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 the sign bit is the only bit demanded by this ashr, then there is no
1347 // need to do it, the shift doesn't change the high bit.
1348 if (DemandedMask.isSignBit())
1349 return UpdateValueUsesWith(I, I->getOperand(0));
1351 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1352 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1354 // Signed shift right.
1355 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1356 // If any of the "high bits" are demanded, we should set the sign bit as
1358 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1359 DemandedMaskIn.set(BitWidth-1);
1360 if (SimplifyDemandedBits(I->getOperand(0),
1362 RHSKnownZero, RHSKnownOne, Depth+1))
1364 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1365 "Bits known to be one AND zero?");
1366 // Compute the new bits that are at the top now.
1367 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1368 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1369 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1371 // Handle the sign bits.
1372 APInt SignBit(APInt::getSignBit(BitWidth));
1373 // Adjust to where it is now in the mask.
1374 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1376 // If the input sign bit is known to be zero, or if none of the top bits
1377 // are demanded, turn this into an unsigned shift right.
1378 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1379 (HighBits & ~DemandedMask) == HighBits) {
1380 // Perform the logical shift right.
1381 Value *NewVal = BinaryOperator::createLShr(
1382 I->getOperand(0), SA, I->getName());
1383 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1384 return UpdateValueUsesWith(I, NewVal);
1385 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1386 RHSKnownOne |= HighBits;
1392 // If the client is only demanding bits that we know, return the known
1394 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1395 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1400 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1401 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1402 /// actually used by the caller. This method analyzes which elements of the
1403 /// operand are undef and returns that information in UndefElts.
1405 /// If the information about demanded elements can be used to simplify the
1406 /// operation, the operation is simplified, then the resultant value is
1407 /// returned. This returns null if no change was made.
1408 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1409 uint64_t &UndefElts,
1411 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1412 assert(VWidth <= 64 && "Vector too wide to analyze!");
1413 uint64_t EltMask = ~0ULL >> (64-VWidth);
1414 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1415 "Invalid DemandedElts!");
1417 if (isa<UndefValue>(V)) {
1418 // If the entire vector is undefined, just return this info.
1419 UndefElts = EltMask;
1421 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1422 UndefElts = EltMask;
1423 return UndefValue::get(V->getType());
1427 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1428 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1429 Constant *Undef = UndefValue::get(EltTy);
1431 std::vector<Constant*> Elts;
1432 for (unsigned i = 0; i != VWidth; ++i)
1433 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1434 Elts.push_back(Undef);
1435 UndefElts |= (1ULL << i);
1436 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1437 Elts.push_back(Undef);
1438 UndefElts |= (1ULL << i);
1439 } else { // Otherwise, defined.
1440 Elts.push_back(CP->getOperand(i));
1443 // If we changed the constant, return it.
1444 Constant *NewCP = ConstantVector::get(Elts);
1445 return NewCP != CP ? NewCP : 0;
1446 } else if (isa<ConstantAggregateZero>(V)) {
1447 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1449 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1450 Constant *Zero = Constant::getNullValue(EltTy);
1451 Constant *Undef = UndefValue::get(EltTy);
1452 std::vector<Constant*> Elts;
1453 for (unsigned i = 0; i != VWidth; ++i)
1454 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1455 UndefElts = DemandedElts ^ EltMask;
1456 return ConstantVector::get(Elts);
1459 if (!V->hasOneUse()) { // Other users may use these bits.
1460 if (Depth != 0) { // Not at the root.
1461 // TODO: Just compute the UndefElts information recursively.
1465 } else if (Depth == 10) { // Limit search depth.
1469 Instruction *I = dyn_cast<Instruction>(V);
1470 if (!I) return false; // Only analyze instructions.
1472 bool MadeChange = false;
1473 uint64_t UndefElts2;
1475 switch (I->getOpcode()) {
1478 case Instruction::InsertElement: {
1479 // If this is a variable index, we don't know which element it overwrites.
1480 // demand exactly the same input as we produce.
1481 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1483 // Note that we can't propagate undef elt info, because we don't know
1484 // which elt is getting updated.
1485 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1486 UndefElts2, Depth+1);
1487 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1491 // If this is inserting an element that isn't demanded, remove this
1493 unsigned IdxNo = Idx->getZExtValue();
1494 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1495 return AddSoonDeadInstToWorklist(*I, 0);
1497 // Otherwise, the element inserted overwrites whatever was there, so the
1498 // input demanded set is simpler than the output set.
1499 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1500 DemandedElts & ~(1ULL << IdxNo),
1501 UndefElts, Depth+1);
1502 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1504 // The inserted element is defined.
1505 UndefElts |= 1ULL << IdxNo;
1508 case Instruction::BitCast: {
1509 // Vector->vector casts only.
1510 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1512 unsigned InVWidth = VTy->getNumElements();
1513 uint64_t InputDemandedElts = 0;
1516 if (VWidth == InVWidth) {
1517 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1518 // elements as are demanded of us.
1520 InputDemandedElts = DemandedElts;
1521 } else if (VWidth > InVWidth) {
1525 // If there are more elements in the result than there are in the source,
1526 // then an input element is live if any of the corresponding output
1527 // elements are live.
1528 Ratio = VWidth/InVWidth;
1529 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1530 if (DemandedElts & (1ULL << OutIdx))
1531 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1537 // If there are more elements in the source than there are in the result,
1538 // then an input element is live if the corresponding output element is
1540 Ratio = InVWidth/VWidth;
1541 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1542 if (DemandedElts & (1ULL << InIdx/Ratio))
1543 InputDemandedElts |= 1ULL << InIdx;
1546 // div/rem demand all inputs, because they don't want divide by zero.
1547 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1548 UndefElts2, Depth+1);
1550 I->setOperand(0, TmpV);
1554 UndefElts = UndefElts2;
1555 if (VWidth > InVWidth) {
1556 assert(0 && "Unimp");
1557 // If there are more elements in the result than there are in the source,
1558 // then an output element is undef if the corresponding input element is
1560 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1561 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1562 UndefElts |= 1ULL << OutIdx;
1563 } else if (VWidth < InVWidth) {
1564 assert(0 && "Unimp");
1565 // If there are more elements in the source than there are in the result,
1566 // then a result element is undef if all of the corresponding input
1567 // elements are undef.
1568 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1569 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1570 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1571 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1575 case Instruction::And:
1576 case Instruction::Or:
1577 case Instruction::Xor:
1578 case Instruction::Add:
1579 case Instruction::Sub:
1580 case Instruction::Mul:
1581 // div/rem demand all inputs, because they don't want divide by zero.
1582 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1583 UndefElts, Depth+1);
1584 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1585 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1586 UndefElts2, Depth+1);
1587 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1589 // Output elements are undefined if both are undefined. Consider things
1590 // like undef&0. The result is known zero, not undef.
1591 UndefElts &= UndefElts2;
1594 case Instruction::Call: {
1595 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1597 switch (II->getIntrinsicID()) {
1600 // Binary vector operations that work column-wise. A dest element is a
1601 // function of the corresponding input elements from the two inputs.
1602 case Intrinsic::x86_sse_sub_ss:
1603 case Intrinsic::x86_sse_mul_ss:
1604 case Intrinsic::x86_sse_min_ss:
1605 case Intrinsic::x86_sse_max_ss:
1606 case Intrinsic::x86_sse2_sub_sd:
1607 case Intrinsic::x86_sse2_mul_sd:
1608 case Intrinsic::x86_sse2_min_sd:
1609 case Intrinsic::x86_sse2_max_sd:
1610 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1611 UndefElts, Depth+1);
1612 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1613 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1614 UndefElts2, Depth+1);
1615 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1617 // If only the low elt is demanded and this is a scalarizable intrinsic,
1618 // scalarize it now.
1619 if (DemandedElts == 1) {
1620 switch (II->getIntrinsicID()) {
1622 case Intrinsic::x86_sse_sub_ss:
1623 case Intrinsic::x86_sse_mul_ss:
1624 case Intrinsic::x86_sse2_sub_sd:
1625 case Intrinsic::x86_sse2_mul_sd:
1626 // TODO: Lower MIN/MAX/ABS/etc
1627 Value *LHS = II->getOperand(1);
1628 Value *RHS = II->getOperand(2);
1629 // Extract the element as scalars.
1630 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1631 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1633 switch (II->getIntrinsicID()) {
1634 default: assert(0 && "Case stmts out of sync!");
1635 case Intrinsic::x86_sse_sub_ss:
1636 case Intrinsic::x86_sse2_sub_sd:
1637 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1638 II->getName()), *II);
1640 case Intrinsic::x86_sse_mul_ss:
1641 case Intrinsic::x86_sse2_mul_sd:
1642 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1643 II->getName()), *II);
1648 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1650 InsertNewInstBefore(New, *II);
1651 AddSoonDeadInstToWorklist(*II, 0);
1656 // Output elements are undefined if both are undefined. Consider things
1657 // like undef&0. The result is known zero, not undef.
1658 UndefElts &= UndefElts2;
1664 return MadeChange ? I : 0;
1667 /// @returns true if the specified compare instruction is
1668 /// true when both operands are equal...
1669 /// @brief Determine if the ICmpInst returns true if both operands are equal
1670 static bool isTrueWhenEqual(ICmpInst &ICI) {
1671 ICmpInst::Predicate pred = ICI.getPredicate();
1672 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1673 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1674 pred == ICmpInst::ICMP_SLE;
1677 /// AssociativeOpt - Perform an optimization on an associative operator. This
1678 /// function is designed to check a chain of associative operators for a
1679 /// potential to apply a certain optimization. Since the optimization may be
1680 /// applicable if the expression was reassociated, this checks the chain, then
1681 /// reassociates the expression as necessary to expose the optimization
1682 /// opportunity. This makes use of a special Functor, which must define
1683 /// 'shouldApply' and 'apply' methods.
1685 template<typename Functor>
1686 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1687 unsigned Opcode = Root.getOpcode();
1688 Value *LHS = Root.getOperand(0);
1690 // Quick check, see if the immediate LHS matches...
1691 if (F.shouldApply(LHS))
1692 return F.apply(Root);
1694 // Otherwise, if the LHS is not of the same opcode as the root, return.
1695 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1696 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1697 // Should we apply this transform to the RHS?
1698 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1700 // If not to the RHS, check to see if we should apply to the LHS...
1701 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1702 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1706 // If the functor wants to apply the optimization to the RHS of LHSI,
1707 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1709 BasicBlock *BB = Root.getParent();
1711 // Now all of the instructions are in the current basic block, go ahead
1712 // and perform the reassociation.
1713 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1715 // First move the selected RHS to the LHS of the root...
1716 Root.setOperand(0, LHSI->getOperand(1));
1718 // Make what used to be the LHS of the root be the user of the root...
1719 Value *ExtraOperand = TmpLHSI->getOperand(1);
1720 if (&Root == TmpLHSI) {
1721 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1724 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1725 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1726 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1727 BasicBlock::iterator ARI = &Root; ++ARI;
1728 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1731 // Now propagate the ExtraOperand down the chain of instructions until we
1733 while (TmpLHSI != LHSI) {
1734 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1735 // Move the instruction to immediately before the chain we are
1736 // constructing to avoid breaking dominance properties.
1737 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1738 BB->getInstList().insert(ARI, NextLHSI);
1741 Value *NextOp = NextLHSI->getOperand(1);
1742 NextLHSI->setOperand(1, ExtraOperand);
1744 ExtraOperand = NextOp;
1747 // Now that the instructions are reassociated, have the functor perform
1748 // the transformation...
1749 return F.apply(Root);
1752 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1758 // AddRHS - Implements: X + X --> X << 1
1761 AddRHS(Value *rhs) : RHS(rhs) {}
1762 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1763 Instruction *apply(BinaryOperator &Add) const {
1764 return BinaryOperator::createShl(Add.getOperand(0),
1765 ConstantInt::get(Add.getType(), 1));
1769 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1771 struct AddMaskingAnd {
1773 AddMaskingAnd(Constant *c) : C2(c) {}
1774 bool shouldApply(Value *LHS) const {
1776 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1777 ConstantExpr::getAnd(C1, C2)->isNullValue();
1779 Instruction *apply(BinaryOperator &Add) const {
1780 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1784 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1786 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1787 if (Constant *SOC = dyn_cast<Constant>(SO))
1788 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1790 return IC->InsertNewInstBefore(CastInst::create(
1791 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1794 // Figure out if the constant is the left or the right argument.
1795 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1796 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1798 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1800 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1801 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1804 Value *Op0 = SO, *Op1 = ConstOperand;
1806 std::swap(Op0, Op1);
1808 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1809 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1810 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1811 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1812 SO->getName()+".cmp");
1814 assert(0 && "Unknown binary instruction type!");
1817 return IC->InsertNewInstBefore(New, I);
1820 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1821 // constant as the other operand, try to fold the binary operator into the
1822 // select arguments. This also works for Cast instructions, which obviously do
1823 // not have a second operand.
1824 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1826 // Don't modify shared select instructions
1827 if (!SI->hasOneUse()) return 0;
1828 Value *TV = SI->getOperand(1);
1829 Value *FV = SI->getOperand(2);
1831 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1832 // Bool selects with constant operands can be folded to logical ops.
1833 if (SI->getType() == Type::Int1Ty) return 0;
1835 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1836 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1838 return new SelectInst(SI->getCondition(), SelectTrueVal,
1845 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1846 /// node as operand #0, see if we can fold the instruction into the PHI (which
1847 /// is only possible if all operands to the PHI are constants).
1848 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1849 PHINode *PN = cast<PHINode>(I.getOperand(0));
1850 unsigned NumPHIValues = PN->getNumIncomingValues();
1851 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1853 // Check to see if all of the operands of the PHI are constants. If there is
1854 // one non-constant value, remember the BB it is. If there is more than one
1855 // or if *it* is a PHI, bail out.
1856 BasicBlock *NonConstBB = 0;
1857 for (unsigned i = 0; i != NumPHIValues; ++i)
1858 if (!isa<Constant>(PN->getIncomingValue(i))) {
1859 if (NonConstBB) return 0; // More than one non-const value.
1860 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1861 NonConstBB = PN->getIncomingBlock(i);
1863 // If the incoming non-constant value is in I's block, we have an infinite
1865 if (NonConstBB == I.getParent())
1869 // If there is exactly one non-constant value, we can insert a copy of the
1870 // operation in that block. However, if this is a critical edge, we would be
1871 // inserting the computation one some other paths (e.g. inside a loop). Only
1872 // do this if the pred block is unconditionally branching into the phi block.
1874 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1875 if (!BI || !BI->isUnconditional()) return 0;
1878 // Okay, we can do the transformation: create the new PHI node.
1879 PHINode *NewPN = new PHINode(I.getType(), "");
1880 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1881 InsertNewInstBefore(NewPN, *PN);
1882 NewPN->takeName(PN);
1884 // Next, add all of the operands to the PHI.
1885 if (I.getNumOperands() == 2) {
1886 Constant *C = cast<Constant>(I.getOperand(1));
1887 for (unsigned i = 0; i != NumPHIValues; ++i) {
1889 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1890 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1891 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1893 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1895 assert(PN->getIncomingBlock(i) == NonConstBB);
1896 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1897 InV = BinaryOperator::create(BO->getOpcode(),
1898 PN->getIncomingValue(i), C, "phitmp",
1899 NonConstBB->getTerminator());
1900 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1901 InV = CmpInst::create(CI->getOpcode(),
1903 PN->getIncomingValue(i), C, "phitmp",
1904 NonConstBB->getTerminator());
1906 assert(0 && "Unknown binop!");
1908 AddToWorkList(cast<Instruction>(InV));
1910 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1913 CastInst *CI = cast<CastInst>(&I);
1914 const Type *RetTy = CI->getType();
1915 for (unsigned i = 0; i != NumPHIValues; ++i) {
1917 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1918 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1920 assert(PN->getIncomingBlock(i) == NonConstBB);
1921 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1922 I.getType(), "phitmp",
1923 NonConstBB->getTerminator());
1924 AddToWorkList(cast<Instruction>(InV));
1926 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1929 return ReplaceInstUsesWith(I, NewPN);
1932 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1933 bool Changed = SimplifyCommutative(I);
1934 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1936 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1937 // X + undef -> undef
1938 if (isa<UndefValue>(RHS))
1939 return ReplaceInstUsesWith(I, RHS);
1942 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1943 if (RHSC->isNullValue())
1944 return ReplaceInstUsesWith(I, LHS);
1945 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1946 if (CFP->isExactlyValue(-0.0))
1947 return ReplaceInstUsesWith(I, LHS);
1950 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1951 // X + (signbit) --> X ^ signbit
1952 const APInt& Val = CI->getValue();
1953 uint32_t BitWidth = Val.getBitWidth();
1954 if (Val == APInt::getSignBit(BitWidth))
1955 return BinaryOperator::createXor(LHS, RHS);
1957 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1958 // (X & 254)+1 -> (X&254)|1
1959 if (!isa<VectorType>(I.getType())) {
1960 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1961 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1962 KnownZero, KnownOne))
1967 if (isa<PHINode>(LHS))
1968 if (Instruction *NV = FoldOpIntoPhi(I))
1971 ConstantInt *XorRHS = 0;
1973 if (isa<ConstantInt>(RHSC) &&
1974 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1975 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1976 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1978 uint32_t Size = TySizeBits / 2;
1979 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1980 APInt CFF80Val(-C0080Val);
1982 if (TySizeBits > Size) {
1983 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1984 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1985 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1986 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1987 // This is a sign extend if the top bits are known zero.
1988 if (!MaskedValueIsZero(XorLHS,
1989 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1990 Size = 0; // Not a sign ext, but can't be any others either.
1995 C0080Val = APIntOps::lshr(C0080Val, Size);
1996 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1997 } while (Size >= 1);
1999 // FIXME: This shouldn't be necessary. When the backends can handle types
2000 // with funny bit widths then this whole cascade of if statements should
2001 // be removed. It is just here to get the size of the "middle" type back
2002 // up to something that the back ends can handle.
2003 const Type *MiddleType = 0;
2006 case 32: MiddleType = Type::Int32Ty; break;
2007 case 16: MiddleType = Type::Int16Ty; break;
2008 case 8: MiddleType = Type::Int8Ty; break;
2011 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2012 InsertNewInstBefore(NewTrunc, I);
2013 return new SExtInst(NewTrunc, I.getType(), I.getName());
2019 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2020 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2022 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2023 if (RHSI->getOpcode() == Instruction::Sub)
2024 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2025 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2027 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2028 if (LHSI->getOpcode() == Instruction::Sub)
2029 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2030 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2035 if (Value *V = dyn_castNegVal(LHS))
2036 return BinaryOperator::createSub(RHS, V);
2039 if (!isa<Constant>(RHS))
2040 if (Value *V = dyn_castNegVal(RHS))
2041 return BinaryOperator::createSub(LHS, V);
2045 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2046 if (X == RHS) // X*C + X --> X * (C+1)
2047 return BinaryOperator::createMul(RHS, AddOne(C2));
2049 // X*C1 + X*C2 --> X * (C1+C2)
2051 if (X == dyn_castFoldableMul(RHS, C1))
2052 return BinaryOperator::createMul(X, Add(C1, C2));
2055 // X + X*C --> X * (C+1)
2056 if (dyn_castFoldableMul(RHS, C2) == LHS)
2057 return BinaryOperator::createMul(LHS, AddOne(C2));
2059 // X + ~X --> -1 since ~X = -X-1
2060 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2061 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2064 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2065 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2066 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2069 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2071 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2072 return BinaryOperator::createSub(SubOne(CRHS), X);
2074 // (X & FF00) + xx00 -> (X+xx00) & FF00
2075 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2076 Constant *Anded = And(CRHS, C2);
2077 if (Anded == CRHS) {
2078 // See if all bits from the first bit set in the Add RHS up are included
2079 // in the mask. First, get the rightmost bit.
2080 const APInt& AddRHSV = CRHS->getValue();
2082 // Form a mask of all bits from the lowest bit added through the top.
2083 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2085 // See if the and mask includes all of these bits.
2086 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2088 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2089 // Okay, the xform is safe. Insert the new add pronto.
2090 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2091 LHS->getName()), I);
2092 return BinaryOperator::createAnd(NewAdd, C2);
2097 // Try to fold constant add into select arguments.
2098 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2099 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2103 // add (cast *A to intptrtype) B ->
2104 // cast (GEP (cast *A to sbyte*) B) ->
2107 CastInst *CI = dyn_cast<CastInst>(LHS);
2110 CI = dyn_cast<CastInst>(RHS);
2113 if (CI && CI->getType()->isSized() &&
2114 (CI->getType()->getPrimitiveSizeInBits() ==
2115 TD->getIntPtrType()->getPrimitiveSizeInBits())
2116 && isa<PointerType>(CI->getOperand(0)->getType())) {
2117 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2118 PointerType::get(Type::Int8Ty), I);
2119 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2120 return new PtrToIntInst(I2, CI->getType());
2124 return Changed ? &I : 0;
2127 // isSignBit - Return true if the value represented by the constant only has the
2128 // highest order bit set.
2129 static bool isSignBit(ConstantInt *CI) {
2130 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2131 return CI->getValue() == APInt::getSignBit(NumBits);
2134 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2135 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2137 if (Op0 == Op1) // sub X, X -> 0
2138 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2140 // If this is a 'B = x-(-A)', change to B = x+A...
2141 if (Value *V = dyn_castNegVal(Op1))
2142 return BinaryOperator::createAdd(Op0, V);
2144 if (isa<UndefValue>(Op0))
2145 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2146 if (isa<UndefValue>(Op1))
2147 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2149 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2150 // Replace (-1 - A) with (~A)...
2151 if (C->isAllOnesValue())
2152 return BinaryOperator::createNot(Op1);
2154 // C - ~X == X + (1+C)
2156 if (match(Op1, m_Not(m_Value(X))))
2157 return BinaryOperator::createAdd(X, AddOne(C));
2159 // -(X >>u 31) -> (X >>s 31)
2160 // -(X >>s 31) -> (X >>u 31)
2162 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2163 if (SI->getOpcode() == Instruction::LShr) {
2164 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2165 // Check to see if we are shifting out everything but the sign bit.
2166 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2167 SI->getType()->getPrimitiveSizeInBits()-1) {
2168 // Ok, the transformation is safe. Insert AShr.
2169 return BinaryOperator::create(Instruction::AShr,
2170 SI->getOperand(0), CU, SI->getName());
2174 else if (SI->getOpcode() == Instruction::AShr) {
2175 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2176 // Check to see if we are shifting out everything but the sign bit.
2177 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2178 SI->getType()->getPrimitiveSizeInBits()-1) {
2179 // Ok, the transformation is safe. Insert LShr.
2180 return BinaryOperator::createLShr(
2181 SI->getOperand(0), CU, SI->getName());
2187 // Try to fold constant sub into select arguments.
2188 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2189 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2192 if (isa<PHINode>(Op0))
2193 if (Instruction *NV = FoldOpIntoPhi(I))
2197 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2198 if (Op1I->getOpcode() == Instruction::Add &&
2199 !Op0->getType()->isFPOrFPVector()) {
2200 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2201 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2202 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2203 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2204 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2205 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2206 // C1-(X+C2) --> (C1-C2)-X
2207 return BinaryOperator::createSub(Subtract(CI1, CI2),
2208 Op1I->getOperand(0));
2212 if (Op1I->hasOneUse()) {
2213 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2214 // is not used by anyone else...
2216 if (Op1I->getOpcode() == Instruction::Sub &&
2217 !Op1I->getType()->isFPOrFPVector()) {
2218 // Swap the two operands of the subexpr...
2219 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2220 Op1I->setOperand(0, IIOp1);
2221 Op1I->setOperand(1, IIOp0);
2223 // Create the new top level add instruction...
2224 return BinaryOperator::createAdd(Op0, Op1);
2227 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2229 if (Op1I->getOpcode() == Instruction::And &&
2230 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2231 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2234 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2235 return BinaryOperator::createAnd(Op0, NewNot);
2238 // 0 - (X sdiv C) -> (X sdiv -C)
2239 if (Op1I->getOpcode() == Instruction::SDiv)
2240 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2242 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2243 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2244 ConstantExpr::getNeg(DivRHS));
2246 // X - X*C --> X * (1-C)
2247 ConstantInt *C2 = 0;
2248 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2249 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2250 return BinaryOperator::createMul(Op0, CP1);
2255 if (!Op0->getType()->isFPOrFPVector())
2256 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2257 if (Op0I->getOpcode() == Instruction::Add) {
2258 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2259 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2260 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2261 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2262 } else if (Op0I->getOpcode() == Instruction::Sub) {
2263 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2264 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2268 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2269 if (X == Op1) // X*C - X --> X * (C-1)
2270 return BinaryOperator::createMul(Op1, SubOne(C1));
2272 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2273 if (X == dyn_castFoldableMul(Op1, C2))
2274 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2279 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2280 /// comparison only checks the sign bit. If it only checks the sign bit, set
2281 /// TrueIfSigned if the result of the comparison is true when the input value is
2283 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2284 bool &TrueIfSigned) {
2286 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2287 TrueIfSigned = true;
2288 return RHS->isZero();
2289 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2290 TrueIfSigned = true;
2291 return RHS->isAllOnesValue();
2292 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2293 TrueIfSigned = false;
2294 return RHS->isAllOnesValue();
2295 case ICmpInst::ICMP_UGT:
2296 // True if LHS u> RHS and RHS == high-bit-mask - 1
2297 TrueIfSigned = true;
2298 return RHS->getValue() ==
2299 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2300 case ICmpInst::ICMP_UGE:
2301 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2302 TrueIfSigned = true;
2303 return RHS->getValue() ==
2304 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2310 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2311 bool Changed = SimplifyCommutative(I);
2312 Value *Op0 = I.getOperand(0);
2314 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2315 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2317 // Simplify mul instructions with a constant RHS...
2318 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2319 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2321 // ((X << C1)*C2) == (X * (C2 << C1))
2322 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2323 if (SI->getOpcode() == Instruction::Shl)
2324 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2325 return BinaryOperator::createMul(SI->getOperand(0),
2326 ConstantExpr::getShl(CI, ShOp));
2329 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2330 if (CI->equalsInt(1)) // X * 1 == X
2331 return ReplaceInstUsesWith(I, Op0);
2332 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2333 return BinaryOperator::createNeg(Op0, I.getName());
2335 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2336 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2337 return BinaryOperator::createShl(Op0,
2338 ConstantInt::get(Op0->getType(), Val.logBase2()));
2340 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2341 if (Op1F->isNullValue())
2342 return ReplaceInstUsesWith(I, Op1);
2344 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2345 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2346 if (Op1F->getValue() == 1.0)
2347 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2350 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2351 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2352 isa<ConstantInt>(Op0I->getOperand(1))) {
2353 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2354 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2356 InsertNewInstBefore(Add, I);
2357 Value *C1C2 = ConstantExpr::getMul(Op1,
2358 cast<Constant>(Op0I->getOperand(1)));
2359 return BinaryOperator::createAdd(Add, C1C2);
2363 // Try to fold constant mul into select arguments.
2364 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2365 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2368 if (isa<PHINode>(Op0))
2369 if (Instruction *NV = FoldOpIntoPhi(I))
2373 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2374 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2375 return BinaryOperator::createMul(Op0v, Op1v);
2377 // If one of the operands of the multiply is a cast from a boolean value, then
2378 // we know the bool is either zero or one, so this is a 'masking' multiply.
2379 // See if we can simplify things based on how the boolean was originally
2381 CastInst *BoolCast = 0;
2382 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2383 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2386 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2387 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2390 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2391 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2392 const Type *SCOpTy = SCIOp0->getType();
2395 // If the icmp is true iff the sign bit of X is set, then convert this
2396 // multiply into a shift/and combination.
2397 if (isa<ConstantInt>(SCIOp1) &&
2398 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2400 // Shift the X value right to turn it into "all signbits".
2401 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2402 SCOpTy->getPrimitiveSizeInBits()-1);
2404 InsertNewInstBefore(
2405 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2406 BoolCast->getOperand(0)->getName()+
2409 // If the multiply type is not the same as the source type, sign extend
2410 // or truncate to the multiply type.
2411 if (I.getType() != V->getType()) {
2412 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2413 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2414 Instruction::CastOps opcode =
2415 (SrcBits == DstBits ? Instruction::BitCast :
2416 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2417 V = InsertCastBefore(opcode, V, I.getType(), I);
2420 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2421 return BinaryOperator::createAnd(V, OtherOp);
2426 return Changed ? &I : 0;
2429 /// This function implements the transforms on div instructions that work
2430 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2431 /// used by the visitors to those instructions.
2432 /// @brief Transforms common to all three div instructions
2433 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2434 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2437 if (isa<UndefValue>(Op0))
2438 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2440 // X / undef -> undef
2441 if (isa<UndefValue>(Op1))
2442 return ReplaceInstUsesWith(I, Op1);
2444 // Handle cases involving: div X, (select Cond, Y, Z)
2445 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2446 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2447 // same basic block, then we replace the select with Y, and the condition
2448 // of the select with false (if the cond value is in the same BB). If the
2449 // select has uses other than the div, this allows them to be simplified
2450 // also. Note that div X, Y is just as good as div X, 0 (undef)
2451 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2452 if (ST->isNullValue()) {
2453 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2454 if (CondI && CondI->getParent() == I.getParent())
2455 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2456 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2457 I.setOperand(1, SI->getOperand(2));
2459 UpdateValueUsesWith(SI, SI->getOperand(2));
2463 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2464 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2465 if (ST->isNullValue()) {
2466 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2467 if (CondI && CondI->getParent() == I.getParent())
2468 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2469 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2470 I.setOperand(1, SI->getOperand(1));
2472 UpdateValueUsesWith(SI, SI->getOperand(1));
2480 /// This function implements the transforms common to both integer division
2481 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2482 /// division instructions.
2483 /// @brief Common integer divide transforms
2484 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2485 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2487 if (Instruction *Common = commonDivTransforms(I))
2490 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2492 if (RHS->equalsInt(1))
2493 return ReplaceInstUsesWith(I, Op0);
2495 // (X / C1) / C2 -> X / (C1*C2)
2496 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2497 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2498 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2499 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2500 Multiply(RHS, LHSRHS));
2503 if (!RHS->isZero()) { // avoid X udiv 0
2504 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2505 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2507 if (isa<PHINode>(Op0))
2508 if (Instruction *NV = FoldOpIntoPhi(I))
2513 // 0 / X == 0, we don't need to preserve faults!
2514 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2515 if (LHS->equalsInt(0))
2516 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2521 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2522 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2524 // Handle the integer div common cases
2525 if (Instruction *Common = commonIDivTransforms(I))
2528 // X udiv C^2 -> X >> C
2529 // Check to see if this is an unsigned division with an exact power of 2,
2530 // if so, convert to a right shift.
2531 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2532 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2533 return BinaryOperator::createLShr(Op0,
2534 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2537 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2538 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2539 if (RHSI->getOpcode() == Instruction::Shl &&
2540 isa<ConstantInt>(RHSI->getOperand(0))) {
2541 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2542 if (C1.isPowerOf2()) {
2543 Value *N = RHSI->getOperand(1);
2544 const Type *NTy = N->getType();
2545 if (uint32_t C2 = C1.logBase2()) {
2546 Constant *C2V = ConstantInt::get(NTy, C2);
2547 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2549 return BinaryOperator::createLShr(Op0, N);
2554 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2555 // where C1&C2 are powers of two.
2556 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2557 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2558 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2559 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2560 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2561 // Compute the shift amounts
2562 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2563 // Construct the "on true" case of the select
2564 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2565 Instruction *TSI = BinaryOperator::createLShr(
2566 Op0, TC, SI->getName()+".t");
2567 TSI = InsertNewInstBefore(TSI, I);
2569 // Construct the "on false" case of the select
2570 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2571 Instruction *FSI = BinaryOperator::createLShr(
2572 Op0, FC, SI->getName()+".f");
2573 FSI = InsertNewInstBefore(FSI, I);
2575 // construct the select instruction and return it.
2576 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2582 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2583 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2585 // Handle the integer div common cases
2586 if (Instruction *Common = commonIDivTransforms(I))
2589 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2591 if (RHS->isAllOnesValue())
2592 return BinaryOperator::createNeg(Op0);
2595 if (Value *LHSNeg = dyn_castNegVal(Op0))
2596 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2599 // If the sign bits of both operands are zero (i.e. we can prove they are
2600 // unsigned inputs), turn this into a udiv.
2601 if (I.getType()->isInteger()) {
2602 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2603 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2604 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2611 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2612 return commonDivTransforms(I);
2615 /// GetFactor - If we can prove that the specified value is at least a multiple
2616 /// of some factor, return that factor.
2617 static Constant *GetFactor(Value *V) {
2618 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2621 // Unless we can be tricky, we know this is a multiple of 1.
2622 Constant *Result = ConstantInt::get(V->getType(), 1);
2624 Instruction *I = dyn_cast<Instruction>(V);
2625 if (!I) return Result;
2627 if (I->getOpcode() == Instruction::Mul) {
2628 // Handle multiplies by a constant, etc.
2629 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2630 GetFactor(I->getOperand(1)));
2631 } else if (I->getOpcode() == Instruction::Shl) {
2632 // (X<<C) -> X * (1 << C)
2633 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2634 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2635 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2637 } else if (I->getOpcode() == Instruction::And) {
2638 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2639 // X & 0xFFF0 is known to be a multiple of 16.
2640 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2641 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2642 return ConstantExpr::getShl(Result,
2643 ConstantInt::get(Result->getType(), Zeros));
2645 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2646 // Only handle int->int casts.
2647 if (!CI->isIntegerCast())
2649 Value *Op = CI->getOperand(0);
2650 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2655 /// This function implements the transforms on rem instructions that work
2656 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2657 /// is used by the visitors to those instructions.
2658 /// @brief Transforms common to all three rem instructions
2659 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2660 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2662 // 0 % X == 0, we don't need to preserve faults!
2663 if (Constant *LHS = dyn_cast<Constant>(Op0))
2664 if (LHS->isNullValue())
2665 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2667 if (isa<UndefValue>(Op0)) // undef % X -> 0
2668 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2669 if (isa<UndefValue>(Op1))
2670 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2672 // Handle cases involving: rem X, (select Cond, Y, Z)
2673 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2674 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2675 // the same basic block, then we replace the select with Y, and the
2676 // condition of the select with false (if the cond value is in the same
2677 // BB). If the select has uses other than the div, this allows them to be
2679 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2680 if (ST->isNullValue()) {
2681 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2682 if (CondI && CondI->getParent() == I.getParent())
2683 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2684 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2685 I.setOperand(1, SI->getOperand(2));
2687 UpdateValueUsesWith(SI, SI->getOperand(2));
2690 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2691 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2692 if (ST->isNullValue()) {
2693 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2694 if (CondI && CondI->getParent() == I.getParent())
2695 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2696 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2697 I.setOperand(1, SI->getOperand(1));
2699 UpdateValueUsesWith(SI, SI->getOperand(1));
2707 /// This function implements the transforms common to both integer remainder
2708 /// instructions (urem and srem). It is called by the visitors to those integer
2709 /// remainder instructions.
2710 /// @brief Common integer remainder transforms
2711 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2712 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2714 if (Instruction *common = commonRemTransforms(I))
2717 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2718 // X % 0 == undef, we don't need to preserve faults!
2719 if (RHS->equalsInt(0))
2720 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2722 if (RHS->equalsInt(1)) // X % 1 == 0
2723 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2725 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2726 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2727 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2729 } else if (isa<PHINode>(Op0I)) {
2730 if (Instruction *NV = FoldOpIntoPhi(I))
2733 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2734 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2735 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2742 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2743 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2745 if (Instruction *common = commonIRemTransforms(I))
2748 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2749 // X urem C^2 -> X and C
2750 // Check to see if this is an unsigned remainder with an exact power of 2,
2751 // if so, convert to a bitwise and.
2752 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2753 if (C->getValue().isPowerOf2())
2754 return BinaryOperator::createAnd(Op0, SubOne(C));
2757 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2758 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2759 if (RHSI->getOpcode() == Instruction::Shl &&
2760 isa<ConstantInt>(RHSI->getOperand(0))) {
2761 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2762 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2763 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2765 return BinaryOperator::createAnd(Op0, Add);
2770 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2771 // where C1&C2 are powers of two.
2772 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2773 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2774 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2775 // STO == 0 and SFO == 0 handled above.
2776 if ((STO->getValue().isPowerOf2()) &&
2777 (SFO->getValue().isPowerOf2())) {
2778 Value *TrueAnd = InsertNewInstBefore(
2779 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2780 Value *FalseAnd = InsertNewInstBefore(
2781 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2782 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2790 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2791 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2793 if (Instruction *common = commonIRemTransforms(I))
2796 if (Value *RHSNeg = dyn_castNegVal(Op1))
2797 if (!isa<ConstantInt>(RHSNeg) ||
2798 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2800 AddUsesToWorkList(I);
2801 I.setOperand(1, RHSNeg);
2805 // If the top bits of both operands are zero (i.e. we can prove they are
2806 // unsigned inputs), turn this into a urem.
2807 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2808 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2809 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2810 return BinaryOperator::createURem(Op0, Op1, I.getName());
2816 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2817 return commonRemTransforms(I);
2820 // isMaxValueMinusOne - return true if this is Max-1
2821 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2822 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2824 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2825 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2828 // isMinValuePlusOne - return true if this is Min+1
2829 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2831 return C->getValue() == 1; // unsigned
2833 // Calculate 1111111111000000000000
2834 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2835 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2838 // isOneBitSet - Return true if there is exactly one bit set in the specified
2840 static bool isOneBitSet(const ConstantInt *CI) {
2841 return CI->getValue().isPowerOf2();
2844 // isHighOnes - Return true if the constant is of the form 1+0+.
2845 // This is the same as lowones(~X).
2846 static bool isHighOnes(const ConstantInt *CI) {
2847 return (~CI->getValue() + 1).isPowerOf2();
2850 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2851 /// are carefully arranged to allow folding of expressions such as:
2853 /// (A < B) | (A > B) --> (A != B)
2855 /// Note that this is only valid if the first and second predicates have the
2856 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2858 /// Three bits are used to represent the condition, as follows:
2863 /// <=> Value Definition
2864 /// 000 0 Always false
2871 /// 111 7 Always true
2873 static unsigned getICmpCode(const ICmpInst *ICI) {
2874 switch (ICI->getPredicate()) {
2876 case ICmpInst::ICMP_UGT: return 1; // 001
2877 case ICmpInst::ICMP_SGT: return 1; // 001
2878 case ICmpInst::ICMP_EQ: return 2; // 010
2879 case ICmpInst::ICMP_UGE: return 3; // 011
2880 case ICmpInst::ICMP_SGE: return 3; // 011
2881 case ICmpInst::ICMP_ULT: return 4; // 100
2882 case ICmpInst::ICMP_SLT: return 4; // 100
2883 case ICmpInst::ICMP_NE: return 5; // 101
2884 case ICmpInst::ICMP_ULE: return 6; // 110
2885 case ICmpInst::ICMP_SLE: return 6; // 110
2888 assert(0 && "Invalid ICmp predicate!");
2893 /// getICmpValue - This is the complement of getICmpCode, which turns an
2894 /// opcode and two operands into either a constant true or false, or a brand
2895 /// new /// ICmp instruction. The sign is passed in to determine which kind
2896 /// of predicate to use in new icmp instructions.
2897 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2899 default: assert(0 && "Illegal ICmp code!");
2900 case 0: return ConstantInt::getFalse();
2903 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2905 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2906 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2909 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2911 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2914 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2916 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2917 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2920 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2922 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2923 case 7: return ConstantInt::getTrue();
2927 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2928 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2929 (ICmpInst::isSignedPredicate(p1) &&
2930 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2931 (ICmpInst::isSignedPredicate(p2) &&
2932 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2936 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2937 struct FoldICmpLogical {
2940 ICmpInst::Predicate pred;
2941 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2942 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2943 pred(ICI->getPredicate()) {}
2944 bool shouldApply(Value *V) const {
2945 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2946 if (PredicatesFoldable(pred, ICI->getPredicate()))
2947 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2948 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2951 Instruction *apply(Instruction &Log) const {
2952 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2953 if (ICI->getOperand(0) != LHS) {
2954 assert(ICI->getOperand(1) == LHS);
2955 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2958 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2959 unsigned LHSCode = getICmpCode(ICI);
2960 unsigned RHSCode = getICmpCode(RHSICI);
2962 switch (Log.getOpcode()) {
2963 case Instruction::And: Code = LHSCode & RHSCode; break;
2964 case Instruction::Or: Code = LHSCode | RHSCode; break;
2965 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2966 default: assert(0 && "Illegal logical opcode!"); return 0;
2969 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2970 ICmpInst::isSignedPredicate(ICI->getPredicate());
2972 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2973 if (Instruction *I = dyn_cast<Instruction>(RV))
2975 // Otherwise, it's a constant boolean value...
2976 return IC.ReplaceInstUsesWith(Log, RV);
2979 } // end anonymous namespace
2981 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2982 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2983 // guaranteed to be a binary operator.
2984 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2986 ConstantInt *AndRHS,
2987 BinaryOperator &TheAnd) {
2988 Value *X = Op->getOperand(0);
2989 Constant *Together = 0;
2991 Together = And(AndRHS, OpRHS);
2993 switch (Op->getOpcode()) {
2994 case Instruction::Xor:
2995 if (Op->hasOneUse()) {
2996 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2997 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2998 InsertNewInstBefore(And, TheAnd);
3000 return BinaryOperator::createXor(And, Together);
3003 case Instruction::Or:
3004 if (Together == AndRHS) // (X | C) & C --> C
3005 return ReplaceInstUsesWith(TheAnd, AndRHS);
3007 if (Op->hasOneUse() && Together != OpRHS) {
3008 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3009 Instruction *Or = BinaryOperator::createOr(X, Together);
3010 InsertNewInstBefore(Or, TheAnd);
3012 return BinaryOperator::createAnd(Or, AndRHS);
3015 case Instruction::Add:
3016 if (Op->hasOneUse()) {
3017 // Adding a one to a single bit bit-field should be turned into an XOR
3018 // of the bit. First thing to check is to see if this AND is with a
3019 // single bit constant.
3020 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3022 // If there is only one bit set...
3023 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3024 // Ok, at this point, we know that we are masking the result of the
3025 // ADD down to exactly one bit. If the constant we are adding has
3026 // no bits set below this bit, then we can eliminate the ADD.
3027 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3029 // Check to see if any bits below the one bit set in AndRHSV are set.
3030 if ((AddRHS & (AndRHSV-1)) == 0) {
3031 // If not, the only thing that can effect the output of the AND is
3032 // the bit specified by AndRHSV. If that bit is set, the effect of
3033 // the XOR is to toggle the bit. If it is clear, then the ADD has
3035 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3036 TheAnd.setOperand(0, X);
3039 // Pull the XOR out of the AND.
3040 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3041 InsertNewInstBefore(NewAnd, TheAnd);
3042 NewAnd->takeName(Op);
3043 return BinaryOperator::createXor(NewAnd, AndRHS);
3050 case Instruction::Shl: {
3051 // We know that the AND will not produce any of the bits shifted in, so if
3052 // the anded constant includes them, clear them now!
3054 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3055 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3056 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3057 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3059 if (CI->getValue() == ShlMask) {
3060 // Masking out bits that the shift already masks
3061 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3062 } else if (CI != AndRHS) { // Reducing bits set in and.
3063 TheAnd.setOperand(1, CI);
3068 case Instruction::LShr:
3070 // We know that the AND will not produce any of the bits shifted in, so if
3071 // the anded constant includes them, clear them now! This only applies to
3072 // unsigned shifts, because a signed shr may bring in set bits!
3074 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3075 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3076 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3077 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3079 if (CI->getValue() == ShrMask) {
3080 // Masking out bits that the shift already masks.
3081 return ReplaceInstUsesWith(TheAnd, Op);
3082 } else if (CI != AndRHS) {
3083 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3088 case Instruction::AShr:
3090 // See if this is shifting in some sign extension, then masking it out
3092 if (Op->hasOneUse()) {
3093 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3094 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3095 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3096 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3097 if (C == AndRHS) { // Masking out bits shifted in.
3098 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3099 // Make the argument unsigned.
3100 Value *ShVal = Op->getOperand(0);
3101 ShVal = InsertNewInstBefore(
3102 BinaryOperator::createLShr(ShVal, OpRHS,
3103 Op->getName()), TheAnd);
3104 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3113 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3114 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3115 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3116 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3117 /// insert new instructions.
3118 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3119 bool isSigned, bool Inside,
3121 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3122 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3123 "Lo is not <= Hi in range emission code!");
3126 if (Lo == Hi) // Trivially false.
3127 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3129 // V >= Min && V < Hi --> V < Hi
3130 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3131 ICmpInst::Predicate pred = (isSigned ?
3132 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3133 return new ICmpInst(pred, V, Hi);
3136 // Emit V-Lo <u Hi-Lo
3137 Constant *NegLo = ConstantExpr::getNeg(Lo);
3138 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3139 InsertNewInstBefore(Add, IB);
3140 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3141 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3144 if (Lo == Hi) // Trivially true.
3145 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3147 // V < Min || V >= Hi -> V > Hi-1
3148 Hi = SubOne(cast<ConstantInt>(Hi));
3149 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3150 ICmpInst::Predicate pred = (isSigned ?
3151 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3152 return new ICmpInst(pred, V, Hi);
3155 // Emit V-Lo >u Hi-1-Lo
3156 // Note that Hi has already had one subtracted from it, above.
3157 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3158 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3159 InsertNewInstBefore(Add, IB);
3160 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3161 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3164 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3165 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3166 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3167 // not, since all 1s are not contiguous.
3168 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3169 const APInt& V = Val->getValue();
3170 uint32_t BitWidth = Val->getType()->getBitWidth();
3171 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3173 // look for the first zero bit after the run of ones
3174 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3175 // look for the first non-zero bit
3176 ME = V.getActiveBits();
3180 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3181 /// where isSub determines whether the operator is a sub. If we can fold one of
3182 /// the following xforms:
3184 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3185 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3186 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3188 /// return (A +/- B).
3190 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3191 ConstantInt *Mask, bool isSub,
3193 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3194 if (!LHSI || LHSI->getNumOperands() != 2 ||
3195 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3197 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3199 switch (LHSI->getOpcode()) {
3201 case Instruction::And:
3202 if (And(N, Mask) == Mask) {
3203 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3204 if ((Mask->getValue().countLeadingZeros() +
3205 Mask->getValue().countPopulation()) ==
3206 Mask->getValue().getBitWidth())
3209 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3210 // part, we don't need any explicit masks to take them out of A. If that
3211 // is all N is, ignore it.
3212 uint32_t MB = 0, ME = 0;
3213 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3214 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3215 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3216 if (MaskedValueIsZero(RHS, Mask))
3221 case Instruction::Or:
3222 case Instruction::Xor:
3223 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3224 if ((Mask->getValue().countLeadingZeros() +
3225 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3226 && And(N, Mask)->isZero())
3233 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3235 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3236 return InsertNewInstBefore(New, I);
3239 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3240 bool Changed = SimplifyCommutative(I);
3241 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3243 if (isa<UndefValue>(Op1)) // X & undef -> 0
3244 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3248 return ReplaceInstUsesWith(I, Op1);
3250 // See if we can simplify any instructions used by the instruction whose sole
3251 // purpose is to compute bits we don't care about.
3252 if (!isa<VectorType>(I.getType())) {
3253 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3254 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3255 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3256 KnownZero, KnownOne))
3259 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3260 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3261 return ReplaceInstUsesWith(I, I.getOperand(0));
3262 } else if (isa<ConstantAggregateZero>(Op1)) {
3263 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3267 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3268 const APInt& AndRHSMask = AndRHS->getValue();
3269 APInt NotAndRHS(~AndRHSMask);
3271 // Optimize a variety of ((val OP C1) & C2) combinations...
3272 if (isa<BinaryOperator>(Op0)) {
3273 Instruction *Op0I = cast<Instruction>(Op0);
3274 Value *Op0LHS = Op0I->getOperand(0);
3275 Value *Op0RHS = Op0I->getOperand(1);
3276 switch (Op0I->getOpcode()) {
3277 case Instruction::Xor:
3278 case Instruction::Or:
3279 // If the mask is only needed on one incoming arm, push it up.
3280 if (Op0I->hasOneUse()) {
3281 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3282 // Not masking anything out for the LHS, move to RHS.
3283 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3284 Op0RHS->getName()+".masked");
3285 InsertNewInstBefore(NewRHS, I);
3286 return BinaryOperator::create(
3287 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3289 if (!isa<Constant>(Op0RHS) &&
3290 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3291 // Not masking anything out for the RHS, move to LHS.
3292 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3293 Op0LHS->getName()+".masked");
3294 InsertNewInstBefore(NewLHS, I);
3295 return BinaryOperator::create(
3296 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3301 case Instruction::Add:
3302 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3303 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3304 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3305 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3306 return BinaryOperator::createAnd(V, AndRHS);
3307 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3308 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3311 case Instruction::Sub:
3312 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3313 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3314 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3315 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3316 return BinaryOperator::createAnd(V, AndRHS);
3320 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3321 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3323 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3324 // If this is an integer truncation or change from signed-to-unsigned, and
3325 // if the source is an and/or with immediate, transform it. This
3326 // frequently occurs for bitfield accesses.
3327 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3328 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3329 CastOp->getNumOperands() == 2)
3330 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3331 if (CastOp->getOpcode() == Instruction::And) {
3332 // Change: and (cast (and X, C1) to T), C2
3333 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3334 // This will fold the two constants together, which may allow
3335 // other simplifications.
3336 Instruction *NewCast = CastInst::createTruncOrBitCast(
3337 CastOp->getOperand(0), I.getType(),
3338 CastOp->getName()+".shrunk");
3339 NewCast = InsertNewInstBefore(NewCast, I);
3340 // trunc_or_bitcast(C1)&C2
3341 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3342 C3 = ConstantExpr::getAnd(C3, AndRHS);
3343 return BinaryOperator::createAnd(NewCast, C3);
3344 } else if (CastOp->getOpcode() == Instruction::Or) {
3345 // Change: and (cast (or X, C1) to T), C2
3346 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3347 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3348 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3349 return ReplaceInstUsesWith(I, AndRHS);
3354 // Try to fold constant and into select arguments.
3355 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3356 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3358 if (isa<PHINode>(Op0))
3359 if (Instruction *NV = FoldOpIntoPhi(I))
3363 Value *Op0NotVal = dyn_castNotVal(Op0);
3364 Value *Op1NotVal = dyn_castNotVal(Op1);
3366 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3367 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3369 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3370 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3371 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3372 I.getName()+".demorgan");
3373 InsertNewInstBefore(Or, I);
3374 return BinaryOperator::createNot(Or);
3378 Value *A = 0, *B = 0, *C = 0, *D = 0;
3379 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3380 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3381 return ReplaceInstUsesWith(I, Op1);
3383 // (A|B) & ~(A&B) -> A^B
3384 if (match(Op1, 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 (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3391 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3392 return ReplaceInstUsesWith(I, Op0);
3394 // ~(A&B) & (A|B) -> A^B
3395 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3396 if ((A == C && B == D) || (A == D && B == C))
3397 return BinaryOperator::createXor(A, B);
3401 if (Op0->hasOneUse() &&
3402 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3403 if (A == Op1) { // (A^B)&A -> A&(A^B)
3404 I.swapOperands(); // Simplify below
3405 std::swap(Op0, Op1);
3406 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3407 cast<BinaryOperator>(Op0)->swapOperands();
3408 I.swapOperands(); // Simplify below
3409 std::swap(Op0, Op1);
3412 if (Op1->hasOneUse() &&
3413 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3414 if (B == Op0) { // B&(A^B) -> B&(B^A)
3415 cast<BinaryOperator>(Op1)->swapOperands();
3418 if (A == Op0) { // A&(A^B) -> A & ~B
3419 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3420 InsertNewInstBefore(NotB, I);
3421 return BinaryOperator::createAnd(A, NotB);
3426 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3427 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3428 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3431 Value *LHSVal, *RHSVal;
3432 ConstantInt *LHSCst, *RHSCst;
3433 ICmpInst::Predicate LHSCC, RHSCC;
3434 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3435 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3436 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3437 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3438 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3439 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3440 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3441 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3442 // Ensure that the larger constant is on the RHS.
3443 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3444 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3445 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3446 ICmpInst *LHS = cast<ICmpInst>(Op0);
3447 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3448 std::swap(LHS, RHS);
3449 std::swap(LHSCst, RHSCst);
3450 std::swap(LHSCC, RHSCC);
3453 // At this point, we know we have have two icmp instructions
3454 // comparing a value against two constants and and'ing the result
3455 // together. Because of the above check, we know that we only have
3456 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3457 // (from the FoldICmpLogical check above), that the two constants
3458 // are not equal and that the larger constant is on the RHS
3459 assert(LHSCst != RHSCst && "Compares not folded above?");
3462 default: assert(0 && "Unknown integer condition code!");
3463 case ICmpInst::ICMP_EQ:
3465 default: assert(0 && "Unknown integer condition code!");
3466 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3467 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3468 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3469 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3470 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3471 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3472 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3473 return ReplaceInstUsesWith(I, LHS);
3475 case ICmpInst::ICMP_NE:
3477 default: assert(0 && "Unknown integer condition code!");
3478 case ICmpInst::ICMP_ULT:
3479 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3480 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3481 break; // (X != 13 & X u< 15) -> no change
3482 case ICmpInst::ICMP_SLT:
3483 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3484 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3485 break; // (X != 13 & X s< 15) -> no change
3486 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3487 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3488 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3489 return ReplaceInstUsesWith(I, RHS);
3490 case ICmpInst::ICMP_NE:
3491 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3492 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3493 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3494 LHSVal->getName()+".off");
3495 InsertNewInstBefore(Add, I);
3496 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3497 ConstantInt::get(Add->getType(), 1));
3499 break; // (X != 13 & X != 15) -> no change
3502 case ICmpInst::ICMP_ULT:
3504 default: assert(0 && "Unknown integer condition code!");
3505 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3506 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3507 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3508 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3510 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3511 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3512 return ReplaceInstUsesWith(I, LHS);
3513 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3517 case ICmpInst::ICMP_SLT:
3519 default: assert(0 && "Unknown integer condition code!");
3520 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3521 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3522 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3523 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3525 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3526 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3527 return ReplaceInstUsesWith(I, LHS);
3528 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3532 case ICmpInst::ICMP_UGT:
3534 default: assert(0 && "Unknown integer condition code!");
3535 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3536 return ReplaceInstUsesWith(I, LHS);
3537 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3538 return ReplaceInstUsesWith(I, RHS);
3539 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3541 case ICmpInst::ICMP_NE:
3542 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3543 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3544 break; // (X u> 13 & X != 15) -> no change
3545 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3546 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3548 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3552 case ICmpInst::ICMP_SGT:
3554 default: assert(0 && "Unknown integer condition code!");
3555 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3556 return ReplaceInstUsesWith(I, LHS);
3557 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3558 return ReplaceInstUsesWith(I, RHS);
3559 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3561 case ICmpInst::ICMP_NE:
3562 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3563 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3564 break; // (X s> 13 & X != 15) -> no change
3565 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3566 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3568 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3576 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3577 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3578 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3579 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3580 const Type *SrcTy = Op0C->getOperand(0)->getType();
3581 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3582 // Only do this if the casts both really cause code to be generated.
3583 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3585 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3587 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3588 Op1C->getOperand(0),
3590 InsertNewInstBefore(NewOp, I);
3591 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3595 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3596 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3597 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3598 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3599 SI0->getOperand(1) == SI1->getOperand(1) &&
3600 (SI0->hasOneUse() || SI1->hasOneUse())) {
3601 Instruction *NewOp =
3602 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3604 SI0->getName()), I);
3605 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3606 SI1->getOperand(1));
3610 return Changed ? &I : 0;
3613 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3614 /// in the result. If it does, and if the specified byte hasn't been filled in
3615 /// yet, fill it in and return false.
3616 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3617 Instruction *I = dyn_cast<Instruction>(V);
3618 if (I == 0) return true;
3620 // If this is an or instruction, it is an inner node of the bswap.
3621 if (I->getOpcode() == Instruction::Or)
3622 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3623 CollectBSwapParts(I->getOperand(1), ByteValues);
3625 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3626 // If this is a shift by a constant int, and it is "24", then its operand
3627 // defines a byte. We only handle unsigned types here.
3628 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3629 // Not shifting the entire input by N-1 bytes?
3630 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3631 8*(ByteValues.size()-1))
3635 if (I->getOpcode() == Instruction::Shl) {
3636 // X << 24 defines the top byte with the lowest of the input bytes.
3637 DestNo = ByteValues.size()-1;
3639 // X >>u 24 defines the low byte with the highest of the input bytes.
3643 // If the destination byte value is already defined, the values are or'd
3644 // together, which isn't a bswap (unless it's an or of the same bits).
3645 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3647 ByteValues[DestNo] = I->getOperand(0);
3651 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3653 Value *Shift = 0, *ShiftLHS = 0;
3654 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3655 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3656 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3658 Instruction *SI = cast<Instruction>(Shift);
3660 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3661 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3662 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3665 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3667 if (AndAmt->getValue().getActiveBits() > 64)
3669 uint64_t AndAmtVal = AndAmt->getZExtValue();
3670 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3671 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3673 // Unknown mask for bswap.
3674 if (DestByte == ByteValues.size()) return true;
3676 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3678 if (SI->getOpcode() == Instruction::Shl)
3679 SrcByte = DestByte - ShiftBytes;
3681 SrcByte = DestByte + ShiftBytes;
3683 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3684 if (SrcByte != ByteValues.size()-DestByte-1)
3687 // If the destination byte value is already defined, the values are or'd
3688 // together, which isn't a bswap (unless it's an or of the same bits).
3689 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3691 ByteValues[DestByte] = SI->getOperand(0);
3695 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3696 /// If so, insert the new bswap intrinsic and return it.
3697 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3698 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3699 if (!ITy || ITy->getBitWidth() % 16)
3700 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3702 /// ByteValues - For each byte of the result, we keep track of which value
3703 /// defines each byte.
3704 SmallVector<Value*, 8> ByteValues;
3705 ByteValues.resize(ITy->getBitWidth()/8);
3707 // Try to find all the pieces corresponding to the bswap.
3708 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3709 CollectBSwapParts(I.getOperand(1), ByteValues))
3712 // Check to see if all of the bytes come from the same value.
3713 Value *V = ByteValues[0];
3714 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3716 // Check to make sure that all of the bytes come from the same value.
3717 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3718 if (ByteValues[i] != V)
3720 const Type *Tys[] = { ITy, ITy };
3721 Module *M = I.getParent()->getParent()->getParent();
3722 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 2);
3723 return new CallInst(F, V);
3727 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3728 bool Changed = SimplifyCommutative(I);
3729 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3731 if (isa<UndefValue>(Op1)) // X | undef -> -1
3732 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3736 return ReplaceInstUsesWith(I, Op0);
3738 // See if we can simplify any instructions used by the instruction whose sole
3739 // purpose is to compute bits we don't care about.
3740 if (!isa<VectorType>(I.getType())) {
3741 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3742 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3743 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3744 KnownZero, KnownOne))
3746 } else if (isa<ConstantAggregateZero>(Op1)) {
3747 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3748 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3749 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3750 return ReplaceInstUsesWith(I, I.getOperand(1));
3756 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3757 ConstantInt *C1 = 0; Value *X = 0;
3758 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3759 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3760 Instruction *Or = BinaryOperator::createOr(X, RHS);
3761 InsertNewInstBefore(Or, I);
3763 return BinaryOperator::createAnd(Or,
3764 ConstantInt::get(RHS->getValue() | C1->getValue()));
3767 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3768 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3769 Instruction *Or = BinaryOperator::createOr(X, RHS);
3770 InsertNewInstBefore(Or, I);
3772 return BinaryOperator::createXor(Or,
3773 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3776 // Try to fold constant and into select arguments.
3777 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3778 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3780 if (isa<PHINode>(Op0))
3781 if (Instruction *NV = FoldOpIntoPhi(I))
3785 Value *A = 0, *B = 0;
3786 ConstantInt *C1 = 0, *C2 = 0;
3788 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3789 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3790 return ReplaceInstUsesWith(I, Op1);
3791 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3792 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3793 return ReplaceInstUsesWith(I, Op0);
3795 // (A | B) | C and A | (B | C) -> bswap if possible.
3796 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3797 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3798 match(Op1, m_Or(m_Value(), m_Value())) ||
3799 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3800 match(Op1, m_Shift(m_Value(), m_Value())))) {
3801 if (Instruction *BSwap = MatchBSwap(I))
3805 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3806 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3807 MaskedValueIsZero(Op1, C1->getValue())) {
3808 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3809 InsertNewInstBefore(NOr, I);
3811 return BinaryOperator::createXor(NOr, C1);
3814 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3815 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3816 MaskedValueIsZero(Op0, C1->getValue())) {
3817 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3818 InsertNewInstBefore(NOr, I);
3820 return BinaryOperator::createXor(NOr, C1);
3824 Value *C = 0, *D = 0;
3825 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3826 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3827 Value *V1 = 0, *V2 = 0, *V3 = 0;
3828 C1 = dyn_cast<ConstantInt>(C);
3829 C2 = dyn_cast<ConstantInt>(D);
3830 if (C1 && C2) { // (A & C1)|(B & C2)
3831 // If we have: ((V + N) & C1) | (V & C2)
3832 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3833 // replace with V+N.
3834 if (C1->getValue() == ~C2->getValue()) {
3835 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3836 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3837 // Add commutes, try both ways.
3838 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3839 return ReplaceInstUsesWith(I, A);
3840 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3841 return ReplaceInstUsesWith(I, A);
3843 // Or commutes, try both ways.
3844 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3845 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3846 // Add commutes, try both ways.
3847 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3848 return ReplaceInstUsesWith(I, B);
3849 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3850 return ReplaceInstUsesWith(I, B);
3853 V1 = 0; V2 = 0; V3 = 0;
3856 // Check to see if we have any common things being and'ed. If so, find the
3857 // terms for V1 & (V2|V3).
3858 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3859 if (A == B) // (A & C)|(A & D) == A & (C|D)
3860 V1 = A, V2 = C, V3 = D;
3861 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3862 V1 = A, V2 = B, V3 = C;
3863 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3864 V1 = C, V2 = A, V3 = D;
3865 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3866 V1 = C, V2 = A, V3 = B;
3870 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3871 return BinaryOperator::createAnd(V1, Or);
3874 // (V1 & V3)|(V2 & ~V3) -> ((V1 ^ V2) & V3) ^ V2
3875 if (isOnlyUse(Op0) && isOnlyUse(Op1)) {
3876 // Try all combination of terms to find V3 and ~V3.
3877 if (A->hasOneUse() && match(A, m_Not(m_Value(V3)))) {
3883 if (B->hasOneUse() && match(B, m_Not(m_Value(V3)))) {
3889 if (C->hasOneUse() && match(C, m_Not(m_Value(V3)))) {
3895 if (D->hasOneUse() && match(D, m_Not(m_Value(V3)))) {
3902 A = InsertNewInstBefore(BinaryOperator::createXor(V1, V2, "tmp"), I);
3903 A = InsertNewInstBefore(BinaryOperator::createAnd(A, V3, "tmp"), I);
3904 return BinaryOperator::createXor(A, V2);
3910 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3911 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3912 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3913 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3914 SI0->getOperand(1) == SI1->getOperand(1) &&
3915 (SI0->hasOneUse() || SI1->hasOneUse())) {
3916 Instruction *NewOp =
3917 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3919 SI0->getName()), I);
3920 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3921 SI1->getOperand(1));
3925 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3926 if (A == Op1) // ~A | A == -1
3927 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3931 // Note, A is still live here!
3932 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3934 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3936 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3937 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3938 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3939 I.getName()+".demorgan"), I);
3940 return BinaryOperator::createNot(And);
3944 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3945 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3946 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3949 Value *LHSVal, *RHSVal;
3950 ConstantInt *LHSCst, *RHSCst;
3951 ICmpInst::Predicate LHSCC, RHSCC;
3952 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3953 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3954 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3955 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3956 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3957 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3958 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3959 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3960 // We can't fold (ugt x, C) | (sgt x, C2).
3961 PredicatesFoldable(LHSCC, RHSCC)) {
3962 // Ensure that the larger constant is on the RHS.
3963 ICmpInst *LHS = cast<ICmpInst>(Op0);
3965 if (ICmpInst::isSignedPredicate(LHSCC))
3966 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3968 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3971 std::swap(LHS, RHS);
3972 std::swap(LHSCst, RHSCst);
3973 std::swap(LHSCC, RHSCC);
3976 // At this point, we know we have have two icmp instructions
3977 // comparing a value against two constants and or'ing the result
3978 // together. Because of the above check, we know that we only have
3979 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3980 // FoldICmpLogical check above), that the two constants are not
3982 assert(LHSCst != RHSCst && "Compares not folded above?");
3985 default: assert(0 && "Unknown integer condition code!");
3986 case ICmpInst::ICMP_EQ:
3988 default: assert(0 && "Unknown integer condition code!");
3989 case ICmpInst::ICMP_EQ:
3990 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3991 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3992 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3993 LHSVal->getName()+".off");
3994 InsertNewInstBefore(Add, I);
3995 AddCST = Subtract(AddOne(RHSCst), LHSCst);
3996 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3998 break; // (X == 13 | X == 15) -> no change
3999 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
4000 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
4002 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
4003 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
4004 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
4005 return ReplaceInstUsesWith(I, RHS);
4008 case ICmpInst::ICMP_NE:
4010 default: assert(0 && "Unknown integer condition code!");
4011 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
4012 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
4013 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
4014 return ReplaceInstUsesWith(I, LHS);
4015 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
4016 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
4017 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
4018 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4021 case ICmpInst::ICMP_ULT:
4023 default: assert(0 && "Unknown integer condition code!");
4024 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
4026 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
4027 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4029 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4031 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4032 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4033 return ReplaceInstUsesWith(I, RHS);
4034 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4038 case ICmpInst::ICMP_SLT:
4040 default: assert(0 && "Unknown integer condition code!");
4041 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4043 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4044 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4046 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4048 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4049 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4050 return ReplaceInstUsesWith(I, RHS);
4051 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4055 case ICmpInst::ICMP_UGT:
4057 default: assert(0 && "Unknown integer condition code!");
4058 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4059 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4060 return ReplaceInstUsesWith(I, LHS);
4061 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4063 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4064 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4065 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4066 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4070 case ICmpInst::ICMP_SGT:
4072 default: assert(0 && "Unknown integer condition code!");
4073 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4074 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4075 return ReplaceInstUsesWith(I, LHS);
4076 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4078 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4079 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4080 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4081 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4089 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4090 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4091 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4092 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4093 const Type *SrcTy = Op0C->getOperand(0)->getType();
4094 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4095 // Only do this if the casts both really cause code to be generated.
4096 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4098 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4100 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4101 Op1C->getOperand(0),
4103 InsertNewInstBefore(NewOp, I);
4104 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4109 return Changed ? &I : 0;
4112 // XorSelf - Implements: X ^ X --> 0
4115 XorSelf(Value *rhs) : RHS(rhs) {}
4116 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4117 Instruction *apply(BinaryOperator &Xor) const {
4123 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4124 bool Changed = SimplifyCommutative(I);
4125 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4127 if (isa<UndefValue>(Op1))
4128 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4130 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4131 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4132 assert(Result == &I && "AssociativeOpt didn't work?");
4133 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4136 // See if we can simplify any instructions used by the instruction whose sole
4137 // purpose is to compute bits we don't care about.
4138 if (!isa<VectorType>(I.getType())) {
4139 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4140 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4141 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4142 KnownZero, KnownOne))
4144 } else if (isa<ConstantAggregateZero>(Op1)) {
4145 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4148 // Is this a ~ operation?
4149 if (Value *NotOp = dyn_castNotVal(&I)) {
4150 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4151 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4152 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4153 if (Op0I->getOpcode() == Instruction::And ||
4154 Op0I->getOpcode() == Instruction::Or) {
4155 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4156 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4158 BinaryOperator::createNot(Op0I->getOperand(1),
4159 Op0I->getOperand(1)->getName()+".not");
4160 InsertNewInstBefore(NotY, I);
4161 if (Op0I->getOpcode() == Instruction::And)
4162 return BinaryOperator::createOr(Op0NotVal, NotY);
4164 return BinaryOperator::createAnd(Op0NotVal, NotY);
4171 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4172 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
4173 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4174 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
4175 return new ICmpInst(ICI->getInversePredicate(),
4176 ICI->getOperand(0), ICI->getOperand(1));
4178 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4179 // ~(c-X) == X-c-1 == X+(-c-1)
4180 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4181 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4182 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4183 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4184 ConstantInt::get(I.getType(), 1));
4185 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4188 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4189 if (Op0I->getOpcode() == Instruction::Add) {
4190 // ~(X-c) --> (-c-1)-X
4191 if (RHS->isAllOnesValue()) {
4192 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4193 return BinaryOperator::createSub(
4194 ConstantExpr::getSub(NegOp0CI,
4195 ConstantInt::get(I.getType(), 1)),
4196 Op0I->getOperand(0));
4197 } else if (RHS->getValue().isSignBit()) {
4198 // (X + C) ^ signbit -> (X + C + signbit)
4199 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4200 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4203 } else if (Op0I->getOpcode() == Instruction::Or) {
4204 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4205 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4206 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4207 // Anything in both C1 and C2 is known to be zero, remove it from
4209 Constant *CommonBits = And(Op0CI, RHS);
4210 NewRHS = ConstantExpr::getAnd(NewRHS,
4211 ConstantExpr::getNot(CommonBits));
4212 AddToWorkList(Op0I);
4213 I.setOperand(0, Op0I->getOperand(0));
4214 I.setOperand(1, NewRHS);
4220 // Try to fold constant and into select arguments.
4221 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4222 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4224 if (isa<PHINode>(Op0))
4225 if (Instruction *NV = FoldOpIntoPhi(I))
4229 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4231 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4233 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4235 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4238 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4241 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4242 if (A == Op0) { // B^(B|A) == (A|B)^B
4243 Op1I->swapOperands();
4245 std::swap(Op0, Op1);
4246 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4247 I.swapOperands(); // Simplified below.
4248 std::swap(Op0, Op1);
4250 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4251 if (Op0 == A) // A^(A^B) == B
4252 return ReplaceInstUsesWith(I, B);
4253 else if (Op0 == B) // A^(B^A) == B
4254 return ReplaceInstUsesWith(I, A);
4255 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4256 if (A == Op0) { // A^(A&B) -> A^(B&A)
4257 Op1I->swapOperands();
4260 if (B == Op0) { // A^(B&A) -> (B&A)^A
4261 I.swapOperands(); // Simplified below.
4262 std::swap(Op0, Op1);
4267 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4270 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4271 if (A == Op1) // (B|A)^B == (A|B)^B
4273 if (B == Op1) { // (A|B)^B == A & ~B
4275 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4276 return BinaryOperator::createAnd(A, NotB);
4278 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4279 if (Op1 == A) // (A^B)^A == B
4280 return ReplaceInstUsesWith(I, B);
4281 else if (Op1 == B) // (B^A)^A == B
4282 return ReplaceInstUsesWith(I, A);
4283 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4284 if (A == Op1) // (A&B)^A -> (B&A)^A
4286 if (B == Op1 && // (B&A)^A == ~B & A
4287 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4289 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4290 return BinaryOperator::createAnd(N, Op1);
4295 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4296 if (Op0I && Op1I && Op0I->isShift() &&
4297 Op0I->getOpcode() == Op1I->getOpcode() &&
4298 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4299 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4300 Instruction *NewOp =
4301 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4302 Op1I->getOperand(0),
4303 Op0I->getName()), I);
4304 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4305 Op1I->getOperand(1));
4309 Value *A, *B, *C, *D;
4310 // (A & B)^(A | B) -> A ^ B
4311 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4312 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4313 if ((A == C && B == D) || (A == D && B == C))
4314 return BinaryOperator::createXor(A, B);
4316 // (A | B)^(A & B) -> A ^ B
4317 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4318 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4319 if ((A == C && B == D) || (A == D && B == C))
4320 return BinaryOperator::createXor(A, B);
4324 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4325 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4326 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4327 // (X & Y)^(X & Y) -> (Y^Z) & X
4328 Value *X = 0, *Y = 0, *Z = 0;
4330 X = A, Y = B, Z = D;
4332 X = A, Y = B, Z = C;
4334 X = B, Y = A, Z = D;
4336 X = B, Y = A, Z = C;
4339 Instruction *NewOp =
4340 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4341 return BinaryOperator::createAnd(NewOp, X);
4346 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4347 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4348 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4351 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4352 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4353 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4354 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4355 const Type *SrcTy = Op0C->getOperand(0)->getType();
4356 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4357 // Only do this if the casts both really cause code to be generated.
4358 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4360 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4362 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4363 Op1C->getOperand(0),
4365 InsertNewInstBefore(NewOp, I);
4366 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4370 return Changed ? &I : 0;
4373 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4374 /// overflowed for this type.
4375 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4376 ConstantInt *In2, bool IsSigned = false) {
4377 Result = cast<ConstantInt>(Add(In1, In2));
4380 if (In2->getValue().isNegative())
4381 return Result->getValue().sgt(In1->getValue());
4383 return Result->getValue().slt(In1->getValue());
4385 return Result->getValue().ult(In1->getValue());
4388 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4389 /// code necessary to compute the offset from the base pointer (without adding
4390 /// in the base pointer). Return the result as a signed integer of intptr size.
4391 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4392 TargetData &TD = IC.getTargetData();
4393 gep_type_iterator GTI = gep_type_begin(GEP);
4394 const Type *IntPtrTy = TD.getIntPtrType();
4395 Value *Result = Constant::getNullValue(IntPtrTy);
4397 // Build a mask for high order bits.
4398 unsigned IntPtrWidth = TD.getPointerSize()*8;
4399 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4401 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4402 Value *Op = GEP->getOperand(i);
4403 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4404 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4405 if (OpC->isZero()) continue;
4407 // Handle a struct index, which adds its field offset to the pointer.
4408 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4409 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4411 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4412 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4414 Result = IC.InsertNewInstBefore(
4415 BinaryOperator::createAdd(Result,
4416 ConstantInt::get(IntPtrTy, Size),
4417 GEP->getName()+".offs"), I);
4421 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4422 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4423 Scale = ConstantExpr::getMul(OC, Scale);
4424 if (Constant *RC = dyn_cast<Constant>(Result))
4425 Result = ConstantExpr::getAdd(RC, Scale);
4427 // Emit an add instruction.
4428 Result = IC.InsertNewInstBefore(
4429 BinaryOperator::createAdd(Result, Scale,
4430 GEP->getName()+".offs"), I);
4434 // Convert to correct type.
4435 if (Op->getType() != IntPtrTy) {
4436 if (Constant *OpC = dyn_cast<Constant>(Op))
4437 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4439 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4440 Op->getName()+".c"), I);
4443 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4444 if (Constant *OpC = dyn_cast<Constant>(Op))
4445 Op = ConstantExpr::getMul(OpC, Scale);
4446 else // We'll let instcombine(mul) convert this to a shl if possible.
4447 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4448 GEP->getName()+".idx"), I);
4451 // Emit an add instruction.
4452 if (isa<Constant>(Op) && isa<Constant>(Result))
4453 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4454 cast<Constant>(Result));
4456 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4457 GEP->getName()+".offs"), I);
4462 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4463 /// else. At this point we know that the GEP is on the LHS of the comparison.
4464 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4465 ICmpInst::Predicate Cond,
4467 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4469 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4470 if (isa<PointerType>(CI->getOperand(0)->getType()))
4471 RHS = CI->getOperand(0);
4473 Value *PtrBase = GEPLHS->getOperand(0);
4474 if (PtrBase == RHS) {
4475 // As an optimization, we don't actually have to compute the actual value of
4476 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4477 // each index is zero or not.
4478 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4479 Instruction *InVal = 0;
4480 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4481 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4483 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4484 if (isa<UndefValue>(C)) // undef index -> undef.
4485 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4486 if (C->isNullValue())
4488 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4489 EmitIt = false; // This is indexing into a zero sized array?
4490 } else if (isa<ConstantInt>(C))
4491 return ReplaceInstUsesWith(I, // No comparison is needed here.
4492 ConstantInt::get(Type::Int1Ty,
4493 Cond == ICmpInst::ICMP_NE));
4498 new ICmpInst(Cond, GEPLHS->getOperand(i),
4499 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4503 InVal = InsertNewInstBefore(InVal, I);
4504 InsertNewInstBefore(Comp, I);
4505 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4506 InVal = BinaryOperator::createOr(InVal, Comp);
4507 else // True if all are equal
4508 InVal = BinaryOperator::createAnd(InVal, Comp);
4516 // No comparison is needed here, all indexes = 0
4517 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4518 Cond == ICmpInst::ICMP_EQ));
4521 // Only lower this if the icmp is the only user of the GEP or if we expect
4522 // the result to fold to a constant!
4523 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4524 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4525 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4526 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4527 Constant::getNullValue(Offset->getType()));
4529 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4530 // If the base pointers are different, but the indices are the same, just
4531 // compare the base pointer.
4532 if (PtrBase != GEPRHS->getOperand(0)) {
4533 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4534 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4535 GEPRHS->getOperand(0)->getType();
4537 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4538 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4539 IndicesTheSame = false;
4543 // If all indices are the same, just compare the base pointers.
4545 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4546 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4548 // Otherwise, the base pointers are different and the indices are
4549 // different, bail out.
4553 // If one of the GEPs has all zero indices, recurse.
4554 bool AllZeros = true;
4555 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4556 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4557 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4562 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4563 ICmpInst::getSwappedPredicate(Cond), I);
4565 // If the other GEP has all zero indices, recurse.
4567 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4568 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4569 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4574 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4576 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4577 // If the GEPs only differ by one index, compare it.
4578 unsigned NumDifferences = 0; // Keep track of # differences.
4579 unsigned DiffOperand = 0; // The operand that differs.
4580 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4581 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4582 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4583 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4584 // Irreconcilable differences.
4588 if (NumDifferences++) break;
4593 if (NumDifferences == 0) // SAME GEP?
4594 return ReplaceInstUsesWith(I, // No comparison is needed here.
4595 ConstantInt::get(Type::Int1Ty,
4596 Cond == ICmpInst::ICMP_EQ));
4597 else if (NumDifferences == 1) {
4598 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4599 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4600 // Make sure we do a signed comparison here.
4601 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4605 // Only lower this if the icmp is the only user of the GEP or if we expect
4606 // the result to fold to a constant!
4607 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4608 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4609 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4610 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4611 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4612 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4618 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4619 bool Changed = SimplifyCompare(I);
4620 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4622 // Fold trivial predicates.
4623 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4624 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4625 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4626 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4628 // Simplify 'fcmp pred X, X'
4630 switch (I.getPredicate()) {
4631 default: assert(0 && "Unknown predicate!");
4632 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4633 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4634 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4635 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4636 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4637 case FCmpInst::FCMP_OLT: // True if ordered and less than
4638 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4639 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4641 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4642 case FCmpInst::FCMP_ULT: // True if unordered or less than
4643 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4644 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4645 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4646 I.setPredicate(FCmpInst::FCMP_UNO);
4647 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4650 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4651 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4652 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4653 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4654 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4655 I.setPredicate(FCmpInst::FCMP_ORD);
4656 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4661 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4662 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4664 // Handle fcmp with constant RHS
4665 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4666 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4667 switch (LHSI->getOpcode()) {
4668 case Instruction::PHI:
4669 if (Instruction *NV = FoldOpIntoPhi(I))
4672 case Instruction::Select:
4673 // If either operand of the select is a constant, we can fold the
4674 // comparison into the select arms, which will cause one to be
4675 // constant folded and the select turned into a bitwise or.
4676 Value *Op1 = 0, *Op2 = 0;
4677 if (LHSI->hasOneUse()) {
4678 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4679 // Fold the known value into the constant operand.
4680 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4681 // Insert a new FCmp of the other select operand.
4682 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4683 LHSI->getOperand(2), RHSC,
4685 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4686 // Fold the known value into the constant operand.
4687 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4688 // Insert a new FCmp of the other select operand.
4689 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4690 LHSI->getOperand(1), RHSC,
4696 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4701 return Changed ? &I : 0;
4704 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4705 bool Changed = SimplifyCompare(I);
4706 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4707 const Type *Ty = Op0->getType();
4711 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4712 isTrueWhenEqual(I)));
4714 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4715 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4717 // icmp of GlobalValues can never equal each other as long as they aren't
4718 // external weak linkage type.
4719 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4720 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4721 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4722 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4723 !isTrueWhenEqual(I)));
4725 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4726 // addresses never equal each other! We already know that Op0 != Op1.
4727 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4728 isa<ConstantPointerNull>(Op0)) &&
4729 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4730 isa<ConstantPointerNull>(Op1)))
4731 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4732 !isTrueWhenEqual(I)));
4734 // icmp's with boolean values can always be turned into bitwise operations
4735 if (Ty == Type::Int1Ty) {
4736 switch (I.getPredicate()) {
4737 default: assert(0 && "Invalid icmp instruction!");
4738 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4739 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4740 InsertNewInstBefore(Xor, I);
4741 return BinaryOperator::createNot(Xor);
4743 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4744 return BinaryOperator::createXor(Op0, Op1);
4746 case ICmpInst::ICMP_UGT:
4747 case ICmpInst::ICMP_SGT:
4748 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4750 case ICmpInst::ICMP_ULT:
4751 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4752 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4753 InsertNewInstBefore(Not, I);
4754 return BinaryOperator::createAnd(Not, Op1);
4756 case ICmpInst::ICMP_UGE:
4757 case ICmpInst::ICMP_SGE:
4758 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4760 case ICmpInst::ICMP_ULE:
4761 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4762 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4763 InsertNewInstBefore(Not, I);
4764 return BinaryOperator::createOr(Not, Op1);
4769 // See if we are doing a comparison between a constant and an instruction that
4770 // can be folded into the comparison.
4771 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4772 switch (I.getPredicate()) {
4774 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4775 if (CI->isMinValue(false))
4776 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4777 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4778 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4779 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4780 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4781 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4782 if (CI->isMinValue(true))
4783 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4784 ConstantInt::getAllOnesValue(Op0->getType()));
4788 case ICmpInst::ICMP_SLT:
4789 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4790 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4791 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4792 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4793 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4794 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4797 case ICmpInst::ICMP_UGT:
4798 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4799 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4800 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4801 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4802 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4803 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4805 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4806 if (CI->isMaxValue(true))
4807 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4808 ConstantInt::getNullValue(Op0->getType()));
4811 case ICmpInst::ICMP_SGT:
4812 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4813 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4814 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4815 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4816 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4817 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4820 case ICmpInst::ICMP_ULE:
4821 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4822 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4823 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4824 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4825 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4826 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4829 case ICmpInst::ICMP_SLE:
4830 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4831 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4832 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4833 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4834 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4835 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4838 case ICmpInst::ICMP_UGE:
4839 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4840 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4841 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4842 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4843 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4844 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4847 case ICmpInst::ICMP_SGE:
4848 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4849 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4850 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4851 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4852 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4853 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4857 // If we still have a icmp le or icmp ge instruction, turn it into the
4858 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4859 // already been handled above, this requires little checking.
4861 switch (I.getPredicate()) {
4863 case ICmpInst::ICMP_ULE:
4864 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4865 case ICmpInst::ICMP_SLE:
4866 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4867 case ICmpInst::ICMP_UGE:
4868 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4869 case ICmpInst::ICMP_SGE:
4870 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4873 // See if we can fold the comparison based on bits known to be zero or one
4874 // in the input. If this comparison is a normal comparison, it demands all
4875 // bits, if it is a sign bit comparison, it only demands the sign bit.
4878 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4880 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4881 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4882 if (SimplifyDemandedBits(Op0,
4883 isSignBit ? APInt::getSignBit(BitWidth)
4884 : APInt::getAllOnesValue(BitWidth),
4885 KnownZero, KnownOne, 0))
4888 // Given the known and unknown bits, compute a range that the LHS could be
4890 if ((KnownOne | KnownZero) != 0) {
4891 // Compute the Min, Max and RHS values based on the known bits. For the
4892 // EQ and NE we use unsigned values.
4893 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4894 const APInt& RHSVal = CI->getValue();
4895 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4896 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4899 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4902 switch (I.getPredicate()) { // LE/GE have been folded already.
4903 default: assert(0 && "Unknown icmp opcode!");
4904 case ICmpInst::ICMP_EQ:
4905 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4906 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4908 case ICmpInst::ICMP_NE:
4909 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4910 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4912 case ICmpInst::ICMP_ULT:
4913 if (Max.ult(RHSVal))
4914 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4915 if (Min.uge(RHSVal))
4916 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4918 case ICmpInst::ICMP_UGT:
4919 if (Min.ugt(RHSVal))
4920 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4921 if (Max.ule(RHSVal))
4922 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4924 case ICmpInst::ICMP_SLT:
4925 if (Max.slt(RHSVal))
4926 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4927 if (Min.sgt(RHSVal))
4928 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4930 case ICmpInst::ICMP_SGT:
4931 if (Min.sgt(RHSVal))
4932 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4933 if (Max.sle(RHSVal))
4934 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4939 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4940 // instruction, see if that instruction also has constants so that the
4941 // instruction can be folded into the icmp
4942 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4943 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4947 // Handle icmp with constant (but not simple integer constant) RHS
4948 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4949 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4950 switch (LHSI->getOpcode()) {
4951 case Instruction::GetElementPtr:
4952 if (RHSC->isNullValue()) {
4953 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4954 bool isAllZeros = true;
4955 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4956 if (!isa<Constant>(LHSI->getOperand(i)) ||
4957 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4962 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
4963 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4967 case Instruction::PHI:
4968 if (Instruction *NV = FoldOpIntoPhi(I))
4971 case Instruction::Select: {
4972 // If either operand of the select is a constant, we can fold the
4973 // comparison into the select arms, which will cause one to be
4974 // constant folded and the select turned into a bitwise or.
4975 Value *Op1 = 0, *Op2 = 0;
4976 if (LHSI->hasOneUse()) {
4977 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4978 // Fold the known value into the constant operand.
4979 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4980 // Insert a new ICmp of the other select operand.
4981 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4982 LHSI->getOperand(2), RHSC,
4984 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4985 // Fold the known value into the constant operand.
4986 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4987 // Insert a new ICmp of the other select operand.
4988 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4989 LHSI->getOperand(1), RHSC,
4995 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4998 case Instruction::Malloc:
4999 // If we have (malloc != null), and if the malloc has a single use, we
5000 // can assume it is successful and remove the malloc.
5001 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
5002 AddToWorkList(LHSI);
5003 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
5004 !isTrueWhenEqual(I)));
5010 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
5011 if (User *GEP = dyn_castGetElementPtr(Op0))
5012 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
5014 if (User *GEP = dyn_castGetElementPtr(Op1))
5015 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
5016 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
5019 // Test to see if the operands of the icmp are casted versions of other
5020 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
5022 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5023 if (isa<PointerType>(Op0->getType()) &&
5024 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5025 // We keep moving the cast from the left operand over to the right
5026 // operand, where it can often be eliminated completely.
5027 Op0 = CI->getOperand(0);
5029 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5030 // so eliminate it as well.
5031 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5032 Op1 = CI2->getOperand(0);
5034 // If Op1 is a constant, we can fold the cast into the constant.
5035 if (Op0->getType() != Op1->getType())
5036 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5037 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5039 // Otherwise, cast the RHS right before the icmp
5040 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5042 return new ICmpInst(I.getPredicate(), Op0, Op1);
5046 if (isa<CastInst>(Op0)) {
5047 // Handle the special case of: icmp (cast bool to X), <cst>
5048 // This comes up when you have code like
5051 // For generality, we handle any zero-extension of any operand comparison
5052 // with a constant or another cast from the same type.
5053 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5054 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5058 if (I.isEquality()) {
5059 Value *A, *B, *C, *D;
5060 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5061 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5062 Value *OtherVal = A == Op1 ? B : A;
5063 return new ICmpInst(I.getPredicate(), OtherVal,
5064 Constant::getNullValue(A->getType()));
5067 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5068 // A^c1 == C^c2 --> A == C^(c1^c2)
5069 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5070 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5071 if (Op1->hasOneUse()) {
5072 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5073 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5074 return new ICmpInst(I.getPredicate(), A,
5075 InsertNewInstBefore(Xor, I));
5078 // A^B == A^D -> B == D
5079 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5080 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5081 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5082 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5086 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5087 (A == Op0 || B == Op0)) {
5088 // A == (A^B) -> B == 0
5089 Value *OtherVal = A == Op0 ? B : A;
5090 return new ICmpInst(I.getPredicate(), OtherVal,
5091 Constant::getNullValue(A->getType()));
5093 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5094 // (A-B) == A -> B == 0
5095 return new ICmpInst(I.getPredicate(), B,
5096 Constant::getNullValue(B->getType()));
5098 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5099 // A == (A-B) -> B == 0
5100 return new ICmpInst(I.getPredicate(), B,
5101 Constant::getNullValue(B->getType()));
5104 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5105 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5106 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5107 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5108 Value *X = 0, *Y = 0, *Z = 0;
5111 X = B; Y = D; Z = A;
5112 } else if (A == D) {
5113 X = B; Y = C; Z = A;
5114 } else if (B == C) {
5115 X = A; Y = D; Z = B;
5116 } else if (B == D) {
5117 X = A; Y = C; Z = B;
5120 if (X) { // Build (X^Y) & Z
5121 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5122 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5123 I.setOperand(0, Op1);
5124 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5129 return Changed ? &I : 0;
5133 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5134 /// and CmpRHS are both known to be integer constants.
5135 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5136 ConstantInt *DivRHS) {
5137 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5138 const APInt &CmpRHSV = CmpRHS->getValue();
5140 // FIXME: If the operand types don't match the type of the divide
5141 // then don't attempt this transform. The code below doesn't have the
5142 // logic to deal with a signed divide and an unsigned compare (and
5143 // vice versa). This is because (x /s C1) <s C2 produces different
5144 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5145 // (x /u C1) <u C2. Simply casting the operands and result won't
5146 // work. :( The if statement below tests that condition and bails
5148 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5149 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5151 if (DivRHS->isZero())
5152 return 0; // The ProdOV computation fails on divide by zero.
5154 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5155 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5156 // C2 (CI). By solving for X we can turn this into a range check
5157 // instead of computing a divide.
5158 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5160 // Determine if the product overflows by seeing if the product is
5161 // not equal to the divide. Make sure we do the same kind of divide
5162 // as in the LHS instruction that we're folding.
5163 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5164 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5166 // Get the ICmp opcode
5167 ICmpInst::Predicate Pred = ICI.getPredicate();
5169 // Figure out the interval that is being checked. For example, a comparison
5170 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5171 // Compute this interval based on the constants involved and the signedness of
5172 // the compare/divide. This computes a half-open interval, keeping track of
5173 // whether either value in the interval overflows. After analysis each
5174 // overflow variable is set to 0 if it's corresponding bound variable is valid
5175 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5176 int LoOverflow = 0, HiOverflow = 0;
5177 ConstantInt *LoBound = 0, *HiBound = 0;
5180 if (!DivIsSigned) { // udiv
5181 // e.g. X/5 op 3 --> [15, 20)
5183 HiOverflow = LoOverflow = ProdOV;
5185 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5186 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5187 if (CmpRHSV == 0) { // (X / pos) op 0
5188 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5189 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5191 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5192 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5193 HiOverflow = LoOverflow = ProdOV;
5195 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5196 } else { // (X / pos) op neg
5197 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5198 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5199 LoOverflow = AddWithOverflow(LoBound, Prod,
5200 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5201 HiBound = AddOne(Prod);
5202 HiOverflow = ProdOV ? -1 : 0;
5204 } else { // Divisor is < 0.
5205 if (CmpRHSV == 0) { // (X / neg) op 0
5206 // e.g. X/-5 op 0 --> [-4, 5)
5207 LoBound = AddOne(DivRHS);
5208 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5209 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5210 HiOverflow = 1; // [INTMIN+1, overflow)
5211 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5213 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5214 // e.g. X/-5 op 3 --> [-19, -14)
5215 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5217 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5218 HiBound = AddOne(Prod);
5219 } else { // (X / neg) op neg
5220 // e.g. X/-5 op -3 --> [15, 20)
5222 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5223 HiBound = Subtract(Prod, DivRHS);
5226 // Dividing by a negative swaps the condition. LT <-> GT
5227 Pred = ICmpInst::getSwappedPredicate(Pred);
5230 Value *X = DivI->getOperand(0);
5232 default: assert(0 && "Unhandled icmp opcode!");
5233 case ICmpInst::ICMP_EQ:
5234 if (LoOverflow && HiOverflow)
5235 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5236 else if (HiOverflow)
5237 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5238 ICmpInst::ICMP_UGE, X, LoBound);
5239 else if (LoOverflow)
5240 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5241 ICmpInst::ICMP_ULT, X, HiBound);
5243 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5244 case ICmpInst::ICMP_NE:
5245 if (LoOverflow && HiOverflow)
5246 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5247 else if (HiOverflow)
5248 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5249 ICmpInst::ICMP_ULT, X, LoBound);
5250 else if (LoOverflow)
5251 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5252 ICmpInst::ICMP_UGE, X, HiBound);
5254 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5255 case ICmpInst::ICMP_ULT:
5256 case ICmpInst::ICMP_SLT:
5257 if (LoOverflow == +1) // Low bound is greater than input range.
5258 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5259 if (LoOverflow == -1) // Low bound is less than input range.
5260 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5261 return new ICmpInst(Pred, X, LoBound);
5262 case ICmpInst::ICMP_UGT:
5263 case ICmpInst::ICMP_SGT:
5264 if (HiOverflow == +1) // High bound greater than input range.
5265 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5266 else if (HiOverflow == -1) // High bound less than input range.
5267 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5268 if (Pred == ICmpInst::ICMP_UGT)
5269 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5271 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5276 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5278 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5281 const APInt &RHSV = RHS->getValue();
5283 switch (LHSI->getOpcode()) {
5284 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5285 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5286 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5288 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5289 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5290 Value *CompareVal = LHSI->getOperand(0);
5292 // If the sign bit of the XorCST is not set, there is no change to
5293 // the operation, just stop using the Xor.
5294 if (!XorCST->getValue().isNegative()) {
5295 ICI.setOperand(0, CompareVal);
5296 AddToWorkList(LHSI);
5300 // Was the old condition true if the operand is positive?
5301 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5303 // If so, the new one isn't.
5304 isTrueIfPositive ^= true;
5306 if (isTrueIfPositive)
5307 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5309 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5313 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5314 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5315 LHSI->getOperand(0)->hasOneUse()) {
5316 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5318 // If the LHS is an AND of a truncating cast, we can widen the
5319 // and/compare to be the input width without changing the value
5320 // produced, eliminating a cast.
5321 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5322 // We can do this transformation if either the AND constant does not
5323 // have its sign bit set or if it is an equality comparison.
5324 // Extending a relational comparison when we're checking the sign
5325 // bit would not work.
5326 if (Cast->hasOneUse() &&
5327 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5328 RHSV.isPositive())) {
5330 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5331 APInt NewCST = AndCST->getValue();
5332 NewCST.zext(BitWidth);
5334 NewCI.zext(BitWidth);
5335 Instruction *NewAnd =
5336 BinaryOperator::createAnd(Cast->getOperand(0),
5337 ConstantInt::get(NewCST),LHSI->getName());
5338 InsertNewInstBefore(NewAnd, ICI);
5339 return new ICmpInst(ICI.getPredicate(), NewAnd,
5340 ConstantInt::get(NewCI));
5344 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5345 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5346 // happens a LOT in code produced by the C front-end, for bitfield
5348 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5349 if (Shift && !Shift->isShift())
5353 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5354 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5355 const Type *AndTy = AndCST->getType(); // Type of the and.
5357 // We can fold this as long as we can't shift unknown bits
5358 // into the mask. This can only happen with signed shift
5359 // rights, as they sign-extend.
5361 bool CanFold = Shift->isLogicalShift();
5363 // To test for the bad case of the signed shr, see if any
5364 // of the bits shifted in could be tested after the mask.
5365 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5366 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5368 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5369 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5370 AndCST->getValue()) == 0)
5376 if (Shift->getOpcode() == Instruction::Shl)
5377 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5379 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5381 // Check to see if we are shifting out any of the bits being
5383 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5384 // If we shifted bits out, the fold is not going to work out.
5385 // As a special case, check to see if this means that the
5386 // result is always true or false now.
5387 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5388 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5389 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5390 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5392 ICI.setOperand(1, NewCst);
5393 Constant *NewAndCST;
5394 if (Shift->getOpcode() == Instruction::Shl)
5395 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5397 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5398 LHSI->setOperand(1, NewAndCST);
5399 LHSI->setOperand(0, Shift->getOperand(0));
5400 AddToWorkList(Shift); // Shift is dead.
5401 AddUsesToWorkList(ICI);
5407 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5408 // preferable because it allows the C<<Y expression to be hoisted out
5409 // of a loop if Y is invariant and X is not.
5410 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5411 ICI.isEquality() && !Shift->isArithmeticShift() &&
5412 isa<Instruction>(Shift->getOperand(0))) {
5415 if (Shift->getOpcode() == Instruction::LShr) {
5416 NS = BinaryOperator::createShl(AndCST,
5417 Shift->getOperand(1), "tmp");
5419 // Insert a logical shift.
5420 NS = BinaryOperator::createLShr(AndCST,
5421 Shift->getOperand(1), "tmp");
5423 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5425 // Compute X & (C << Y).
5426 Instruction *NewAnd =
5427 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5428 InsertNewInstBefore(NewAnd, ICI);
5430 ICI.setOperand(0, NewAnd);
5436 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5437 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5440 uint32_t TypeBits = RHSV.getBitWidth();
5442 // Check that the shift amount is in range. If not, don't perform
5443 // undefined shifts. When the shift is visited it will be
5445 if (ShAmt->uge(TypeBits))
5448 if (ICI.isEquality()) {
5449 // If we are comparing against bits always shifted out, the
5450 // comparison cannot succeed.
5452 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5453 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5454 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5455 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5456 return ReplaceInstUsesWith(ICI, Cst);
5459 if (LHSI->hasOneUse()) {
5460 // Otherwise strength reduce the shift into an and.
5461 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5463 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5466 BinaryOperator::createAnd(LHSI->getOperand(0),
5467 Mask, LHSI->getName()+".mask");
5468 Value *And = InsertNewInstBefore(AndI, ICI);
5469 return new ICmpInst(ICI.getPredicate(), And,
5470 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5474 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5475 bool TrueIfSigned = false;
5476 if (LHSI->hasOneUse() &&
5477 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5478 // (X << 31) <s 0 --> (X&1) != 0
5479 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5480 (TypeBits-ShAmt->getZExtValue()-1));
5482 BinaryOperator::createAnd(LHSI->getOperand(0),
5483 Mask, LHSI->getName()+".mask");
5484 Value *And = InsertNewInstBefore(AndI, ICI);
5486 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5487 And, Constant::getNullValue(And->getType()));
5492 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5493 case Instruction::AShr: {
5494 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5497 if (ICI.isEquality()) {
5498 // Check that the shift amount is in range. If not, don't perform
5499 // undefined shifts. When the shift is visited it will be
5501 uint32_t TypeBits = RHSV.getBitWidth();
5502 if (ShAmt->uge(TypeBits))
5504 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5506 // If we are comparing against bits always shifted out, the
5507 // comparison cannot succeed.
5508 APInt Comp = RHSV << ShAmtVal;
5509 if (LHSI->getOpcode() == Instruction::LShr)
5510 Comp = Comp.lshr(ShAmtVal);
5512 Comp = Comp.ashr(ShAmtVal);
5514 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5515 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5516 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5517 return ReplaceInstUsesWith(ICI, Cst);
5520 if (LHSI->hasOneUse() || RHSV == 0) {
5521 // Otherwise strength reduce the shift into an and.
5522 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5523 Constant *Mask = ConstantInt::get(Val);
5526 BinaryOperator::createAnd(LHSI->getOperand(0),
5527 Mask, LHSI->getName()+".mask");
5528 Value *And = InsertNewInstBefore(AndI, ICI);
5529 return new ICmpInst(ICI.getPredicate(), And,
5530 ConstantExpr::getShl(RHS, ShAmt));
5536 case Instruction::SDiv:
5537 case Instruction::UDiv:
5538 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5539 // Fold this div into the comparison, producing a range check.
5540 // Determine, based on the divide type, what the range is being
5541 // checked. If there is an overflow on the low or high side, remember
5542 // it, otherwise compute the range [low, hi) bounding the new value.
5543 // See: InsertRangeTest above for the kinds of replacements possible.
5544 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5545 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5551 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5552 if (ICI.isEquality()) {
5553 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5555 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5556 // the second operand is a constant, simplify a bit.
5557 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5558 switch (BO->getOpcode()) {
5559 case Instruction::SRem:
5560 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5561 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5562 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5563 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5564 Instruction *NewRem =
5565 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5567 InsertNewInstBefore(NewRem, ICI);
5568 return new ICmpInst(ICI.getPredicate(), NewRem,
5569 Constant::getNullValue(BO->getType()));
5573 case Instruction::Add:
5574 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5575 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5576 if (BO->hasOneUse())
5577 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5578 Subtract(RHS, BOp1C));
5579 } else if (RHSV == 0) {
5580 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5581 // efficiently invertible, or if the add has just this one use.
5582 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5584 if (Value *NegVal = dyn_castNegVal(BOp1))
5585 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5586 else if (Value *NegVal = dyn_castNegVal(BOp0))
5587 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5588 else if (BO->hasOneUse()) {
5589 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5590 InsertNewInstBefore(Neg, ICI);
5592 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5596 case Instruction::Xor:
5597 // For the xor case, we can xor two constants together, eliminating
5598 // the explicit xor.
5599 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5600 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5601 ConstantExpr::getXor(RHS, BOC));
5604 case Instruction::Sub:
5605 // Replace (([sub|xor] A, B) != 0) with (A != B)
5607 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5611 case Instruction::Or:
5612 // If bits are being or'd in that are not present in the constant we
5613 // are comparing against, then the comparison could never succeed!
5614 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5615 Constant *NotCI = ConstantExpr::getNot(RHS);
5616 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5617 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5622 case Instruction::And:
5623 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5624 // If bits are being compared against that are and'd out, then the
5625 // comparison can never succeed!
5626 if ((RHSV & ~BOC->getValue()) != 0)
5627 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5630 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5631 if (RHS == BOC && RHSV.isPowerOf2())
5632 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5633 ICmpInst::ICMP_NE, LHSI,
5634 Constant::getNullValue(RHS->getType()));
5636 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5637 if (isSignBit(BOC)) {
5638 Value *X = BO->getOperand(0);
5639 Constant *Zero = Constant::getNullValue(X->getType());
5640 ICmpInst::Predicate pred = isICMP_NE ?
5641 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5642 return new ICmpInst(pred, X, Zero);
5645 // ((X & ~7) == 0) --> X < 8
5646 if (RHSV == 0 && isHighOnes(BOC)) {
5647 Value *X = BO->getOperand(0);
5648 Constant *NegX = ConstantExpr::getNeg(BOC);
5649 ICmpInst::Predicate pred = isICMP_NE ?
5650 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5651 return new ICmpInst(pred, X, NegX);
5656 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5657 // Handle icmp {eq|ne} <intrinsic>, intcst.
5658 if (II->getIntrinsicID() == Intrinsic::bswap) {
5660 ICI.setOperand(0, II->getOperand(1));
5661 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5665 } else { // Not a ICMP_EQ/ICMP_NE
5666 // If the LHS is a cast from an integral value of the same size,
5667 // then since we know the RHS is a constant, try to simlify.
5668 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5669 Value *CastOp = Cast->getOperand(0);
5670 const Type *SrcTy = CastOp->getType();
5671 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5672 if (SrcTy->isInteger() &&
5673 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5674 // If this is an unsigned comparison, try to make the comparison use
5675 // smaller constant values.
5676 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5677 // X u< 128 => X s> -1
5678 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5679 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5680 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5681 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5682 // X u> 127 => X s< 0
5683 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5684 Constant::getNullValue(SrcTy));
5692 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5693 /// We only handle extending casts so far.
5695 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5696 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5697 Value *LHSCIOp = LHSCI->getOperand(0);
5698 const Type *SrcTy = LHSCIOp->getType();
5699 const Type *DestTy = LHSCI->getType();
5702 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5703 // integer type is the same size as the pointer type.
5704 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5705 getTargetData().getPointerSizeInBits() ==
5706 cast<IntegerType>(DestTy)->getBitWidth()) {
5708 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5709 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5710 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5711 RHSOp = RHSC->getOperand(0);
5712 // If the pointer types don't match, insert a bitcast.
5713 if (LHSCIOp->getType() != RHSOp->getType())
5714 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5715 LHSCIOp->getType(), ICI);
5719 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5722 // The code below only handles extension cast instructions, so far.
5724 if (LHSCI->getOpcode() != Instruction::ZExt &&
5725 LHSCI->getOpcode() != Instruction::SExt)
5728 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5729 bool isSignedCmp = ICI.isSignedPredicate();
5731 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5732 // Not an extension from the same type?
5733 RHSCIOp = CI->getOperand(0);
5734 if (RHSCIOp->getType() != LHSCIOp->getType())
5737 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5738 // and the other is a zext), then we can't handle this.
5739 if (CI->getOpcode() != LHSCI->getOpcode())
5742 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5743 // then we can't handle this.
5744 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5747 // Okay, just insert a compare of the reduced operands now!
5748 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5751 // If we aren't dealing with a constant on the RHS, exit early
5752 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5756 // Compute the constant that would happen if we truncated to SrcTy then
5757 // reextended to DestTy.
5758 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5759 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5761 // If the re-extended constant didn't change...
5763 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5764 // For example, we might have:
5765 // %A = sext short %X to uint
5766 // %B = icmp ugt uint %A, 1330
5767 // It is incorrect to transform this into
5768 // %B = icmp ugt short %X, 1330
5769 // because %A may have negative value.
5771 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5772 // OR operation is EQ/NE.
5773 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5774 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5779 // The re-extended constant changed so the constant cannot be represented
5780 // in the shorter type. Consequently, we cannot emit a simple comparison.
5782 // First, handle some easy cases. We know the result cannot be equal at this
5783 // point so handle the ICI.isEquality() cases
5784 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5785 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5786 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5787 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5789 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5790 // should have been folded away previously and not enter in here.
5793 // We're performing a signed comparison.
5794 if (cast<ConstantInt>(CI)->getValue().isNegative())
5795 Result = ConstantInt::getFalse(); // X < (small) --> false
5797 Result = ConstantInt::getTrue(); // X < (large) --> true
5799 // We're performing an unsigned comparison.
5801 // We're performing an unsigned comp with a sign extended value.
5802 // This is true if the input is >= 0. [aka >s -1]
5803 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5804 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5805 NegOne, ICI.getName()), ICI);
5807 // Unsigned extend & unsigned compare -> always true.
5808 Result = ConstantInt::getTrue();
5812 // Finally, return the value computed.
5813 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5814 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5815 return ReplaceInstUsesWith(ICI, Result);
5817 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5818 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5819 "ICmp should be folded!");
5820 if (Constant *CI = dyn_cast<Constant>(Result))
5821 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5823 return BinaryOperator::createNot(Result);
5827 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5828 return commonShiftTransforms(I);
5831 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5832 return commonShiftTransforms(I);
5835 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5836 return commonShiftTransforms(I);
5839 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5840 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5841 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5843 // shl X, 0 == X and shr X, 0 == X
5844 // shl 0, X == 0 and shr 0, X == 0
5845 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5846 Op0 == Constant::getNullValue(Op0->getType()))
5847 return ReplaceInstUsesWith(I, Op0);
5849 if (isa<UndefValue>(Op0)) {
5850 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5851 return ReplaceInstUsesWith(I, Op0);
5852 else // undef << X -> 0, undef >>u X -> 0
5853 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5855 if (isa<UndefValue>(Op1)) {
5856 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5857 return ReplaceInstUsesWith(I, Op0);
5858 else // X << undef, X >>u undef -> 0
5859 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5862 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5863 if (I.getOpcode() == Instruction::AShr)
5864 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5865 if (CSI->isAllOnesValue())
5866 return ReplaceInstUsesWith(I, CSI);
5868 // Try to fold constant and into select arguments.
5869 if (isa<Constant>(Op0))
5870 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5871 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5874 // See if we can turn a signed shr into an unsigned shr.
5875 if (I.isArithmeticShift()) {
5876 if (MaskedValueIsZero(Op0,
5877 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5878 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5882 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5883 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5888 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5889 BinaryOperator &I) {
5890 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5892 // See if we can simplify any instructions used by the instruction whose sole
5893 // purpose is to compute bits we don't care about.
5894 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5895 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5896 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5897 KnownZero, KnownOne))
5900 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5901 // of a signed value.
5903 if (Op1->uge(TypeBits)) {
5904 if (I.getOpcode() != Instruction::AShr)
5905 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5907 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5912 // ((X*C1) << C2) == (X * (C1 << C2))
5913 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5914 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5915 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5916 return BinaryOperator::createMul(BO->getOperand(0),
5917 ConstantExpr::getShl(BOOp, Op1));
5919 // Try to fold constant and into select arguments.
5920 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5921 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5923 if (isa<PHINode>(Op0))
5924 if (Instruction *NV = FoldOpIntoPhi(I))
5927 if (Op0->hasOneUse()) {
5928 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5929 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5932 switch (Op0BO->getOpcode()) {
5934 case Instruction::Add:
5935 case Instruction::And:
5936 case Instruction::Or:
5937 case Instruction::Xor: {
5938 // These operators commute.
5939 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5940 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5941 match(Op0BO->getOperand(1),
5942 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5943 Instruction *YS = BinaryOperator::createShl(
5944 Op0BO->getOperand(0), Op1,
5946 InsertNewInstBefore(YS, I); // (Y << C)
5948 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5949 Op0BO->getOperand(1)->getName());
5950 InsertNewInstBefore(X, I); // (X + (Y << C))
5951 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5952 return BinaryOperator::createAnd(X, ConstantInt::get(
5953 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5956 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5957 Value *Op0BOOp1 = Op0BO->getOperand(1);
5958 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5960 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5961 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5963 Instruction *YS = BinaryOperator::createShl(
5964 Op0BO->getOperand(0), Op1,
5966 InsertNewInstBefore(YS, I); // (Y << C)
5968 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5969 V1->getName()+".mask");
5970 InsertNewInstBefore(XM, I); // X & (CC << C)
5972 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5977 case Instruction::Sub: {
5978 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5979 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5980 match(Op0BO->getOperand(0),
5981 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5982 Instruction *YS = BinaryOperator::createShl(
5983 Op0BO->getOperand(1), Op1,
5985 InsertNewInstBefore(YS, I); // (Y << C)
5987 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5988 Op0BO->getOperand(0)->getName());
5989 InsertNewInstBefore(X, I); // (X + (Y << C))
5990 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5991 return BinaryOperator::createAnd(X, ConstantInt::get(
5992 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5995 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5996 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5997 match(Op0BO->getOperand(0),
5998 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5999 m_ConstantInt(CC))) && V2 == Op1 &&
6000 cast<BinaryOperator>(Op0BO->getOperand(0))
6001 ->getOperand(0)->hasOneUse()) {
6002 Instruction *YS = BinaryOperator::createShl(
6003 Op0BO->getOperand(1), Op1,
6005 InsertNewInstBefore(YS, I); // (Y << C)
6007 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
6008 V1->getName()+".mask");
6009 InsertNewInstBefore(XM, I); // X & (CC << C)
6011 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
6019 // If the operand is an bitwise operator with a constant RHS, and the
6020 // shift is the only use, we can pull it out of the shift.
6021 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
6022 bool isValid = true; // Valid only for And, Or, Xor
6023 bool highBitSet = false; // Transform if high bit of constant set?
6025 switch (Op0BO->getOpcode()) {
6026 default: isValid = false; break; // Do not perform transform!
6027 case Instruction::Add:
6028 isValid = isLeftShift;
6030 case Instruction::Or:
6031 case Instruction::Xor:
6034 case Instruction::And:
6039 // If this is a signed shift right, and the high bit is modified
6040 // by the logical operation, do not perform the transformation.
6041 // The highBitSet boolean indicates the value of the high bit of
6042 // the constant which would cause it to be modified for this
6045 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
6046 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6050 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6052 Instruction *NewShift =
6053 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6054 InsertNewInstBefore(NewShift, I);
6055 NewShift->takeName(Op0BO);
6057 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6064 // Find out if this is a shift of a shift by a constant.
6065 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6066 if (ShiftOp && !ShiftOp->isShift())
6069 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6070 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6071 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6072 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6073 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6074 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6075 Value *X = ShiftOp->getOperand(0);
6077 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6078 if (AmtSum > TypeBits)
6081 const IntegerType *Ty = cast<IntegerType>(I.getType());
6083 // Check for (X << c1) << c2 and (X >> c1) >> c2
6084 if (I.getOpcode() == ShiftOp->getOpcode()) {
6085 return BinaryOperator::create(I.getOpcode(), X,
6086 ConstantInt::get(Ty, AmtSum));
6087 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6088 I.getOpcode() == Instruction::AShr) {
6089 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6090 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6091 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6092 I.getOpcode() == Instruction::LShr) {
6093 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6094 Instruction *Shift =
6095 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6096 InsertNewInstBefore(Shift, I);
6098 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6099 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6102 // Okay, if we get here, one shift must be left, and the other shift must be
6103 // right. See if the amounts are equal.
6104 if (ShiftAmt1 == ShiftAmt2) {
6105 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6106 if (I.getOpcode() == Instruction::Shl) {
6107 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6108 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6110 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6111 if (I.getOpcode() == Instruction::LShr) {
6112 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6113 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6115 // We can simplify ((X << C) >>s C) into a trunc + sext.
6116 // NOTE: we could do this for any C, but that would make 'unusual' integer
6117 // types. For now, just stick to ones well-supported by the code
6119 const Type *SExtType = 0;
6120 switch (Ty->getBitWidth() - ShiftAmt1) {
6127 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6132 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6133 InsertNewInstBefore(NewTrunc, I);
6134 return new SExtInst(NewTrunc, Ty);
6136 // Otherwise, we can't handle it yet.
6137 } else if (ShiftAmt1 < ShiftAmt2) {
6138 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6140 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6141 if (I.getOpcode() == Instruction::Shl) {
6142 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6143 ShiftOp->getOpcode() == Instruction::AShr);
6144 Instruction *Shift =
6145 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6146 InsertNewInstBefore(Shift, I);
6148 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6149 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6152 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6153 if (I.getOpcode() == Instruction::LShr) {
6154 assert(ShiftOp->getOpcode() == Instruction::Shl);
6155 Instruction *Shift =
6156 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6157 InsertNewInstBefore(Shift, I);
6159 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6160 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6163 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6165 assert(ShiftAmt2 < ShiftAmt1);
6166 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6168 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6169 if (I.getOpcode() == Instruction::Shl) {
6170 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6171 ShiftOp->getOpcode() == Instruction::AShr);
6172 Instruction *Shift =
6173 BinaryOperator::create(ShiftOp->getOpcode(), X,
6174 ConstantInt::get(Ty, ShiftDiff));
6175 InsertNewInstBefore(Shift, I);
6177 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6178 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6181 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6182 if (I.getOpcode() == Instruction::LShr) {
6183 assert(ShiftOp->getOpcode() == Instruction::Shl);
6184 Instruction *Shift =
6185 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6186 InsertNewInstBefore(Shift, I);
6188 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6189 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6192 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6199 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6200 /// expression. If so, decompose it, returning some value X, such that Val is
6203 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6205 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6206 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6207 Offset = CI->getZExtValue();
6209 return ConstantInt::get(Type::Int32Ty, 0);
6210 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
6211 if (I->getNumOperands() == 2) {
6212 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6213 if (I->getOpcode() == Instruction::Shl) {
6214 // This is a value scaled by '1 << the shift amt'.
6215 Scale = 1U << CUI->getZExtValue();
6217 return I->getOperand(0);
6218 } else if (I->getOpcode() == Instruction::Mul) {
6219 // This value is scaled by 'CUI'.
6220 Scale = CUI->getZExtValue();
6222 return I->getOperand(0);
6223 } else if (I->getOpcode() == Instruction::Add) {
6224 // We have X+C. Check to see if we really have (X*C2)+C1,
6225 // where C1 is divisible by C2.
6228 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6229 Offset += CUI->getZExtValue();
6230 if (SubScale > 1 && (Offset % SubScale == 0)) {
6239 // Otherwise, we can't look past this.
6246 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6247 /// try to eliminate the cast by moving the type information into the alloc.
6248 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6249 AllocationInst &AI) {
6250 const PointerType *PTy = cast<PointerType>(CI.getType());
6252 // Remove any uses of AI that are dead.
6253 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6255 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6256 Instruction *User = cast<Instruction>(*UI++);
6257 if (isInstructionTriviallyDead(User)) {
6258 while (UI != E && *UI == User)
6259 ++UI; // If this instruction uses AI more than once, don't break UI.
6262 DOUT << "IC: DCE: " << *User;
6263 EraseInstFromFunction(*User);
6267 // Get the type really allocated and the type casted to.
6268 const Type *AllocElTy = AI.getAllocatedType();
6269 const Type *CastElTy = PTy->getElementType();
6270 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6272 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6273 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6274 if (CastElTyAlign < AllocElTyAlign) return 0;
6276 // If the allocation has multiple uses, only promote it if we are strictly
6277 // increasing the alignment of the resultant allocation. If we keep it the
6278 // same, we open the door to infinite loops of various kinds.
6279 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6281 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
6282 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
6283 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6285 // See if we can satisfy the modulus by pulling a scale out of the array
6287 unsigned ArraySizeScale;
6289 Value *NumElements = // See if the array size is a decomposable linear expr.
6290 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6292 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6294 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6295 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6297 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6302 // If the allocation size is constant, form a constant mul expression
6303 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6304 if (isa<ConstantInt>(NumElements))
6305 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6306 // otherwise multiply the amount and the number of elements
6307 else if (Scale != 1) {
6308 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6309 Amt = InsertNewInstBefore(Tmp, AI);
6313 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6314 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6315 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6316 Amt = InsertNewInstBefore(Tmp, AI);
6319 AllocationInst *New;
6320 if (isa<MallocInst>(AI))
6321 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6323 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6324 InsertNewInstBefore(New, AI);
6327 // If the allocation has multiple uses, insert a cast and change all things
6328 // that used it to use the new cast. This will also hack on CI, but it will
6330 if (!AI.hasOneUse()) {
6331 AddUsesToWorkList(AI);
6332 // New is the allocation instruction, pointer typed. AI is the original
6333 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6334 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6335 InsertNewInstBefore(NewCast, AI);
6336 AI.replaceAllUsesWith(NewCast);
6338 return ReplaceInstUsesWith(CI, New);
6341 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6342 /// and return it as type Ty without inserting any new casts and without
6343 /// changing the computed value. This is used by code that tries to decide
6344 /// whether promoting or shrinking integer operations to wider or smaller types
6345 /// will allow us to eliminate a truncate or extend.
6347 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6348 /// extension operation if Ty is larger.
6349 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6350 int &NumCastsRemoved) {
6351 // We can always evaluate constants in another type.
6352 if (isa<ConstantInt>(V))
6355 Instruction *I = dyn_cast<Instruction>(V);
6356 if (!I) return false;
6358 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6360 switch (I->getOpcode()) {
6361 case Instruction::Add:
6362 case Instruction::Sub:
6363 case Instruction::And:
6364 case Instruction::Or:
6365 case Instruction::Xor:
6366 if (!I->hasOneUse()) return false;
6367 // These operators can all arbitrarily be extended or truncated.
6368 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
6369 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
6371 case Instruction::Shl:
6372 if (!I->hasOneUse()) return false;
6373 // If we are truncating the result of this SHL, and if it's a shift of a
6374 // constant amount, we can always perform a SHL in a smaller type.
6375 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6376 uint32_t BitWidth = Ty->getBitWidth();
6377 if (BitWidth < OrigTy->getBitWidth() &&
6378 CI->getLimitedValue(BitWidth) < BitWidth)
6379 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6382 case Instruction::LShr:
6383 if (!I->hasOneUse()) return false;
6384 // If this is a truncate of a logical shr, we can truncate it to a smaller
6385 // lshr iff we know that the bits we would otherwise be shifting in are
6387 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6388 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6389 uint32_t BitWidth = Ty->getBitWidth();
6390 if (BitWidth < OrigBitWidth &&
6391 MaskedValueIsZero(I->getOperand(0),
6392 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6393 CI->getLimitedValue(BitWidth) < BitWidth) {
6394 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6398 case Instruction::Trunc:
6399 case Instruction::ZExt:
6400 case Instruction::SExt:
6401 // If this is a cast from the destination type, we can trivially eliminate
6402 // it, and this will remove a cast overall.
6403 if (I->getOperand(0)->getType() == Ty) {
6404 // If the first operand is itself a cast, and is eliminable, do not count
6405 // this as an eliminable cast. We would prefer to eliminate those two
6407 if (isa<CastInst>(I->getOperand(0)))
6415 // TODO: Can handle more cases here.
6422 /// EvaluateInDifferentType - Given an expression that
6423 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6424 /// evaluate the expression.
6425 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6427 if (Constant *C = dyn_cast<Constant>(V))
6428 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6430 // Otherwise, it must be an instruction.
6431 Instruction *I = cast<Instruction>(V);
6432 Instruction *Res = 0;
6433 switch (I->getOpcode()) {
6434 case Instruction::Add:
6435 case Instruction::Sub:
6436 case Instruction::And:
6437 case Instruction::Or:
6438 case Instruction::Xor:
6439 case Instruction::AShr:
6440 case Instruction::LShr:
6441 case Instruction::Shl: {
6442 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6443 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6444 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6445 LHS, RHS, I->getName());
6448 case Instruction::Trunc:
6449 case Instruction::ZExt:
6450 case Instruction::SExt:
6451 case Instruction::BitCast:
6452 // If the source type of the cast is the type we're trying for then we can
6453 // just return the source. There's no need to insert it because its not new.
6454 if (I->getOperand(0)->getType() == Ty)
6455 return I->getOperand(0);
6457 // Some other kind of cast, which shouldn't happen, so just ..
6460 // TODO: Can handle more cases here.
6461 assert(0 && "Unreachable!");
6465 return InsertNewInstBefore(Res, *I);
6468 /// @brief Implement the transforms common to all CastInst visitors.
6469 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6470 Value *Src = CI.getOperand(0);
6472 // Casting undef to anything results in undef so might as just replace it and
6473 // get rid of the cast.
6474 if (isa<UndefValue>(Src)) // cast undef -> undef
6475 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6477 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6478 // eliminate it now.
6479 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6480 if (Instruction::CastOps opc =
6481 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6482 // The first cast (CSrc) is eliminable so we need to fix up or replace
6483 // the second cast (CI). CSrc will then have a good chance of being dead.
6484 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6488 // If we are casting a select then fold the cast into the select
6489 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6490 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6493 // If we are casting a PHI then fold the cast into the PHI
6494 if (isa<PHINode>(Src))
6495 if (Instruction *NV = FoldOpIntoPhi(CI))
6501 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6502 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6503 Value *Src = CI.getOperand(0);
6505 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6506 // If casting the result of a getelementptr instruction with no offset, turn
6507 // this into a cast of the original pointer!
6508 if (GEP->hasAllZeroIndices()) {
6509 // Changing the cast operand is usually not a good idea but it is safe
6510 // here because the pointer operand is being replaced with another
6511 // pointer operand so the opcode doesn't need to change.
6513 CI.setOperand(0, GEP->getOperand(0));
6517 // If the GEP has a single use, and the base pointer is a bitcast, and the
6518 // GEP computes a constant offset, see if we can convert these three
6519 // instructions into fewer. This typically happens with unions and other
6520 // non-type-safe code.
6521 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6522 if (GEP->hasAllConstantIndices()) {
6523 // We are guaranteed to get a constant from EmitGEPOffset.
6524 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6525 int64_t Offset = OffsetV->getSExtValue();
6527 // Get the base pointer input of the bitcast, and the type it points to.
6528 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6529 const Type *GEPIdxTy =
6530 cast<PointerType>(OrigBase->getType())->getElementType();
6531 if (GEPIdxTy->isSized()) {
6532 SmallVector<Value*, 8> NewIndices;
6534 // Start with the index over the outer type. Note that the type size
6535 // might be zero (even if the offset isn't zero) if the indexed type
6536 // is something like [0 x {int, int}]
6537 const Type *IntPtrTy = TD->getIntPtrType();
6538 int64_t FirstIdx = 0;
6539 if (int64_t TySize = TD->getTypeSize(GEPIdxTy)) {
6540 FirstIdx = Offset/TySize;
6543 // Handle silly modulus not returning values values [0..TySize).
6547 assert(Offset >= 0);
6549 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6552 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6554 // Index into the types. If we fail, set OrigBase to null.
6556 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6557 const StructLayout *SL = TD->getStructLayout(STy);
6558 if (Offset < (int64_t)SL->getSizeInBytes()) {
6559 unsigned Elt = SL->getElementContainingOffset(Offset);
6560 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6562 Offset -= SL->getElementOffset(Elt);
6563 GEPIdxTy = STy->getElementType(Elt);
6565 // Otherwise, we can't index into this, bail out.
6569 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6570 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6571 if (uint64_t EltSize = TD->getTypeSize(STy->getElementType())) {
6572 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6575 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6577 GEPIdxTy = STy->getElementType();
6579 // Otherwise, we can't index into this, bail out.
6585 // If we were able to index down into an element, create the GEP
6586 // and bitcast the result. This eliminates one bitcast, potentially
6588 Instruction *NGEP = new GetElementPtrInst(OrigBase, &NewIndices[0],
6589 NewIndices.size(), "");
6590 InsertNewInstBefore(NGEP, CI);
6591 NGEP->takeName(GEP);
6593 if (isa<BitCastInst>(CI))
6594 return new BitCastInst(NGEP, CI.getType());
6595 assert(isa<PtrToIntInst>(CI));
6596 return new PtrToIntInst(NGEP, CI.getType());
6603 return commonCastTransforms(CI);
6608 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6609 /// integer types. This function implements the common transforms for all those
6611 /// @brief Implement the transforms common to CastInst with integer operands
6612 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6613 if (Instruction *Result = commonCastTransforms(CI))
6616 Value *Src = CI.getOperand(0);
6617 const Type *SrcTy = Src->getType();
6618 const Type *DestTy = CI.getType();
6619 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6620 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6622 // See if we can simplify any instructions used by the LHS whose sole
6623 // purpose is to compute bits we don't care about.
6624 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6625 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6626 KnownZero, KnownOne))
6629 // If the source isn't an instruction or has more than one use then we
6630 // can't do anything more.
6631 Instruction *SrcI = dyn_cast<Instruction>(Src);
6632 if (!SrcI || !Src->hasOneUse())
6635 // Attempt to propagate the cast into the instruction for int->int casts.
6636 int NumCastsRemoved = 0;
6637 if (!isa<BitCastInst>(CI) &&
6638 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6640 // If this cast is a truncate, evaluting in a different type always
6641 // eliminates the cast, so it is always a win. If this is a noop-cast
6642 // this just removes a noop cast which isn't pointful, but simplifies
6643 // the code. If this is a zero-extension, we need to do an AND to
6644 // maintain the clear top-part of the computation, so we require that
6645 // the input have eliminated at least one cast. If this is a sign
6646 // extension, we insert two new casts (to do the extension) so we
6647 // require that two casts have been eliminated.
6649 switch (CI.getOpcode()) {
6651 // All the others use floating point so we shouldn't actually
6652 // get here because of the check above.
6653 assert(0 && "Unknown cast type");
6654 case Instruction::Trunc:
6657 case Instruction::ZExt:
6658 DoXForm = NumCastsRemoved >= 1;
6660 case Instruction::SExt:
6661 DoXForm = NumCastsRemoved >= 2;
6663 case Instruction::BitCast:
6669 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6670 CI.getOpcode() == Instruction::SExt);
6671 assert(Res->getType() == DestTy);
6672 switch (CI.getOpcode()) {
6673 default: assert(0 && "Unknown cast type!");
6674 case Instruction::Trunc:
6675 case Instruction::BitCast:
6676 // Just replace this cast with the result.
6677 return ReplaceInstUsesWith(CI, Res);
6678 case Instruction::ZExt: {
6679 // We need to emit an AND to clear the high bits.
6680 assert(SrcBitSize < DestBitSize && "Not a zext?");
6681 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6683 return BinaryOperator::createAnd(Res, C);
6685 case Instruction::SExt:
6686 // We need to emit a cast to truncate, then a cast to sext.
6687 return CastInst::create(Instruction::SExt,
6688 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6694 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6695 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6697 switch (SrcI->getOpcode()) {
6698 case Instruction::Add:
6699 case Instruction::Mul:
6700 case Instruction::And:
6701 case Instruction::Or:
6702 case Instruction::Xor:
6703 // If we are discarding information, rewrite.
6704 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6705 // Don't insert two casts if they cannot be eliminated. We allow
6706 // two casts to be inserted if the sizes are the same. This could
6707 // only be converting signedness, which is a noop.
6708 if (DestBitSize == SrcBitSize ||
6709 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6710 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6711 Instruction::CastOps opcode = CI.getOpcode();
6712 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6713 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6714 return BinaryOperator::create(
6715 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6719 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6720 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6721 SrcI->getOpcode() == Instruction::Xor &&
6722 Op1 == ConstantInt::getTrue() &&
6723 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6724 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6725 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6728 case Instruction::SDiv:
6729 case Instruction::UDiv:
6730 case Instruction::SRem:
6731 case Instruction::URem:
6732 // If we are just changing the sign, rewrite.
6733 if (DestBitSize == SrcBitSize) {
6734 // Don't insert two casts if they cannot be eliminated. We allow
6735 // two casts to be inserted if the sizes are the same. This could
6736 // only be converting signedness, which is a noop.
6737 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6738 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6739 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6741 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6743 return BinaryOperator::create(
6744 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6749 case Instruction::Shl:
6750 // Allow changing the sign of the source operand. Do not allow
6751 // changing the size of the shift, UNLESS the shift amount is a
6752 // constant. We must not change variable sized shifts to a smaller
6753 // size, because it is undefined to shift more bits out than exist
6755 if (DestBitSize == SrcBitSize ||
6756 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6757 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6758 Instruction::BitCast : Instruction::Trunc);
6759 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6760 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6761 return BinaryOperator::createShl(Op0c, Op1c);
6764 case Instruction::AShr:
6765 // If this is a signed shr, and if all bits shifted in are about to be
6766 // truncated off, turn it into an unsigned shr to allow greater
6768 if (DestBitSize < SrcBitSize &&
6769 isa<ConstantInt>(Op1)) {
6770 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6771 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6772 // Insert the new logical shift right.
6773 return BinaryOperator::createLShr(Op0, Op1);
6781 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6782 if (Instruction *Result = commonIntCastTransforms(CI))
6785 Value *Src = CI.getOperand(0);
6786 const Type *Ty = CI.getType();
6787 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6788 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6790 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6791 switch (SrcI->getOpcode()) {
6793 case Instruction::LShr:
6794 // We can shrink lshr to something smaller if we know the bits shifted in
6795 // are already zeros.
6796 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6797 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6799 // Get a mask for the bits shifting in.
6800 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6801 Value* SrcIOp0 = SrcI->getOperand(0);
6802 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6803 if (ShAmt >= DestBitWidth) // All zeros.
6804 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6806 // Okay, we can shrink this. Truncate the input, then return a new
6808 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6809 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6811 return BinaryOperator::createLShr(V1, V2);
6813 } else { // This is a variable shr.
6815 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6816 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6817 // loop-invariant and CSE'd.
6818 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6819 Value *One = ConstantInt::get(SrcI->getType(), 1);
6821 Value *V = InsertNewInstBefore(
6822 BinaryOperator::createShl(One, SrcI->getOperand(1),
6824 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6825 SrcI->getOperand(0),
6827 Value *Zero = Constant::getNullValue(V->getType());
6828 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6838 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6839 // If one of the common conversion will work ..
6840 if (Instruction *Result = commonIntCastTransforms(CI))
6843 Value *Src = CI.getOperand(0);
6845 // If this is a cast of a cast
6846 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6847 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6848 // types and if the sizes are just right we can convert this into a logical
6849 // 'and' which will be much cheaper than the pair of casts.
6850 if (isa<TruncInst>(CSrc)) {
6851 // Get the sizes of the types involved
6852 Value *A = CSrc->getOperand(0);
6853 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6854 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6855 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6856 // If we're actually extending zero bits and the trunc is a no-op
6857 if (MidSize < DstSize && SrcSize == DstSize) {
6858 // Replace both of the casts with an And of the type mask.
6859 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6860 Constant *AndConst = ConstantInt::get(AndValue);
6862 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6863 // Unfortunately, if the type changed, we need to cast it back.
6864 if (And->getType() != CI.getType()) {
6865 And->setName(CSrc->getName()+".mask");
6866 InsertNewInstBefore(And, CI);
6867 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6874 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6875 // If we are just checking for a icmp eq of a single bit and zext'ing it
6876 // to an integer, then shift the bit to the appropriate place and then
6877 // cast to integer to avoid the comparison.
6878 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6879 const APInt &Op1CV = Op1C->getValue();
6881 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
6882 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
6883 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6884 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6885 Value *In = ICI->getOperand(0);
6886 Value *Sh = ConstantInt::get(In->getType(),
6887 In->getType()->getPrimitiveSizeInBits()-1);
6888 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
6889 In->getName()+".lobit"),
6891 if (In->getType() != CI.getType())
6892 In = CastInst::createIntegerCast(In, CI.getType(),
6893 false/*ZExt*/, "tmp", &CI);
6895 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
6896 Constant *One = ConstantInt::get(In->getType(), 1);
6897 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
6898 In->getName()+".not"),
6902 return ReplaceInstUsesWith(CI, In);
6907 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
6908 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6909 // zext (X == 1) to i32 --> X iff X has only the low bit set.
6910 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
6911 // zext (X != 0) to i32 --> X iff X has only the low bit set.
6912 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
6913 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
6914 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6915 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
6916 // This only works for EQ and NE
6917 ICI->isEquality()) {
6918 // If Op1C some other power of two, convert:
6919 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6920 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6921 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6922 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
6924 APInt KnownZeroMask(~KnownZero);
6925 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6926 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
6927 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6928 // (X&4) == 2 --> false
6929 // (X&4) != 2 --> true
6930 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6931 Res = ConstantExpr::getZExt(Res, CI.getType());
6932 return ReplaceInstUsesWith(CI, Res);
6935 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6936 Value *In = ICI->getOperand(0);
6938 // Perform a logical shr by shiftamt.
6939 // Insert the shift to put the result in the low bit.
6940 In = InsertNewInstBefore(
6941 BinaryOperator::createLShr(In,
6942 ConstantInt::get(In->getType(), ShiftAmt),
6943 In->getName()+".lobit"), CI);
6946 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6947 Constant *One = ConstantInt::get(In->getType(), 1);
6948 In = BinaryOperator::createXor(In, One, "tmp");
6949 InsertNewInstBefore(cast<Instruction>(In), CI);
6952 if (CI.getType() == In->getType())
6953 return ReplaceInstUsesWith(CI, In);
6955 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6963 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
6964 if (Instruction *I = commonIntCastTransforms(CI))
6967 Value *Src = CI.getOperand(0);
6969 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
6970 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
6971 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6972 // If we are just checking for a icmp eq of a single bit and zext'ing it
6973 // to an integer, then shift the bit to the appropriate place and then
6974 // cast to integer to avoid the comparison.
6975 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6976 const APInt &Op1CV = Op1C->getValue();
6978 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
6979 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
6980 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6981 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6982 Value *In = ICI->getOperand(0);
6983 Value *Sh = ConstantInt::get(In->getType(),
6984 In->getType()->getPrimitiveSizeInBits()-1);
6985 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
6986 In->getName()+".lobit"),
6988 if (In->getType() != CI.getType())
6989 In = CastInst::createIntegerCast(In, CI.getType(),
6990 true/*SExt*/, "tmp", &CI);
6992 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
6993 In = InsertNewInstBefore(BinaryOperator::createNot(In,
6994 In->getName()+".not"), CI);
6996 return ReplaceInstUsesWith(CI, In);
7004 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
7005 return commonCastTransforms(CI);
7008 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
7009 return commonCastTransforms(CI);
7012 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
7013 return commonCastTransforms(CI);
7016 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
7017 return commonCastTransforms(CI);
7020 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7021 return commonCastTransforms(CI);
7024 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7025 return commonCastTransforms(CI);
7028 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7029 return commonPointerCastTransforms(CI);
7032 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7033 return commonCastTransforms(CI);
7036 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7037 // If the operands are integer typed then apply the integer transforms,
7038 // otherwise just apply the common ones.
7039 Value *Src = CI.getOperand(0);
7040 const Type *SrcTy = Src->getType();
7041 const Type *DestTy = CI.getType();
7043 if (SrcTy->isInteger() && DestTy->isInteger()) {
7044 if (Instruction *Result = commonIntCastTransforms(CI))
7046 } else if (isa<PointerType>(SrcTy)) {
7047 if (Instruction *I = commonPointerCastTransforms(CI))
7050 if (Instruction *Result = commonCastTransforms(CI))
7055 // Get rid of casts from one type to the same type. These are useless and can
7056 // be replaced by the operand.
7057 if (DestTy == Src->getType())
7058 return ReplaceInstUsesWith(CI, Src);
7060 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7061 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7062 const Type *DstElTy = DstPTy->getElementType();
7063 const Type *SrcElTy = SrcPTy->getElementType();
7065 // If we are casting a malloc or alloca to a pointer to a type of the same
7066 // size, rewrite the allocation instruction to allocate the "right" type.
7067 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7068 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7071 // If the source and destination are pointers, and this cast is equivalent
7072 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7073 // This can enhance SROA and other transforms that want type-safe pointers.
7074 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7075 unsigned NumZeros = 0;
7076 while (SrcElTy != DstElTy &&
7077 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7078 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7079 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7083 // If we found a path from the src to dest, create the getelementptr now.
7084 if (SrcElTy == DstElTy) {
7085 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7086 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
7090 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7091 if (SVI->hasOneUse()) {
7092 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7093 // a bitconvert to a vector with the same # elts.
7094 if (isa<VectorType>(DestTy) &&
7095 cast<VectorType>(DestTy)->getNumElements() ==
7096 SVI->getType()->getNumElements()) {
7098 // If either of the operands is a cast from CI.getType(), then
7099 // evaluating the shuffle in the casted destination's type will allow
7100 // us to eliminate at least one cast.
7101 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7102 Tmp->getOperand(0)->getType() == DestTy) ||
7103 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7104 Tmp->getOperand(0)->getType() == DestTy)) {
7105 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7106 SVI->getOperand(0), DestTy, &CI);
7107 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7108 SVI->getOperand(1), DestTy, &CI);
7109 // Return a new shuffle vector. Use the same element ID's, as we
7110 // know the vector types match #elts.
7111 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7119 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7121 /// %D = select %cond, %C, %A
7123 /// %C = select %cond, %B, 0
7126 /// Assuming that the specified instruction is an operand to the select, return
7127 /// a bitmask indicating which operands of this instruction are foldable if they
7128 /// equal the other incoming value of the select.
7130 static unsigned GetSelectFoldableOperands(Instruction *I) {
7131 switch (I->getOpcode()) {
7132 case Instruction::Add:
7133 case Instruction::Mul:
7134 case Instruction::And:
7135 case Instruction::Or:
7136 case Instruction::Xor:
7137 return 3; // Can fold through either operand.
7138 case Instruction::Sub: // Can only fold on the amount subtracted.
7139 case Instruction::Shl: // Can only fold on the shift amount.
7140 case Instruction::LShr:
7141 case Instruction::AShr:
7144 return 0; // Cannot fold
7148 /// GetSelectFoldableConstant - For the same transformation as the previous
7149 /// function, return the identity constant that goes into the select.
7150 static Constant *GetSelectFoldableConstant(Instruction *I) {
7151 switch (I->getOpcode()) {
7152 default: assert(0 && "This cannot happen!"); abort();
7153 case Instruction::Add:
7154 case Instruction::Sub:
7155 case Instruction::Or:
7156 case Instruction::Xor:
7157 case Instruction::Shl:
7158 case Instruction::LShr:
7159 case Instruction::AShr:
7160 return Constant::getNullValue(I->getType());
7161 case Instruction::And:
7162 return Constant::getAllOnesValue(I->getType());
7163 case Instruction::Mul:
7164 return ConstantInt::get(I->getType(), 1);
7168 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7169 /// have the same opcode and only one use each. Try to simplify this.
7170 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7172 if (TI->getNumOperands() == 1) {
7173 // If this is a non-volatile load or a cast from the same type,
7176 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7179 return 0; // unknown unary op.
7182 // Fold this by inserting a select from the input values.
7183 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7184 FI->getOperand(0), SI.getName()+".v");
7185 InsertNewInstBefore(NewSI, SI);
7186 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7190 // Only handle binary operators here.
7191 if (!isa<BinaryOperator>(TI))
7194 // Figure out if the operations have any operands in common.
7195 Value *MatchOp, *OtherOpT, *OtherOpF;
7197 if (TI->getOperand(0) == FI->getOperand(0)) {
7198 MatchOp = TI->getOperand(0);
7199 OtherOpT = TI->getOperand(1);
7200 OtherOpF = FI->getOperand(1);
7201 MatchIsOpZero = true;
7202 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7203 MatchOp = TI->getOperand(1);
7204 OtherOpT = TI->getOperand(0);
7205 OtherOpF = FI->getOperand(0);
7206 MatchIsOpZero = false;
7207 } else if (!TI->isCommutative()) {
7209 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7210 MatchOp = TI->getOperand(0);
7211 OtherOpT = TI->getOperand(1);
7212 OtherOpF = FI->getOperand(0);
7213 MatchIsOpZero = true;
7214 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7215 MatchOp = TI->getOperand(1);
7216 OtherOpT = TI->getOperand(0);
7217 OtherOpF = FI->getOperand(1);
7218 MatchIsOpZero = true;
7223 // If we reach here, they do have operations in common.
7224 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7225 OtherOpF, SI.getName()+".v");
7226 InsertNewInstBefore(NewSI, SI);
7228 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7230 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7232 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7234 assert(0 && "Shouldn't get here");
7238 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7239 Value *CondVal = SI.getCondition();
7240 Value *TrueVal = SI.getTrueValue();
7241 Value *FalseVal = SI.getFalseValue();
7243 // select true, X, Y -> X
7244 // select false, X, Y -> Y
7245 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7246 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7248 // select C, X, X -> X
7249 if (TrueVal == FalseVal)
7250 return ReplaceInstUsesWith(SI, TrueVal);
7252 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7253 return ReplaceInstUsesWith(SI, FalseVal);
7254 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7255 return ReplaceInstUsesWith(SI, TrueVal);
7256 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7257 if (isa<Constant>(TrueVal))
7258 return ReplaceInstUsesWith(SI, TrueVal);
7260 return ReplaceInstUsesWith(SI, FalseVal);
7263 if (SI.getType() == Type::Int1Ty) {
7264 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7265 if (C->getZExtValue()) {
7266 // Change: A = select B, true, C --> A = or B, C
7267 return BinaryOperator::createOr(CondVal, FalseVal);
7269 // Change: A = select B, false, C --> A = and !B, C
7271 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7272 "not."+CondVal->getName()), SI);
7273 return BinaryOperator::createAnd(NotCond, FalseVal);
7275 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7276 if (C->getZExtValue() == false) {
7277 // Change: A = select B, C, false --> A = and B, C
7278 return BinaryOperator::createAnd(CondVal, TrueVal);
7280 // Change: A = select B, C, true --> A = or !B, C
7282 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7283 "not."+CondVal->getName()), SI);
7284 return BinaryOperator::createOr(NotCond, TrueVal);
7289 // Selecting between two integer constants?
7290 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7291 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7292 // select C, 1, 0 -> zext C to int
7293 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7294 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7295 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7296 // select C, 0, 1 -> zext !C to int
7298 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7299 "not."+CondVal->getName()), SI);
7300 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7303 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7305 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7307 // (x <s 0) ? -1 : 0 -> ashr x, 31
7308 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7309 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7310 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7311 // The comparison constant and the result are not neccessarily the
7312 // same width. Make an all-ones value by inserting a AShr.
7313 Value *X = IC->getOperand(0);
7314 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7315 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7316 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7318 InsertNewInstBefore(SRA, SI);
7320 // Finally, convert to the type of the select RHS. We figure out
7321 // if this requires a SExt, Trunc or BitCast based on the sizes.
7322 Instruction::CastOps opc = Instruction::BitCast;
7323 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7324 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7325 if (SRASize < SISize)
7326 opc = Instruction::SExt;
7327 else if (SRASize > SISize)
7328 opc = Instruction::Trunc;
7329 return CastInst::create(opc, SRA, SI.getType());
7334 // If one of the constants is zero (we know they can't both be) and we
7335 // have an icmp instruction with zero, and we have an 'and' with the
7336 // non-constant value, eliminate this whole mess. This corresponds to
7337 // cases like this: ((X & 27) ? 27 : 0)
7338 if (TrueValC->isZero() || FalseValC->isZero())
7339 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7340 cast<Constant>(IC->getOperand(1))->isNullValue())
7341 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7342 if (ICA->getOpcode() == Instruction::And &&
7343 isa<ConstantInt>(ICA->getOperand(1)) &&
7344 (ICA->getOperand(1) == TrueValC ||
7345 ICA->getOperand(1) == FalseValC) &&
7346 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7347 // Okay, now we know that everything is set up, we just don't
7348 // know whether we have a icmp_ne or icmp_eq and whether the
7349 // true or false val is the zero.
7350 bool ShouldNotVal = !TrueValC->isZero();
7351 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7354 V = InsertNewInstBefore(BinaryOperator::create(
7355 Instruction::Xor, V, ICA->getOperand(1)), SI);
7356 return ReplaceInstUsesWith(SI, V);
7361 // See if we are selecting two values based on a comparison of the two values.
7362 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7363 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7364 // Transform (X == Y) ? X : Y -> Y
7365 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7366 return ReplaceInstUsesWith(SI, FalseVal);
7367 // Transform (X != Y) ? X : Y -> X
7368 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7369 return ReplaceInstUsesWith(SI, TrueVal);
7370 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7372 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7373 // Transform (X == Y) ? Y : X -> X
7374 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7375 return ReplaceInstUsesWith(SI, FalseVal);
7376 // Transform (X != Y) ? Y : X -> Y
7377 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7378 return ReplaceInstUsesWith(SI, TrueVal);
7379 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7383 // See if we are selecting two values based on a comparison of the two values.
7384 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7385 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7386 // Transform (X == Y) ? X : Y -> Y
7387 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7388 return ReplaceInstUsesWith(SI, FalseVal);
7389 // Transform (X != Y) ? X : Y -> X
7390 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7391 return ReplaceInstUsesWith(SI, TrueVal);
7392 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7394 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7395 // Transform (X == Y) ? Y : X -> X
7396 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7397 return ReplaceInstUsesWith(SI, FalseVal);
7398 // Transform (X != Y) ? Y : X -> Y
7399 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7400 return ReplaceInstUsesWith(SI, TrueVal);
7401 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7405 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7406 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7407 if (TI->hasOneUse() && FI->hasOneUse()) {
7408 Instruction *AddOp = 0, *SubOp = 0;
7410 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7411 if (TI->getOpcode() == FI->getOpcode())
7412 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7415 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7416 // even legal for FP.
7417 if (TI->getOpcode() == Instruction::Sub &&
7418 FI->getOpcode() == Instruction::Add) {
7419 AddOp = FI; SubOp = TI;
7420 } else if (FI->getOpcode() == Instruction::Sub &&
7421 TI->getOpcode() == Instruction::Add) {
7422 AddOp = TI; SubOp = FI;
7426 Value *OtherAddOp = 0;
7427 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7428 OtherAddOp = AddOp->getOperand(1);
7429 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7430 OtherAddOp = AddOp->getOperand(0);
7434 // So at this point we know we have (Y -> OtherAddOp):
7435 // select C, (add X, Y), (sub X, Z)
7436 Value *NegVal; // Compute -Z
7437 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7438 NegVal = ConstantExpr::getNeg(C);
7440 NegVal = InsertNewInstBefore(
7441 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7444 Value *NewTrueOp = OtherAddOp;
7445 Value *NewFalseOp = NegVal;
7447 std::swap(NewTrueOp, NewFalseOp);
7448 Instruction *NewSel =
7449 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7451 NewSel = InsertNewInstBefore(NewSel, SI);
7452 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7457 // See if we can fold the select into one of our operands.
7458 if (SI.getType()->isInteger()) {
7459 // See the comment above GetSelectFoldableOperands for a description of the
7460 // transformation we are doing here.
7461 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7462 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7463 !isa<Constant>(FalseVal))
7464 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7465 unsigned OpToFold = 0;
7466 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7468 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7473 Constant *C = GetSelectFoldableConstant(TVI);
7474 Instruction *NewSel =
7475 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7476 InsertNewInstBefore(NewSel, SI);
7477 NewSel->takeName(TVI);
7478 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7479 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7481 assert(0 && "Unknown instruction!!");
7486 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7487 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7488 !isa<Constant>(TrueVal))
7489 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7490 unsigned OpToFold = 0;
7491 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7493 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7498 Constant *C = GetSelectFoldableConstant(FVI);
7499 Instruction *NewSel =
7500 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7501 InsertNewInstBefore(NewSel, SI);
7502 NewSel->takeName(FVI);
7503 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7504 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7506 assert(0 && "Unknown instruction!!");
7511 if (BinaryOperator::isNot(CondVal)) {
7512 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7513 SI.setOperand(1, FalseVal);
7514 SI.setOperand(2, TrueVal);
7521 /// GetKnownAlignment - If the specified pointer has an alignment that we can
7522 /// determine, return it, otherwise return 0.
7523 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
7524 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7525 unsigned Align = GV->getAlignment();
7526 if (Align == 0 && TD)
7527 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7529 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7530 unsigned Align = AI->getAlignment();
7531 if (Align == 0 && TD) {
7532 if (isa<AllocaInst>(AI))
7533 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7534 else if (isa<MallocInst>(AI)) {
7535 // Malloc returns maximally aligned memory.
7536 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7539 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7542 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7546 } else if (isa<BitCastInst>(V) ||
7547 (isa<ConstantExpr>(V) &&
7548 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7549 User *CI = cast<User>(V);
7550 if (isa<PointerType>(CI->getOperand(0)->getType()))
7551 return GetKnownAlignment(CI->getOperand(0), TD);
7553 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7554 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7555 if (BaseAlignment == 0) return 0;
7557 // If all indexes are zero, it is just the alignment of the base pointer.
7558 bool AllZeroOperands = true;
7559 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7560 if (!isa<Constant>(GEPI->getOperand(i)) ||
7561 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7562 AllZeroOperands = false;
7565 if (AllZeroOperands)
7566 return BaseAlignment;
7568 // Otherwise, if the base alignment is >= the alignment we expect for the
7569 // base pointer type, then we know that the resultant pointer is aligned at
7570 // least as much as its type requires.
7573 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7574 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7575 if (TD->getABITypeAlignment(PtrTy->getElementType())
7577 const Type *GEPTy = GEPI->getType();
7578 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7579 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7587 /// visitCallInst - CallInst simplification. This mostly only handles folding
7588 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7589 /// the heavy lifting.
7591 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7592 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7593 if (!II) return visitCallSite(&CI);
7595 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7597 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7598 bool Changed = false;
7600 // memmove/cpy/set of zero bytes is a noop.
7601 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7602 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7604 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7605 if (CI->getZExtValue() == 1) {
7606 // Replace the instruction with just byte operations. We would
7607 // transform other cases to loads/stores, but we don't know if
7608 // alignment is sufficient.
7612 // If we have a memmove and the source operation is a constant global,
7613 // then the source and dest pointers can't alias, so we can change this
7614 // into a call to memcpy.
7615 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7616 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7617 if (GVSrc->isConstant()) {
7618 Module *M = CI.getParent()->getParent()->getParent();
7620 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7622 Name = "llvm.memcpy.i32";
7624 Name = "llvm.memcpy.i64";
7625 Constant *MemCpy = M->getOrInsertFunction(Name,
7626 CI.getCalledFunction()->getFunctionType());
7627 CI.setOperand(0, MemCpy);
7632 // If we can determine a pointer alignment that is bigger than currently
7633 // set, update the alignment.
7634 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7635 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7636 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7637 unsigned Align = std::min(Alignment1, Alignment2);
7638 if (MI->getAlignment()->getZExtValue() < Align) {
7639 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7642 } else if (isa<MemSetInst>(MI)) {
7643 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7644 if (MI->getAlignment()->getZExtValue() < Alignment) {
7645 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7650 if (Changed) return II;
7652 switch (II->getIntrinsicID()) {
7654 case Intrinsic::ppc_altivec_lvx:
7655 case Intrinsic::ppc_altivec_lvxl:
7656 case Intrinsic::x86_sse_loadu_ps:
7657 case Intrinsic::x86_sse2_loadu_pd:
7658 case Intrinsic::x86_sse2_loadu_dq:
7659 // Turn PPC lvx -> load if the pointer is known aligned.
7660 // Turn X86 loadups -> load if the pointer is known aligned.
7661 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7662 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7663 PointerType::get(II->getType()), CI);
7664 return new LoadInst(Ptr);
7667 case Intrinsic::ppc_altivec_stvx:
7668 case Intrinsic::ppc_altivec_stvxl:
7669 // Turn stvx -> store if the pointer is known aligned.
7670 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7671 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7672 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7674 return new StoreInst(II->getOperand(1), Ptr);
7677 case Intrinsic::x86_sse_storeu_ps:
7678 case Intrinsic::x86_sse2_storeu_pd:
7679 case Intrinsic::x86_sse2_storeu_dq:
7680 case Intrinsic::x86_sse2_storel_dq:
7681 // Turn X86 storeu -> store if the pointer is known aligned.
7682 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7683 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7684 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7686 return new StoreInst(II->getOperand(2), Ptr);
7690 case Intrinsic::x86_sse_cvttss2si: {
7691 // These intrinsics only demands the 0th element of its input vector. If
7692 // we can simplify the input based on that, do so now.
7694 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7696 II->setOperand(1, V);
7702 case Intrinsic::ppc_altivec_vperm:
7703 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7704 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7705 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7707 // Check that all of the elements are integer constants or undefs.
7708 bool AllEltsOk = true;
7709 for (unsigned i = 0; i != 16; ++i) {
7710 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7711 !isa<UndefValue>(Mask->getOperand(i))) {
7718 // Cast the input vectors to byte vectors.
7719 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7720 II->getOperand(1), Mask->getType(), CI);
7721 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7722 II->getOperand(2), Mask->getType(), CI);
7723 Value *Result = UndefValue::get(Op0->getType());
7725 // Only extract each element once.
7726 Value *ExtractedElts[32];
7727 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7729 for (unsigned i = 0; i != 16; ++i) {
7730 if (isa<UndefValue>(Mask->getOperand(i)))
7732 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7733 Idx &= 31; // Match the hardware behavior.
7735 if (ExtractedElts[Idx] == 0) {
7737 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7738 InsertNewInstBefore(Elt, CI);
7739 ExtractedElts[Idx] = Elt;
7742 // Insert this value into the result vector.
7743 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7744 InsertNewInstBefore(cast<Instruction>(Result), CI);
7746 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7751 case Intrinsic::stackrestore: {
7752 // If the save is right next to the restore, remove the restore. This can
7753 // happen when variable allocas are DCE'd.
7754 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7755 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7756 BasicBlock::iterator BI = SS;
7758 return EraseInstFromFunction(CI);
7762 // If the stack restore is in a return/unwind block and if there are no
7763 // allocas or calls between the restore and the return, nuke the restore.
7764 TerminatorInst *TI = II->getParent()->getTerminator();
7765 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7766 BasicBlock::iterator BI = II;
7767 bool CannotRemove = false;
7768 for (++BI; &*BI != TI; ++BI) {
7769 if (isa<AllocaInst>(BI) ||
7770 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7771 CannotRemove = true;
7776 return EraseInstFromFunction(CI);
7783 return visitCallSite(II);
7786 // InvokeInst simplification
7788 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7789 return visitCallSite(&II);
7792 // visitCallSite - Improvements for call and invoke instructions.
7794 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7795 bool Changed = false;
7797 // If the callee is a constexpr cast of a function, attempt to move the cast
7798 // to the arguments of the call/invoke.
7799 if (transformConstExprCastCall(CS)) return 0;
7801 Value *Callee = CS.getCalledValue();
7803 if (Function *CalleeF = dyn_cast<Function>(Callee))
7804 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7805 Instruction *OldCall = CS.getInstruction();
7806 // If the call and callee calling conventions don't match, this call must
7807 // be unreachable, as the call is undefined.
7808 new StoreInst(ConstantInt::getTrue(),
7809 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7810 if (!OldCall->use_empty())
7811 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7812 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7813 return EraseInstFromFunction(*OldCall);
7817 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7818 // This instruction is not reachable, just remove it. We insert a store to
7819 // undef so that we know that this code is not reachable, despite the fact
7820 // that we can't modify the CFG here.
7821 new StoreInst(ConstantInt::getTrue(),
7822 UndefValue::get(PointerType::get(Type::Int1Ty)),
7823 CS.getInstruction());
7825 if (!CS.getInstruction()->use_empty())
7826 CS.getInstruction()->
7827 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7829 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7830 // Don't break the CFG, insert a dummy cond branch.
7831 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7832 ConstantInt::getTrue(), II);
7834 return EraseInstFromFunction(*CS.getInstruction());
7837 const PointerType *PTy = cast<PointerType>(Callee->getType());
7838 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7839 if (FTy->isVarArg()) {
7840 // See if we can optimize any arguments passed through the varargs area of
7842 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7843 E = CS.arg_end(); I != E; ++I)
7844 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7845 // If this cast does not effect the value passed through the varargs
7846 // area, we can eliminate the use of the cast.
7847 Value *Op = CI->getOperand(0);
7848 if (CI->isLosslessCast()) {
7855 return Changed ? CS.getInstruction() : 0;
7858 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7859 // attempt to move the cast to the arguments of the call/invoke.
7861 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7862 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7863 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7864 if (CE->getOpcode() != Instruction::BitCast ||
7865 !isa<Function>(CE->getOperand(0)))
7867 Function *Callee = cast<Function>(CE->getOperand(0));
7868 Instruction *Caller = CS.getInstruction();
7870 // Okay, this is a cast from a function to a different type. Unless doing so
7871 // would cause a type conversion of one of our arguments, change this call to
7872 // be a direct call with arguments casted to the appropriate types.
7874 const FunctionType *FT = Callee->getFunctionType();
7875 const Type *OldRetTy = Caller->getType();
7877 const FunctionType *ActualFT =
7878 cast<FunctionType>(cast<PointerType>(CE->getType())->getElementType());
7880 // If the parameter attributes don't match up, don't do the xform. We don't
7881 // want to lose an sret attribute or something.
7882 if (FT->getParamAttrs() != ActualFT->getParamAttrs())
7885 // Check to see if we are changing the return type...
7886 if (OldRetTy != FT->getReturnType()) {
7887 if (Callee->isDeclaration() && !Caller->use_empty() &&
7888 // Conversion is ok if changing from pointer to int of same size.
7889 !(isa<PointerType>(FT->getReturnType()) &&
7890 TD->getIntPtrType() == OldRetTy))
7891 return false; // Cannot transform this return value.
7893 // If the callsite is an invoke instruction, and the return value is used by
7894 // a PHI node in a successor, we cannot change the return type of the call
7895 // because there is no place to put the cast instruction (without breaking
7896 // the critical edge). Bail out in this case.
7897 if (!Caller->use_empty())
7898 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7899 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7901 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7902 if (PN->getParent() == II->getNormalDest() ||
7903 PN->getParent() == II->getUnwindDest())
7907 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7908 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7910 CallSite::arg_iterator AI = CS.arg_begin();
7911 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7912 const Type *ParamTy = FT->getParamType(i);
7913 const Type *ActTy = (*AI)->getType();
7914 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7915 //Some conversions are safe even if we do not have a body.
7916 //Either we can cast directly, or we can upconvert the argument
7917 bool isConvertible = ActTy == ParamTy ||
7918 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7919 (ParamTy->isInteger() && ActTy->isInteger() &&
7920 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7921 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7922 && c->getValue().isStrictlyPositive());
7923 if (Callee->isDeclaration() && !isConvertible) return false;
7925 // Most other conversions can be done if we have a body, even if these
7926 // lose information, e.g. int->short.
7927 // Some conversions cannot be done at all, e.g. float to pointer.
7928 // Logic here parallels CastInst::getCastOpcode (the design there
7929 // requires legality checks like this be done before calling it).
7930 if (ParamTy->isInteger()) {
7931 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7932 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7935 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
7936 !isa<PointerType>(ActTy))
7938 } else if (ParamTy->isFloatingPoint()) {
7939 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7940 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7943 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
7945 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
7946 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7947 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
7950 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
7952 } else if (isa<PointerType>(ParamTy)) {
7953 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
7960 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7961 Callee->isDeclaration())
7962 return false; // Do not delete arguments unless we have a function body...
7964 // Okay, we decided that this is a safe thing to do: go ahead and start
7965 // inserting cast instructions as necessary...
7966 std::vector<Value*> Args;
7967 Args.reserve(NumActualArgs);
7969 AI = CS.arg_begin();
7970 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7971 const Type *ParamTy = FT->getParamType(i);
7972 if ((*AI)->getType() == ParamTy) {
7973 Args.push_back(*AI);
7975 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7976 false, ParamTy, false);
7977 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7978 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7982 // If the function takes more arguments than the call was taking, add them
7984 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7985 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7987 // If we are removing arguments to the function, emit an obnoxious warning...
7988 if (FT->getNumParams() < NumActualArgs)
7989 if (!FT->isVarArg()) {
7990 cerr << "WARNING: While resolving call to function '"
7991 << Callee->getName() << "' arguments were dropped!\n";
7993 // Add all of the arguments in their promoted form to the arg list...
7994 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7995 const Type *PTy = getPromotedType((*AI)->getType());
7996 if (PTy != (*AI)->getType()) {
7997 // Must promote to pass through va_arg area!
7998 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8000 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8001 InsertNewInstBefore(Cast, *Caller);
8002 Args.push_back(Cast);
8004 Args.push_back(*AI);
8009 if (FT->getReturnType() == Type::VoidTy)
8010 Caller->setName(""); // Void type should not have a name.
8013 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8014 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8015 &Args[0], Args.size(), Caller->getName(), Caller);
8016 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
8018 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
8019 if (cast<CallInst>(Caller)->isTailCall())
8020 cast<CallInst>(NC)->setTailCall();
8021 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8024 // Insert a cast of the return type as necessary.
8026 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
8027 if (NV->getType() != Type::VoidTy) {
8028 const Type *CallerTy = Caller->getType();
8029 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8031 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
8033 // If this is an invoke instruction, we should insert it after the first
8034 // non-phi, instruction in the normal successor block.
8035 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8036 BasicBlock::iterator I = II->getNormalDest()->begin();
8037 while (isa<PHINode>(I)) ++I;
8038 InsertNewInstBefore(NC, *I);
8040 // Otherwise, it's a call, just insert cast right after the call instr
8041 InsertNewInstBefore(NC, *Caller);
8043 AddUsersToWorkList(*Caller);
8045 NV = UndefValue::get(Caller->getType());
8049 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8050 Caller->replaceAllUsesWith(NV);
8051 Caller->eraseFromParent();
8052 RemoveFromWorkList(Caller);
8056 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8057 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8058 /// and a single binop.
8059 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8060 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8061 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8062 isa<CmpInst>(FirstInst));
8063 unsigned Opc = FirstInst->getOpcode();
8064 Value *LHSVal = FirstInst->getOperand(0);
8065 Value *RHSVal = FirstInst->getOperand(1);
8067 const Type *LHSType = LHSVal->getType();
8068 const Type *RHSType = RHSVal->getType();
8070 // Scan to see if all operands are the same opcode, all have one use, and all
8071 // kill their operands (i.e. the operands have one use).
8072 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8073 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8074 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8075 // Verify type of the LHS matches so we don't fold cmp's of different
8076 // types or GEP's with different index types.
8077 I->getOperand(0)->getType() != LHSType ||
8078 I->getOperand(1)->getType() != RHSType)
8081 // If they are CmpInst instructions, check their predicates
8082 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8083 if (cast<CmpInst>(I)->getPredicate() !=
8084 cast<CmpInst>(FirstInst)->getPredicate())
8087 // Keep track of which operand needs a phi node.
8088 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8089 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8092 // Otherwise, this is safe to transform, determine if it is profitable.
8094 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8095 // Indexes are often folded into load/store instructions, so we don't want to
8096 // hide them behind a phi.
8097 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8100 Value *InLHS = FirstInst->getOperand(0);
8101 Value *InRHS = FirstInst->getOperand(1);
8102 PHINode *NewLHS = 0, *NewRHS = 0;
8104 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8105 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8106 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8107 InsertNewInstBefore(NewLHS, PN);
8112 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8113 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8114 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8115 InsertNewInstBefore(NewRHS, PN);
8119 // Add all operands to the new PHIs.
8120 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8122 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8123 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8126 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8127 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8131 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8132 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8133 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8134 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8137 assert(isa<GetElementPtrInst>(FirstInst));
8138 return new GetElementPtrInst(LHSVal, RHSVal);
8142 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8143 /// of the block that defines it. This means that it must be obvious the value
8144 /// of the load is not changed from the point of the load to the end of the
8147 /// Finally, it is safe, but not profitable, to sink a load targetting a
8148 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8150 static bool isSafeToSinkLoad(LoadInst *L) {
8151 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8153 for (++BBI; BBI != E; ++BBI)
8154 if (BBI->mayWriteToMemory())
8157 // Check for non-address taken alloca. If not address-taken already, it isn't
8158 // profitable to do this xform.
8159 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8160 bool isAddressTaken = false;
8161 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8163 if (isa<LoadInst>(UI)) continue;
8164 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8165 // If storing TO the alloca, then the address isn't taken.
8166 if (SI->getOperand(1) == AI) continue;
8168 isAddressTaken = true;
8172 if (!isAddressTaken)
8180 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8181 // operator and they all are only used by the PHI, PHI together their
8182 // inputs, and do the operation once, to the result of the PHI.
8183 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8184 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8186 // Scan the instruction, looking for input operations that can be folded away.
8187 // If all input operands to the phi are the same instruction (e.g. a cast from
8188 // the same type or "+42") we can pull the operation through the PHI, reducing
8189 // code size and simplifying code.
8190 Constant *ConstantOp = 0;
8191 const Type *CastSrcTy = 0;
8192 bool isVolatile = false;
8193 if (isa<CastInst>(FirstInst)) {
8194 CastSrcTy = FirstInst->getOperand(0)->getType();
8195 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8196 // Can fold binop, compare or shift here if the RHS is a constant,
8197 // otherwise call FoldPHIArgBinOpIntoPHI.
8198 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8199 if (ConstantOp == 0)
8200 return FoldPHIArgBinOpIntoPHI(PN);
8201 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8202 isVolatile = LI->isVolatile();
8203 // We can't sink the load if the loaded value could be modified between the
8204 // load and the PHI.
8205 if (LI->getParent() != PN.getIncomingBlock(0) ||
8206 !isSafeToSinkLoad(LI))
8208 } else if (isa<GetElementPtrInst>(FirstInst)) {
8209 if (FirstInst->getNumOperands() == 2)
8210 return FoldPHIArgBinOpIntoPHI(PN);
8211 // Can't handle general GEPs yet.
8214 return 0; // Cannot fold this operation.
8217 // Check to see if all arguments are the same operation.
8218 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8219 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8220 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8221 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8224 if (I->getOperand(0)->getType() != CastSrcTy)
8225 return 0; // Cast operation must match.
8226 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8227 // We can't sink the load if the loaded value could be modified between
8228 // the load and the PHI.
8229 if (LI->isVolatile() != isVolatile ||
8230 LI->getParent() != PN.getIncomingBlock(i) ||
8231 !isSafeToSinkLoad(LI))
8233 } else if (I->getOperand(1) != ConstantOp) {
8238 // Okay, they are all the same operation. Create a new PHI node of the
8239 // correct type, and PHI together all of the LHS's of the instructions.
8240 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8241 PN.getName()+".in");
8242 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8244 Value *InVal = FirstInst->getOperand(0);
8245 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8247 // Add all operands to the new PHI.
8248 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8249 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8250 if (NewInVal != InVal)
8252 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8257 // The new PHI unions all of the same values together. This is really
8258 // common, so we handle it intelligently here for compile-time speed.
8262 InsertNewInstBefore(NewPN, PN);
8266 // Insert and return the new operation.
8267 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8268 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8269 else if (isa<LoadInst>(FirstInst))
8270 return new LoadInst(PhiVal, "", isVolatile);
8271 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8272 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8273 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8274 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8275 PhiVal, ConstantOp);
8277 assert(0 && "Unknown operation");
8281 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8283 static bool DeadPHICycle(PHINode *PN,
8284 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8285 if (PN->use_empty()) return true;
8286 if (!PN->hasOneUse()) return false;
8288 // Remember this node, and if we find the cycle, return.
8289 if (!PotentiallyDeadPHIs.insert(PN))
8292 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8293 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8298 // PHINode simplification
8300 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8301 // If LCSSA is around, don't mess with Phi nodes
8302 if (MustPreserveLCSSA) return 0;
8304 if (Value *V = PN.hasConstantValue())
8305 return ReplaceInstUsesWith(PN, V);
8307 // If all PHI operands are the same operation, pull them through the PHI,
8308 // reducing code size.
8309 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8310 PN.getIncomingValue(0)->hasOneUse())
8311 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8314 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8315 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8316 // PHI)... break the cycle.
8317 if (PN.hasOneUse()) {
8318 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8319 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8320 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8321 PotentiallyDeadPHIs.insert(&PN);
8322 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8323 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8326 // If this phi has a single use, and if that use just computes a value for
8327 // the next iteration of a loop, delete the phi. This occurs with unused
8328 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8329 // common case here is good because the only other things that catch this
8330 // are induction variable analysis (sometimes) and ADCE, which is only run
8332 if (PHIUser->hasOneUse() &&
8333 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8334 PHIUser->use_back() == &PN) {
8335 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8342 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8343 Instruction *InsertPoint,
8345 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8346 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8347 // We must cast correctly to the pointer type. Ensure that we
8348 // sign extend the integer value if it is smaller as this is
8349 // used for address computation.
8350 Instruction::CastOps opcode =
8351 (VTySize < PtrSize ? Instruction::SExt :
8352 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8353 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8357 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8358 Value *PtrOp = GEP.getOperand(0);
8359 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8360 // If so, eliminate the noop.
8361 if (GEP.getNumOperands() == 1)
8362 return ReplaceInstUsesWith(GEP, PtrOp);
8364 if (isa<UndefValue>(GEP.getOperand(0)))
8365 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8367 bool HasZeroPointerIndex = false;
8368 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8369 HasZeroPointerIndex = C->isNullValue();
8371 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8372 return ReplaceInstUsesWith(GEP, PtrOp);
8374 // Eliminate unneeded casts for indices.
8375 bool MadeChange = false;
8377 gep_type_iterator GTI = gep_type_begin(GEP);
8378 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8379 if (isa<SequentialType>(*GTI)) {
8380 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8381 if (CI->getOpcode() == Instruction::ZExt ||
8382 CI->getOpcode() == Instruction::SExt) {
8383 const Type *SrcTy = CI->getOperand(0)->getType();
8384 // We can eliminate a cast from i32 to i64 iff the target
8385 // is a 32-bit pointer target.
8386 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8388 GEP.setOperand(i, CI->getOperand(0));
8392 // If we are using a wider index than needed for this platform, shrink it
8393 // to what we need. If the incoming value needs a cast instruction,
8394 // insert it. This explicit cast can make subsequent optimizations more
8396 Value *Op = GEP.getOperand(i);
8397 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
8398 if (Constant *C = dyn_cast<Constant>(Op)) {
8399 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8402 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8404 GEP.setOperand(i, Op);
8409 if (MadeChange) return &GEP;
8411 // If this GEP instruction doesn't move the pointer, and if the input operand
8412 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8413 // real input to the dest type.
8414 if (GEP.hasAllZeroIndices() && isa<BitCastInst>(GEP.getOperand(0)))
8415 return new BitCastInst(cast<BitCastInst>(GEP.getOperand(0))->getOperand(0),
8418 // Combine Indices - If the source pointer to this getelementptr instruction
8419 // is a getelementptr instruction, combine the indices of the two
8420 // getelementptr instructions into a single instruction.
8422 SmallVector<Value*, 8> SrcGEPOperands;
8423 if (User *Src = dyn_castGetElementPtr(PtrOp))
8424 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8426 if (!SrcGEPOperands.empty()) {
8427 // Note that if our source is a gep chain itself that we wait for that
8428 // chain to be resolved before we perform this transformation. This
8429 // avoids us creating a TON of code in some cases.
8431 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8432 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8433 return 0; // Wait until our source is folded to completion.
8435 SmallVector<Value*, 8> Indices;
8437 // Find out whether the last index in the source GEP is a sequential idx.
8438 bool EndsWithSequential = false;
8439 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8440 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8441 EndsWithSequential = !isa<StructType>(*I);
8443 // Can we combine the two pointer arithmetics offsets?
8444 if (EndsWithSequential) {
8445 // Replace: gep (gep %P, long B), long A, ...
8446 // With: T = long A+B; gep %P, T, ...
8448 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8449 if (SO1 == Constant::getNullValue(SO1->getType())) {
8451 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8454 // If they aren't the same type, convert both to an integer of the
8455 // target's pointer size.
8456 if (SO1->getType() != GO1->getType()) {
8457 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8458 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8459 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8460 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8462 unsigned PS = TD->getPointerSize();
8463 if (TD->getTypeSize(SO1->getType()) == PS) {
8464 // Convert GO1 to SO1's type.
8465 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8467 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8468 // Convert SO1 to GO1's type.
8469 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8471 const Type *PT = TD->getIntPtrType();
8472 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8473 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8477 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8478 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8480 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8481 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8485 // Recycle the GEP we already have if possible.
8486 if (SrcGEPOperands.size() == 2) {
8487 GEP.setOperand(0, SrcGEPOperands[0]);
8488 GEP.setOperand(1, Sum);
8491 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8492 SrcGEPOperands.end()-1);
8493 Indices.push_back(Sum);
8494 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8496 } else if (isa<Constant>(*GEP.idx_begin()) &&
8497 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8498 SrcGEPOperands.size() != 1) {
8499 // Otherwise we can do the fold if the first index of the GEP is a zero
8500 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8501 SrcGEPOperands.end());
8502 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8505 if (!Indices.empty())
8506 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
8507 Indices.size(), GEP.getName());
8509 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8510 // GEP of global variable. If all of the indices for this GEP are
8511 // constants, we can promote this to a constexpr instead of an instruction.
8513 // Scan for nonconstants...
8514 SmallVector<Constant*, 8> Indices;
8515 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8516 for (; I != E && isa<Constant>(*I); ++I)
8517 Indices.push_back(cast<Constant>(*I));
8519 if (I == E) { // If they are all constants...
8520 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8521 &Indices[0],Indices.size());
8523 // Replace all uses of the GEP with the new constexpr...
8524 return ReplaceInstUsesWith(GEP, CE);
8526 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8527 if (!isa<PointerType>(X->getType())) {
8528 // Not interesting. Source pointer must be a cast from pointer.
8529 } else if (HasZeroPointerIndex) {
8530 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8531 // into : GEP [10 x ubyte]* X, long 0, ...
8533 // This occurs when the program declares an array extern like "int X[];"
8535 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8536 const PointerType *XTy = cast<PointerType>(X->getType());
8537 if (const ArrayType *XATy =
8538 dyn_cast<ArrayType>(XTy->getElementType()))
8539 if (const ArrayType *CATy =
8540 dyn_cast<ArrayType>(CPTy->getElementType()))
8541 if (CATy->getElementType() == XATy->getElementType()) {
8542 // At this point, we know that the cast source type is a pointer
8543 // to an array of the same type as the destination pointer
8544 // array. Because the array type is never stepped over (there
8545 // is a leading zero) we can fold the cast into this GEP.
8546 GEP.setOperand(0, X);
8549 } else if (GEP.getNumOperands() == 2) {
8550 // Transform things like:
8551 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8552 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8553 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8554 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8555 if (isa<ArrayType>(SrcElTy) &&
8556 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8557 TD->getTypeSize(ResElTy)) {
8558 Value *V = InsertNewInstBefore(
8559 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8560 GEP.getOperand(1), GEP.getName()), GEP);
8561 // V and GEP are both pointer types --> BitCast
8562 return new BitCastInst(V, GEP.getType());
8565 // Transform things like:
8566 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8567 // (where tmp = 8*tmp2) into:
8568 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8570 if (isa<ArrayType>(SrcElTy) &&
8571 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8572 uint64_t ArrayEltSize =
8573 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8575 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8576 // allow either a mul, shift, or constant here.
8578 ConstantInt *Scale = 0;
8579 if (ArrayEltSize == 1) {
8580 NewIdx = GEP.getOperand(1);
8581 Scale = ConstantInt::get(NewIdx->getType(), 1);
8582 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8583 NewIdx = ConstantInt::get(CI->getType(), 1);
8585 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8586 if (Inst->getOpcode() == Instruction::Shl &&
8587 isa<ConstantInt>(Inst->getOperand(1))) {
8588 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8589 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8590 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8591 NewIdx = Inst->getOperand(0);
8592 } else if (Inst->getOpcode() == Instruction::Mul &&
8593 isa<ConstantInt>(Inst->getOperand(1))) {
8594 Scale = cast<ConstantInt>(Inst->getOperand(1));
8595 NewIdx = Inst->getOperand(0);
8599 // If the index will be to exactly the right offset with the scale taken
8600 // out, perform the transformation.
8601 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8602 if (isa<ConstantInt>(Scale))
8603 Scale = ConstantInt::get(Scale->getType(),
8604 Scale->getZExtValue() / ArrayEltSize);
8605 if (Scale->getZExtValue() != 1) {
8606 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8608 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8609 NewIdx = InsertNewInstBefore(Sc, GEP);
8612 // Insert the new GEP instruction.
8613 Instruction *NewGEP =
8614 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8615 NewIdx, GEP.getName());
8616 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8617 // The NewGEP must be pointer typed, so must the old one -> BitCast
8618 return new BitCastInst(NewGEP, GEP.getType());
8627 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8628 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8629 if (AI.isArrayAllocation()) // Check C != 1
8630 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8632 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8633 AllocationInst *New = 0;
8635 // Create and insert the replacement instruction...
8636 if (isa<MallocInst>(AI))
8637 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8639 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8640 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8643 InsertNewInstBefore(New, AI);
8645 // Scan to the end of the allocation instructions, to skip over a block of
8646 // allocas if possible...
8648 BasicBlock::iterator It = New;
8649 while (isa<AllocationInst>(*It)) ++It;
8651 // Now that I is pointing to the first non-allocation-inst in the block,
8652 // insert our getelementptr instruction...
8654 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8655 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8656 New->getName()+".sub", It);
8658 // Now make everything use the getelementptr instead of the original
8660 return ReplaceInstUsesWith(AI, V);
8661 } else if (isa<UndefValue>(AI.getArraySize())) {
8662 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8665 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8666 // Note that we only do this for alloca's, because malloc should allocate and
8667 // return a unique pointer, even for a zero byte allocation.
8668 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8669 TD->getTypeSize(AI.getAllocatedType()) == 0)
8670 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8675 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8676 Value *Op = FI.getOperand(0);
8678 // free undef -> unreachable.
8679 if (isa<UndefValue>(Op)) {
8680 // Insert a new store to null because we cannot modify the CFG here.
8681 new StoreInst(ConstantInt::getTrue(),
8682 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8683 return EraseInstFromFunction(FI);
8686 // If we have 'free null' delete the instruction. This can happen in stl code
8687 // when lots of inlining happens.
8688 if (isa<ConstantPointerNull>(Op))
8689 return EraseInstFromFunction(FI);
8691 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8692 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
8693 FI.setOperand(0, CI->getOperand(0));
8697 // Change free (gep X, 0,0,0,0) into free(X)
8698 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
8699 if (GEPI->hasAllZeroIndices()) {
8700 AddToWorkList(GEPI);
8701 FI.setOperand(0, GEPI->getOperand(0));
8706 // Change free(malloc) into nothing, if the malloc has a single use.
8707 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
8708 if (MI->hasOneUse()) {
8709 EraseInstFromFunction(FI);
8710 return EraseInstFromFunction(*MI);
8717 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8718 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8719 User *CI = cast<User>(LI.getOperand(0));
8720 Value *CastOp = CI->getOperand(0);
8722 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8723 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8724 const Type *SrcPTy = SrcTy->getElementType();
8726 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8727 isa<VectorType>(DestPTy)) {
8728 // If the source is an array, the code below will not succeed. Check to
8729 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8731 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8732 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8733 if (ASrcTy->getNumElements() != 0) {
8735 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8736 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8737 SrcTy = cast<PointerType>(CastOp->getType());
8738 SrcPTy = SrcTy->getElementType();
8741 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8742 isa<VectorType>(SrcPTy)) &&
8743 // Do not allow turning this into a load of an integer, which is then
8744 // casted to a pointer, this pessimizes pointer analysis a lot.
8745 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8746 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8747 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8749 // Okay, we are casting from one integer or pointer type to another of
8750 // the same size. Instead of casting the pointer before the load, cast
8751 // the result of the loaded value.
8752 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8754 LI.isVolatile()),LI);
8755 // Now cast the result of the load.
8756 return new BitCastInst(NewLoad, LI.getType());
8763 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8764 /// from this value cannot trap. If it is not obviously safe to load from the
8765 /// specified pointer, we do a quick local scan of the basic block containing
8766 /// ScanFrom, to determine if the address is already accessed.
8767 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8768 // If it is an alloca or global variable, it is always safe to load from.
8769 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8771 // Otherwise, be a little bit agressive by scanning the local block where we
8772 // want to check to see if the pointer is already being loaded or stored
8773 // from/to. If so, the previous load or store would have already trapped,
8774 // so there is no harm doing an extra load (also, CSE will later eliminate
8775 // the load entirely).
8776 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8781 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8782 if (LI->getOperand(0) == V) return true;
8783 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8784 if (SI->getOperand(1) == V) return true;
8790 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8791 Value *Op = LI.getOperand(0);
8793 // Attempt to improve the alignment.
8794 unsigned KnownAlign = GetKnownAlignment(Op, TD);
8795 if (KnownAlign > LI.getAlignment())
8796 LI.setAlignment(KnownAlign);
8798 // load (cast X) --> cast (load X) iff safe
8799 if (isa<CastInst>(Op))
8800 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8803 // None of the following transforms are legal for volatile loads.
8804 if (LI.isVolatile()) return 0;
8806 if (&LI.getParent()->front() != &LI) {
8807 BasicBlock::iterator BBI = &LI; --BBI;
8808 // If the instruction immediately before this is a store to the same
8809 // address, do a simple form of store->load forwarding.
8810 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8811 if (SI->getOperand(1) == LI.getOperand(0))
8812 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8813 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8814 if (LIB->getOperand(0) == LI.getOperand(0))
8815 return ReplaceInstUsesWith(LI, LIB);
8818 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8819 if (isa<ConstantPointerNull>(GEPI->getOperand(0))) {
8820 // Insert a new store to null instruction before the load to indicate
8821 // that this code is not reachable. We do this instead of inserting
8822 // an unreachable instruction directly because we cannot modify the
8824 new StoreInst(UndefValue::get(LI.getType()),
8825 Constant::getNullValue(Op->getType()), &LI);
8826 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8829 if (Constant *C = dyn_cast<Constant>(Op)) {
8830 // load null/undef -> undef
8831 if ((C->isNullValue() || isa<UndefValue>(C))) {
8832 // Insert a new store to null instruction before the load to indicate that
8833 // this code is not reachable. We do this instead of inserting an
8834 // unreachable instruction directly because we cannot modify the CFG.
8835 new StoreInst(UndefValue::get(LI.getType()),
8836 Constant::getNullValue(Op->getType()), &LI);
8837 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8840 // Instcombine load (constant global) into the value loaded.
8841 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8842 if (GV->isConstant() && !GV->isDeclaration())
8843 return ReplaceInstUsesWith(LI, GV->getInitializer());
8845 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8846 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8847 if (CE->getOpcode() == Instruction::GetElementPtr) {
8848 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8849 if (GV->isConstant() && !GV->isDeclaration())
8851 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8852 return ReplaceInstUsesWith(LI, V);
8853 if (CE->getOperand(0)->isNullValue()) {
8854 // Insert a new store to null instruction before the load to indicate
8855 // that this code is not reachable. We do this instead of inserting
8856 // an unreachable instruction directly because we cannot modify the
8858 new StoreInst(UndefValue::get(LI.getType()),
8859 Constant::getNullValue(Op->getType()), &LI);
8860 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8863 } else if (CE->isCast()) {
8864 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8869 if (Op->hasOneUse()) {
8870 // Change select and PHI nodes to select values instead of addresses: this
8871 // helps alias analysis out a lot, allows many others simplifications, and
8872 // exposes redundancy in the code.
8874 // Note that we cannot do the transformation unless we know that the
8875 // introduced loads cannot trap! Something like this is valid as long as
8876 // the condition is always false: load (select bool %C, int* null, int* %G),
8877 // but it would not be valid if we transformed it to load from null
8880 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8881 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8882 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8883 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8884 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8885 SI->getOperand(1)->getName()+".val"), LI);
8886 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8887 SI->getOperand(2)->getName()+".val"), LI);
8888 return new SelectInst(SI->getCondition(), V1, V2);
8891 // load (select (cond, null, P)) -> load P
8892 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8893 if (C->isNullValue()) {
8894 LI.setOperand(0, SI->getOperand(2));
8898 // load (select (cond, P, null)) -> load P
8899 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8900 if (C->isNullValue()) {
8901 LI.setOperand(0, SI->getOperand(1));
8909 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8911 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8912 User *CI = cast<User>(SI.getOperand(1));
8913 Value *CastOp = CI->getOperand(0);
8915 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8916 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8917 const Type *SrcPTy = SrcTy->getElementType();
8919 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8920 // If the source is an array, the code below will not succeed. Check to
8921 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8923 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8924 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8925 if (ASrcTy->getNumElements() != 0) {
8927 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8928 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8929 SrcTy = cast<PointerType>(CastOp->getType());
8930 SrcPTy = SrcTy->getElementType();
8933 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8934 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8935 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8937 // Okay, we are casting from one integer or pointer type to another of
8938 // the same size. Instead of casting the pointer before
8939 // the store, cast the value to be stored.
8941 Value *SIOp0 = SI.getOperand(0);
8942 Instruction::CastOps opcode = Instruction::BitCast;
8943 const Type* CastSrcTy = SIOp0->getType();
8944 const Type* CastDstTy = SrcPTy;
8945 if (isa<PointerType>(CastDstTy)) {
8946 if (CastSrcTy->isInteger())
8947 opcode = Instruction::IntToPtr;
8948 } else if (isa<IntegerType>(CastDstTy)) {
8949 if (isa<PointerType>(SIOp0->getType()))
8950 opcode = Instruction::PtrToInt;
8952 if (Constant *C = dyn_cast<Constant>(SIOp0))
8953 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8955 NewCast = IC.InsertNewInstBefore(
8956 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8958 return new StoreInst(NewCast, CastOp);
8965 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8966 Value *Val = SI.getOperand(0);
8967 Value *Ptr = SI.getOperand(1);
8969 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8970 EraseInstFromFunction(SI);
8975 // If the RHS is an alloca with a single use, zapify the store, making the
8977 if (Ptr->hasOneUse()) {
8978 if (isa<AllocaInst>(Ptr)) {
8979 EraseInstFromFunction(SI);
8984 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8985 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8986 GEP->getOperand(0)->hasOneUse()) {
8987 EraseInstFromFunction(SI);
8993 // Attempt to improve the alignment.
8994 unsigned KnownAlign = GetKnownAlignment(Ptr, TD);
8995 if (KnownAlign > SI.getAlignment())
8996 SI.setAlignment(KnownAlign);
8998 // Do really simple DSE, to catch cases where there are several consequtive
8999 // stores to the same location, separated by a few arithmetic operations. This
9000 // situation often occurs with bitfield accesses.
9001 BasicBlock::iterator BBI = &SI;
9002 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9006 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9007 // Prev store isn't volatile, and stores to the same location?
9008 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9011 EraseInstFromFunction(*PrevSI);
9017 // If this is a load, we have to stop. However, if the loaded value is from
9018 // the pointer we're loading and is producing the pointer we're storing,
9019 // then *this* store is dead (X = load P; store X -> P).
9020 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9021 if (LI == Val && LI->getOperand(0) == Ptr) {
9022 EraseInstFromFunction(SI);
9026 // Otherwise, this is a load from some other location. Stores before it
9031 // Don't skip over loads or things that can modify memory.
9032 if (BBI->mayWriteToMemory())
9037 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9039 // store X, null -> turns into 'unreachable' in SimplifyCFG
9040 if (isa<ConstantPointerNull>(Ptr)) {
9041 if (!isa<UndefValue>(Val)) {
9042 SI.setOperand(0, UndefValue::get(Val->getType()));
9043 if (Instruction *U = dyn_cast<Instruction>(Val))
9044 AddToWorkList(U); // Dropped a use.
9047 return 0; // Do not modify these!
9050 // store undef, Ptr -> noop
9051 if (isa<UndefValue>(Val)) {
9052 EraseInstFromFunction(SI);
9057 // If the pointer destination is a cast, see if we can fold the cast into the
9059 if (isa<CastInst>(Ptr))
9060 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9062 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9064 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9068 // If this store is the last instruction in the basic block, and if the block
9069 // ends with an unconditional branch, try to move it to the successor block.
9071 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9072 if (BI->isUnconditional())
9073 if (SimplifyStoreAtEndOfBlock(SI))
9074 return 0; // xform done!
9079 /// SimplifyStoreAtEndOfBlock - Turn things like:
9080 /// if () { *P = v1; } else { *P = v2 }
9081 /// into a phi node with a store in the successor.
9083 /// Simplify things like:
9084 /// *P = v1; if () { *P = v2; }
9085 /// into a phi node with a store in the successor.
9087 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9088 BasicBlock *StoreBB = SI.getParent();
9090 // Check to see if the successor block has exactly two incoming edges. If
9091 // so, see if the other predecessor contains a store to the same location.
9092 // if so, insert a PHI node (if needed) and move the stores down.
9093 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9095 // Determine whether Dest has exactly two predecessors and, if so, compute
9096 // the other predecessor.
9097 pred_iterator PI = pred_begin(DestBB);
9098 BasicBlock *OtherBB = 0;
9102 if (PI == pred_end(DestBB))
9105 if (*PI != StoreBB) {
9110 if (++PI != pred_end(DestBB))
9114 // Verify that the other block ends in a branch and is not otherwise empty.
9115 BasicBlock::iterator BBI = OtherBB->getTerminator();
9116 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9117 if (!OtherBr || BBI == OtherBB->begin())
9120 // If the other block ends in an unconditional branch, check for the 'if then
9121 // else' case. there is an instruction before the branch.
9122 StoreInst *OtherStore = 0;
9123 if (OtherBr->isUnconditional()) {
9124 // If this isn't a store, or isn't a store to the same location, bail out.
9126 OtherStore = dyn_cast<StoreInst>(BBI);
9127 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9130 // Otherwise, the other block ended with a conditional branch. If one of the
9131 // destinations is StoreBB, then we have the if/then case.
9132 if (OtherBr->getSuccessor(0) != StoreBB &&
9133 OtherBr->getSuccessor(1) != StoreBB)
9136 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9137 // if/then triangle. See if there is a store to the same ptr as SI that
9138 // lives in OtherBB.
9140 // Check to see if we find the matching store.
9141 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9142 if (OtherStore->getOperand(1) != SI.getOperand(1))
9146 // If we find something that may be using the stored value, or if we run
9147 // out of instructions, we can't do the xform.
9148 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9149 BBI == OtherBB->begin())
9153 // In order to eliminate the store in OtherBr, we have to
9154 // make sure nothing reads the stored value in StoreBB.
9155 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9156 // FIXME: This should really be AA driven.
9157 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9162 // Insert a PHI node now if we need it.
9163 Value *MergedVal = OtherStore->getOperand(0);
9164 if (MergedVal != SI.getOperand(0)) {
9165 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9166 PN->reserveOperandSpace(2);
9167 PN->addIncoming(SI.getOperand(0), SI.getParent());
9168 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9169 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9172 // Advance to a place where it is safe to insert the new store and
9174 BBI = DestBB->begin();
9175 while (isa<PHINode>(BBI)) ++BBI;
9176 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9177 OtherStore->isVolatile()), *BBI);
9179 // Nuke the old stores.
9180 EraseInstFromFunction(SI);
9181 EraseInstFromFunction(*OtherStore);
9187 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9188 // Change br (not X), label True, label False to: br X, label False, True
9190 BasicBlock *TrueDest;
9191 BasicBlock *FalseDest;
9192 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9193 !isa<Constant>(X)) {
9194 // Swap Destinations and condition...
9196 BI.setSuccessor(0, FalseDest);
9197 BI.setSuccessor(1, TrueDest);
9201 // Cannonicalize fcmp_one -> fcmp_oeq
9202 FCmpInst::Predicate FPred; Value *Y;
9203 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9204 TrueDest, FalseDest)))
9205 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9206 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9207 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9208 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9209 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9210 NewSCC->takeName(I);
9211 // Swap Destinations and condition...
9212 BI.setCondition(NewSCC);
9213 BI.setSuccessor(0, FalseDest);
9214 BI.setSuccessor(1, TrueDest);
9215 RemoveFromWorkList(I);
9216 I->eraseFromParent();
9217 AddToWorkList(NewSCC);
9221 // Cannonicalize icmp_ne -> icmp_eq
9222 ICmpInst::Predicate IPred;
9223 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9224 TrueDest, FalseDest)))
9225 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9226 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9227 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9228 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9229 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9230 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9231 NewSCC->takeName(I);
9232 // Swap Destinations and condition...
9233 BI.setCondition(NewSCC);
9234 BI.setSuccessor(0, FalseDest);
9235 BI.setSuccessor(1, TrueDest);
9236 RemoveFromWorkList(I);
9237 I->eraseFromParent();;
9238 AddToWorkList(NewSCC);
9245 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9246 Value *Cond = SI.getCondition();
9247 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9248 if (I->getOpcode() == Instruction::Add)
9249 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9250 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9251 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9252 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9254 SI.setOperand(0, I->getOperand(0));
9262 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9263 /// is to leave as a vector operation.
9264 static bool CheapToScalarize(Value *V, bool isConstant) {
9265 if (isa<ConstantAggregateZero>(V))
9267 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9268 if (isConstant) return true;
9269 // If all elts are the same, we can extract.
9270 Constant *Op0 = C->getOperand(0);
9271 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9272 if (C->getOperand(i) != Op0)
9276 Instruction *I = dyn_cast<Instruction>(V);
9277 if (!I) return false;
9279 // Insert element gets simplified to the inserted element or is deleted if
9280 // this is constant idx extract element and its a constant idx insertelt.
9281 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9282 isa<ConstantInt>(I->getOperand(2)))
9284 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9286 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9287 if (BO->hasOneUse() &&
9288 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9289 CheapToScalarize(BO->getOperand(1), isConstant)))
9291 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9292 if (CI->hasOneUse() &&
9293 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9294 CheapToScalarize(CI->getOperand(1), isConstant)))
9300 /// Read and decode a shufflevector mask.
9302 /// It turns undef elements into values that are larger than the number of
9303 /// elements in the input.
9304 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9305 unsigned NElts = SVI->getType()->getNumElements();
9306 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9307 return std::vector<unsigned>(NElts, 0);
9308 if (isa<UndefValue>(SVI->getOperand(2)))
9309 return std::vector<unsigned>(NElts, 2*NElts);
9311 std::vector<unsigned> Result;
9312 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9313 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9314 if (isa<UndefValue>(CP->getOperand(i)))
9315 Result.push_back(NElts*2); // undef -> 8
9317 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9321 /// FindScalarElement - Given a vector and an element number, see if the scalar
9322 /// value is already around as a register, for example if it were inserted then
9323 /// extracted from the vector.
9324 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9325 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9326 const VectorType *PTy = cast<VectorType>(V->getType());
9327 unsigned Width = PTy->getNumElements();
9328 if (EltNo >= Width) // Out of range access.
9329 return UndefValue::get(PTy->getElementType());
9331 if (isa<UndefValue>(V))
9332 return UndefValue::get(PTy->getElementType());
9333 else if (isa<ConstantAggregateZero>(V))
9334 return Constant::getNullValue(PTy->getElementType());
9335 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9336 return CP->getOperand(EltNo);
9337 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9338 // If this is an insert to a variable element, we don't know what it is.
9339 if (!isa<ConstantInt>(III->getOperand(2)))
9341 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9343 // If this is an insert to the element we are looking for, return the
9346 return III->getOperand(1);
9348 // Otherwise, the insertelement doesn't modify the value, recurse on its
9350 return FindScalarElement(III->getOperand(0), EltNo);
9351 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9352 unsigned InEl = getShuffleMask(SVI)[EltNo];
9354 return FindScalarElement(SVI->getOperand(0), InEl);
9355 else if (InEl < Width*2)
9356 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9358 return UndefValue::get(PTy->getElementType());
9361 // Otherwise, we don't know.
9365 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9367 // If vector val is undef, replace extract with scalar undef.
9368 if (isa<UndefValue>(EI.getOperand(0)))
9369 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9371 // If vector val is constant 0, replace extract with scalar 0.
9372 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9373 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9375 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9376 // If vector val is constant with uniform operands, replace EI
9377 // with that operand
9378 Constant *op0 = C->getOperand(0);
9379 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9380 if (C->getOperand(i) != op0) {
9385 return ReplaceInstUsesWith(EI, op0);
9388 // If extracting a specified index from the vector, see if we can recursively
9389 // find a previously computed scalar that was inserted into the vector.
9390 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9391 unsigned IndexVal = IdxC->getZExtValue();
9392 unsigned VectorWidth =
9393 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9395 // If this is extracting an invalid index, turn this into undef, to avoid
9396 // crashing the code below.
9397 if (IndexVal >= VectorWidth)
9398 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9400 // This instruction only demands the single element from the input vector.
9401 // If the input vector has a single use, simplify it based on this use
9403 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9405 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9408 EI.setOperand(0, V);
9413 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9414 return ReplaceInstUsesWith(EI, Elt);
9416 // If the this extractelement is directly using a bitcast from a vector of
9417 // the same number of elements, see if we can find the source element from
9418 // it. In this case, we will end up needing to bitcast the scalars.
9419 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9420 if (const VectorType *VT =
9421 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9422 if (VT->getNumElements() == VectorWidth)
9423 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9424 return new BitCastInst(Elt, EI.getType());
9428 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9429 if (I->hasOneUse()) {
9430 // Push extractelement into predecessor operation if legal and
9431 // profitable to do so
9432 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9433 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9434 if (CheapToScalarize(BO, isConstantElt)) {
9435 ExtractElementInst *newEI0 =
9436 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9437 EI.getName()+".lhs");
9438 ExtractElementInst *newEI1 =
9439 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9440 EI.getName()+".rhs");
9441 InsertNewInstBefore(newEI0, EI);
9442 InsertNewInstBefore(newEI1, EI);
9443 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9445 } else if (isa<LoadInst>(I)) {
9446 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9447 PointerType::get(EI.getType()), EI);
9448 GetElementPtrInst *GEP =
9449 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9450 InsertNewInstBefore(GEP, EI);
9451 return new LoadInst(GEP);
9454 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9455 // Extracting the inserted element?
9456 if (IE->getOperand(2) == EI.getOperand(1))
9457 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9458 // If the inserted and extracted elements are constants, they must not
9459 // be the same value, extract from the pre-inserted value instead.
9460 if (isa<Constant>(IE->getOperand(2)) &&
9461 isa<Constant>(EI.getOperand(1))) {
9462 AddUsesToWorkList(EI);
9463 EI.setOperand(0, IE->getOperand(0));
9466 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9467 // If this is extracting an element from a shufflevector, figure out where
9468 // it came from and extract from the appropriate input element instead.
9469 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9470 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9472 if (SrcIdx < SVI->getType()->getNumElements())
9473 Src = SVI->getOperand(0);
9474 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9475 SrcIdx -= SVI->getType()->getNumElements();
9476 Src = SVI->getOperand(1);
9478 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9480 return new ExtractElementInst(Src, SrcIdx);
9487 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9488 /// elements from either LHS or RHS, return the shuffle mask and true.
9489 /// Otherwise, return false.
9490 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9491 std::vector<Constant*> &Mask) {
9492 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9493 "Invalid CollectSingleShuffleElements");
9494 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9496 if (isa<UndefValue>(V)) {
9497 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9499 } else if (V == LHS) {
9500 for (unsigned i = 0; i != NumElts; ++i)
9501 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9503 } else if (V == RHS) {
9504 for (unsigned i = 0; i != NumElts; ++i)
9505 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9507 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9508 // If this is an insert of an extract from some other vector, include it.
9509 Value *VecOp = IEI->getOperand(0);
9510 Value *ScalarOp = IEI->getOperand(1);
9511 Value *IdxOp = IEI->getOperand(2);
9513 if (!isa<ConstantInt>(IdxOp))
9515 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9517 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9518 // Okay, we can handle this if the vector we are insertinting into is
9520 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9521 // If so, update the mask to reflect the inserted undef.
9522 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9525 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9526 if (isa<ConstantInt>(EI->getOperand(1)) &&
9527 EI->getOperand(0)->getType() == V->getType()) {
9528 unsigned ExtractedIdx =
9529 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9531 // This must be extracting from either LHS or RHS.
9532 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9533 // Okay, we can handle this if the vector we are insertinting into is
9535 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9536 // If so, update the mask to reflect the inserted value.
9537 if (EI->getOperand(0) == LHS) {
9538 Mask[InsertedIdx & (NumElts-1)] =
9539 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9541 assert(EI->getOperand(0) == RHS);
9542 Mask[InsertedIdx & (NumElts-1)] =
9543 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9552 // TODO: Handle shufflevector here!
9557 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9558 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9559 /// that computes V and the LHS value of the shuffle.
9560 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9562 assert(isa<VectorType>(V->getType()) &&
9563 (RHS == 0 || V->getType() == RHS->getType()) &&
9564 "Invalid shuffle!");
9565 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9567 if (isa<UndefValue>(V)) {
9568 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9570 } else if (isa<ConstantAggregateZero>(V)) {
9571 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9573 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9574 // If this is an insert of an extract from some other vector, include it.
9575 Value *VecOp = IEI->getOperand(0);
9576 Value *ScalarOp = IEI->getOperand(1);
9577 Value *IdxOp = IEI->getOperand(2);
9579 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9580 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9581 EI->getOperand(0)->getType() == V->getType()) {
9582 unsigned ExtractedIdx =
9583 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9584 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9586 // Either the extracted from or inserted into vector must be RHSVec,
9587 // otherwise we'd end up with a shuffle of three inputs.
9588 if (EI->getOperand(0) == RHS || RHS == 0) {
9589 RHS = EI->getOperand(0);
9590 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9591 Mask[InsertedIdx & (NumElts-1)] =
9592 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9597 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9598 // Everything but the extracted element is replaced with the RHS.
9599 for (unsigned i = 0; i != NumElts; ++i) {
9600 if (i != InsertedIdx)
9601 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9606 // If this insertelement is a chain that comes from exactly these two
9607 // vectors, return the vector and the effective shuffle.
9608 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9609 return EI->getOperand(0);
9614 // TODO: Handle shufflevector here!
9616 // Otherwise, can't do anything fancy. Return an identity vector.
9617 for (unsigned i = 0; i != NumElts; ++i)
9618 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9622 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9623 Value *VecOp = IE.getOperand(0);
9624 Value *ScalarOp = IE.getOperand(1);
9625 Value *IdxOp = IE.getOperand(2);
9627 // Inserting an undef or into an undefined place, remove this.
9628 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
9629 ReplaceInstUsesWith(IE, VecOp);
9631 // If the inserted element was extracted from some other vector, and if the
9632 // indexes are constant, try to turn this into a shufflevector operation.
9633 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9634 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9635 EI->getOperand(0)->getType() == IE.getType()) {
9636 unsigned NumVectorElts = IE.getType()->getNumElements();
9637 unsigned ExtractedIdx =
9638 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9639 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9641 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9642 return ReplaceInstUsesWith(IE, VecOp);
9644 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9645 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9647 // If we are extracting a value from a vector, then inserting it right
9648 // back into the same place, just use the input vector.
9649 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9650 return ReplaceInstUsesWith(IE, VecOp);
9652 // We could theoretically do this for ANY input. However, doing so could
9653 // turn chains of insertelement instructions into a chain of shufflevector
9654 // instructions, and right now we do not merge shufflevectors. As such,
9655 // only do this in a situation where it is clear that there is benefit.
9656 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9657 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9658 // the values of VecOp, except then one read from EIOp0.
9659 // Build a new shuffle mask.
9660 std::vector<Constant*> Mask;
9661 if (isa<UndefValue>(VecOp))
9662 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9664 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9665 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9668 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9669 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9670 ConstantVector::get(Mask));
9673 // If this insertelement isn't used by some other insertelement, turn it
9674 // (and any insertelements it points to), into one big shuffle.
9675 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9676 std::vector<Constant*> Mask;
9678 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9679 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9680 // We now have a shuffle of LHS, RHS, Mask.
9681 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9690 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9691 Value *LHS = SVI.getOperand(0);
9692 Value *RHS = SVI.getOperand(1);
9693 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9695 bool MadeChange = false;
9697 // Undefined shuffle mask -> undefined value.
9698 if (isa<UndefValue>(SVI.getOperand(2)))
9699 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9701 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9702 // the undef, change them to undefs.
9703 if (isa<UndefValue>(SVI.getOperand(1))) {
9704 // Scan to see if there are any references to the RHS. If so, replace them
9705 // with undef element refs and set MadeChange to true.
9706 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9707 if (Mask[i] >= e && Mask[i] != 2*e) {
9714 // Remap any references to RHS to use LHS.
9715 std::vector<Constant*> Elts;
9716 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9718 Elts.push_back(UndefValue::get(Type::Int32Ty));
9720 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9722 SVI.setOperand(2, ConstantVector::get(Elts));
9726 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9727 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9728 if (LHS == RHS || isa<UndefValue>(LHS)) {
9729 if (isa<UndefValue>(LHS) && LHS == RHS) {
9730 // shuffle(undef,undef,mask) -> undef.
9731 return ReplaceInstUsesWith(SVI, LHS);
9734 // Remap any references to RHS to use LHS.
9735 std::vector<Constant*> Elts;
9736 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9738 Elts.push_back(UndefValue::get(Type::Int32Ty));
9740 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9741 (Mask[i] < e && isa<UndefValue>(LHS)))
9742 Mask[i] = 2*e; // Turn into undef.
9744 Mask[i] &= (e-1); // Force to LHS.
9745 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9748 SVI.setOperand(0, SVI.getOperand(1));
9749 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9750 SVI.setOperand(2, ConstantVector::get(Elts));
9751 LHS = SVI.getOperand(0);
9752 RHS = SVI.getOperand(1);
9756 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9757 bool isLHSID = true, isRHSID = true;
9759 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9760 if (Mask[i] >= e*2) continue; // Ignore undef values.
9761 // Is this an identity shuffle of the LHS value?
9762 isLHSID &= (Mask[i] == i);
9764 // Is this an identity shuffle of the RHS value?
9765 isRHSID &= (Mask[i]-e == i);
9768 // Eliminate identity shuffles.
9769 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9770 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9772 // If the LHS is a shufflevector itself, see if we can combine it with this
9773 // one without producing an unusual shuffle. Here we are really conservative:
9774 // we are absolutely afraid of producing a shuffle mask not in the input
9775 // program, because the code gen may not be smart enough to turn a merged
9776 // shuffle into two specific shuffles: it may produce worse code. As such,
9777 // we only merge two shuffles if the result is one of the two input shuffle
9778 // masks. In this case, merging the shuffles just removes one instruction,
9779 // which we know is safe. This is good for things like turning:
9780 // (splat(splat)) -> splat.
9781 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9782 if (isa<UndefValue>(RHS)) {
9783 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9785 std::vector<unsigned> NewMask;
9786 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9788 NewMask.push_back(2*e);
9790 NewMask.push_back(LHSMask[Mask[i]]);
9792 // If the result mask is equal to the src shuffle or this shuffle mask, do
9794 if (NewMask == LHSMask || NewMask == Mask) {
9795 std::vector<Constant*> Elts;
9796 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9797 if (NewMask[i] >= e*2) {
9798 Elts.push_back(UndefValue::get(Type::Int32Ty));
9800 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9803 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9804 LHSSVI->getOperand(1),
9805 ConstantVector::get(Elts));
9810 return MadeChange ? &SVI : 0;
9816 /// TryToSinkInstruction - Try to move the specified instruction from its
9817 /// current block into the beginning of DestBlock, which can only happen if it's
9818 /// safe to move the instruction past all of the instructions between it and the
9819 /// end of its block.
9820 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9821 assert(I->hasOneUse() && "Invariants didn't hold!");
9823 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9824 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9826 // Do not sink alloca instructions out of the entry block.
9827 if (isa<AllocaInst>(I) && I->getParent() ==
9828 &DestBlock->getParent()->getEntryBlock())
9831 // We can only sink load instructions if there is nothing between the load and
9832 // the end of block that could change the value.
9833 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9834 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9836 if (Scan->mayWriteToMemory())
9840 BasicBlock::iterator InsertPos = DestBlock->begin();
9841 while (isa<PHINode>(InsertPos)) ++InsertPos;
9843 I->moveBefore(InsertPos);
9849 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9850 /// all reachable code to the worklist.
9852 /// This has a couple of tricks to make the code faster and more powerful. In
9853 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9854 /// them to the worklist (this significantly speeds up instcombine on code where
9855 /// many instructions are dead or constant). Additionally, if we find a branch
9856 /// whose condition is a known constant, we only visit the reachable successors.
9858 static void AddReachableCodeToWorklist(BasicBlock *BB,
9859 SmallPtrSet<BasicBlock*, 64> &Visited,
9861 const TargetData *TD) {
9862 std::vector<BasicBlock*> Worklist;
9863 Worklist.push_back(BB);
9865 while (!Worklist.empty()) {
9866 BB = Worklist.back();
9867 Worklist.pop_back();
9869 // We have now visited this block! If we've already been here, ignore it.
9870 if (!Visited.insert(BB)) continue;
9872 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9873 Instruction *Inst = BBI++;
9875 // DCE instruction if trivially dead.
9876 if (isInstructionTriviallyDead(Inst)) {
9878 DOUT << "IC: DCE: " << *Inst;
9879 Inst->eraseFromParent();
9883 // ConstantProp instruction if trivially constant.
9884 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9885 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9886 Inst->replaceAllUsesWith(C);
9888 Inst->eraseFromParent();
9892 IC.AddToWorkList(Inst);
9895 // Recursively visit successors. If this is a branch or switch on a
9896 // constant, only visit the reachable successor.
9897 TerminatorInst *TI = BB->getTerminator();
9898 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9899 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9900 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9901 Worklist.push_back(BI->getSuccessor(!CondVal));
9904 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9905 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9906 // See if this is an explicit destination.
9907 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9908 if (SI->getCaseValue(i) == Cond) {
9909 Worklist.push_back(SI->getSuccessor(i));
9913 // Otherwise it is the default destination.
9914 Worklist.push_back(SI->getSuccessor(0));
9919 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9920 Worklist.push_back(TI->getSuccessor(i));
9924 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9925 bool Changed = false;
9926 TD = &getAnalysis<TargetData>();
9928 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9929 << F.getNameStr() << "\n");
9932 // Do a depth-first traversal of the function, populate the worklist with
9933 // the reachable instructions. Ignore blocks that are not reachable. Keep
9934 // track of which blocks we visit.
9935 SmallPtrSet<BasicBlock*, 64> Visited;
9936 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9938 // Do a quick scan over the function. If we find any blocks that are
9939 // unreachable, remove any instructions inside of them. This prevents
9940 // the instcombine code from having to deal with some bad special cases.
9941 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9942 if (!Visited.count(BB)) {
9943 Instruction *Term = BB->getTerminator();
9944 while (Term != BB->begin()) { // Remove instrs bottom-up
9945 BasicBlock::iterator I = Term; --I;
9947 DOUT << "IC: DCE: " << *I;
9950 if (!I->use_empty())
9951 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9952 I->eraseFromParent();
9957 while (!Worklist.empty()) {
9958 Instruction *I = RemoveOneFromWorkList();
9959 if (I == 0) continue; // skip null values.
9961 // Check to see if we can DCE the instruction.
9962 if (isInstructionTriviallyDead(I)) {
9963 // Add operands to the worklist.
9964 if (I->getNumOperands() < 4)
9965 AddUsesToWorkList(*I);
9968 DOUT << "IC: DCE: " << *I;
9970 I->eraseFromParent();
9971 RemoveFromWorkList(I);
9975 // Instruction isn't dead, see if we can constant propagate it.
9976 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9977 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9979 // Add operands to the worklist.
9980 AddUsesToWorkList(*I);
9981 ReplaceInstUsesWith(*I, C);
9984 I->eraseFromParent();
9985 RemoveFromWorkList(I);
9989 // See if we can trivially sink this instruction to a successor basic block.
9990 if (I->hasOneUse()) {
9991 BasicBlock *BB = I->getParent();
9992 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9993 if (UserParent != BB) {
9994 bool UserIsSuccessor = false;
9995 // See if the user is one of our successors.
9996 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9997 if (*SI == UserParent) {
9998 UserIsSuccessor = true;
10002 // If the user is one of our immediate successors, and if that successor
10003 // only has us as a predecessors (we'd have to split the critical edge
10004 // otherwise), we can keep going.
10005 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10006 next(pred_begin(UserParent)) == pred_end(UserParent))
10007 // Okay, the CFG is simple enough, try to sink this instruction.
10008 Changed |= TryToSinkInstruction(I, UserParent);
10012 // Now that we have an instruction, try combining it to simplify it...
10016 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10017 if (Instruction *Result = visit(*I)) {
10019 // Should we replace the old instruction with a new one?
10021 DOUT << "IC: Old = " << *I
10022 << " New = " << *Result;
10024 // Everything uses the new instruction now.
10025 I->replaceAllUsesWith(Result);
10027 // Push the new instruction and any users onto the worklist.
10028 AddToWorkList(Result);
10029 AddUsersToWorkList(*Result);
10031 // Move the name to the new instruction first.
10032 Result->takeName(I);
10034 // Insert the new instruction into the basic block...
10035 BasicBlock *InstParent = I->getParent();
10036 BasicBlock::iterator InsertPos = I;
10038 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10039 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10042 InstParent->getInstList().insert(InsertPos, Result);
10044 // Make sure that we reprocess all operands now that we reduced their
10046 AddUsesToWorkList(*I);
10048 // Instructions can end up on the worklist more than once. Make sure
10049 // we do not process an instruction that has been deleted.
10050 RemoveFromWorkList(I);
10052 // Erase the old instruction.
10053 InstParent->getInstList().erase(I);
10056 DOUT << "IC: Mod = " << OrigI
10057 << " New = " << *I;
10060 // If the instruction was modified, it's possible that it is now dead.
10061 // if so, remove it.
10062 if (isInstructionTriviallyDead(I)) {
10063 // Make sure we process all operands now that we are reducing their
10065 AddUsesToWorkList(*I);
10067 // Instructions may end up in the worklist more than once. Erase all
10068 // occurrences of this instruction.
10069 RemoveFromWorkList(I);
10070 I->eraseFromParent();
10073 AddUsersToWorkList(*I);
10080 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10085 bool InstCombiner::runOnFunction(Function &F) {
10086 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10088 bool EverMadeChange = false;
10090 // Iterate while there is work to do.
10091 unsigned Iteration = 0;
10092 while (DoOneIteration(F, Iteration++))
10093 EverMadeChange = true;
10094 return EverMadeChange;
10097 FunctionPass *llvm::createInstructionCombiningPass() {
10098 return new InstCombiner();