7957 const Type *PTy = getPromotedType((*AI)->getType());
7958 if (PTy != (*AI)->getType()) {
7959 // Must promote to pass through va_arg area!
7960 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7962 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7963 InsertNewInstBefore(Cast, *Caller);
7964 Args.push_back(Cast);
7966 Args.push_back(*AI);
7971 if (FT->getReturnType() == Type::VoidTy)
7972 Caller->setName(""); // Void type should not have a name.
7975 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7976 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7977 &Args[0], Args.size(), Caller->getName(), Caller);
7978 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7980 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7981 if (cast<CallInst>(Caller)->isTailCall())
7982 cast<CallInst>(NC)->setTailCall();
7983 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7986 // Insert a cast of the return type as necessary.
7988 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7989 if (NV->getType() != Type::VoidTy) {
7990 const Type *CallerTy = Caller->getType();
7991 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7993 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7995 // If this is an invoke instruction, we should insert it after the first
7996 // non-phi, instruction in the normal successor block.
7997 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7998 BasicBlock::iterator I = II->getNormalDest()->begin();
7999 while (isa<PHINode>(I)) ++I;
8000 InsertNewInstBefore(NC, *I);
8002 // Otherwise, it's a call, just insert cast right after the call instr
8003 InsertNewInstBefore(NC, *Caller);
8005 AddUsersToWorkList(*Caller);
8007 NV = UndefValue::get(Caller->getType());
8011 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8012 Caller->replaceAllUsesWith(NV);
8013 Caller->eraseFromParent();
8014 RemoveFromWorkList(Caller);
8018 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8019 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8020 /// and a single binop.
8021 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8022 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8023 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8024 isa<CmpInst>(FirstInst));
8025 unsigned Opc = FirstInst->getOpcode();
8026 Value *LHSVal = FirstInst->getOperand(0);
8027 Value *RHSVal = FirstInst->getOperand(1);
8029 const Type *LHSType = LHSVal->getType();
8030 const Type *RHSType = RHSVal->getType();
8032 // Scan to see if all operands are the same opcode, all have one use, and all
8033 // kill their operands (i.e. the operands have one use).
8034 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8035 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8036 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8037 // Verify type of the LHS matches so we don't fold cmp's of different
8038 // types or GEP's with different index types.
8039 I->getOperand(0)->getType() != LHSType ||
8040 I->getOperand(1)->getType() != RHSType)
8043 // If they are CmpInst instructions, check their predicates
8044 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8045 if (cast<CmpInst>(I)->getPredicate() !=
8046 cast<CmpInst>(FirstInst)->getPredicate())
8049 // Keep track of which operand needs a phi node.
8050 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8051 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8054 // Otherwise, this is safe to transform, determine if it is profitable.
8056 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8057 // Indexes are often folded into load/store instructions, so we don't want to
8058 // hide them behind a phi.
8059 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8062 Value *InLHS = FirstInst->getOperand(0);
8063 Value *InRHS = FirstInst->getOperand(1);
8064 PHINode *NewLHS = 0, *NewRHS = 0;
8066 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8067 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8068 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8069 InsertNewInstBefore(NewLHS, PN);
8074 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8075 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8076 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8077 InsertNewInstBefore(NewRHS, PN);
8081 // Add all operands to the new PHIs.
8082 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8084 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8085 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8088 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8089 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8093 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8094 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8095 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8096 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8099 assert(isa<GetElementPtrInst>(FirstInst));
8100 return new GetElementPtrInst(LHSVal, RHSVal);
8104 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8105 /// of the block that defines it. This means that it must be obvious the value
8106 /// of the load is not changed from the point of the load to the end of the
8109 /// Finally, it is safe, but not profitable, to sink a load targetting a
8110 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8112 static bool isSafeToSinkLoad(LoadInst *L) {
8113 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8115 for (++BBI; BBI != E; ++BBI)
8116 if (BBI->mayWriteToMemory())
8119 // Check for non-address taken alloca. If not address-taken already, it isn't
8120 // profitable to do this xform.
8121 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8122 bool isAddressTaken = false;
8123 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8125 if (isa<LoadInst>(UI)) continue;
8126 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8127 // If storing TO the alloca, then the address isn't taken.
8128 if (SI->getOperand(1) == AI) continue;
8130 isAddressTaken = true;
8134 if (!isAddressTaken)
8142 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8143 // operator and they all are only used by the PHI, PHI together their
8144 // inputs, and do the operation once, to the result of the PHI.
8145 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8146 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8148 // Scan the instruction, looking for input operations that can be folded away.
8149 // If all input operands to the phi are the same instruction (e.g. a cast from
8150 // the same type or "+42") we can pull the operation through the PHI, reducing
8151 // code size and simplifying code.
8152 Constant *ConstantOp = 0;
8153 const Type *CastSrcTy = 0;
8154 bool isVolatile = false;
8155 if (isa<CastInst>(FirstInst)) {
8156 CastSrcTy = FirstInst->getOperand(0)->getType();
8157 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8158 // Can fold binop, compare or shift here if the RHS is a constant,
8159 // otherwise call FoldPHIArgBinOpIntoPHI.
8160 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8161 if (ConstantOp == 0)
8162 return FoldPHIArgBinOpIntoPHI(PN);
8163 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8164 isVolatile = LI->isVolatile();
8165 // We can't sink the load if the loaded value could be modified between the
8166 // load and the PHI.
8167 if (LI->getParent() != PN.getIncomingBlock(0) ||
8168 !isSafeToSinkLoad(LI))
8170 } else if (isa<GetElementPtrInst>(FirstInst)) {
8171 if (FirstInst->getNumOperands() == 2)
8172 return FoldPHIArgBinOpIntoPHI(PN);
8173 // Can't handle general GEPs yet.
8176 return 0; // Cannot fold this operation.
8179 // Check to see if all arguments are the same operation.
8180 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8181 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8182 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8183 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8186 if (I->getOperand(0)->getType() != CastSrcTy)
8187 return 0; // Cast operation must match.
8188 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8189 // We can't sink the load if the loaded value could be modified between
8190 // the load and the PHI.
8191 if (LI->isVolatile() != isVolatile ||
8192 LI->getParent() != PN.getIncomingBlock(i) ||
8193 !isSafeToSinkLoad(LI))
8195 } else if (I->getOperand(1) != ConstantOp) {
8200 // Okay, they are all the same operation. Create a new PHI node of the
8201 // correct type, and PHI together all of the LHS's of the instructions.
8202 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8203 PN.getName()+".in");
8204 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8206 Value *InVal = FirstInst->getOperand(0);
8207 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8209 // Add all operands to the new PHI.
8210 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8211 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8212 if (NewInVal != InVal)
8214 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8219 // The new PHI unions all of the same values together. This is really
8220 // common, so we handle it intelligently here for compile-time speed.
8224 InsertNewInstBefore(NewPN, PN);
8228 // Insert and return the new operation.
8229 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8230 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8231 else if (isa<LoadInst>(FirstInst))
8232 return new LoadInst(PhiVal, "", isVolatile);
8233 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8234 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8235 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8236 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8237 PhiVal, ConstantOp);
8239 assert(0 && "Unknown operation");
8243 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8245 static bool DeadPHICycle(PHINode *PN,
8246 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8247 if (PN->use_empty()) return true;
8248 if (!PN->hasOneUse()) return false;
8250 // Remember this node, and if we find the cycle, return.
8251 if (!PotentiallyDeadPHIs.insert(PN))
8254 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8255 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8260 // PHINode simplification
8262 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8263 // If LCSSA is around, don't mess with Phi nodes
8264 if (MustPreserveLCSSA) return 0;
8266 if (Value *V = PN.hasConstantValue())
8267 return ReplaceInstUsesWith(PN, V);
8269 // If all PHI operands are the same operation, pull them through the PHI,
8270 // reducing code size.
8271 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8272 PN.getIncomingValue(0)->hasOneUse())
8273 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8276 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8277 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8278 // PHI)... break the cycle.
8279 if (PN.hasOneUse()) {
8280 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8281 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8282 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8283 PotentiallyDeadPHIs.insert(&PN);
8284 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8285 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8288 // If this phi has a single use, and if that use just computes a value for
8289 // the next iteration of a loop, delete the phi. This occurs with unused
8290 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8291 // common case here is good because the only other things that catch this
8292 // are induction variable analysis (sometimes) and ADCE, which is only run
8294 if (PHIUser->hasOneUse() &&
8295 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8296 PHIUser->use_back() == &PN) {
8297 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8304 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8305 Instruction *InsertPoint,
8307 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8308 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8309 // We must cast correctly to the pointer type. Ensure that we
8310 // sign extend the integer value if it is smaller as this is
8311 // used for address computation.
8312 Instruction::CastOps opcode =
8313 (VTySize < PtrSize ? Instruction::SExt :
8314 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8315 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8319 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8320 Value *PtrOp = GEP.getOperand(0);
8321 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8322 // If so, eliminate the noop.
8323 if (GEP.getNumOperands() == 1)
8324 return ReplaceInstUsesWith(GEP, PtrOp);
8326 if (isa<UndefValue>(GEP.getOperand(0)))
8327 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8329 bool HasZeroPointerIndex = false;
8330 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8331 HasZeroPointerIndex = C->isNullValue();
8333 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8334 return ReplaceInstUsesWith(GEP, PtrOp);
8336 // Eliminate unneeded casts for indices.
8337 bool MadeChange = false;
8339 gep_type_iterator GTI = gep_type_begin(GEP);
8340 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8341 if (isa<SequentialType>(*GTI)) {
8342 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8343 if (CI->getOpcode() == Instruction::ZExt ||
8344 CI->getOpcode() == Instruction::SExt) {
8345 const Type *SrcTy = CI->getOperand(0)->getType();
8346 // We can eliminate a cast from i32 to i64 iff the target
8347 // is a 32-bit pointer target.
8348 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8350 GEP.setOperand(i, CI->getOperand(0));
8354 // If we are using a wider index than needed for this platform, shrink it
8355 // to what we need. If the incoming value needs a cast instruction,
8356 // insert it. This explicit cast can make subsequent optimizations more
8358 Value *Op = GEP.getOperand(i);
8359 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
8360 if (Constant *C = dyn_cast<Constant>(Op)) {
8361 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8364 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8366 GEP.setOperand(i, Op);
8371 if (MadeChange) return &GEP;
8373 // If this GEP instruction doesn't move the pointer, and if the input operand
8374 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8375 // real input to the dest type.
8376 if (GEP.hasAllZeroIndices() && isa<BitCastInst>(GEP.getOperand(0)))
8377 return new BitCastInst(cast<BitCastInst>(GEP.getOperand(0))->getOperand(0),
8380 // Combine Indices - If the source pointer to this getelementptr instruction
8381 // is a getelementptr instruction, combine the indices of the two
8382 // getelementptr instructions into a single instruction.
8384 SmallVector<Value*, 8> SrcGEPOperands;
8385 if (User *Src = dyn_castGetElementPtr(PtrOp))
8386 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8388 if (!SrcGEPOperands.empty()) {
8389 // Note that if our source is a gep chain itself that we wait for that
8390 // chain to be resolved before we perform this transformation. This
8391 // avoids us creating a TON of code in some cases.
8393 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8394 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8395 return 0; // Wait until our source is folded to completion.
8397 SmallVector<Value*, 8> Indices;
8399 // Find out whether the last index in the source GEP is a sequential idx.
8400 bool EndsWithSequential = false;
8401 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8402 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8403 EndsWithSequential = !isa<StructType>(*I);
8405 // Can we combine the two pointer arithmetics offsets?
8406 if (EndsWithSequential) {
8407 // Replace: gep (gep %P, long B), long A, ...
8408 // With: T = long A+B; gep %P, T, ...
8410 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8411 if (SO1 == Constant::getNullValue(SO1->getType())) {
8413 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8416 // If they aren't the same type, convert both to an integer of the
8417 // target's pointer size.
8418 if (SO1->getType() != GO1->getType()) {
8419 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8420 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8421 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8422 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8424 unsigned PS = TD->getPointerSize();
8425 if (TD->getTypeSize(SO1->getType()) == PS) {
8426 // Convert GO1 to SO1's type.
8427 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8429 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8430 // Convert SO1 to GO1's type.
8431 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8433 const Type *PT = TD->getIntPtrType();
8434 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8435 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8439 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8440 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8442 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8443 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8447 // Recycle the GEP we already have if possible.
8448 if (SrcGEPOperands.size() == 2) {
8449 GEP.setOperand(0, SrcGEPOperands[0]);
8450 GEP.setOperand(1, Sum);
8453 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8454 SrcGEPOperands.end()-1);
8455 Indices.push_back(Sum);
8456 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8458 } else if (isa<Constant>(*GEP.idx_begin()) &&
8459 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8460 SrcGEPOperands.size() != 1) {
8461 // Otherwise we can do the fold if the first index of the GEP is a zero
8462 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8463 SrcGEPOperands.end());
8464 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8467 if (!Indices.empty())
8468 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
8469 Indices.size(), GEP.getName());
8471 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8472 // GEP of global variable. If all of the indices for this GEP are
8473 // constants, we can promote this to a constexpr instead of an instruction.
8475 // Scan for nonconstants...
8476 SmallVector<Constant*, 8> Indices;
8477 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8478 for (; I != E && isa<Constant>(*I); ++I)
8479 Indices.push_back(cast<Constant>(*I));
8481 if (I == E) { // If they are all constants...
8482 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8483 &Indices[0],Indices.size());
8485 // Replace all uses of the GEP with the new constexpr...
8486 return ReplaceInstUsesWith(GEP, CE);
8488 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8489 if (!isa<PointerType>(X->getType())) {
8490 // Not interesting. Source pointer must be a cast from pointer.
8491 } else if (HasZeroPointerIndex) {
8492 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8493 // into : GEP [10 x ubyte]* X, long 0, ...
8495 // This occurs when the program declares an array extern like "int X[];"
8497 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8498 const PointerType *XTy = cast<PointerType>(X->getType());
8499 if (const ArrayType *XATy =
8500 dyn_cast<ArrayType>(XTy->getElementType()))
8501 if (const ArrayType *CATy =
8502 dyn_cast<ArrayType>(CPTy->getElementType()))
8503 if (CATy->getElementType() == XATy->getElementType()) {
8504 // At this point, we know that the cast source type is a pointer
8505 // to an array of the same type as the destination pointer
8506 // array. Because the array type is never stepped over (there
8507 // is a leading zero) we can fold the cast into this GEP.
8508 GEP.setOperand(0, X);
8511 } else if (GEP.getNumOperands() == 2) {
8512 // Transform things like:
8513 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8514 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8515 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8516 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8517 if (isa<ArrayType>(SrcElTy) &&
8518 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8519 TD->getTypeSize(ResElTy)) {
8520 Value *V = InsertNewInstBefore(
8521 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8522 GEP.getOperand(1), GEP.getName()), GEP);
8523 // V and GEP are both pointer types --> BitCast
8524 return new BitCastInst(V, GEP.getType());
8527 // Transform things like:
8528 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8529 // (where tmp = 8*tmp2) into:
8530 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8532 if (isa<ArrayType>(SrcElTy) &&
8533 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8534 uint64_t ArrayEltSize =
8535 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8537 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8538 // allow either a mul, shift, or constant here.
8540 ConstantInt *Scale = 0;
8541 if (ArrayEltSize == 1) {
8542 NewIdx = GEP.getOperand(1);
8543 Scale = ConstantInt::get(NewIdx->getType(), 1);
8544 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8545 NewIdx = ConstantInt::get(CI->getType(), 1);
8547 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8548 if (Inst->getOpcode() == Instruction::Shl &&
8549 isa<ConstantInt>(Inst->getOperand(1))) {
8550 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8551 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8552 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8553 NewIdx = Inst->getOperand(0);
8554 } else if (Inst->getOpcode() == Instruction::Mul &&
8555 isa<ConstantInt>(Inst->getOperand(1))) {
8556 Scale = cast<ConstantInt>(Inst->getOperand(1));
8557 NewIdx = Inst->getOperand(0);
8561 // If the index will be to exactly the right offset with the scale taken
8562 // out, perform the transformation.
8563 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8564 if (isa<ConstantInt>(Scale))
8565 Scale = ConstantInt::get(Scale->getType(),
8566 Scale->getZExtValue() / ArrayEltSize);
8567 if (Scale->getZExtValue() != 1) {
8568 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8570 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8571 NewIdx = InsertNewInstBefore(Sc, GEP);
8574 // Insert the new GEP instruction.
8575 Instruction *NewGEP =
8576 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8577 NewIdx, GEP.getName());
8578 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8579 // The NewGEP must be pointer typed, so must the old one -> BitCast
8580 return new BitCastInst(NewGEP, GEP.getType());
8589 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8590 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8591 if (AI.isArrayAllocation()) // Check C != 1
8592 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8594 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8595 AllocationInst *New = 0;
8597 // Create and insert the replacement instruction...
8598 if (isa<MallocInst>(AI))
8599 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8601 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8602 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8605 InsertNewInstBefore(New, AI);
8607 // Scan to the end of the allocation instructions, to skip over a block of
8608 // allocas if possible...
8610 BasicBlock::iterator It = New;
8611 while (isa<AllocationInst>(*It)) ++It;
8613 // Now that I is pointing to the first non-allocation-inst in the block,
8614 // insert our getelementptr instruction...
8616 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8617 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8618 New->getName()+".sub", It);
8620 // Now make everything use the getelementptr instead of the original
8622 return ReplaceInstUsesWith(AI, V);
8623 } else if (isa<UndefValue>(AI.getArraySize())) {
8624 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8627 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8628 // Note that we only do this for alloca's, because malloc should allocate and
8629 // return a unique pointer, even for a zero byte allocation.
8630 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8631 TD->getTypeSize(AI.getAllocatedType()) == 0)
8632 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8637 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8638 Value *Op = FI.getOperand(0);
8640 // free undef -> unreachable.
8641 if (isa<UndefValue>(Op)) {
8642 // Insert a new store to null because we cannot modify the CFG here.
8643 new StoreInst(ConstantInt::getTrue(),
8644 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8645 return EraseInstFromFunction(FI);
8648 // If we have 'free null' delete the instruction. This can happen in stl code
8649 // when lots of inlining happens.
8650 if (isa<ConstantPointerNull>(Op))
8651 return EraseInstFromFunction(FI);
8653 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8654 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
8655 FI.setOperand(0, CI->getOperand(0));
8659 // Change free (gep X, 0,0,0,0) into free(X)
8660 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
8661 if (GEPI->hasAllZeroIndices()) {
8662 AddToWorkList(GEPI);
8663 FI.setOperand(0, GEPI->getOperand(0));
8668 // Change free(malloc) into nothing, if the malloc has a single use.
8669 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
8670 if (MI->hasOneUse()) {
8671 EraseInstFromFunction(FI);
8672 return EraseInstFromFunction(*MI);
8679 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8680 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8681 User *CI = cast<User>(LI.getOperand(0));
8682 Value *CastOp = CI->getOperand(0);
8684 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8685 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8686 const Type *SrcPTy = SrcTy->getElementType();
8688 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8689 isa<VectorType>(DestPTy)) {
8690 // If the source is an array, the code below will not succeed. Check to
8691 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8693 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8694 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8695 if (ASrcTy->getNumElements() != 0) {
8697 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8698 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8699 SrcTy = cast<PointerType>(CastOp->getType());
8700 SrcPTy = SrcTy->getElementType();
8703 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8704 isa<VectorType>(SrcPTy)) &&
8705 // Do not allow turning this into a load of an integer, which is then
8706 // casted to a pointer, this pessimizes pointer analysis a lot.
8707 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8708 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8709 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8711 // Okay, we are casting from one integer or pointer type to another of
8712 // the same size. Instead of casting the pointer before the load, cast
8713 // the result of the loaded value.
8714 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8716 LI.isVolatile()),LI);
8717 // Now cast the result of the load.
8718 return new BitCastInst(NewLoad, LI.getType());
8725 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8726 /// from this value cannot trap. If it is not obviously safe to load from the
8727 /// specified pointer, we do a quick local scan of the basic block containing
8728 /// ScanFrom, to determine if the address is already accessed.
8729 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8730 // If it is an alloca or global variable, it is always safe to load from.
8731 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8733 // Otherwise, be a little bit agressive by scanning the local block where we
8734 // want to check to see if the pointer is already being loaded or stored
8735 // from/to. If so, the previous load or store would have already trapped,
8736 // so there is no harm doing an extra load (also, CSE will later eliminate
8737 // the load entirely).
8738 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8743 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8744 if (LI->getOperand(0) == V) return true;
8745 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8746 if (SI->getOperand(1) == V) return true;
8752 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8753 Value *Op = LI.getOperand(0);
8755 // load (cast X) --> cast (load X) iff safe
8756 if (isa<CastInst>(Op))
8757 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8760 // None of the following transforms are legal for volatile loads.
8761 if (LI.isVolatile()) return 0;
8763 if (&LI.getParent()->front() != &LI) {
8764 BasicBlock::iterator BBI = &LI; --BBI;
8765 // If the instruction immediately before this is a store to the same
8766 // address, do a simple form of store->load forwarding.
8767 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8768 if (SI->getOperand(1) == LI.getOperand(0))
8769 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8770 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8771 if (LIB->getOperand(0) == LI.getOperand(0))
8772 return ReplaceInstUsesWith(LI, LIB);
8775 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8776 if (isa<ConstantPointerNull>(GEPI->getOperand(0))) {
8777 // Insert a new store to null instruction before the load to indicate
8778 // that this code is not reachable. We do this instead of inserting
8779 // an unreachable instruction directly because we cannot modify the
8781 new StoreInst(UndefValue::get(LI.getType()),
8782 Constant::getNullValue(Op->getType()), &LI);
8783 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8786 if (Constant *C = dyn_cast<Constant>(Op)) {
8787 // load null/undef -> undef
8788 if ((C->isNullValue() || isa<UndefValue>(C))) {
8789 // Insert a new store to null instruction before the load to indicate that
8790 // this code is not reachable. We do this instead of inserting an
8791 // unreachable instruction directly because we cannot modify the CFG.
8792 new StoreInst(UndefValue::get(LI.getType()),
8793 Constant::getNullValue(Op->getType()), &LI);
8794 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8797 // Instcombine load (constant global) into the value loaded.
8798 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8799 if (GV->isConstant() && !GV->isDeclaration())
8800 return ReplaceInstUsesWith(LI, GV->getInitializer());
8802 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8803 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8804 if (CE->getOpcode() == Instruction::GetElementPtr) {
8805 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8806 if (GV->isConstant() && !GV->isDeclaration())
8808 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8809 return ReplaceInstUsesWith(LI, V);
8810 if (CE->getOperand(0)->isNullValue()) {
8811 // Insert a new store to null instruction before the load to indicate
8812 // that this code is not reachable. We do this instead of inserting
8813 // an unreachable instruction directly because we cannot modify the
8815 new StoreInst(UndefValue::get(LI.getType()),
8816 Constant::getNullValue(Op->getType()), &LI);
8817 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8820 } else if (CE->isCast()) {
8821 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8826 if (Op->hasOneUse()) {
8827 // Change select and PHI nodes to select values instead of addresses: this
8828 // helps alias analysis out a lot, allows many others simplifications, and
8829 // exposes redundancy in the code.
8831 // Note that we cannot do the transformation unless we know that the
8832 // introduced loads cannot trap! Something like this is valid as long as
8833 // the condition is always false: load (select bool %C, int* null, int* %G),
8834 // but it would not be valid if we transformed it to load from null
8837 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8838 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8839 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8840 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8841 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8842 SI->getOperand(1)->getName()+".val"), LI);
8843 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8844 SI->getOperand(2)->getName()+".val"), LI);
8845 return new SelectInst(SI->getCondition(), V1, V2);
8848 // load (select (cond, null, P)) -> load P
8849 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8850 if (C->isNullValue()) {
8851 LI.setOperand(0, SI->getOperand(2));
8855 // load (select (cond, P, null)) -> load P
8856 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8857 if (C->isNullValue()) {
8858 LI.setOperand(0, SI->getOperand(1));
8866 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8868 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8869 User *CI = cast<User>(SI.getOperand(1));
8870 Value *CastOp = CI->getOperand(0);
8872 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8873 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8874 const Type *SrcPTy = SrcTy->getElementType();
8876 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8877 // If the source is an array, the code below will not succeed. Check to
8878 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8880 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8881 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8882 if (ASrcTy->getNumElements() != 0) {
8884 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8885 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8886 SrcTy = cast<PointerType>(CastOp->getType());
8887 SrcPTy = SrcTy->getElementType();
8890 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8891 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8892 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8894 // Okay, we are casting from one integer or pointer type to another of
8895 // the same size. Instead of casting the pointer before
8896 // the store, cast the value to be stored.
8898 Value *SIOp0 = SI.getOperand(0);
8899 Instruction::CastOps opcode = Instruction::BitCast;
8900 const Type* CastSrcTy = SIOp0->getType();
8901 const Type* CastDstTy = SrcPTy;
8902 if (isa<PointerType>(CastDstTy)) {
8903 if (CastSrcTy->isInteger())
8904 opcode = Instruction::IntToPtr;
8905 } else if (isa<IntegerType>(CastDstTy)) {
8906 if (isa<PointerType>(SIOp0->getType()))
8907 opcode = Instruction::PtrToInt;
8909 if (Constant *C = dyn_cast<Constant>(SIOp0))
8910 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8912 NewCast = IC.InsertNewInstBefore(
8913 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8915 return new StoreInst(NewCast, CastOp);
8922 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8923 Value *Val = SI.getOperand(0);
8924 Value *Ptr = SI.getOperand(1);
8926 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8927 EraseInstFromFunction(SI);
8932 // If the RHS is an alloca with a single use, zapify the store, making the
8934 if (Ptr->hasOneUse()) {
8935 if (isa<AllocaInst>(Ptr)) {
8936 EraseInstFromFunction(SI);
8941 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8942 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8943 GEP->getOperand(0)->hasOneUse()) {
8944 EraseInstFromFunction(SI);
8950 // Do really simple DSE, to catch cases where there are several consequtive
8951 // stores to the same location, separated by a few arithmetic operations. This
8952 // situation often occurs with bitfield accesses.
8953 BasicBlock::iterator BBI = &SI;
8954 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8958 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8959 // Prev store isn't volatile, and stores to the same location?
8960 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8963 EraseInstFromFunction(*PrevSI);
8969 // If this is a load, we have to stop. However, if the loaded value is from
8970 // the pointer we're loading and is producing the pointer we're storing,
8971 // then *this* store is dead (X = load P; store X -> P).
8972 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8973 if (LI == Val && LI->getOperand(0) == Ptr) {
8974 EraseInstFromFunction(SI);
8978 // Otherwise, this is a load from some other location. Stores before it
8983 // Don't skip over loads or things that can modify memory.
8984 if (BBI->mayWriteToMemory())
8989 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8991 // store X, null -> turns into 'unreachable' in SimplifyCFG
8992 if (isa<ConstantPointerNull>(Ptr)) {
8993 if (!isa<UndefValue>(Val)) {
8994 SI.setOperand(0, UndefValue::get(Val->getType()));
8995 if (Instruction *U = dyn_cast<Instruction>(Val))
8996 AddToWorkList(U); // Dropped a use.
8999 return 0; // Do not modify these!
9002 // store undef, Ptr -> noop
9003 if (isa<UndefValue>(Val)) {
9004 EraseInstFromFunction(SI);
9009 // If the pointer destination is a cast, see if we can fold the cast into the
9011 if (isa<CastInst>(Ptr))
9012 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9014 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9016 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9020 // If this store is the last instruction in the basic block, and if the block
9021 // ends with an unconditional branch, try to move it to the successor block.
9023 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9024 if (BI->isUnconditional())
9025 if (SimplifyStoreAtEndOfBlock(SI))
9026 return 0; // xform done!
9031 /// SimplifyStoreAtEndOfBlock - Turn things like:
9032 /// if () { *P = v1; } else { *P = v2 }
9033 /// into a phi node with a store in the successor.
9035 /// Simplify things like:
9036 /// *P = v1; if () { *P = v2; }
9037 /// into a phi node with a store in the successor.
9039 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9040 BasicBlock *StoreBB = SI.getParent();
9042 // Check to see if the successor block has exactly two incoming edges. If
9043 // so, see if the other predecessor contains a store to the same location.
9044 // if so, insert a PHI node (if needed) and move the stores down.
9045 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9047 // Determine whether Dest has exactly two predecessors and, if so, compute
9048 // the other predecessor.
9049 pred_iterator PI = pred_begin(DestBB);
9050 BasicBlock *OtherBB = 0;
9054 if (PI == pred_end(DestBB))
9057 if (*PI != StoreBB) {
9062 if (++PI != pred_end(DestBB))
9066 // Verify that the other block ends in a branch and is not otherwise empty.
9067 BasicBlock::iterator BBI = OtherBB->getTerminator();
9068 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9069 if (!OtherBr || BBI == OtherBB->begin())
9072 // If the other block ends in an unconditional branch, check for the 'if then
9073 // else' case. there is an instruction before the branch.
9074 StoreInst *OtherStore = 0;
9075 if (OtherBr->isUnconditional()) {
9076 // If this isn't a store, or isn't a store to the same location, bail out.
9078 OtherStore = dyn_cast<StoreInst>(BBI);
9079 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9082 // Otherwise, the other block ended with a conditional branch. If one of the
9083 // destinations is StoreBB, then we have the if/then case.
9084 if (OtherBr->getSuccessor(0) != StoreBB &&
9085 OtherBr->getSuccessor(1) != StoreBB)
9088 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9089 // if/then triangle. See if there is a store to the same ptr as SI that
9090 // lives in OtherBB.
9092 // Check to see if we find the matching store.
9093 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9094 if (OtherStore->getOperand(1) != SI.getOperand(1))
9098 // If we find something that may be using the stored value, or if we run
9099 // out of instructions, we can't do the xform.
9100 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9101 BBI == OtherBB->begin())
9105 // In order to eliminate the store in OtherBr, we have to
9106 // make sure nothing reads the stored value in StoreBB.
9107 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9108 // FIXME: This should really be AA driven.
9109 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9114 // Insert a PHI node now if we need it.
9115 Value *MergedVal = OtherStore->getOperand(0);
9116 if (MergedVal != SI.getOperand(0)) {
9117 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9118 PN->reserveOperandSpace(2);
9119 PN->addIncoming(SI.getOperand(0), SI.getParent());
9120 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9121 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9124 // Advance to a place where it is safe to insert the new store and
9126 BBI = DestBB->begin();
9127 while (isa<PHINode>(BBI)) ++BBI;
9128 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9129 OtherStore->isVolatile()), *BBI);
9131 // Nuke the old stores.
9132 EraseInstFromFunction(SI);
9133 EraseInstFromFunction(*OtherStore);
9139 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9140 // Change br (not X), label True, label False to: br X, label False, True
9142 BasicBlock *TrueDest;
9143 BasicBlock *FalseDest;
9144 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9145 !isa<Constant>(X)) {
9146 // Swap Destinations and condition...
9148 BI.setSuccessor(0, FalseDest);
9149 BI.setSuccessor(1, TrueDest);
9153 // Cannonicalize fcmp_one -> fcmp_oeq
9154 FCmpInst::Predicate FPred; Value *Y;
9155 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9156 TrueDest, FalseDest)))
9157 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9158 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9159 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9160 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9161 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9162 NewSCC->takeName(I);
9163 // Swap Destinations and condition...
9164 BI.setCondition(NewSCC);
9165 BI.setSuccessor(0, FalseDest);
9166 BI.setSuccessor(1, TrueDest);
9167 RemoveFromWorkList(I);
9168 I->eraseFromParent();
9169 AddToWorkList(NewSCC);
9173 // Cannonicalize icmp_ne -> icmp_eq
9174 ICmpInst::Predicate IPred;
9175 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9176 TrueDest, FalseDest)))
9177 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9178 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9179 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9180 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9181 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9182 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9183 NewSCC->takeName(I);
9184 // Swap Destinations and condition...
9185 BI.setCondition(NewSCC);
9186 BI.setSuccessor(0, FalseDest);
9187 BI.setSuccessor(1, TrueDest);
9188 RemoveFromWorkList(I);
9189 I->eraseFromParent();;
9190 AddToWorkList(NewSCC);
9197 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9198 Value *Cond = SI.getCondition();
9199 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9200 if (I->getOpcode() == Instruction::Add)
9201 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9202 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9203 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9204 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9206 SI.setOperand(0, I->getOperand(0));
9214 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9215 /// is to leave as a vector operation.
9216 static bool CheapToScalarize(Value *V, bool isConstant) {
9217 if (isa<ConstantAggregateZero>(V))
9219 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9220 if (isConstant) return true;
9221 // If all elts are the same, we can extract.
9222 Constant *Op0 = C->getOperand(0);
9223 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9224 if (C->getOperand(i) != Op0)
9228 Instruction *I = dyn_cast<Instruction>(V);
9229 if (!I) return false;
9231 // Insert element gets simplified to the inserted element or is deleted if
9232 // this is constant idx extract element and its a constant idx insertelt.
9233 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9234 isa<ConstantInt>(I->getOperand(2)))
9236 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9238 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9239 if (BO->hasOneUse() &&
9240 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9241 CheapToScalarize(BO->getOperand(1), isConstant)))
9243 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9244 if (CI->hasOneUse() &&
9245 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9246 CheapToScalarize(CI->getOperand(1), isConstant)))
9252 /// Read and decode a shufflevector mask.
9254 /// It turns undef elements into values that are larger than the number of
9255 /// elements in the input.
9256 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9257 unsigned NElts = SVI->getType()->getNumElements();
9258 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9259 return std::vector<unsigned>(NElts, 0);
9260 if (isa<UndefValue>(SVI->getOperand(2)))
9261 return std::vector<unsigned>(NElts, 2*NElts);
9263 std::vector<unsigned> Result;
9264 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9265 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9266 if (isa<UndefValue>(CP->getOperand(i)))
9267 Result.push_back(NElts*2); // undef -> 8
9269 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9273 /// FindScalarElement - Given a vector and an element number, see if the scalar
9274 /// value is already around as a register, for example if it were inserted then
9275 /// extracted from the vector.
9276 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9277 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9278 const VectorType *PTy = cast<VectorType>(V->getType());
9279 unsigned Width = PTy->getNumElements();
9280 if (EltNo >= Width) // Out of range access.
9281 return UndefValue::get(PTy->getElementType());
9283 if (isa<UndefValue>(V))
9284 return UndefValue::get(PTy->getElementType());
9285 else if (isa<ConstantAggregateZero>(V))
9286 return Constant::getNullValue(PTy->getElementType());
9287 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9288 return CP->getOperand(EltNo);
9289 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9290 // If this is an insert to a variable element, we don't know what it is.
9291 if (!isa<ConstantInt>(III->getOperand(2)))
9293 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9295 // If this is an insert to the element we are looking for, return the
9298 return III->getOperand(1);
9300 // Otherwise, the insertelement doesn't modify the value, recurse on its
9302 return FindScalarElement(III->getOperand(0), EltNo);
9303 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9304 unsigned InEl = getShuffleMask(SVI)[EltNo];
9306 return FindScalarElement(SVI->getOperand(0), InEl);
9307 else if (InEl < Width*2)
9308 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9310 return UndefValue::get(PTy->getElementType());
9313 // Otherwise, we don't know.
9317 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9319 // If packed val is undef, replace extract with scalar undef.
9320 if (isa<UndefValue>(EI.getOperand(0)))
9321 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9323 // If packed val is constant 0, replace extract with scalar 0.
9324 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9325 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9327 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9328 // If packed val is constant with uniform operands, replace EI
9329 // with that operand
9330 Constant *op0 = C->getOperand(0);
9331 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9332 if (C->getOperand(i) != op0) {
9337 return ReplaceInstUsesWith(EI, op0);
9340 // If extracting a specified index from the vector, see if we can recursively
9341 // find a previously computed scalar that was inserted into the vector.
9342 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9343 unsigned IndexVal = IdxC->getZExtValue();
9344 unsigned VectorWidth =
9345 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9347 // If this is extracting an invalid index, turn this into undef, to avoid
9348 // crashing the code below.
9349 if (IndexVal >= VectorWidth)
9350 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9352 // This instruction only demands the single element from the input vector.
9353 // If the input vector has a single use, simplify it based on this use
9355 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9357 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9360 EI.setOperand(0, V);
9365 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9366 return ReplaceInstUsesWith(EI, Elt);
9368 // If the this extractelement is directly using a bitcast from a vector of
9369 // the same number of elements, see if we can find the source element from
9370 // it. In this case, we will end up needing to bitcast the scalars.
9371 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9372 if (const VectorType *VT =
9373 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9374 if (VT->getNumElements() == VectorWidth)
9375 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9376 return new BitCastInst(Elt, EI.getType());
9380 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9381 if (I->hasOneUse()) {
9382 // Push extractelement into predecessor operation if legal and
9383 // profitable to do so
9384 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9385 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9386 if (CheapToScalarize(BO, isConstantElt)) {
9387 ExtractElementInst *newEI0 =
9388 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9389 EI.getName()+".lhs");
9390 ExtractElementInst *newEI1 =
9391 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9392 EI.getName()+".rhs");
9393 InsertNewInstBefore(newEI0, EI);
9394 InsertNewInstBefore(newEI1, EI);
9395 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9397 } else if (isa<LoadInst>(I)) {
9398 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9399 PointerType::get(EI.getType()), EI);
9400 GetElementPtrInst *GEP =
9401 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9402 InsertNewInstBefore(GEP, EI);
9403 return new LoadInst(GEP);
9406 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9407 // Extracting the inserted element?
9408 if (IE->getOperand(2) == EI.getOperand(1))
9409 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9410 // If the inserted and extracted elements are constants, they must not
9411 // be the same value, extract from the pre-inserted value instead.
9412 if (isa<Constant>(IE->getOperand(2)) &&
9413 isa<Constant>(EI.getOperand(1))) {
9414 AddUsesToWorkList(EI);
9415 EI.setOperand(0, IE->getOperand(0));
9418 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9419 // If this is extracting an element from a shufflevector, figure out where
9420 // it came from and extract from the appropriate input element instead.
9421 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9422 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9424 if (SrcIdx < SVI->getType()->getNumElements())
9425 Src = SVI->getOperand(0);
9426 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9427 SrcIdx -= SVI->getType()->getNumElements();
9428 Src = SVI->getOperand(1);
9430 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9432 return new ExtractElementInst(Src, SrcIdx);
9439 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9440 /// elements from either LHS or RHS, return the shuffle mask and true.
9441 /// Otherwise, return false.
9442 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9443 std::vector<Constant*> &Mask) {
9444 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9445 "Invalid CollectSingleShuffleElements");
9446 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9448 if (isa<UndefValue>(V)) {
9449 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9451 } else if (V == LHS) {
9452 for (unsigned i = 0; i != NumElts; ++i)
9453 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9455 } else if (V == RHS) {
9456 for (unsigned i = 0; i != NumElts; ++i)
9457 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9459 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9460 // If this is an insert of an extract from some other vector, include it.
9461 Value *VecOp = IEI->getOperand(0);
9462 Value *ScalarOp = IEI->getOperand(1);
9463 Value *IdxOp = IEI->getOperand(2);
9465 if (!isa<ConstantInt>(IdxOp))
9467 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9469 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9470 // Okay, we can handle this if the vector we are insertinting into is
9472 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9473 // If so, update the mask to reflect the inserted undef.
9474 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9477 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9478 if (isa<ConstantInt>(EI->getOperand(1)) &&
9479 EI->getOperand(0)->getType() == V->getType()) {
9480 unsigned ExtractedIdx =
9481 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9483 // This must be extracting from either LHS or RHS.
9484 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9485 // Okay, we can handle this if the vector we are insertinting into is
9487 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9488 // If so, update the mask to reflect the inserted value.
9489 if (EI->getOperand(0) == LHS) {
9490 Mask[InsertedIdx & (NumElts-1)] =
9491 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9493 assert(EI->getOperand(0) == RHS);
9494 Mask[InsertedIdx & (NumElts-1)] =
9495 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9504 // TODO: Handle shufflevector here!
9509 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9510 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9511 /// that computes V and the LHS value of the shuffle.
9512 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9514 assert(isa<VectorType>(V->getType()) &&
9515 (RHS == 0 || V->getType() == RHS->getType()) &&
9516 "Invalid shuffle!");
9517 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9519 if (isa<UndefValue>(V)) {
9520 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9522 } else if (isa<ConstantAggregateZero>(V)) {
9523 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9525 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9526 // If this is an insert of an extract from some other vector, include it.
9527 Value *VecOp = IEI->getOperand(0);
9528 Value *ScalarOp = IEI->getOperand(1);
9529 Value *IdxOp = IEI->getOperand(2);
9531 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9532 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9533 EI->getOperand(0)->getType() == V->getType()) {
9534 unsigned ExtractedIdx =
9535 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9536 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9538 // Either the extracted from or inserted into vector must be RHSVec,
9539 // otherwise we'd end up with a shuffle of three inputs.
9540 if (EI->getOperand(0) == RHS || RHS == 0) {
9541 RHS = EI->getOperand(0);
9542 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9543 Mask[InsertedIdx & (NumElts-1)] =
9544 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9549 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9550 // Everything but the extracted element is replaced with the RHS.
9551 for (unsigned i = 0; i != NumElts; ++i) {
9552 if (i != InsertedIdx)
9553 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9558 // If this insertelement is a chain that comes from exactly these two
9559 // vectors, return the vector and the effective shuffle.
9560 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9561 return EI->getOperand(0);
9566 // TODO: Handle shufflevector here!
9568 // Otherwise, can't do anything fancy. Return an identity vector.
9569 for (unsigned i = 0; i != NumElts; ++i)
9570 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9574 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9575 Value *VecOp = IE.getOperand(0);
9576 Value *ScalarOp = IE.getOperand(1);
9577 Value *IdxOp = IE.getOperand(2);
9579 // Inserting an undef or into an undefined place, remove this.
9580 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
9581 ReplaceInstUsesWith(IE, VecOp);
9583 // If the inserted element was extracted from some other vector, and if the
9584 // indexes are constant, try to turn this into a shufflevector operation.
9585 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9586 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9587 EI->getOperand(0)->getType() == IE.getType()) {
9588 unsigned NumVectorElts = IE.getType()->getNumElements();
9589 unsigned ExtractedIdx =
9590 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9591 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9593 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9594 return ReplaceInstUsesWith(IE, VecOp);
9596 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9597 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9599 // If we are extracting a value from a vector, then inserting it right
9600 // back into the same place, just use the input vector.
9601 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9602 return ReplaceInstUsesWith(IE, VecOp);
9604 // We could theoretically do this for ANY input. However, doing so could
9605 // turn chains of insertelement instructions into a chain of shufflevector
9606 // instructions, and right now we do not merge shufflevectors. As such,
9607 // only do this in a situation where it is clear that there is benefit.
9608 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9609 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9610 // the values of VecOp, except then one read from EIOp0.
9611 // Build a new shuffle mask.
9612 std::vector<Constant*> Mask;
9613 if (isa<UndefValue>(VecOp))
9614 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9616 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9617 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9620 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9621 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9622 ConstantVector::get(Mask));
9625 // If this insertelement isn't used by some other insertelement, turn it
9626 // (and any insertelements it points to), into one big shuffle.
9627 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9628 std::vector<Constant*> Mask;
9630 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9631 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9632 // We now have a shuffle of LHS, RHS, Mask.
9633 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9642 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9643 Value *LHS = SVI.getOperand(0);
9644 Value *RHS = SVI.getOperand(1);
9645 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9647 bool MadeChange = false;
9649 // Undefined shuffle mask -> undefined value.
9650 if (isa<UndefValue>(SVI.getOperand(2)))
9651 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9653 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9654 // the undef, change them to undefs.
9655 if (isa<UndefValue>(SVI.getOperand(1))) {
9656 // Scan to see if there are any references to the RHS. If so, replace them
9657 // with undef element refs and set MadeChange to true.
9658 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9659 if (Mask[i] >= e && Mask[i] != 2*e) {
9666 // Remap any references to RHS to use LHS.
9667 std::vector<Constant*> Elts;
9668 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9670 Elts.push_back(UndefValue::get(Type::Int32Ty));
9672 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9674 SVI.setOperand(2, ConstantVector::get(Elts));
9678 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9679 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9680 if (LHS == RHS || isa<UndefValue>(LHS)) {
9681 if (isa<UndefValue>(LHS) && LHS == RHS) {
9682 // shuffle(undef,undef,mask) -> undef.
9683 return ReplaceInstUsesWith(SVI, LHS);
9686 // Remap any references to RHS to use LHS.
9687 std::vector<Constant*> Elts;
9688 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9690 Elts.push_back(UndefValue::get(Type::Int32Ty));
9692 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9693 (Mask[i] < e && isa<UndefValue>(LHS)))
9694 Mask[i] = 2*e; // Turn into undef.
9696 Mask[i] &= (e-1); // Force to LHS.
9697 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9700 SVI.setOperand(0, SVI.getOperand(1));
9701 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9702 SVI.setOperand(2, ConstantVector::get(Elts));
9703 LHS = SVI.getOperand(0);
9704 RHS = SVI.getOperand(1);
9708 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9709 bool isLHSID = true, isRHSID = true;
9711 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9712 if (Mask[i] >= e*2) continue; // Ignore undef values.
9713 // Is this an identity shuffle of the LHS value?
9714 isLHSID &= (Mask[i] == i);
9716 // Is this an identity shuffle of the RHS value?
9717 isRHSID &= (Mask[i]-e == i);
9720 // Eliminate identity shuffles.
9721 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9722 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9724 // If the LHS is a shufflevector itself, see if we can combine it with this
9725 // one without producing an unusual shuffle. Here we are really conservative:
9726 // we are absolutely afraid of producing a shuffle mask not in the input
9727 // program, because the code gen may not be smart enough to turn a merged
9728 // shuffle into two specific shuffles: it may produce worse code. As such,
9729 // we only merge two shuffles if the result is one of the two input shuffle
9730 // masks. In this case, merging the shuffles just removes one instruction,
9731 // which we know is safe. This is good for things like turning:
9732 // (splat(splat)) -> splat.
9733 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9734 if (isa<UndefValue>(RHS)) {
9735 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9737 std::vector<unsigned> NewMask;
9738 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9740 NewMask.push_back(2*e);
9742 NewMask.push_back(LHSMask[Mask[i]]);
9744 // If the result mask is equal to the src shuffle or this shuffle mask, do
9746 if (NewMask == LHSMask || NewMask == Mask) {
9747 std::vector<Constant*> Elts;
9748 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9749 if (NewMask[i] >= e*2) {
9750 Elts.push_back(UndefValue::get(Type::Int32Ty));
9752 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9755 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9756 LHSSVI->getOperand(1),
9757 ConstantVector::get(Elts));
9762 return MadeChange ? &SVI : 0;
9768 /// TryToSinkInstruction - Try to move the specified instruction from its
9769 /// current block into the beginning of DestBlock, which can only happen if it's
9770 /// safe to move the instruction past all of the instructions between it and the
9771 /// end of its block.
9772 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9773 assert(I->hasOneUse() && "Invariants didn't hold!");
9775 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9776 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9778 // Do not sink alloca instructions out of the entry block.
9779 if (isa<AllocaInst>(I) && I->getParent() ==
9780 &DestBlock->getParent()->getEntryBlock())
9783 // We can only sink load instructions if there is nothing between the load and
9784 // the end of block that could change the value.
9785 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9786 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9788 if (Scan->mayWriteToMemory())
9792 BasicBlock::iterator InsertPos = DestBlock->begin();
9793 while (isa<PHINode>(InsertPos)) ++InsertPos;
9795 I->moveBefore(InsertPos);
9801 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9802 /// all reachable code to the worklist.
9804 /// This has a couple of tricks to make the code faster and more powerful. In
9805 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9806 /// them to the worklist (this significantly speeds up instcombine on code where
9807 /// many instructions are dead or constant). Additionally, if we find a branch
9808 /// whose condition is a known constant, we only visit the reachable successors.
9810 static void AddReachableCodeToWorklist(BasicBlock *BB,
9811 SmallPtrSet<BasicBlock*, 64> &Visited,
9813 const TargetData *TD) {
9814 std::vector<BasicBlock*> Worklist;
9815 Worklist.push_back(BB);
9817 while (!Worklist.empty()) {
9818 BB = Worklist.back();
9819 Worklist.pop_back();
9821 // We have now visited this block! If we've already been here, ignore it.
9822 if (!Visited.insert(BB)) continue;
9824 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9825 Instruction *Inst = BBI++;
9827 // DCE instruction if trivially dead.
9828 if (isInstructionTriviallyDead(Inst)) {
9830 DOUT << "IC: DCE: " << *Inst;
9831 Inst->eraseFromParent();
9835 // ConstantProp instruction if trivially constant.
9836 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9837 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9838 Inst->replaceAllUsesWith(C);
9840 Inst->eraseFromParent();
9844 IC.AddToWorkList(Inst);
9847 // Recursively visit successors. If this is a branch or switch on a
9848 // constant, only visit the reachable successor.
9849 TerminatorInst *TI = BB->getTerminator();
9850 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9851 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9852 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9853 Worklist.push_back(BI->getSuccessor(!CondVal));
9856 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9857 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9858 // See if this is an explicit destination.
9859 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9860 if (SI->getCaseValue(i) == Cond) {
9861 Worklist.push_back(SI->getSuccessor(i));
9865 // Otherwise it is the default destination.
9866 Worklist.push_back(SI->getSuccessor(0));
9871 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9872 Worklist.push_back(TI->getSuccessor(i));
9876 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9877 bool Changed = false;
9878 TD = &getAnalysis<TargetData>();
9880 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9881 << F.getNameStr() << "\n");
9884 // Do a depth-first traversal of the function, populate the worklist with
9885 // the reachable instructions. Ignore blocks that are not reachable. Keep
9886 // track of which blocks we visit.
9887 SmallPtrSet<BasicBlock*, 64> Visited;
9888 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9890 // Do a quick scan over the function. If we find any blocks that are
9891 // unreachable, remove any instructions inside of them. This prevents
9892 // the instcombine code from having to deal with some bad special cases.
9893 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9894 if (!Visited.count(BB)) {
9895 Instruction *Term = BB->getTerminator();
9896 while (Term != BB->begin()) { // Remove instrs bottom-up
9897 BasicBlock::iterator I = Term; --I;
9899 DOUT << "IC: DCE: " << *I;
9902 if (!I->use_empty())
9903 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9904 I->eraseFromParent();
9909 while (!Worklist.empty()) {
9910 Instruction *I = RemoveOneFromWorkList();
9911 if (I == 0) continue; // skip null values.
9913 // Check to see if we can DCE the instruction.
9914 if (isInstructionTriviallyDead(I)) {
9915 // Add operands to the worklist.
9916 if (I->getNumOperands() < 4)
9917 AddUsesToWorkList(*I);
9920 DOUT << "IC: DCE: " << *I;
9922 I->eraseFromParent();
9923 RemoveFromWorkList(I);
9927 // Instruction isn't dead, see if we can constant propagate it.
9928 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9929 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9931 // Add operands to the worklist.
9932 AddUsesToWorkList(*I);
9933 ReplaceInstUsesWith(*I, C);
9936 I->eraseFromParent();
9937 RemoveFromWorkList(I);
9941 // See if we can trivially sink this instruction to a successor basic block.
9942 if (I->hasOneUse()) {
9943 BasicBlock *BB = I->getParent();
9944 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9945 if (UserParent != BB) {
9946 bool UserIsSuccessor = false;
9947 // See if the user is one of our successors.
9948 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9949 if (*SI == UserParent) {
9950 UserIsSuccessor = true;
9954 // If the user is one of our immediate successors, and if that successor
9955 // only has us as a predecessors (we'd have to split the critical edge
9956 // otherwise), we can keep going.
9957 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9958 next(pred_begin(UserParent)) == pred_end(UserParent))
9959 // Okay, the CFG is simple enough, try to sink this instruction.
9960 Changed |= TryToSinkInstruction(I, UserParent);
9964 // Now that we have an instruction, try combining it to simplify it...
9968 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
9969 if (Instruction *Result = visit(*I)) {
9971 // Should we replace the old instruction with a new one?
9973 DOUT << "IC: Old = " << *I
9974 << " New = " << *Result;
9976 // Everything uses the new instruction now.
9977 I->replaceAllUsesWith(Result);
9979 // Push the new instruction and any users onto the worklist.
9980 AddToWorkList(Result);
9981 AddUsersToWorkList(*Result);
9983 // Move the name to the new instruction first.
9984 Result->takeName(I);
9986 // Insert the new instruction into the basic block...
9987 BasicBlock *InstParent = I->getParent();
9988 BasicBlock::iterator InsertPos = I;
9990 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9991 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9994 InstParent->getInstList().insert(InsertPos, Result);
9996 // Make sure that we reprocess all operands now that we reduced their
9998 AddUsesToWorkList(*I);
10000 // Instructions can end up on the worklist more than once. Make sure
10001 // we do not process an instruction that has been deleted.
10002 RemoveFromWorkList(I);
10004 // Erase the old instruction.
10005 InstParent->getInstList().erase(I);
10008 DOUT << "IC: Mod = " << OrigI
10009 << " New = " << *I;
10012 // If the instruction was modified, it's possible that it is now dead.
10013 // if so, remove it.
10014 if (isInstructionTriviallyDead(I)) {
10015 // Make sure we process all operands now that we are reducing their
10017 AddUsesToWorkList(*I);
10019 // Instructions may end up in the worklist more than once. Erase all
10020 // occurrences of this instruction.
10021 RemoveFromWorkList(I);
10022 I->eraseFromParent();
10025 AddUsersToWorkList(*I);
10032 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10037 bool InstCombiner::runOnFunction(Function &F) {
10038 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10040 bool EverMadeChange = false;
10042 // Iterate while there is work to do.
10043 unsigned Iteration = 0;
10044 while (DoOneIteration(F, Iteration++))
10045 EverMadeChange = true;
10046 return EverMadeChange;
10049 FunctionPass *llvm::createInstructionCombiningPass() {
10050 return new InstCombiner();