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,
875 APInt &Min, APInt &Max) {
876 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
877 assert(KnownZero.getBitWidth() == BitWidth &&
878 KnownOne.getBitWidth() == BitWidth &&
879 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
880 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
881 APInt UnknownBits = ~(KnownZero|KnownOne);
883 // The minimum value is when the unknown bits are all zeros.
885 // The maximum value is when the unknown bits are all ones.
886 Max = KnownOne|UnknownBits;
889 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
890 /// value based on the demanded bits. When this function is called, it is known
891 /// that only the bits set in DemandedMask of the result of V are ever used
892 /// downstream. Consequently, depending on the mask and V, it may be possible
893 /// to replace V with a constant or one of its operands. In such cases, this
894 /// function does the replacement and returns true. In all other cases, it
895 /// returns false after analyzing the expression and setting KnownOne and known
896 /// to be one in the expression. KnownZero contains all the bits that are known
897 /// to be zero in the expression. These are provided to potentially allow the
898 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
899 /// the expression. KnownOne and KnownZero always follow the invariant that
900 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
901 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
902 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
903 /// and KnownOne must all be the same.
904 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
905 APInt& KnownZero, APInt& KnownOne,
907 assert(V != 0 && "Null pointer of Value???");
908 assert(Depth <= 6 && "Limit Search Depth");
909 uint32_t BitWidth = DemandedMask.getBitWidth();
910 const IntegerType *VTy = cast<IntegerType>(V->getType());
911 assert(VTy->getBitWidth() == BitWidth &&
912 KnownZero.getBitWidth() == BitWidth &&
913 KnownOne.getBitWidth() == BitWidth &&
914 "Value *V, DemandedMask, KnownZero and KnownOne \
915 must have same BitWidth");
916 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
917 // We know all of the bits for a constant!
918 KnownOne = CI->getValue() & DemandedMask;
919 KnownZero = ~KnownOne & DemandedMask;
925 if (!V->hasOneUse()) { // Other users may use these bits.
926 if (Depth != 0) { // Not at the root.
927 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
928 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
931 // If this is the root being simplified, allow it to have multiple uses,
932 // just set the DemandedMask to all bits.
933 DemandedMask = APInt::getAllOnesValue(BitWidth);
934 } else if (DemandedMask == 0) { // Not demanding any bits from V.
935 if (V != UndefValue::get(VTy))
936 return UpdateValueUsesWith(V, UndefValue::get(VTy));
938 } else if (Depth == 6) { // Limit search depth.
942 Instruction *I = dyn_cast<Instruction>(V);
943 if (!I) return false; // Only analyze instructions.
945 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
946 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
947 switch (I->getOpcode()) {
949 case Instruction::And:
950 // If either the LHS or the RHS are Zero, the result is zero.
951 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
952 RHSKnownZero, RHSKnownOne, Depth+1))
954 assert((RHSKnownZero & RHSKnownOne) == 0 &&
955 "Bits known to be one AND zero?");
957 // If something is known zero on the RHS, the bits aren't demanded on the
959 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
960 LHSKnownZero, LHSKnownOne, Depth+1))
962 assert((LHSKnownZero & LHSKnownOne) == 0 &&
963 "Bits known to be one AND zero?");
965 // If all of the demanded bits are known 1 on one side, return the other.
966 // These bits cannot contribute to the result of the 'and'.
967 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
968 (DemandedMask & ~LHSKnownZero))
969 return UpdateValueUsesWith(I, I->getOperand(0));
970 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
971 (DemandedMask & ~RHSKnownZero))
972 return UpdateValueUsesWith(I, I->getOperand(1));
974 // If all of the demanded bits in the inputs are known zeros, return zero.
975 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
976 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
978 // If the RHS is a constant, see if we can simplify it.
979 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
980 return UpdateValueUsesWith(I, I);
982 // Output known-1 bits are only known if set in both the LHS & RHS.
983 RHSKnownOne &= LHSKnownOne;
984 // Output known-0 are known to be clear if zero in either the LHS | RHS.
985 RHSKnownZero |= LHSKnownZero;
987 case Instruction::Or:
988 // If either the LHS or the RHS are One, the result is One.
989 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
990 RHSKnownZero, RHSKnownOne, Depth+1))
992 assert((RHSKnownZero & RHSKnownOne) == 0 &&
993 "Bits known to be one AND zero?");
994 // If something is known one on the RHS, the bits aren't demanded on the
996 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
997 LHSKnownZero, LHSKnownOne, Depth+1))
999 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1000 "Bits known to be one AND zero?");
1002 // If all of the demanded bits are known zero on one side, return the other.
1003 // These bits cannot contribute to the result of the 'or'.
1004 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1005 (DemandedMask & ~LHSKnownOne))
1006 return UpdateValueUsesWith(I, I->getOperand(0));
1007 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1008 (DemandedMask & ~RHSKnownOne))
1009 return UpdateValueUsesWith(I, I->getOperand(1));
1011 // If all of the potentially set bits on one side are known to be set on
1012 // the other side, just use the 'other' side.
1013 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1014 (DemandedMask & (~RHSKnownZero)))
1015 return UpdateValueUsesWith(I, I->getOperand(0));
1016 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1017 (DemandedMask & (~LHSKnownZero)))
1018 return UpdateValueUsesWith(I, I->getOperand(1));
1020 // If the RHS is a constant, see if we can simplify it.
1021 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1022 return UpdateValueUsesWith(I, I);
1024 // Output known-0 bits are only known if clear in both the LHS & RHS.
1025 RHSKnownZero &= LHSKnownZero;
1026 // Output known-1 are known to be set if set in either the LHS | RHS.
1027 RHSKnownOne |= LHSKnownOne;
1029 case Instruction::Xor: {
1030 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1031 RHSKnownZero, RHSKnownOne, Depth+1))
1033 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1034 "Bits known to be one AND zero?");
1035 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1036 LHSKnownZero, LHSKnownOne, Depth+1))
1038 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1039 "Bits known to be one AND zero?");
1041 // If all of the demanded bits are known zero on one side, return the other.
1042 // These bits cannot contribute to the result of the 'xor'.
1043 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1044 return UpdateValueUsesWith(I, I->getOperand(0));
1045 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1046 return UpdateValueUsesWith(I, I->getOperand(1));
1048 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1049 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1050 (RHSKnownOne & LHSKnownOne);
1051 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1052 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1053 (RHSKnownOne & LHSKnownZero);
1055 // If all of the demanded bits are known to be zero on one side or the
1056 // other, turn this into an *inclusive* or.
1057 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1058 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1060 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1062 InsertNewInstBefore(Or, *I);
1063 return UpdateValueUsesWith(I, Or);
1066 // If all of the demanded bits on one side are known, and all of the set
1067 // bits on that side are also known to be set on the other side, turn this
1068 // into an AND, as we know the bits will be cleared.
1069 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1070 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1072 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1073 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1075 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1076 InsertNewInstBefore(And, *I);
1077 return UpdateValueUsesWith(I, And);
1081 // If the RHS is a constant, see if we can simplify it.
1082 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1083 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1084 return UpdateValueUsesWith(I, I);
1086 RHSKnownZero = KnownZeroOut;
1087 RHSKnownOne = KnownOneOut;
1090 case Instruction::Select:
1091 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1092 RHSKnownZero, RHSKnownOne, Depth+1))
1094 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1095 LHSKnownZero, LHSKnownOne, Depth+1))
1097 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1098 "Bits known to be one AND zero?");
1099 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1100 "Bits known to be one AND zero?");
1102 // If the operands are constants, see if we can simplify them.
1103 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1104 return UpdateValueUsesWith(I, I);
1105 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1106 return UpdateValueUsesWith(I, I);
1108 // Only known if known in both the LHS and RHS.
1109 RHSKnownOne &= LHSKnownOne;
1110 RHSKnownZero &= LHSKnownZero;
1112 case Instruction::Trunc: {
1114 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1115 DemandedMask.zext(truncBf);
1116 RHSKnownZero.zext(truncBf);
1117 RHSKnownOne.zext(truncBf);
1118 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1119 RHSKnownZero, RHSKnownOne, Depth+1))
1121 DemandedMask.trunc(BitWidth);
1122 RHSKnownZero.trunc(BitWidth);
1123 RHSKnownOne.trunc(BitWidth);
1124 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1125 "Bits known to be one AND zero?");
1128 case Instruction::BitCast:
1129 if (!I->getOperand(0)->getType()->isInteger())
1132 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1133 RHSKnownZero, RHSKnownOne, Depth+1))
1135 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1136 "Bits known to be one AND zero?");
1138 case Instruction::ZExt: {
1139 // Compute the bits in the result that are not present in the input.
1140 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1141 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1143 DemandedMask.trunc(SrcBitWidth);
1144 RHSKnownZero.trunc(SrcBitWidth);
1145 RHSKnownOne.trunc(SrcBitWidth);
1146 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1147 RHSKnownZero, RHSKnownOne, Depth+1))
1149 DemandedMask.zext(BitWidth);
1150 RHSKnownZero.zext(BitWidth);
1151 RHSKnownOne.zext(BitWidth);
1152 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1153 "Bits known to be one AND zero?");
1154 // The top bits are known to be zero.
1155 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1158 case Instruction::SExt: {
1159 // Compute the bits in the result that are not present in the input.
1160 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1161 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1163 APInt InputDemandedBits = DemandedMask &
1164 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1166 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1167 // If any of the sign extended bits are demanded, we know that the sign
1169 if ((NewBits & DemandedMask) != 0)
1170 InputDemandedBits.set(SrcBitWidth-1);
1172 InputDemandedBits.trunc(SrcBitWidth);
1173 RHSKnownZero.trunc(SrcBitWidth);
1174 RHSKnownOne.trunc(SrcBitWidth);
1175 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1176 RHSKnownZero, RHSKnownOne, Depth+1))
1178 InputDemandedBits.zext(BitWidth);
1179 RHSKnownZero.zext(BitWidth);
1180 RHSKnownOne.zext(BitWidth);
1181 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1182 "Bits known to be one AND zero?");
1184 // If the sign bit of the input is known set or clear, then we know the
1185 // top bits of the result.
1187 // If the input sign bit is known zero, or if the NewBits are not demanded
1188 // convert this into a zero extension.
1189 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1191 // Convert to ZExt cast
1192 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1193 return UpdateValueUsesWith(I, NewCast);
1194 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1195 RHSKnownOne |= NewBits;
1199 case Instruction::Add: {
1200 // Figure out what the input bits are. If the top bits of the and result
1201 // are not demanded, then the add doesn't demand them from its input
1203 uint32_t NLZ = DemandedMask.countLeadingZeros();
1205 // If there is a constant on the RHS, there are a variety of xformations
1207 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1208 // If null, this should be simplified elsewhere. Some of the xforms here
1209 // won't work if the RHS is zero.
1213 // If the top bit of the output is demanded, demand everything from the
1214 // input. Otherwise, we demand all the input bits except NLZ top bits.
1215 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1217 // Find information about known zero/one bits in the input.
1218 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1219 LHSKnownZero, LHSKnownOne, Depth+1))
1222 // If the RHS of the add has bits set that can't affect the input, reduce
1224 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1225 return UpdateValueUsesWith(I, I);
1227 // Avoid excess work.
1228 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1231 // Turn it into OR if input bits are zero.
1232 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1234 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1236 InsertNewInstBefore(Or, *I);
1237 return UpdateValueUsesWith(I, Or);
1240 // We can say something about the output known-zero and known-one bits,
1241 // depending on potential carries from the input constant and the
1242 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1243 // bits set and the RHS constant is 0x01001, then we know we have a known
1244 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1246 // To compute this, we first compute the potential carry bits. These are
1247 // the bits which may be modified. I'm not aware of a better way to do
1249 const APInt& RHSVal = RHS->getValue();
1250 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1252 // Now that we know which bits have carries, compute the known-1/0 sets.
1254 // Bits are known one if they are known zero in one operand and one in the
1255 // other, and there is no input carry.
1256 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1257 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1259 // Bits are known zero if they are known zero in both operands and there
1260 // is no input carry.
1261 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1263 // If the high-bits of this ADD are not demanded, then it does not demand
1264 // the high bits of its LHS or RHS.
1265 if (DemandedMask[BitWidth-1] == 0) {
1266 // Right fill the mask of bits for this ADD to demand the most
1267 // significant bit and all those below it.
1268 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1269 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1270 LHSKnownZero, LHSKnownOne, Depth+1))
1272 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1273 LHSKnownZero, LHSKnownOne, Depth+1))
1279 case Instruction::Sub:
1280 // If the high-bits of this SUB are not demanded, then it does not demand
1281 // the high bits of its LHS or RHS.
1282 if (DemandedMask[BitWidth-1] == 0) {
1283 // Right fill the mask of bits for this SUB to demand the most
1284 // significant bit and all those below it.
1285 uint32_t NLZ = DemandedMask.countLeadingZeros();
1286 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1287 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1288 LHSKnownZero, LHSKnownOne, Depth+1))
1290 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1291 LHSKnownZero, LHSKnownOne, Depth+1))
1295 case Instruction::Shl:
1296 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1297 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1298 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1299 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1300 RHSKnownZero, RHSKnownOne, Depth+1))
1302 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1303 "Bits known to be one AND zero?");
1304 RHSKnownZero <<= ShiftAmt;
1305 RHSKnownOne <<= ShiftAmt;
1306 // low bits known zero.
1308 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1311 case Instruction::LShr:
1312 // For a logical shift right
1313 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1314 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1316 // Unsigned shift right.
1317 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1318 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1319 RHSKnownZero, RHSKnownOne, Depth+1))
1321 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1322 "Bits known to be one AND zero?");
1323 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1324 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1326 // Compute the new bits that are at the top now.
1327 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1328 RHSKnownZero |= HighBits; // high bits known zero.
1332 case Instruction::AShr:
1333 // If this is an arithmetic shift right and only the low-bit is set, we can
1334 // always convert this into a logical shr, even if the shift amount is
1335 // variable. The low bit of the shift cannot be an input sign bit unless
1336 // the shift amount is >= the size of the datatype, which is undefined.
1337 if (DemandedMask == 1) {
1338 // Perform the logical shift right.
1339 Value *NewVal = BinaryOperator::createLShr(
1340 I->getOperand(0), I->getOperand(1), I->getName());
1341 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1342 return UpdateValueUsesWith(I, NewVal);
1345 // If the sign bit is the only bit demanded by this ashr, then there is no
1346 // need to do it, the shift doesn't change the high bit.
1347 if (DemandedMask.isSignBit())
1348 return UpdateValueUsesWith(I, I->getOperand(0));
1350 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1351 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1353 // Signed shift right.
1354 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1355 // If any of the "high bits" are demanded, we should set the sign bit as
1357 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1358 DemandedMaskIn.set(BitWidth-1);
1359 if (SimplifyDemandedBits(I->getOperand(0),
1361 RHSKnownZero, RHSKnownOne, Depth+1))
1363 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1364 "Bits known to be one AND zero?");
1365 // Compute the new bits that are at the top now.
1366 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1367 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1368 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1370 // Handle the sign bits.
1371 APInt SignBit(APInt::getSignBit(BitWidth));
1372 // Adjust to where it is now in the mask.
1373 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1375 // If the input sign bit is known to be zero, or if none of the top bits
1376 // are demanded, turn this into an unsigned shift right.
1377 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1378 (HighBits & ~DemandedMask) == HighBits) {
1379 // Perform the logical shift right.
1380 Value *NewVal = BinaryOperator::createLShr(
1381 I->getOperand(0), SA, I->getName());
1382 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1383 return UpdateValueUsesWith(I, NewVal);
1384 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1385 RHSKnownOne |= HighBits;
1391 // If the client is only demanding bits that we know, return the known
1393 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1394 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1399 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1400 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1401 /// actually used by the caller. This method analyzes which elements of the
1402 /// operand are undef and returns that information in UndefElts.
1404 /// If the information about demanded elements can be used to simplify the
1405 /// operation, the operation is simplified, then the resultant value is
1406 /// returned. This returns null if no change was made.
1407 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1408 uint64_t &UndefElts,
1410 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1411 assert(VWidth <= 64 && "Vector too wide to analyze!");
1412 uint64_t EltMask = ~0ULL >> (64-VWidth);
1413 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1414 "Invalid DemandedElts!");
1416 if (isa<UndefValue>(V)) {
1417 // If the entire vector is undefined, just return this info.
1418 UndefElts = EltMask;
1420 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1421 UndefElts = EltMask;
1422 return UndefValue::get(V->getType());
1426 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1427 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1428 Constant *Undef = UndefValue::get(EltTy);
1430 std::vector<Constant*> Elts;
1431 for (unsigned i = 0; i != VWidth; ++i)
1432 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1433 Elts.push_back(Undef);
1434 UndefElts |= (1ULL << i);
1435 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1436 Elts.push_back(Undef);
1437 UndefElts |= (1ULL << i);
1438 } else { // Otherwise, defined.
1439 Elts.push_back(CP->getOperand(i));
1442 // If we changed the constant, return it.
1443 Constant *NewCP = ConstantVector::get(Elts);
1444 return NewCP != CP ? NewCP : 0;
1445 } else if (isa<ConstantAggregateZero>(V)) {
1446 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1448 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1449 Constant *Zero = Constant::getNullValue(EltTy);
1450 Constant *Undef = UndefValue::get(EltTy);
1451 std::vector<Constant*> Elts;
1452 for (unsigned i = 0; i != VWidth; ++i)
1453 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1454 UndefElts = DemandedElts ^ EltMask;
1455 return ConstantVector::get(Elts);
1458 if (!V->hasOneUse()) { // Other users may use these bits.
1459 if (Depth != 0) { // Not at the root.
1460 // TODO: Just compute the UndefElts information recursively.
1464 } else if (Depth == 10) { // Limit search depth.
1468 Instruction *I = dyn_cast<Instruction>(V);
1469 if (!I) return false; // Only analyze instructions.
1471 bool MadeChange = false;
1472 uint64_t UndefElts2;
1474 switch (I->getOpcode()) {
1477 case Instruction::InsertElement: {
1478 // If this is a variable index, we don't know which element it overwrites.
1479 // demand exactly the same input as we produce.
1480 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1482 // Note that we can't propagate undef elt info, because we don't know
1483 // which elt is getting updated.
1484 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1485 UndefElts2, Depth+1);
1486 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1490 // If this is inserting an element that isn't demanded, remove this
1492 unsigned IdxNo = Idx->getZExtValue();
1493 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1494 return AddSoonDeadInstToWorklist(*I, 0);
1496 // Otherwise, the element inserted overwrites whatever was there, so the
1497 // input demanded set is simpler than the output set.
1498 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1499 DemandedElts & ~(1ULL << IdxNo),
1500 UndefElts, Depth+1);
1501 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1503 // The inserted element is defined.
1504 UndefElts |= 1ULL << IdxNo;
1507 case Instruction::BitCast: {
1508 // Vector->vector casts only.
1509 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1511 unsigned InVWidth = VTy->getNumElements();
1512 uint64_t InputDemandedElts = 0;
1515 if (VWidth == InVWidth) {
1516 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1517 // elements as are demanded of us.
1519 InputDemandedElts = DemandedElts;
1520 } else if (VWidth > InVWidth) {
1524 // If there are more elements in the result than there are in the source,
1525 // then an input element is live if any of the corresponding output
1526 // elements are live.
1527 Ratio = VWidth/InVWidth;
1528 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1529 if (DemandedElts & (1ULL << OutIdx))
1530 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1536 // If there are more elements in the source than there are in the result,
1537 // then an input element is live if the corresponding output element is
1539 Ratio = InVWidth/VWidth;
1540 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1541 if (DemandedElts & (1ULL << InIdx/Ratio))
1542 InputDemandedElts |= 1ULL << InIdx;
1545 // div/rem demand all inputs, because they don't want divide by zero.
1546 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1547 UndefElts2, Depth+1);
1549 I->setOperand(0, TmpV);
1553 UndefElts = UndefElts2;
1554 if (VWidth > InVWidth) {
1555 assert(0 && "Unimp");
1556 // If there are more elements in the result than there are in the source,
1557 // then an output element is undef if the corresponding input element is
1559 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1560 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1561 UndefElts |= 1ULL << OutIdx;
1562 } else if (VWidth < InVWidth) {
1563 assert(0 && "Unimp");
1564 // If there are more elements in the source than there are in the result,
1565 // then a result element is undef if all of the corresponding input
1566 // elements are undef.
1567 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1568 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1569 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1570 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1574 case Instruction::And:
1575 case Instruction::Or:
1576 case Instruction::Xor:
1577 case Instruction::Add:
1578 case Instruction::Sub:
1579 case Instruction::Mul:
1580 // div/rem demand all inputs, because they don't want divide by zero.
1581 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1582 UndefElts, Depth+1);
1583 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1584 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1585 UndefElts2, Depth+1);
1586 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1588 // Output elements are undefined if both are undefined. Consider things
1589 // like undef&0. The result is known zero, not undef.
1590 UndefElts &= UndefElts2;
1593 case Instruction::Call: {
1594 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1596 switch (II->getIntrinsicID()) {
1599 // Binary vector operations that work column-wise. A dest element is a
1600 // function of the corresponding input elements from the two inputs.
1601 case Intrinsic::x86_sse_sub_ss:
1602 case Intrinsic::x86_sse_mul_ss:
1603 case Intrinsic::x86_sse_min_ss:
1604 case Intrinsic::x86_sse_max_ss:
1605 case Intrinsic::x86_sse2_sub_sd:
1606 case Intrinsic::x86_sse2_mul_sd:
1607 case Intrinsic::x86_sse2_min_sd:
1608 case Intrinsic::x86_sse2_max_sd:
1609 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1610 UndefElts, Depth+1);
1611 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1612 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1613 UndefElts2, Depth+1);
1614 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1616 // If only the low elt is demanded and this is a scalarizable intrinsic,
1617 // scalarize it now.
1618 if (DemandedElts == 1) {
1619 switch (II->getIntrinsicID()) {
1621 case Intrinsic::x86_sse_sub_ss:
1622 case Intrinsic::x86_sse_mul_ss:
1623 case Intrinsic::x86_sse2_sub_sd:
1624 case Intrinsic::x86_sse2_mul_sd:
1625 // TODO: Lower MIN/MAX/ABS/etc
1626 Value *LHS = II->getOperand(1);
1627 Value *RHS = II->getOperand(2);
1628 // Extract the element as scalars.
1629 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1630 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1632 switch (II->getIntrinsicID()) {
1633 default: assert(0 && "Case stmts out of sync!");
1634 case Intrinsic::x86_sse_sub_ss:
1635 case Intrinsic::x86_sse2_sub_sd:
1636 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1637 II->getName()), *II);
1639 case Intrinsic::x86_sse_mul_ss:
1640 case Intrinsic::x86_sse2_mul_sd:
1641 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1642 II->getName()), *II);
1647 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1649 InsertNewInstBefore(New, *II);
1650 AddSoonDeadInstToWorklist(*II, 0);
1655 // Output elements are undefined if both are undefined. Consider things
1656 // like undef&0. The result is known zero, not undef.
1657 UndefElts &= UndefElts2;
1663 return MadeChange ? I : 0;
1666 /// @returns true if the specified compare predicate is
1667 /// true when both operands are equal...
1668 /// @brief Determine if the icmp Predicate is true when both operands are equal
1669 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1670 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1671 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1672 pred == ICmpInst::ICMP_SLE;
1675 /// @returns true if the specified compare instruction is
1676 /// true when both operands are equal...
1677 /// @brief Determine if the ICmpInst returns true when both operands are equal
1678 static bool isTrueWhenEqual(ICmpInst &ICI) {
1679 return isTrueWhenEqual(ICI.getPredicate());
1682 /// AssociativeOpt - Perform an optimization on an associative operator. This
1683 /// function is designed to check a chain of associative operators for a
1684 /// potential to apply a certain optimization. Since the optimization may be
1685 /// applicable if the expression was reassociated, this checks the chain, then
1686 /// reassociates the expression as necessary to expose the optimization
1687 /// opportunity. This makes use of a special Functor, which must define
1688 /// 'shouldApply' and 'apply' methods.
1690 template<typename Functor>
1691 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1692 unsigned Opcode = Root.getOpcode();
1693 Value *LHS = Root.getOperand(0);
1695 // Quick check, see if the immediate LHS matches...
1696 if (F.shouldApply(LHS))
1697 return F.apply(Root);
1699 // Otherwise, if the LHS is not of the same opcode as the root, return.
1700 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1701 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1702 // Should we apply this transform to the RHS?
1703 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1705 // If not to the RHS, check to see if we should apply to the LHS...
1706 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1707 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1711 // If the functor wants to apply the optimization to the RHS of LHSI,
1712 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1714 BasicBlock *BB = Root.getParent();
1716 // Now all of the instructions are in the current basic block, go ahead
1717 // and perform the reassociation.
1718 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1720 // First move the selected RHS to the LHS of the root...
1721 Root.setOperand(0, LHSI->getOperand(1));
1723 // Make what used to be the LHS of the root be the user of the root...
1724 Value *ExtraOperand = TmpLHSI->getOperand(1);
1725 if (&Root == TmpLHSI) {
1726 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1729 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1730 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1731 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1732 BasicBlock::iterator ARI = &Root; ++ARI;
1733 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1736 // Now propagate the ExtraOperand down the chain of instructions until we
1738 while (TmpLHSI != LHSI) {
1739 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1740 // Move the instruction to immediately before the chain we are
1741 // constructing to avoid breaking dominance properties.
1742 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1743 BB->getInstList().insert(ARI, NextLHSI);
1746 Value *NextOp = NextLHSI->getOperand(1);
1747 NextLHSI->setOperand(1, ExtraOperand);
1749 ExtraOperand = NextOp;
1752 // Now that the instructions are reassociated, have the functor perform
1753 // the transformation...
1754 return F.apply(Root);
1757 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1763 // AddRHS - Implements: X + X --> X << 1
1766 AddRHS(Value *rhs) : RHS(rhs) {}
1767 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1768 Instruction *apply(BinaryOperator &Add) const {
1769 return BinaryOperator::createShl(Add.getOperand(0),
1770 ConstantInt::get(Add.getType(), 1));
1774 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1776 struct AddMaskingAnd {
1778 AddMaskingAnd(Constant *c) : C2(c) {}
1779 bool shouldApply(Value *LHS) const {
1781 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1782 ConstantExpr::getAnd(C1, C2)->isNullValue();
1784 Instruction *apply(BinaryOperator &Add) const {
1785 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1789 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1791 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1792 if (Constant *SOC = dyn_cast<Constant>(SO))
1793 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1795 return IC->InsertNewInstBefore(CastInst::create(
1796 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1799 // Figure out if the constant is the left or the right argument.
1800 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1801 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1803 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1805 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1806 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1809 Value *Op0 = SO, *Op1 = ConstOperand;
1811 std::swap(Op0, Op1);
1813 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1814 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1815 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1816 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1817 SO->getName()+".cmp");
1819 assert(0 && "Unknown binary instruction type!");
1822 return IC->InsertNewInstBefore(New, I);
1825 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1826 // constant as the other operand, try to fold the binary operator into the
1827 // select arguments. This also works for Cast instructions, which obviously do
1828 // not have a second operand.
1829 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1831 // Don't modify shared select instructions
1832 if (!SI->hasOneUse()) return 0;
1833 Value *TV = SI->getOperand(1);
1834 Value *FV = SI->getOperand(2);
1836 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1837 // Bool selects with constant operands can be folded to logical ops.
1838 if (SI->getType() == Type::Int1Ty) return 0;
1840 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1841 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1843 return new SelectInst(SI->getCondition(), SelectTrueVal,
1850 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1851 /// node as operand #0, see if we can fold the instruction into the PHI (which
1852 /// is only possible if all operands to the PHI are constants).
1853 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1854 PHINode *PN = cast<PHINode>(I.getOperand(0));
1855 unsigned NumPHIValues = PN->getNumIncomingValues();
1856 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1858 // Check to see if all of the operands of the PHI are constants. If there is
1859 // one non-constant value, remember the BB it is. If there is more than one
1860 // or if *it* is a PHI, bail out.
1861 BasicBlock *NonConstBB = 0;
1862 for (unsigned i = 0; i != NumPHIValues; ++i)
1863 if (!isa<Constant>(PN->getIncomingValue(i))) {
1864 if (NonConstBB) return 0; // More than one non-const value.
1865 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1866 NonConstBB = PN->getIncomingBlock(i);
1868 // If the incoming non-constant value is in I's block, we have an infinite
1870 if (NonConstBB == I.getParent())
1874 // If there is exactly one non-constant value, we can insert a copy of the
1875 // operation in that block. However, if this is a critical edge, we would be
1876 // inserting the computation one some other paths (e.g. inside a loop). Only
1877 // do this if the pred block is unconditionally branching into the phi block.
1879 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1880 if (!BI || !BI->isUnconditional()) return 0;
1883 // Okay, we can do the transformation: create the new PHI node.
1884 PHINode *NewPN = new PHINode(I.getType(), "");
1885 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1886 InsertNewInstBefore(NewPN, *PN);
1887 NewPN->takeName(PN);
1889 // Next, add all of the operands to the PHI.
1890 if (I.getNumOperands() == 2) {
1891 Constant *C = cast<Constant>(I.getOperand(1));
1892 for (unsigned i = 0; i != NumPHIValues; ++i) {
1894 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1895 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1896 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1898 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1900 assert(PN->getIncomingBlock(i) == NonConstBB);
1901 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1902 InV = BinaryOperator::create(BO->getOpcode(),
1903 PN->getIncomingValue(i), C, "phitmp",
1904 NonConstBB->getTerminator());
1905 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1906 InV = CmpInst::create(CI->getOpcode(),
1908 PN->getIncomingValue(i), C, "phitmp",
1909 NonConstBB->getTerminator());
1911 assert(0 && "Unknown binop!");
1913 AddToWorkList(cast<Instruction>(InV));
1915 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1918 CastInst *CI = cast<CastInst>(&I);
1919 const Type *RetTy = CI->getType();
1920 for (unsigned i = 0; i != NumPHIValues; ++i) {
1922 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1923 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1925 assert(PN->getIncomingBlock(i) == NonConstBB);
1926 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1927 I.getType(), "phitmp",
1928 NonConstBB->getTerminator());
1929 AddToWorkList(cast<Instruction>(InV));
1931 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1934 return ReplaceInstUsesWith(I, NewPN);
1937 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1938 bool Changed = SimplifyCommutative(I);
1939 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1941 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1942 // X + undef -> undef
1943 if (isa<UndefValue>(RHS))
1944 return ReplaceInstUsesWith(I, RHS);
1947 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1948 if (RHSC->isNullValue())
1949 return ReplaceInstUsesWith(I, LHS);
1950 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1951 if (CFP->isExactlyValue(-0.0))
1952 return ReplaceInstUsesWith(I, LHS);
1955 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1956 // X + (signbit) --> X ^ signbit
1957 const APInt& Val = CI->getValue();
1958 uint32_t BitWidth = Val.getBitWidth();
1959 if (Val == APInt::getSignBit(BitWidth))
1960 return BinaryOperator::createXor(LHS, RHS);
1962 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1963 // (X & 254)+1 -> (X&254)|1
1964 if (!isa<VectorType>(I.getType())) {
1965 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1966 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1967 KnownZero, KnownOne))
1972 if (isa<PHINode>(LHS))
1973 if (Instruction *NV = FoldOpIntoPhi(I))
1976 ConstantInt *XorRHS = 0;
1978 if (isa<ConstantInt>(RHSC) &&
1979 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1980 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1981 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1983 uint32_t Size = TySizeBits / 2;
1984 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1985 APInt CFF80Val(-C0080Val);
1987 if (TySizeBits > Size) {
1988 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1989 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1990 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1991 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1992 // This is a sign extend if the top bits are known zero.
1993 if (!MaskedValueIsZero(XorLHS,
1994 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1995 Size = 0; // Not a sign ext, but can't be any others either.
2000 C0080Val = APIntOps::lshr(C0080Val, Size);
2001 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2002 } while (Size >= 1);
2004 // FIXME: This shouldn't be necessary. When the backends can handle types
2005 // with funny bit widths then this whole cascade of if statements should
2006 // be removed. It is just here to get the size of the "middle" type back
2007 // up to something that the back ends can handle.
2008 const Type *MiddleType = 0;
2011 case 32: MiddleType = Type::Int32Ty; break;
2012 case 16: MiddleType = Type::Int16Ty; break;
2013 case 8: MiddleType = Type::Int8Ty; break;
2016 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2017 InsertNewInstBefore(NewTrunc, I);
2018 return new SExtInst(NewTrunc, I.getType(), I.getName());
2024 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2025 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2027 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2028 if (RHSI->getOpcode() == Instruction::Sub)
2029 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2030 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2032 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2033 if (LHSI->getOpcode() == Instruction::Sub)
2034 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2035 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2040 if (Value *V = dyn_castNegVal(LHS))
2041 return BinaryOperator::createSub(RHS, V);
2044 if (!isa<Constant>(RHS))
2045 if (Value *V = dyn_castNegVal(RHS))
2046 return BinaryOperator::createSub(LHS, V);
2050 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2051 if (X == RHS) // X*C + X --> X * (C+1)
2052 return BinaryOperator::createMul(RHS, AddOne(C2));
2054 // X*C1 + X*C2 --> X * (C1+C2)
2056 if (X == dyn_castFoldableMul(RHS, C1))
2057 return BinaryOperator::createMul(X, Add(C1, C2));
2060 // X + X*C --> X * (C+1)
2061 if (dyn_castFoldableMul(RHS, C2) == LHS)
2062 return BinaryOperator::createMul(LHS, AddOne(C2));
2064 // X + ~X --> -1 since ~X = -X-1
2065 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2066 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2069 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2070 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2071 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2074 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2076 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2077 return BinaryOperator::createSub(SubOne(CRHS), X);
2079 // (X & FF00) + xx00 -> (X+xx00) & FF00
2080 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2081 Constant *Anded = And(CRHS, C2);
2082 if (Anded == CRHS) {
2083 // See if all bits from the first bit set in the Add RHS up are included
2084 // in the mask. First, get the rightmost bit.
2085 const APInt& AddRHSV = CRHS->getValue();
2087 // Form a mask of all bits from the lowest bit added through the top.
2088 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2090 // See if the and mask includes all of these bits.
2091 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2093 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2094 // Okay, the xform is safe. Insert the new add pronto.
2095 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2096 LHS->getName()), I);
2097 return BinaryOperator::createAnd(NewAdd, C2);
2102 // Try to fold constant add into select arguments.
2103 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2104 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2108 // add (cast *A to intptrtype) B ->
2109 // cast (GEP (cast *A to sbyte*) B) ->
2112 CastInst *CI = dyn_cast<CastInst>(LHS);
2115 CI = dyn_cast<CastInst>(RHS);
2118 if (CI && CI->getType()->isSized() &&
2119 (CI->getType()->getPrimitiveSizeInBits() ==
2120 TD->getIntPtrType()->getPrimitiveSizeInBits())
2121 && isa<PointerType>(CI->getOperand(0)->getType())) {
2122 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2123 PointerType::get(Type::Int8Ty), I);
2124 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2125 return new PtrToIntInst(I2, CI->getType());
2129 return Changed ? &I : 0;
2132 // isSignBit - Return true if the value represented by the constant only has the
2133 // highest order bit set.
2134 static bool isSignBit(ConstantInt *CI) {
2135 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2136 return CI->getValue() == APInt::getSignBit(NumBits);
2139 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2140 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2142 if (Op0 == Op1) // sub X, X -> 0
2143 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2145 // If this is a 'B = x-(-A)', change to B = x+A...
2146 if (Value *V = dyn_castNegVal(Op1))
2147 return BinaryOperator::createAdd(Op0, V);
2149 if (isa<UndefValue>(Op0))
2150 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2151 if (isa<UndefValue>(Op1))
2152 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2154 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2155 // Replace (-1 - A) with (~A)...
2156 if (C->isAllOnesValue())
2157 return BinaryOperator::createNot(Op1);
2159 // C - ~X == X + (1+C)
2161 if (match(Op1, m_Not(m_Value(X))))
2162 return BinaryOperator::createAdd(X, AddOne(C));
2164 // -(X >>u 31) -> (X >>s 31)
2165 // -(X >>s 31) -> (X >>u 31)
2167 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2168 if (SI->getOpcode() == Instruction::LShr) {
2169 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2170 // Check to see if we are shifting out everything but the sign bit.
2171 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2172 SI->getType()->getPrimitiveSizeInBits()-1) {
2173 // Ok, the transformation is safe. Insert AShr.
2174 return BinaryOperator::create(Instruction::AShr,
2175 SI->getOperand(0), CU, SI->getName());
2179 else if (SI->getOpcode() == Instruction::AShr) {
2180 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2181 // Check to see if we are shifting out everything but the sign bit.
2182 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2183 SI->getType()->getPrimitiveSizeInBits()-1) {
2184 // Ok, the transformation is safe. Insert LShr.
2185 return BinaryOperator::createLShr(
2186 SI->getOperand(0), CU, SI->getName());
2192 // Try to fold constant sub into select arguments.
2193 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2194 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2197 if (isa<PHINode>(Op0))
2198 if (Instruction *NV = FoldOpIntoPhi(I))
2202 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2203 if (Op1I->getOpcode() == Instruction::Add &&
2204 !Op0->getType()->isFPOrFPVector()) {
2205 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2206 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2207 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2208 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2209 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2210 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2211 // C1-(X+C2) --> (C1-C2)-X
2212 return BinaryOperator::createSub(Subtract(CI1, CI2),
2213 Op1I->getOperand(0));
2217 if (Op1I->hasOneUse()) {
2218 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2219 // is not used by anyone else...
2221 if (Op1I->getOpcode() == Instruction::Sub &&
2222 !Op1I->getType()->isFPOrFPVector()) {
2223 // Swap the two operands of the subexpr...
2224 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2225 Op1I->setOperand(0, IIOp1);
2226 Op1I->setOperand(1, IIOp0);
2228 // Create the new top level add instruction...
2229 return BinaryOperator::createAdd(Op0, Op1);
2232 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2234 if (Op1I->getOpcode() == Instruction::And &&
2235 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2236 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2239 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2240 return BinaryOperator::createAnd(Op0, NewNot);
2243 // 0 - (X sdiv C) -> (X sdiv -C)
2244 if (Op1I->getOpcode() == Instruction::SDiv)
2245 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2247 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2248 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2249 ConstantExpr::getNeg(DivRHS));
2251 // X - X*C --> X * (1-C)
2252 ConstantInt *C2 = 0;
2253 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2254 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2255 return BinaryOperator::createMul(Op0, CP1);
2260 if (!Op0->getType()->isFPOrFPVector())
2261 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2262 if (Op0I->getOpcode() == Instruction::Add) {
2263 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2264 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2265 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2266 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2267 } else if (Op0I->getOpcode() == Instruction::Sub) {
2268 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2269 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2273 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2274 if (X == Op1) // X*C - X --> X * (C-1)
2275 return BinaryOperator::createMul(Op1, SubOne(C1));
2277 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2278 if (X == dyn_castFoldableMul(Op1, C2))
2279 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2284 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2285 /// comparison only checks the sign bit. If it only checks the sign bit, set
2286 /// TrueIfSigned if the result of the comparison is true when the input value is
2288 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2289 bool &TrueIfSigned) {
2291 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2292 TrueIfSigned = true;
2293 return RHS->isZero();
2294 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2295 TrueIfSigned = true;
2296 return RHS->isAllOnesValue();
2297 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2298 TrueIfSigned = false;
2299 return RHS->isAllOnesValue();
2300 case ICmpInst::ICMP_UGT:
2301 // True if LHS u> RHS and RHS == high-bit-mask - 1
2302 TrueIfSigned = true;
2303 return RHS->getValue() ==
2304 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2305 case ICmpInst::ICMP_UGE:
2306 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2307 TrueIfSigned = true;
2308 return RHS->getValue() ==
2309 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2315 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2316 bool Changed = SimplifyCommutative(I);
2317 Value *Op0 = I.getOperand(0);
2319 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2320 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2322 // Simplify mul instructions with a constant RHS...
2323 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2324 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2326 // ((X << C1)*C2) == (X * (C2 << C1))
2327 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2328 if (SI->getOpcode() == Instruction::Shl)
2329 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2330 return BinaryOperator::createMul(SI->getOperand(0),
2331 ConstantExpr::getShl(CI, ShOp));
2334 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2335 if (CI->equalsInt(1)) // X * 1 == X
2336 return ReplaceInstUsesWith(I, Op0);
2337 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2338 return BinaryOperator::createNeg(Op0, I.getName());
2340 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2341 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2342 return BinaryOperator::createShl(Op0,
2343 ConstantInt::get(Op0->getType(), Val.logBase2()));
2345 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2346 if (Op1F->isNullValue())
2347 return ReplaceInstUsesWith(I, Op1);
2349 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2350 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2351 if (Op1F->isExactlyValue(1.0))
2352 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2355 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2356 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2357 isa<ConstantInt>(Op0I->getOperand(1))) {
2358 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2359 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2361 InsertNewInstBefore(Add, I);
2362 Value *C1C2 = ConstantExpr::getMul(Op1,
2363 cast<Constant>(Op0I->getOperand(1)));
2364 return BinaryOperator::createAdd(Add, C1C2);
2368 // Try to fold constant mul into select arguments.
2369 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2370 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2373 if (isa<PHINode>(Op0))
2374 if (Instruction *NV = FoldOpIntoPhi(I))
2378 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2379 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2380 return BinaryOperator::createMul(Op0v, Op1v);
2382 // If one of the operands of the multiply is a cast from a boolean value, then
2383 // we know the bool is either zero or one, so this is a 'masking' multiply.
2384 // See if we can simplify things based on how the boolean was originally
2386 CastInst *BoolCast = 0;
2387 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2388 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2391 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2392 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2395 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2396 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2397 const Type *SCOpTy = SCIOp0->getType();
2400 // If the icmp is true iff the sign bit of X is set, then convert this
2401 // multiply into a shift/and combination.
2402 if (isa<ConstantInt>(SCIOp1) &&
2403 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2405 // Shift the X value right to turn it into "all signbits".
2406 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2407 SCOpTy->getPrimitiveSizeInBits()-1);
2409 InsertNewInstBefore(
2410 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2411 BoolCast->getOperand(0)->getName()+
2414 // If the multiply type is not the same as the source type, sign extend
2415 // or truncate to the multiply type.
2416 if (I.getType() != V->getType()) {
2417 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2418 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2419 Instruction::CastOps opcode =
2420 (SrcBits == DstBits ? Instruction::BitCast :
2421 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2422 V = InsertCastBefore(opcode, V, I.getType(), I);
2425 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2426 return BinaryOperator::createAnd(V, OtherOp);
2431 return Changed ? &I : 0;
2434 /// This function implements the transforms on div instructions that work
2435 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2436 /// used by the visitors to those instructions.
2437 /// @brief Transforms common to all three div instructions
2438 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2439 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2442 if (isa<UndefValue>(Op0))
2443 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2445 // X / undef -> undef
2446 if (isa<UndefValue>(Op1))
2447 return ReplaceInstUsesWith(I, Op1);
2449 // Handle cases involving: div X, (select Cond, Y, Z)
2450 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2451 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2452 // same basic block, then we replace the select with Y, and the condition
2453 // of the select with false (if the cond value is in the same BB). If the
2454 // select has uses other than the div, this allows them to be simplified
2455 // also. Note that div X, Y is just as good as div X, 0 (undef)
2456 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2457 if (ST->isNullValue()) {
2458 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2459 if (CondI && CondI->getParent() == I.getParent())
2460 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2461 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2462 I.setOperand(1, SI->getOperand(2));
2464 UpdateValueUsesWith(SI, SI->getOperand(2));
2468 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2469 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2470 if (ST->isNullValue()) {
2471 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2472 if (CondI && CondI->getParent() == I.getParent())
2473 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2474 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2475 I.setOperand(1, SI->getOperand(1));
2477 UpdateValueUsesWith(SI, SI->getOperand(1));
2485 /// This function implements the transforms common to both integer division
2486 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2487 /// division instructions.
2488 /// @brief Common integer divide transforms
2489 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2490 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2492 if (Instruction *Common = commonDivTransforms(I))
2495 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2497 if (RHS->equalsInt(1))
2498 return ReplaceInstUsesWith(I, Op0);
2500 // (X / C1) / C2 -> X / (C1*C2)
2501 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2502 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2503 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2504 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2505 Multiply(RHS, LHSRHS));
2508 if (!RHS->isZero()) { // avoid X udiv 0
2509 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2510 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2512 if (isa<PHINode>(Op0))
2513 if (Instruction *NV = FoldOpIntoPhi(I))
2518 // 0 / X == 0, we don't need to preserve faults!
2519 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2520 if (LHS->equalsInt(0))
2521 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2526 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2527 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2529 // Handle the integer div common cases
2530 if (Instruction *Common = commonIDivTransforms(I))
2533 // X udiv C^2 -> X >> C
2534 // Check to see if this is an unsigned division with an exact power of 2,
2535 // if so, convert to a right shift.
2536 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2537 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2538 return BinaryOperator::createLShr(Op0,
2539 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2542 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2543 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2544 if (RHSI->getOpcode() == Instruction::Shl &&
2545 isa<ConstantInt>(RHSI->getOperand(0))) {
2546 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2547 if (C1.isPowerOf2()) {
2548 Value *N = RHSI->getOperand(1);
2549 const Type *NTy = N->getType();
2550 if (uint32_t C2 = C1.logBase2()) {
2551 Constant *C2V = ConstantInt::get(NTy, C2);
2552 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2554 return BinaryOperator::createLShr(Op0, N);
2559 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2560 // where C1&C2 are powers of two.
2561 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2562 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2563 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2564 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2565 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2566 // Compute the shift amounts
2567 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2568 // Construct the "on true" case of the select
2569 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2570 Instruction *TSI = BinaryOperator::createLShr(
2571 Op0, TC, SI->getName()+".t");
2572 TSI = InsertNewInstBefore(TSI, I);
2574 // Construct the "on false" case of the select
2575 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2576 Instruction *FSI = BinaryOperator::createLShr(
2577 Op0, FC, SI->getName()+".f");
2578 FSI = InsertNewInstBefore(FSI, I);
2580 // construct the select instruction and return it.
2581 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2587 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2588 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2590 // Handle the integer div common cases
2591 if (Instruction *Common = commonIDivTransforms(I))
2594 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2596 if (RHS->isAllOnesValue())
2597 return BinaryOperator::createNeg(Op0);
2600 if (Value *LHSNeg = dyn_castNegVal(Op0))
2601 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2604 // If the sign bits of both operands are zero (i.e. we can prove they are
2605 // unsigned inputs), turn this into a udiv.
2606 if (I.getType()->isInteger()) {
2607 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2608 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2609 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2616 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2617 return commonDivTransforms(I);
2620 /// GetFactor - If we can prove that the specified value is at least a multiple
2621 /// of some factor, return that factor.
2622 static Constant *GetFactor(Value *V) {
2623 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2626 // Unless we can be tricky, we know this is a multiple of 1.
2627 Constant *Result = ConstantInt::get(V->getType(), 1);
2629 Instruction *I = dyn_cast<Instruction>(V);
2630 if (!I) return Result;
2632 if (I->getOpcode() == Instruction::Mul) {
2633 // Handle multiplies by a constant, etc.
2634 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2635 GetFactor(I->getOperand(1)));
2636 } else if (I->getOpcode() == Instruction::Shl) {
2637 // (X<<C) -> X * (1 << C)
2638 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2639 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2640 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2642 } else if (I->getOpcode() == Instruction::And) {
2643 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2644 // X & 0xFFF0 is known to be a multiple of 16.
2645 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2646 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2647 return ConstantExpr::getShl(Result,
2648 ConstantInt::get(Result->getType(), Zeros));
2650 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2651 // Only handle int->int casts.
2652 if (!CI->isIntegerCast())
2654 Value *Op = CI->getOperand(0);
2655 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2660 /// This function implements the transforms on rem instructions that work
2661 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2662 /// is used by the visitors to those instructions.
2663 /// @brief Transforms common to all three rem instructions
2664 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2665 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2667 // 0 % X == 0, we don't need to preserve faults!
2668 if (Constant *LHS = dyn_cast<Constant>(Op0))
2669 if (LHS->isNullValue())
2670 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2672 if (isa<UndefValue>(Op0)) // undef % X -> 0
2673 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2674 if (isa<UndefValue>(Op1))
2675 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2677 // Handle cases involving: rem X, (select Cond, Y, Z)
2678 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2679 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2680 // the same basic block, then we replace the select with Y, and the
2681 // condition of the select with false (if the cond value is in the same
2682 // BB). If the select has uses other than the div, this allows them to be
2684 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2685 if (ST->isNullValue()) {
2686 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2687 if (CondI && CondI->getParent() == I.getParent())
2688 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2689 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2690 I.setOperand(1, SI->getOperand(2));
2692 UpdateValueUsesWith(SI, SI->getOperand(2));
2695 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2696 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2697 if (ST->isNullValue()) {
2698 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2699 if (CondI && CondI->getParent() == I.getParent())
2700 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2701 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2702 I.setOperand(1, SI->getOperand(1));
2704 UpdateValueUsesWith(SI, SI->getOperand(1));
2712 /// This function implements the transforms common to both integer remainder
2713 /// instructions (urem and srem). It is called by the visitors to those integer
2714 /// remainder instructions.
2715 /// @brief Common integer remainder transforms
2716 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2717 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2719 if (Instruction *common = commonRemTransforms(I))
2722 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2723 // X % 0 == undef, we don't need to preserve faults!
2724 if (RHS->equalsInt(0))
2725 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2727 if (RHS->equalsInt(1)) // X % 1 == 0
2728 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2730 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2731 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2732 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2734 } else if (isa<PHINode>(Op0I)) {
2735 if (Instruction *NV = FoldOpIntoPhi(I))
2738 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2739 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2740 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2747 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2748 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2750 if (Instruction *common = commonIRemTransforms(I))
2753 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2754 // X urem C^2 -> X and C
2755 // Check to see if this is an unsigned remainder with an exact power of 2,
2756 // if so, convert to a bitwise and.
2757 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2758 if (C->getValue().isPowerOf2())
2759 return BinaryOperator::createAnd(Op0, SubOne(C));
2762 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2763 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2764 if (RHSI->getOpcode() == Instruction::Shl &&
2765 isa<ConstantInt>(RHSI->getOperand(0))) {
2766 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2767 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2768 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2770 return BinaryOperator::createAnd(Op0, Add);
2775 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2776 // where C1&C2 are powers of two.
2777 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2778 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2779 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2780 // STO == 0 and SFO == 0 handled above.
2781 if ((STO->getValue().isPowerOf2()) &&
2782 (SFO->getValue().isPowerOf2())) {
2783 Value *TrueAnd = InsertNewInstBefore(
2784 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2785 Value *FalseAnd = InsertNewInstBefore(
2786 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2787 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2795 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2796 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2798 if (Instruction *common = commonIRemTransforms(I))
2801 if (Value *RHSNeg = dyn_castNegVal(Op1))
2802 if (!isa<ConstantInt>(RHSNeg) ||
2803 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2805 AddUsesToWorkList(I);
2806 I.setOperand(1, RHSNeg);
2810 // If the top bits of both operands are zero (i.e. we can prove they are
2811 // unsigned inputs), turn this into a urem.
2812 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2813 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2814 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2815 return BinaryOperator::createURem(Op0, Op1, I.getName());
2821 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2822 return commonRemTransforms(I);
2825 // isMaxValueMinusOne - return true if this is Max-1
2826 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2827 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2829 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2830 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2833 // isMinValuePlusOne - return true if this is Min+1
2834 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2836 return C->getValue() == 1; // unsigned
2838 // Calculate 1111111111000000000000
2839 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2840 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2843 // isOneBitSet - Return true if there is exactly one bit set in the specified
2845 static bool isOneBitSet(const ConstantInt *CI) {
2846 return CI->getValue().isPowerOf2();
2849 // isHighOnes - Return true if the constant is of the form 1+0+.
2850 // This is the same as lowones(~X).
2851 static bool isHighOnes(const ConstantInt *CI) {
2852 return (~CI->getValue() + 1).isPowerOf2();
2855 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2856 /// are carefully arranged to allow folding of expressions such as:
2858 /// (A < B) | (A > B) --> (A != B)
2860 /// Note that this is only valid if the first and second predicates have the
2861 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2863 /// Three bits are used to represent the condition, as follows:
2868 /// <=> Value Definition
2869 /// 000 0 Always false
2876 /// 111 7 Always true
2878 static unsigned getICmpCode(const ICmpInst *ICI) {
2879 switch (ICI->getPredicate()) {
2881 case ICmpInst::ICMP_UGT: return 1; // 001
2882 case ICmpInst::ICMP_SGT: return 1; // 001
2883 case ICmpInst::ICMP_EQ: return 2; // 010
2884 case ICmpInst::ICMP_UGE: return 3; // 011
2885 case ICmpInst::ICMP_SGE: return 3; // 011
2886 case ICmpInst::ICMP_ULT: return 4; // 100
2887 case ICmpInst::ICMP_SLT: return 4; // 100
2888 case ICmpInst::ICMP_NE: return 5; // 101
2889 case ICmpInst::ICMP_ULE: return 6; // 110
2890 case ICmpInst::ICMP_SLE: return 6; // 110
2893 assert(0 && "Invalid ICmp predicate!");
2898 /// getICmpValue - This is the complement of getICmpCode, which turns an
2899 /// opcode and two operands into either a constant true or false, or a brand
2900 /// new /// ICmp instruction. The sign is passed in to determine which kind
2901 /// of predicate to use in new icmp instructions.
2902 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2904 default: assert(0 && "Illegal ICmp code!");
2905 case 0: return ConstantInt::getFalse();
2908 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2910 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2911 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2914 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2916 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2919 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2921 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2922 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2925 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2927 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2928 case 7: return ConstantInt::getTrue();
2932 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2933 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2934 (ICmpInst::isSignedPredicate(p1) &&
2935 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2936 (ICmpInst::isSignedPredicate(p2) &&
2937 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2941 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2942 struct FoldICmpLogical {
2945 ICmpInst::Predicate pred;
2946 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2947 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2948 pred(ICI->getPredicate()) {}
2949 bool shouldApply(Value *V) const {
2950 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2951 if (PredicatesFoldable(pred, ICI->getPredicate()))
2952 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2953 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2956 Instruction *apply(Instruction &Log) const {
2957 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2958 if (ICI->getOperand(0) != LHS) {
2959 assert(ICI->getOperand(1) == LHS);
2960 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2963 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2964 unsigned LHSCode = getICmpCode(ICI);
2965 unsigned RHSCode = getICmpCode(RHSICI);
2967 switch (Log.getOpcode()) {
2968 case Instruction::And: Code = LHSCode & RHSCode; break;
2969 case Instruction::Or: Code = LHSCode | RHSCode; break;
2970 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2971 default: assert(0 && "Illegal logical opcode!"); return 0;
2974 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2975 ICmpInst::isSignedPredicate(ICI->getPredicate());
2977 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2978 if (Instruction *I = dyn_cast<Instruction>(RV))
2980 // Otherwise, it's a constant boolean value...
2981 return IC.ReplaceInstUsesWith(Log, RV);
2984 } // end anonymous namespace
2986 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2987 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2988 // guaranteed to be a binary operator.
2989 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2991 ConstantInt *AndRHS,
2992 BinaryOperator &TheAnd) {
2993 Value *X = Op->getOperand(0);
2994 Constant *Together = 0;
2996 Together = And(AndRHS, OpRHS);
2998 switch (Op->getOpcode()) {
2999 case Instruction::Xor:
3000 if (Op->hasOneUse()) {
3001 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3002 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3003 InsertNewInstBefore(And, TheAnd);
3005 return BinaryOperator::createXor(And, Together);
3008 case Instruction::Or:
3009 if (Together == AndRHS) // (X | C) & C --> C
3010 return ReplaceInstUsesWith(TheAnd, AndRHS);
3012 if (Op->hasOneUse() && Together != OpRHS) {
3013 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3014 Instruction *Or = BinaryOperator::createOr(X, Together);
3015 InsertNewInstBefore(Or, TheAnd);
3017 return BinaryOperator::createAnd(Or, AndRHS);
3020 case Instruction::Add:
3021 if (Op->hasOneUse()) {
3022 // Adding a one to a single bit bit-field should be turned into an XOR
3023 // of the bit. First thing to check is to see if this AND is with a
3024 // single bit constant.
3025 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3027 // If there is only one bit set...
3028 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3029 // Ok, at this point, we know that we are masking the result of the
3030 // ADD down to exactly one bit. If the constant we are adding has
3031 // no bits set below this bit, then we can eliminate the ADD.
3032 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3034 // Check to see if any bits below the one bit set in AndRHSV are set.
3035 if ((AddRHS & (AndRHSV-1)) == 0) {
3036 // If not, the only thing that can effect the output of the AND is
3037 // the bit specified by AndRHSV. If that bit is set, the effect of
3038 // the XOR is to toggle the bit. If it is clear, then the ADD has
3040 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3041 TheAnd.setOperand(0, X);
3044 // Pull the XOR out of the AND.
3045 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3046 InsertNewInstBefore(NewAnd, TheAnd);
3047 NewAnd->takeName(Op);
3048 return BinaryOperator::createXor(NewAnd, AndRHS);
3055 case Instruction::Shl: {
3056 // We know that the AND will not produce any of the bits shifted in, so if
3057 // the anded constant includes them, clear them now!
3059 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3060 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3061 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3062 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3064 if (CI->getValue() == ShlMask) {
3065 // Masking out bits that the shift already masks
3066 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3067 } else if (CI != AndRHS) { // Reducing bits set in and.
3068 TheAnd.setOperand(1, CI);
3073 case Instruction::LShr:
3075 // We know that the AND will not produce any of the bits shifted in, so if
3076 // the anded constant includes them, clear them now! This only applies to
3077 // unsigned shifts, because a signed shr may bring in set bits!
3079 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3080 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3081 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3082 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3084 if (CI->getValue() == ShrMask) {
3085 // Masking out bits that the shift already masks.
3086 return ReplaceInstUsesWith(TheAnd, Op);
3087 } else if (CI != AndRHS) {
3088 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3093 case Instruction::AShr:
3095 // See if this is shifting in some sign extension, then masking it out
3097 if (Op->hasOneUse()) {
3098 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3099 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3100 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3101 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3102 if (C == AndRHS) { // Masking out bits shifted in.
3103 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3104 // Make the argument unsigned.
3105 Value *ShVal = Op->getOperand(0);
3106 ShVal = InsertNewInstBefore(
3107 BinaryOperator::createLShr(ShVal, OpRHS,
3108 Op->getName()), TheAnd);
3109 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3118 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3119 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3120 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3121 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3122 /// insert new instructions.
3123 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3124 bool isSigned, bool Inside,
3126 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3127 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3128 "Lo is not <= Hi in range emission code!");
3131 if (Lo == Hi) // Trivially false.
3132 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3134 // V >= Min && V < Hi --> V < Hi
3135 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3136 ICmpInst::Predicate pred = (isSigned ?
3137 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3138 return new ICmpInst(pred, V, Hi);
3141 // Emit V-Lo <u Hi-Lo
3142 Constant *NegLo = ConstantExpr::getNeg(Lo);
3143 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3144 InsertNewInstBefore(Add, IB);
3145 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3146 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3149 if (Lo == Hi) // Trivially true.
3150 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3152 // V < Min || V >= Hi -> V > Hi-1
3153 Hi = SubOne(cast<ConstantInt>(Hi));
3154 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3155 ICmpInst::Predicate pred = (isSigned ?
3156 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3157 return new ICmpInst(pred, V, Hi);
3160 // Emit V-Lo >u Hi-1-Lo
3161 // Note that Hi has already had one subtracted from it, above.
3162 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3163 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3164 InsertNewInstBefore(Add, IB);
3165 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3166 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3169 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3170 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3171 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3172 // not, since all 1s are not contiguous.
3173 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3174 const APInt& V = Val->getValue();
3175 uint32_t BitWidth = Val->getType()->getBitWidth();
3176 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3178 // look for the first zero bit after the run of ones
3179 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3180 // look for the first non-zero bit
3181 ME = V.getActiveBits();
3185 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3186 /// where isSub determines whether the operator is a sub. If we can fold one of
3187 /// the following xforms:
3189 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3190 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3191 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3193 /// return (A +/- B).
3195 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3196 ConstantInt *Mask, bool isSub,
3198 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3199 if (!LHSI || LHSI->getNumOperands() != 2 ||
3200 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3202 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3204 switch (LHSI->getOpcode()) {
3206 case Instruction::And:
3207 if (And(N, Mask) == Mask) {
3208 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3209 if ((Mask->getValue().countLeadingZeros() +
3210 Mask->getValue().countPopulation()) ==
3211 Mask->getValue().getBitWidth())
3214 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3215 // part, we don't need any explicit masks to take them out of A. If that
3216 // is all N is, ignore it.
3217 uint32_t MB = 0, ME = 0;
3218 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3219 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3220 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3221 if (MaskedValueIsZero(RHS, Mask))
3226 case Instruction::Or:
3227 case Instruction::Xor:
3228 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3229 if ((Mask->getValue().countLeadingZeros() +
3230 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3231 && And(N, Mask)->isZero())
3238 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3240 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3241 return InsertNewInstBefore(New, I);
3244 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3245 bool Changed = SimplifyCommutative(I);
3246 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3248 if (isa<UndefValue>(Op1)) // X & undef -> 0
3249 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3253 return ReplaceInstUsesWith(I, Op1);
3255 // See if we can simplify any instructions used by the instruction whose sole
3256 // purpose is to compute bits we don't care about.
3257 if (!isa<VectorType>(I.getType())) {
3258 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3259 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3260 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3261 KnownZero, KnownOne))
3264 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3265 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3266 return ReplaceInstUsesWith(I, I.getOperand(0));
3267 } else if (isa<ConstantAggregateZero>(Op1)) {
3268 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3272 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3273 const APInt& AndRHSMask = AndRHS->getValue();
3274 APInt NotAndRHS(~AndRHSMask);
3276 // Optimize a variety of ((val OP C1) & C2) combinations...
3277 if (isa<BinaryOperator>(Op0)) {
3278 Instruction *Op0I = cast<Instruction>(Op0);
3279 Value *Op0LHS = Op0I->getOperand(0);
3280 Value *Op0RHS = Op0I->getOperand(1);
3281 switch (Op0I->getOpcode()) {
3282 case Instruction::Xor:
3283 case Instruction::Or:
3284 // If the mask is only needed on one incoming arm, push it up.
3285 if (Op0I->hasOneUse()) {
3286 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3287 // Not masking anything out for the LHS, move to RHS.
3288 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3289 Op0RHS->getName()+".masked");
3290 InsertNewInstBefore(NewRHS, I);
3291 return BinaryOperator::create(
3292 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3294 if (!isa<Constant>(Op0RHS) &&
3295 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3296 // Not masking anything out for the RHS, move to LHS.
3297 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3298 Op0LHS->getName()+".masked");
3299 InsertNewInstBefore(NewLHS, I);
3300 return BinaryOperator::create(
3301 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3306 case Instruction::Add:
3307 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3308 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3309 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3310 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3311 return BinaryOperator::createAnd(V, AndRHS);
3312 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3313 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3316 case Instruction::Sub:
3317 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3318 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3319 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3320 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3321 return BinaryOperator::createAnd(V, AndRHS);
3325 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3326 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3328 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3329 // If this is an integer truncation or change from signed-to-unsigned, and
3330 // if the source is an and/or with immediate, transform it. This
3331 // frequently occurs for bitfield accesses.
3332 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3333 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3334 CastOp->getNumOperands() == 2)
3335 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3336 if (CastOp->getOpcode() == Instruction::And) {
3337 // Change: and (cast (and X, C1) to T), C2
3338 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3339 // This will fold the two constants together, which may allow
3340 // other simplifications.
3341 Instruction *NewCast = CastInst::createTruncOrBitCast(
3342 CastOp->getOperand(0), I.getType(),
3343 CastOp->getName()+".shrunk");
3344 NewCast = InsertNewInstBefore(NewCast, I);
3345 // trunc_or_bitcast(C1)&C2
3346 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3347 C3 = ConstantExpr::getAnd(C3, AndRHS);
3348 return BinaryOperator::createAnd(NewCast, C3);
3349 } else if (CastOp->getOpcode() == Instruction::Or) {
3350 // Change: and (cast (or X, C1) to T), C2
3351 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3352 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3353 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3354 return ReplaceInstUsesWith(I, AndRHS);
3359 // Try to fold constant and into select arguments.
3360 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3361 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3363 if (isa<PHINode>(Op0))
3364 if (Instruction *NV = FoldOpIntoPhi(I))
3368 Value *Op0NotVal = dyn_castNotVal(Op0);
3369 Value *Op1NotVal = dyn_castNotVal(Op1);
3371 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3372 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3374 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3375 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3376 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3377 I.getName()+".demorgan");
3378 InsertNewInstBefore(Or, I);
3379 return BinaryOperator::createNot(Or);
3383 Value *A = 0, *B = 0, *C = 0, *D = 0;
3384 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3385 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3386 return ReplaceInstUsesWith(I, Op1);
3388 // (A|B) & ~(A&B) -> A^B
3389 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3390 if ((A == C && B == D) || (A == D && B == C))
3391 return BinaryOperator::createXor(A, B);
3395 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3396 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3397 return ReplaceInstUsesWith(I, Op0);
3399 // ~(A&B) & (A|B) -> A^B
3400 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3401 if ((A == C && B == D) || (A == D && B == C))
3402 return BinaryOperator::createXor(A, B);
3406 if (Op0->hasOneUse() &&
3407 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3408 if (A == Op1) { // (A^B)&A -> A&(A^B)
3409 I.swapOperands(); // Simplify below
3410 std::swap(Op0, Op1);
3411 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3412 cast<BinaryOperator>(Op0)->swapOperands();
3413 I.swapOperands(); // Simplify below
3414 std::swap(Op0, Op1);
3417 if (Op1->hasOneUse() &&
3418 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3419 if (B == Op0) { // B&(A^B) -> B&(B^A)
3420 cast<BinaryOperator>(Op1)->swapOperands();
3423 if (A == Op0) { // A&(A^B) -> A & ~B
3424 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3425 InsertNewInstBefore(NotB, I);
3426 return BinaryOperator::createAnd(A, NotB);
3431 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3432 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3433 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3436 Value *LHSVal, *RHSVal;
3437 ConstantInt *LHSCst, *RHSCst;
3438 ICmpInst::Predicate LHSCC, RHSCC;
3439 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3440 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3441 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3442 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3443 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3444 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3445 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3446 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3447 // Ensure that the larger constant is on the RHS.
3448 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3449 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3450 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3451 ICmpInst *LHS = cast<ICmpInst>(Op0);
3452 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3453 std::swap(LHS, RHS);
3454 std::swap(LHSCst, RHSCst);
3455 std::swap(LHSCC, RHSCC);
3458 // At this point, we know we have have two icmp instructions
3459 // comparing a value against two constants and and'ing the result
3460 // together. Because of the above check, we know that we only have
3461 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3462 // (from the FoldICmpLogical check above), that the two constants
3463 // are not equal and that the larger constant is on the RHS
3464 assert(LHSCst != RHSCst && "Compares not folded above?");
3467 default: assert(0 && "Unknown integer condition code!");
3468 case ICmpInst::ICMP_EQ:
3470 default: assert(0 && "Unknown integer condition code!");
3471 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3472 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3473 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3474 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3475 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3476 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3477 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3478 return ReplaceInstUsesWith(I, LHS);
3480 case ICmpInst::ICMP_NE:
3482 default: assert(0 && "Unknown integer condition code!");
3483 case ICmpInst::ICMP_ULT:
3484 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3485 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3486 break; // (X != 13 & X u< 15) -> no change
3487 case ICmpInst::ICMP_SLT:
3488 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3489 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3490 break; // (X != 13 & X s< 15) -> no change
3491 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3492 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3493 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3494 return ReplaceInstUsesWith(I, RHS);
3495 case ICmpInst::ICMP_NE:
3496 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3497 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3498 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3499 LHSVal->getName()+".off");
3500 InsertNewInstBefore(Add, I);
3501 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3502 ConstantInt::get(Add->getType(), 1));
3504 break; // (X != 13 & X != 15) -> no change
3507 case ICmpInst::ICMP_ULT:
3509 default: assert(0 && "Unknown integer condition code!");
3510 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3511 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3512 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3513 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3515 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3516 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3517 return ReplaceInstUsesWith(I, LHS);
3518 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3522 case ICmpInst::ICMP_SLT:
3524 default: assert(0 && "Unknown integer condition code!");
3525 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3526 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3527 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3528 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3530 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3531 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3532 return ReplaceInstUsesWith(I, LHS);
3533 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3537 case ICmpInst::ICMP_UGT:
3539 default: assert(0 && "Unknown integer condition code!");
3540 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3541 return ReplaceInstUsesWith(I, LHS);
3542 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3543 return ReplaceInstUsesWith(I, RHS);
3544 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3546 case ICmpInst::ICMP_NE:
3547 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3548 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3549 break; // (X u> 13 & X != 15) -> no change
3550 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3551 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3553 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3557 case ICmpInst::ICMP_SGT:
3559 default: assert(0 && "Unknown integer condition code!");
3560 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3561 return ReplaceInstUsesWith(I, LHS);
3562 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3563 return ReplaceInstUsesWith(I, RHS);
3564 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3566 case ICmpInst::ICMP_NE:
3567 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3568 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3569 break; // (X s> 13 & X != 15) -> no change
3570 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3571 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3573 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3581 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3582 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3583 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3584 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3585 const Type *SrcTy = Op0C->getOperand(0)->getType();
3586 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3587 // Only do this if the casts both really cause code to be generated.
3588 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3590 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3592 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3593 Op1C->getOperand(0),
3595 InsertNewInstBefore(NewOp, I);
3596 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3600 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3601 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3602 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3603 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3604 SI0->getOperand(1) == SI1->getOperand(1) &&
3605 (SI0->hasOneUse() || SI1->hasOneUse())) {
3606 Instruction *NewOp =
3607 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3609 SI0->getName()), I);
3610 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3611 SI1->getOperand(1));
3615 return Changed ? &I : 0;
3618 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3619 /// in the result. If it does, and if the specified byte hasn't been filled in
3620 /// yet, fill it in and return false.
3621 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3622 Instruction *I = dyn_cast<Instruction>(V);
3623 if (I == 0) return true;
3625 // If this is an or instruction, it is an inner node of the bswap.
3626 if (I->getOpcode() == Instruction::Or)
3627 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3628 CollectBSwapParts(I->getOperand(1), ByteValues);
3630 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3631 // If this is a shift by a constant int, and it is "24", then its operand
3632 // defines a byte. We only handle unsigned types here.
3633 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3634 // Not shifting the entire input by N-1 bytes?
3635 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3636 8*(ByteValues.size()-1))
3640 if (I->getOpcode() == Instruction::Shl) {
3641 // X << 24 defines the top byte with the lowest of the input bytes.
3642 DestNo = ByteValues.size()-1;
3644 // X >>u 24 defines the low byte with the highest of the input bytes.
3648 // If the destination byte value is already defined, the values are or'd
3649 // together, which isn't a bswap (unless it's an or of the same bits).
3650 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3652 ByteValues[DestNo] = I->getOperand(0);
3656 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3658 Value *Shift = 0, *ShiftLHS = 0;
3659 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3660 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3661 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3663 Instruction *SI = cast<Instruction>(Shift);
3665 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3666 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3667 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3670 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3672 if (AndAmt->getValue().getActiveBits() > 64)
3674 uint64_t AndAmtVal = AndAmt->getZExtValue();
3675 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3676 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3678 // Unknown mask for bswap.
3679 if (DestByte == ByteValues.size()) return true;
3681 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3683 if (SI->getOpcode() == Instruction::Shl)
3684 SrcByte = DestByte - ShiftBytes;
3686 SrcByte = DestByte + ShiftBytes;
3688 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3689 if (SrcByte != ByteValues.size()-DestByte-1)
3692 // If the destination byte value is already defined, the values are or'd
3693 // together, which isn't a bswap (unless it's an or of the same bits).
3694 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3696 ByteValues[DestByte] = SI->getOperand(0);
3700 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3701 /// If so, insert the new bswap intrinsic and return it.
3702 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3703 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3704 if (!ITy || ITy->getBitWidth() % 16)
3705 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3707 /// ByteValues - For each byte of the result, we keep track of which value
3708 /// defines each byte.
3709 SmallVector<Value*, 8> ByteValues;
3710 ByteValues.resize(ITy->getBitWidth()/8);
3712 // Try to find all the pieces corresponding to the bswap.
3713 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3714 CollectBSwapParts(I.getOperand(1), ByteValues))
3717 // Check to see if all of the bytes come from the same value.
3718 Value *V = ByteValues[0];
3719 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3721 // Check to make sure that all of the bytes come from the same value.
3722 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3723 if (ByteValues[i] != V)
3725 const Type *Tys[] = { ITy };
3726 Module *M = I.getParent()->getParent()->getParent();
3727 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3728 return new CallInst(F, V);
3732 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3733 bool Changed = SimplifyCommutative(I);
3734 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3736 if (isa<UndefValue>(Op1)) // X | undef -> -1
3737 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3741 return ReplaceInstUsesWith(I, Op0);
3743 // See if we can simplify any instructions used by the instruction whose sole
3744 // purpose is to compute bits we don't care about.
3745 if (!isa<VectorType>(I.getType())) {
3746 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3747 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3748 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3749 KnownZero, KnownOne))
3751 } else if (isa<ConstantAggregateZero>(Op1)) {
3752 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3753 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3754 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3755 return ReplaceInstUsesWith(I, I.getOperand(1));
3761 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3762 ConstantInt *C1 = 0; Value *X = 0;
3763 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3764 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3765 Instruction *Or = BinaryOperator::createOr(X, RHS);
3766 InsertNewInstBefore(Or, I);
3768 return BinaryOperator::createAnd(Or,
3769 ConstantInt::get(RHS->getValue() | C1->getValue()));
3772 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3773 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3774 Instruction *Or = BinaryOperator::createOr(X, RHS);
3775 InsertNewInstBefore(Or, I);
3777 return BinaryOperator::createXor(Or,
3778 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3781 // Try to fold constant and into select arguments.
3782 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3783 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3785 if (isa<PHINode>(Op0))
3786 if (Instruction *NV = FoldOpIntoPhi(I))
3790 Value *A = 0, *B = 0;
3791 ConstantInt *C1 = 0, *C2 = 0;
3793 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3794 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3795 return ReplaceInstUsesWith(I, Op1);
3796 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3797 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3798 return ReplaceInstUsesWith(I, Op0);
3800 // (A | B) | C and A | (B | C) -> bswap if possible.
3801 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3802 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3803 match(Op1, m_Or(m_Value(), m_Value())) ||
3804 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3805 match(Op1, m_Shift(m_Value(), m_Value())))) {
3806 if (Instruction *BSwap = MatchBSwap(I))
3810 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3811 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3812 MaskedValueIsZero(Op1, C1->getValue())) {
3813 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3814 InsertNewInstBefore(NOr, I);
3816 return BinaryOperator::createXor(NOr, C1);
3819 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3820 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3821 MaskedValueIsZero(Op0, C1->getValue())) {
3822 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3823 InsertNewInstBefore(NOr, I);
3825 return BinaryOperator::createXor(NOr, C1);
3829 Value *C = 0, *D = 0;
3830 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3831 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3832 Value *V1 = 0, *V2 = 0, *V3 = 0;
3833 C1 = dyn_cast<ConstantInt>(C);
3834 C2 = dyn_cast<ConstantInt>(D);
3835 if (C1 && C2) { // (A & C1)|(B & C2)
3836 // If we have: ((V + N) & C1) | (V & C2)
3837 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3838 // replace with V+N.
3839 if (C1->getValue() == ~C2->getValue()) {
3840 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3841 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3842 // Add commutes, try both ways.
3843 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3844 return ReplaceInstUsesWith(I, A);
3845 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3846 return ReplaceInstUsesWith(I, A);
3848 // Or commutes, try both ways.
3849 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3850 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3851 // Add commutes, try both ways.
3852 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3853 return ReplaceInstUsesWith(I, B);
3854 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3855 return ReplaceInstUsesWith(I, B);
3858 V1 = 0; V2 = 0; V3 = 0;
3861 // Check to see if we have any common things being and'ed. If so, find the
3862 // terms for V1 & (V2|V3).
3863 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3864 if (A == B) // (A & C)|(A & D) == A & (C|D)
3865 V1 = A, V2 = C, V3 = D;
3866 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3867 V1 = A, V2 = B, V3 = C;
3868 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3869 V1 = C, V2 = A, V3 = D;
3870 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3871 V1 = C, V2 = A, V3 = B;
3875 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3876 return BinaryOperator::createAnd(V1, Or);
3881 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3882 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3883 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3884 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3885 SI0->getOperand(1) == SI1->getOperand(1) &&
3886 (SI0->hasOneUse() || SI1->hasOneUse())) {
3887 Instruction *NewOp =
3888 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3890 SI0->getName()), I);
3891 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3892 SI1->getOperand(1));
3896 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3897 if (A == Op1) // ~A | A == -1
3898 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3902 // Note, A is still live here!
3903 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3905 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3907 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3908 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3909 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3910 I.getName()+".demorgan"), I);
3911 return BinaryOperator::createNot(And);
3915 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3916 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3917 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3920 Value *LHSVal, *RHSVal;
3921 ConstantInt *LHSCst, *RHSCst;
3922 ICmpInst::Predicate LHSCC, RHSCC;
3923 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3924 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3925 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3926 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3927 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3928 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3929 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3930 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3931 // We can't fold (ugt x, C) | (sgt x, C2).
3932 PredicatesFoldable(LHSCC, RHSCC)) {
3933 // Ensure that the larger constant is on the RHS.
3934 ICmpInst *LHS = cast<ICmpInst>(Op0);
3936 if (ICmpInst::isSignedPredicate(LHSCC))
3937 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3939 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3942 std::swap(LHS, RHS);
3943 std::swap(LHSCst, RHSCst);
3944 std::swap(LHSCC, RHSCC);
3947 // At this point, we know we have have two icmp instructions
3948 // comparing a value against two constants and or'ing the result
3949 // together. Because of the above check, we know that we only have
3950 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3951 // FoldICmpLogical check above), that the two constants are not
3953 assert(LHSCst != RHSCst && "Compares not folded above?");
3956 default: assert(0 && "Unknown integer condition code!");
3957 case ICmpInst::ICMP_EQ:
3959 default: assert(0 && "Unknown integer condition code!");
3960 case ICmpInst::ICMP_EQ:
3961 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3962 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3963 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3964 LHSVal->getName()+".off");
3965 InsertNewInstBefore(Add, I);
3966 AddCST = Subtract(AddOne(RHSCst), LHSCst);
3967 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3969 break; // (X == 13 | X == 15) -> no change
3970 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3971 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3973 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3974 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3975 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3976 return ReplaceInstUsesWith(I, RHS);
3979 case ICmpInst::ICMP_NE:
3981 default: assert(0 && "Unknown integer condition code!");
3982 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3983 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3984 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3985 return ReplaceInstUsesWith(I, LHS);
3986 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3987 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3988 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3989 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3992 case ICmpInst::ICMP_ULT:
3994 default: assert(0 && "Unknown integer condition code!");
3995 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3997 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3998 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4000 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4002 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4003 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4004 return ReplaceInstUsesWith(I, RHS);
4005 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4009 case ICmpInst::ICMP_SLT:
4011 default: assert(0 && "Unknown integer condition code!");
4012 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4014 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4015 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4017 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4019 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4020 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4021 return ReplaceInstUsesWith(I, RHS);
4022 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4026 case ICmpInst::ICMP_UGT:
4028 default: assert(0 && "Unknown integer condition code!");
4029 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4030 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4031 return ReplaceInstUsesWith(I, LHS);
4032 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4034 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4035 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4036 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4037 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4041 case ICmpInst::ICMP_SGT:
4043 default: assert(0 && "Unknown integer condition code!");
4044 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4045 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4046 return ReplaceInstUsesWith(I, LHS);
4047 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4049 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4050 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4051 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4052 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4060 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4061 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4062 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4063 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4064 const Type *SrcTy = Op0C->getOperand(0)->getType();
4065 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4066 // Only do this if the casts both really cause code to be generated.
4067 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4069 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4071 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4072 Op1C->getOperand(0),
4074 InsertNewInstBefore(NewOp, I);
4075 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4080 return Changed ? &I : 0;
4083 // XorSelf - Implements: X ^ X --> 0
4086 XorSelf(Value *rhs) : RHS(rhs) {}
4087 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4088 Instruction *apply(BinaryOperator &Xor) const {
4094 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4095 bool Changed = SimplifyCommutative(I);
4096 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4098 if (isa<UndefValue>(Op1))
4099 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4101 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4102 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4103 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4104 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4107 // See if we can simplify any instructions used by the instruction whose sole
4108 // purpose is to compute bits we don't care about.
4109 if (!isa<VectorType>(I.getType())) {
4110 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4111 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4112 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4113 KnownZero, KnownOne))
4115 } else if (isa<ConstantAggregateZero>(Op1)) {
4116 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4119 // Is this a ~ operation?
4120 if (Value *NotOp = dyn_castNotVal(&I)) {
4121 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4122 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4123 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4124 if (Op0I->getOpcode() == Instruction::And ||
4125 Op0I->getOpcode() == Instruction::Or) {
4126 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4127 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4129 BinaryOperator::createNot(Op0I->getOperand(1),
4130 Op0I->getOperand(1)->getName()+".not");
4131 InsertNewInstBefore(NotY, I);
4132 if (Op0I->getOpcode() == Instruction::And)
4133 return BinaryOperator::createOr(Op0NotVal, NotY);
4135 return BinaryOperator::createAnd(Op0NotVal, NotY);
4142 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4143 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4144 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4145 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4146 return new ICmpInst(ICI->getInversePredicate(),
4147 ICI->getOperand(0), ICI->getOperand(1));
4149 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4150 return new FCmpInst(FCI->getInversePredicate(),
4151 FCI->getOperand(0), FCI->getOperand(1));
4154 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4155 // ~(c-X) == X-c-1 == X+(-c-1)
4156 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4157 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4158 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4159 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4160 ConstantInt::get(I.getType(), 1));
4161 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4164 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4165 if (Op0I->getOpcode() == Instruction::Add) {
4166 // ~(X-c) --> (-c-1)-X
4167 if (RHS->isAllOnesValue()) {
4168 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4169 return BinaryOperator::createSub(
4170 ConstantExpr::getSub(NegOp0CI,
4171 ConstantInt::get(I.getType(), 1)),
4172 Op0I->getOperand(0));
4173 } else if (RHS->getValue().isSignBit()) {
4174 // (X + C) ^ signbit -> (X + C + signbit)
4175 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4176 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4179 } else if (Op0I->getOpcode() == Instruction::Or) {
4180 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4181 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4182 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4183 // Anything in both C1 and C2 is known to be zero, remove it from
4185 Constant *CommonBits = And(Op0CI, RHS);
4186 NewRHS = ConstantExpr::getAnd(NewRHS,
4187 ConstantExpr::getNot(CommonBits));
4188 AddToWorkList(Op0I);
4189 I.setOperand(0, Op0I->getOperand(0));
4190 I.setOperand(1, NewRHS);
4196 // Try to fold constant and into select arguments.
4197 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4198 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4200 if (isa<PHINode>(Op0))
4201 if (Instruction *NV = FoldOpIntoPhi(I))
4205 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4207 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4209 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4211 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4214 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4217 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4218 if (A == Op0) { // B^(B|A) == (A|B)^B
4219 Op1I->swapOperands();
4221 std::swap(Op0, Op1);
4222 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4223 I.swapOperands(); // Simplified below.
4224 std::swap(Op0, Op1);
4226 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4227 if (Op0 == A) // A^(A^B) == B
4228 return ReplaceInstUsesWith(I, B);
4229 else if (Op0 == B) // A^(B^A) == B
4230 return ReplaceInstUsesWith(I, A);
4231 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4232 if (A == Op0) { // A^(A&B) -> A^(B&A)
4233 Op1I->swapOperands();
4236 if (B == Op0) { // A^(B&A) -> (B&A)^A
4237 I.swapOperands(); // Simplified below.
4238 std::swap(Op0, Op1);
4243 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4246 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4247 if (A == Op1) // (B|A)^B == (A|B)^B
4249 if (B == Op1) { // (A|B)^B == A & ~B
4251 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4252 return BinaryOperator::createAnd(A, NotB);
4254 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4255 if (Op1 == A) // (A^B)^A == B
4256 return ReplaceInstUsesWith(I, B);
4257 else if (Op1 == B) // (B^A)^A == B
4258 return ReplaceInstUsesWith(I, A);
4259 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4260 if (A == Op1) // (A&B)^A -> (B&A)^A
4262 if (B == Op1 && // (B&A)^A == ~B & A
4263 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4265 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4266 return BinaryOperator::createAnd(N, Op1);
4271 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4272 if (Op0I && Op1I && Op0I->isShift() &&
4273 Op0I->getOpcode() == Op1I->getOpcode() &&
4274 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4275 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4276 Instruction *NewOp =
4277 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4278 Op1I->getOperand(0),
4279 Op0I->getName()), I);
4280 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4281 Op1I->getOperand(1));
4285 Value *A, *B, *C, *D;
4286 // (A & B)^(A | B) -> A ^ B
4287 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4288 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4289 if ((A == C && B == D) || (A == D && B == C))
4290 return BinaryOperator::createXor(A, B);
4292 // (A | B)^(A & B) -> A ^ B
4293 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4294 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4295 if ((A == C && B == D) || (A == D && B == C))
4296 return BinaryOperator::createXor(A, B);
4300 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4301 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4302 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4303 // (X & Y)^(X & Y) -> (Y^Z) & X
4304 Value *X = 0, *Y = 0, *Z = 0;
4306 X = A, Y = B, Z = D;
4308 X = A, Y = B, Z = C;
4310 X = B, Y = A, Z = D;
4312 X = B, Y = A, Z = C;
4315 Instruction *NewOp =
4316 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4317 return BinaryOperator::createAnd(NewOp, X);
4322 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4323 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4324 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4327 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4328 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4329 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4330 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4331 const Type *SrcTy = Op0C->getOperand(0)->getType();
4332 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4333 // Only do this if the casts both really cause code to be generated.
4334 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4336 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4338 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4339 Op1C->getOperand(0),
4341 InsertNewInstBefore(NewOp, I);
4342 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4346 return Changed ? &I : 0;
4349 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4350 /// overflowed for this type.
4351 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4352 ConstantInt *In2, bool IsSigned = false) {
4353 Result = cast<ConstantInt>(Add(In1, In2));
4356 if (In2->getValue().isNegative())
4357 return Result->getValue().sgt(In1->getValue());
4359 return Result->getValue().slt(In1->getValue());
4361 return Result->getValue().ult(In1->getValue());
4364 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4365 /// code necessary to compute the offset from the base pointer (without adding
4366 /// in the base pointer). Return the result as a signed integer of intptr size.
4367 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4368 TargetData &TD = IC.getTargetData();
4369 gep_type_iterator GTI = gep_type_begin(GEP);
4370 const Type *IntPtrTy = TD.getIntPtrType();
4371 Value *Result = Constant::getNullValue(IntPtrTy);
4373 // Build a mask for high order bits.
4374 unsigned IntPtrWidth = TD.getPointerSize()*8;
4375 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4377 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4378 Value *Op = GEP->getOperand(i);
4379 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4380 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4381 if (OpC->isZero()) continue;
4383 // Handle a struct index, which adds its field offset to the pointer.
4384 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4385 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4387 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4388 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4390 Result = IC.InsertNewInstBefore(
4391 BinaryOperator::createAdd(Result,
4392 ConstantInt::get(IntPtrTy, Size),
4393 GEP->getName()+".offs"), I);
4397 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4398 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4399 Scale = ConstantExpr::getMul(OC, Scale);
4400 if (Constant *RC = dyn_cast<Constant>(Result))
4401 Result = ConstantExpr::getAdd(RC, Scale);
4403 // Emit an add instruction.
4404 Result = IC.InsertNewInstBefore(
4405 BinaryOperator::createAdd(Result, Scale,
4406 GEP->getName()+".offs"), I);
4410 // Convert to correct type.
4411 if (Op->getType() != IntPtrTy) {
4412 if (Constant *OpC = dyn_cast<Constant>(Op))
4413 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4415 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4416 Op->getName()+".c"), I);
4419 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4420 if (Constant *OpC = dyn_cast<Constant>(Op))
4421 Op = ConstantExpr::getMul(OpC, Scale);
4422 else // We'll let instcombine(mul) convert this to a shl if possible.
4423 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4424 GEP->getName()+".idx"), I);
4427 // Emit an add instruction.
4428 if (isa<Constant>(Op) && isa<Constant>(Result))
4429 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4430 cast<Constant>(Result));
4432 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4433 GEP->getName()+".offs"), I);
4438 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4439 /// else. At this point we know that the GEP is on the LHS of the comparison.
4440 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4441 ICmpInst::Predicate Cond,
4443 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4445 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4446 if (isa<PointerType>(CI->getOperand(0)->getType()))
4447 RHS = CI->getOperand(0);
4449 Value *PtrBase = GEPLHS->getOperand(0);
4450 if (PtrBase == RHS) {
4451 // As an optimization, we don't actually have to compute the actual value of
4452 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4453 // each index is zero or not.
4454 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4455 Instruction *InVal = 0;
4456 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4457 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4459 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4460 if (isa<UndefValue>(C)) // undef index -> undef.
4461 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4462 if (C->isNullValue())
4464 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4465 EmitIt = false; // This is indexing into a zero sized array?
4466 } else if (isa<ConstantInt>(C))
4467 return ReplaceInstUsesWith(I, // No comparison is needed here.
4468 ConstantInt::get(Type::Int1Ty,
4469 Cond == ICmpInst::ICMP_NE));
4474 new ICmpInst(Cond, GEPLHS->getOperand(i),
4475 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4479 InVal = InsertNewInstBefore(InVal, I);
4480 InsertNewInstBefore(Comp, I);
4481 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4482 InVal = BinaryOperator::createOr(InVal, Comp);
4483 else // True if all are equal
4484 InVal = BinaryOperator::createAnd(InVal, Comp);
4492 // No comparison is needed here, all indexes = 0
4493 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4494 Cond == ICmpInst::ICMP_EQ));
4497 // Only lower this if the icmp is the only user of the GEP or if we expect
4498 // the result to fold to a constant!
4499 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4500 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4501 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4502 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4503 Constant::getNullValue(Offset->getType()));
4505 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4506 // If the base pointers are different, but the indices are the same, just
4507 // compare the base pointer.
4508 if (PtrBase != GEPRHS->getOperand(0)) {
4509 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4510 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4511 GEPRHS->getOperand(0)->getType();
4513 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4514 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4515 IndicesTheSame = false;
4519 // If all indices are the same, just compare the base pointers.
4521 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4522 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4524 // Otherwise, the base pointers are different and the indices are
4525 // different, bail out.
4529 // If one of the GEPs has all zero indices, recurse.
4530 bool AllZeros = true;
4531 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4532 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4533 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4538 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4539 ICmpInst::getSwappedPredicate(Cond), I);
4541 // If the other GEP has all zero indices, recurse.
4543 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4544 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4545 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4550 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4552 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4553 // If the GEPs only differ by one index, compare it.
4554 unsigned NumDifferences = 0; // Keep track of # differences.
4555 unsigned DiffOperand = 0; // The operand that differs.
4556 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4557 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4558 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4559 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4560 // Irreconcilable differences.
4564 if (NumDifferences++) break;
4569 if (NumDifferences == 0) // SAME GEP?
4570 return ReplaceInstUsesWith(I, // No comparison is needed here.
4571 ConstantInt::get(Type::Int1Ty,
4572 isTrueWhenEqual(Cond)));
4574 else if (NumDifferences == 1) {
4575 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4576 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4577 // Make sure we do a signed comparison here.
4578 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4582 // Only lower this if the icmp is the only user of the GEP or if we expect
4583 // the result to fold to a constant!
4584 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4585 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4586 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4587 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4588 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4589 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4595 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4596 bool Changed = SimplifyCompare(I);
4597 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4599 // Fold trivial predicates.
4600 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4601 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4602 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4603 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4605 // Simplify 'fcmp pred X, X'
4607 switch (I.getPredicate()) {
4608 default: assert(0 && "Unknown predicate!");
4609 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4610 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4611 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4612 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4613 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4614 case FCmpInst::FCMP_OLT: // True if ordered and less than
4615 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4616 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4618 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4619 case FCmpInst::FCMP_ULT: // True if unordered or less than
4620 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4621 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4622 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4623 I.setPredicate(FCmpInst::FCMP_UNO);
4624 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4627 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4628 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4629 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4630 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4631 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4632 I.setPredicate(FCmpInst::FCMP_ORD);
4633 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4638 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4639 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4641 // Handle fcmp with constant RHS
4642 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4643 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4644 switch (LHSI->getOpcode()) {
4645 case Instruction::PHI:
4646 if (Instruction *NV = FoldOpIntoPhi(I))
4649 case Instruction::Select:
4650 // If either operand of the select is a constant, we can fold the
4651 // comparison into the select arms, which will cause one to be
4652 // constant folded and the select turned into a bitwise or.
4653 Value *Op1 = 0, *Op2 = 0;
4654 if (LHSI->hasOneUse()) {
4655 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4656 // Fold the known value into the constant operand.
4657 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4658 // Insert a new FCmp of the other select operand.
4659 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4660 LHSI->getOperand(2), RHSC,
4662 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4663 // Fold the known value into the constant operand.
4664 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4665 // Insert a new FCmp of the other select operand.
4666 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4667 LHSI->getOperand(1), RHSC,
4673 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4678 return Changed ? &I : 0;
4681 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4682 bool Changed = SimplifyCompare(I);
4683 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4684 const Type *Ty = Op0->getType();
4688 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4689 isTrueWhenEqual(I)));
4691 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4692 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4694 // icmp of GlobalValues can never equal each other as long as they aren't
4695 // external weak linkage type.
4696 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4697 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4698 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4699 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4700 !isTrueWhenEqual(I)));
4702 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4703 // addresses never equal each other! We already know that Op0 != Op1.
4704 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4705 isa<ConstantPointerNull>(Op0)) &&
4706 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4707 isa<ConstantPointerNull>(Op1)))
4708 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4709 !isTrueWhenEqual(I)));
4711 // icmp's with boolean values can always be turned into bitwise operations
4712 if (Ty == Type::Int1Ty) {
4713 switch (I.getPredicate()) {
4714 default: assert(0 && "Invalid icmp instruction!");
4715 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4716 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4717 InsertNewInstBefore(Xor, I);
4718 return BinaryOperator::createNot(Xor);
4720 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4721 return BinaryOperator::createXor(Op0, Op1);
4723 case ICmpInst::ICMP_UGT:
4724 case ICmpInst::ICMP_SGT:
4725 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4727 case ICmpInst::ICMP_ULT:
4728 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4729 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4730 InsertNewInstBefore(Not, I);
4731 return BinaryOperator::createAnd(Not, Op1);
4733 case ICmpInst::ICMP_UGE:
4734 case ICmpInst::ICMP_SGE:
4735 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4737 case ICmpInst::ICMP_ULE:
4738 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4739 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4740 InsertNewInstBefore(Not, I);
4741 return BinaryOperator::createOr(Not, Op1);
4746 // See if we are doing a comparison between a constant and an instruction that
4747 // can be folded into the comparison.
4748 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4749 switch (I.getPredicate()) {
4751 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4752 if (CI->isMinValue(false))
4753 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4754 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4755 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4756 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4757 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4758 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4759 if (CI->isMinValue(true))
4760 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4761 ConstantInt::getAllOnesValue(Op0->getType()));
4765 case ICmpInst::ICMP_SLT:
4766 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4767 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4768 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4769 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4770 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4771 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4774 case ICmpInst::ICMP_UGT:
4775 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4776 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4777 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4778 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4779 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4780 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4782 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4783 if (CI->isMaxValue(true))
4784 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4785 ConstantInt::getNullValue(Op0->getType()));
4788 case ICmpInst::ICMP_SGT:
4789 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4790 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4791 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4792 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4793 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4794 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4797 case ICmpInst::ICMP_ULE:
4798 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4799 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4800 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4801 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4802 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4803 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4806 case ICmpInst::ICMP_SLE:
4807 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4808 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4809 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4810 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4811 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4812 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4815 case ICmpInst::ICMP_UGE:
4816 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4817 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4818 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4819 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4820 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4821 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4824 case ICmpInst::ICMP_SGE:
4825 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4826 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4827 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4828 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4829 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4830 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4834 // If we still have a icmp le or icmp ge instruction, turn it into the
4835 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4836 // already been handled above, this requires little checking.
4838 switch (I.getPredicate()) {
4840 case ICmpInst::ICMP_ULE:
4841 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4842 case ICmpInst::ICMP_SLE:
4843 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4844 case ICmpInst::ICMP_UGE:
4845 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4846 case ICmpInst::ICMP_SGE:
4847 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4850 // See if we can fold the comparison based on bits known to be zero or one
4851 // in the input. If this comparison is a normal comparison, it demands all
4852 // bits, if it is a sign bit comparison, it only demands the sign bit.
4855 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4857 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4858 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4859 if (SimplifyDemandedBits(Op0,
4860 isSignBit ? APInt::getSignBit(BitWidth)
4861 : APInt::getAllOnesValue(BitWidth),
4862 KnownZero, KnownOne, 0))
4865 // Given the known and unknown bits, compute a range that the LHS could be
4867 if ((KnownOne | KnownZero) != 0) {
4868 // Compute the Min, Max and RHS values based on the known bits. For the
4869 // EQ and NE we use unsigned values.
4870 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4871 const APInt& RHSVal = CI->getValue();
4872 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4873 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4876 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4879 switch (I.getPredicate()) { // LE/GE have been folded already.
4880 default: assert(0 && "Unknown icmp opcode!");
4881 case ICmpInst::ICMP_EQ:
4882 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4883 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4885 case ICmpInst::ICMP_NE:
4886 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4887 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4889 case ICmpInst::ICMP_ULT:
4890 if (Max.ult(RHSVal))
4891 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4892 if (Min.uge(RHSVal))
4893 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4895 case ICmpInst::ICMP_UGT:
4896 if (Min.ugt(RHSVal))
4897 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4898 if (Max.ule(RHSVal))
4899 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4901 case ICmpInst::ICMP_SLT:
4902 if (Max.slt(RHSVal))
4903 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4904 if (Min.sgt(RHSVal))
4905 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4907 case ICmpInst::ICMP_SGT:
4908 if (Min.sgt(RHSVal))
4909 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4910 if (Max.sle(RHSVal))
4911 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4916 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4917 // instruction, see if that instruction also has constants so that the
4918 // instruction can be folded into the icmp
4919 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4920 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4924 // Handle icmp with constant (but not simple integer constant) RHS
4925 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4926 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4927 switch (LHSI->getOpcode()) {
4928 case Instruction::GetElementPtr:
4929 if (RHSC->isNullValue()) {
4930 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4931 bool isAllZeros = true;
4932 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4933 if (!isa<Constant>(LHSI->getOperand(i)) ||
4934 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4939 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
4940 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4944 case Instruction::PHI:
4945 if (Instruction *NV = FoldOpIntoPhi(I))
4948 case Instruction::Select: {
4949 // If either operand of the select is a constant, we can fold the
4950 // comparison into the select arms, which will cause one to be
4951 // constant folded and the select turned into a bitwise or.
4952 Value *Op1 = 0, *Op2 = 0;
4953 if (LHSI->hasOneUse()) {
4954 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4955 // Fold the known value into the constant operand.
4956 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4957 // Insert a new ICmp of the other select operand.
4958 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4959 LHSI->getOperand(2), RHSC,
4961 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4962 // Fold the known value into the constant operand.
4963 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4964 // Insert a new ICmp of the other select operand.
4965 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4966 LHSI->getOperand(1), RHSC,
4972 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4975 case Instruction::Malloc:
4976 // If we have (malloc != null), and if the malloc has a single use, we
4977 // can assume it is successful and remove the malloc.
4978 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
4979 AddToWorkList(LHSI);
4980 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4981 !isTrueWhenEqual(I)));
4987 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4988 if (User *GEP = dyn_castGetElementPtr(Op0))
4989 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
4991 if (User *GEP = dyn_castGetElementPtr(Op1))
4992 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
4993 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
4996 // Test to see if the operands of the icmp are casted versions of other
4997 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4999 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5000 if (isa<PointerType>(Op0->getType()) &&
5001 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
5002 // We keep moving the cast from the left operand over to the right
5003 // operand, where it can often be eliminated completely.
5004 Op0 = CI->getOperand(0);
5006 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5007 // so eliminate it as well.
5008 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5009 Op1 = CI2->getOperand(0);
5011 // If Op1 is a constant, we can fold the cast into the constant.
5012 if (Op0->getType() != Op1->getType())
5013 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5014 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5016 // Otherwise, cast the RHS right before the icmp
5017 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5019 return new ICmpInst(I.getPredicate(), Op0, Op1);
5023 if (isa<CastInst>(Op0)) {
5024 // Handle the special case of: icmp (cast bool to X), <cst>
5025 // This comes up when you have code like
5028 // For generality, we handle any zero-extension of any operand comparison
5029 // with a constant or another cast from the same type.
5030 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5031 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5035 if (I.isEquality()) {
5036 Value *A, *B, *C, *D;
5037 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5038 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5039 Value *OtherVal = A == Op1 ? B : A;
5040 return new ICmpInst(I.getPredicate(), OtherVal,
5041 Constant::getNullValue(A->getType()));
5044 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5045 // A^c1 == C^c2 --> A == C^(c1^c2)
5046 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5047 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5048 if (Op1->hasOneUse()) {
5049 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5050 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5051 return new ICmpInst(I.getPredicate(), A,
5052 InsertNewInstBefore(Xor, I));
5055 // A^B == A^D -> B == D
5056 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5057 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5058 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5059 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5063 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5064 (A == Op0 || B == Op0)) {
5065 // A == (A^B) -> B == 0
5066 Value *OtherVal = A == Op0 ? B : A;
5067 return new ICmpInst(I.getPredicate(), OtherVal,
5068 Constant::getNullValue(A->getType()));
5070 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5071 // (A-B) == A -> B == 0
5072 return new ICmpInst(I.getPredicate(), B,
5073 Constant::getNullValue(B->getType()));
5075 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5076 // A == (A-B) -> B == 0
5077 return new ICmpInst(I.getPredicate(), B,
5078 Constant::getNullValue(B->getType()));
5081 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5082 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5083 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5084 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5085 Value *X = 0, *Y = 0, *Z = 0;
5088 X = B; Y = D; Z = A;
5089 } else if (A == D) {
5090 X = B; Y = C; Z = A;
5091 } else if (B == C) {
5092 X = A; Y = D; Z = B;
5093 } else if (B == D) {
5094 X = A; Y = C; Z = B;
5097 if (X) { // Build (X^Y) & Z
5098 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5099 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5100 I.setOperand(0, Op1);
5101 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5106 return Changed ? &I : 0;
5110 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5111 /// and CmpRHS are both known to be integer constants.
5112 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5113 ConstantInt *DivRHS) {
5114 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5115 const APInt &CmpRHSV = CmpRHS->getValue();
5117 // FIXME: If the operand types don't match the type of the divide
5118 // then don't attempt this transform. The code below doesn't have the
5119 // logic to deal with a signed divide and an unsigned compare (and
5120 // vice versa). This is because (x /s C1) <s C2 produces different
5121 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5122 // (x /u C1) <u C2. Simply casting the operands and result won't
5123 // work. :( The if statement below tests that condition and bails
5125 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5126 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5128 if (DivRHS->isZero())
5129 return 0; // The ProdOV computation fails on divide by zero.
5131 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5132 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5133 // C2 (CI). By solving for X we can turn this into a range check
5134 // instead of computing a divide.
5135 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5137 // Determine if the product overflows by seeing if the product is
5138 // not equal to the divide. Make sure we do the same kind of divide
5139 // as in the LHS instruction that we're folding.
5140 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5141 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5143 // Get the ICmp opcode
5144 ICmpInst::Predicate Pred = ICI.getPredicate();
5146 // Figure out the interval that is being checked. For example, a comparison
5147 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5148 // Compute this interval based on the constants involved and the signedness of
5149 // the compare/divide. This computes a half-open interval, keeping track of
5150 // whether either value in the interval overflows. After analysis each
5151 // overflow variable is set to 0 if it's corresponding bound variable is valid
5152 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5153 int LoOverflow = 0, HiOverflow = 0;
5154 ConstantInt *LoBound = 0, *HiBound = 0;
5157 if (!DivIsSigned) { // udiv
5158 // e.g. X/5 op 3 --> [15, 20)
5160 HiOverflow = LoOverflow = ProdOV;
5162 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5163 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5164 if (CmpRHSV == 0) { // (X / pos) op 0
5165 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5166 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5168 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5169 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5170 HiOverflow = LoOverflow = ProdOV;
5172 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5173 } else { // (X / pos) op neg
5174 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5175 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5176 LoOverflow = AddWithOverflow(LoBound, Prod,
5177 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5178 HiBound = AddOne(Prod);
5179 HiOverflow = ProdOV ? -1 : 0;
5181 } else { // Divisor is < 0.
5182 if (CmpRHSV == 0) { // (X / neg) op 0
5183 // e.g. X/-5 op 0 --> [-4, 5)
5184 LoBound = AddOne(DivRHS);
5185 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5186 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5187 HiOverflow = 1; // [INTMIN+1, overflow)
5188 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5190 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5191 // e.g. X/-5 op 3 --> [-19, -14)
5192 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5194 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5195 HiBound = AddOne(Prod);
5196 } else { // (X / neg) op neg
5197 // e.g. X/-5 op -3 --> [15, 20)
5199 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5200 HiBound = Subtract(Prod, DivRHS);
5203 // Dividing by a negative swaps the condition. LT <-> GT
5204 Pred = ICmpInst::getSwappedPredicate(Pred);
5207 Value *X = DivI->getOperand(0);
5209 default: assert(0 && "Unhandled icmp opcode!");
5210 case ICmpInst::ICMP_EQ:
5211 if (LoOverflow && HiOverflow)
5212 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5213 else if (HiOverflow)
5214 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5215 ICmpInst::ICMP_UGE, X, LoBound);
5216 else if (LoOverflow)
5217 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5218 ICmpInst::ICMP_ULT, X, HiBound);
5220 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5221 case ICmpInst::ICMP_NE:
5222 if (LoOverflow && HiOverflow)
5223 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5224 else if (HiOverflow)
5225 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5226 ICmpInst::ICMP_ULT, X, LoBound);
5227 else if (LoOverflow)
5228 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5229 ICmpInst::ICMP_UGE, X, HiBound);
5231 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5232 case ICmpInst::ICMP_ULT:
5233 case ICmpInst::ICMP_SLT:
5234 if (LoOverflow == +1) // Low bound is greater than input range.
5235 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5236 if (LoOverflow == -1) // Low bound is less than input range.
5237 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5238 return new ICmpInst(Pred, X, LoBound);
5239 case ICmpInst::ICMP_UGT:
5240 case ICmpInst::ICMP_SGT:
5241 if (HiOverflow == +1) // High bound greater than input range.
5242 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5243 else if (HiOverflow == -1) // High bound less than input range.
5244 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5245 if (Pred == ICmpInst::ICMP_UGT)
5246 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5248 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5253 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5255 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5258 const APInt &RHSV = RHS->getValue();
5260 switch (LHSI->getOpcode()) {
5261 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5262 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5263 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5265 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5266 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5267 Value *CompareVal = LHSI->getOperand(0);
5269 // If the sign bit of the XorCST is not set, there is no change to
5270 // the operation, just stop using the Xor.
5271 if (!XorCST->getValue().isNegative()) {
5272 ICI.setOperand(0, CompareVal);
5273 AddToWorkList(LHSI);
5277 // Was the old condition true if the operand is positive?
5278 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5280 // If so, the new one isn't.
5281 isTrueIfPositive ^= true;
5283 if (isTrueIfPositive)
5284 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5286 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5290 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5291 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5292 LHSI->getOperand(0)->hasOneUse()) {
5293 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5295 // If the LHS is an AND of a truncating cast, we can widen the
5296 // and/compare to be the input width without changing the value
5297 // produced, eliminating a cast.
5298 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5299 // We can do this transformation if either the AND constant does not
5300 // have its sign bit set or if it is an equality comparison.
5301 // Extending a relational comparison when we're checking the sign
5302 // bit would not work.
5303 if (Cast->hasOneUse() &&
5304 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5305 RHSV.isPositive())) {
5307 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5308 APInt NewCST = AndCST->getValue();
5309 NewCST.zext(BitWidth);
5311 NewCI.zext(BitWidth);
5312 Instruction *NewAnd =
5313 BinaryOperator::createAnd(Cast->getOperand(0),
5314 ConstantInt::get(NewCST),LHSI->getName());
5315 InsertNewInstBefore(NewAnd, ICI);
5316 return new ICmpInst(ICI.getPredicate(), NewAnd,
5317 ConstantInt::get(NewCI));
5321 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5322 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5323 // happens a LOT in code produced by the C front-end, for bitfield
5325 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5326 if (Shift && !Shift->isShift())
5330 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5331 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5332 const Type *AndTy = AndCST->getType(); // Type of the and.
5334 // We can fold this as long as we can't shift unknown bits
5335 // into the mask. This can only happen with signed shift
5336 // rights, as they sign-extend.
5338 bool CanFold = Shift->isLogicalShift();
5340 // To test for the bad case of the signed shr, see if any
5341 // of the bits shifted in could be tested after the mask.
5342 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5343 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5345 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5346 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5347 AndCST->getValue()) == 0)
5353 if (Shift->getOpcode() == Instruction::Shl)
5354 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5356 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5358 // Check to see if we are shifting out any of the bits being
5360 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5361 // If we shifted bits out, the fold is not going to work out.
5362 // As a special case, check to see if this means that the
5363 // result is always true or false now.
5364 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5365 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5366 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5367 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5369 ICI.setOperand(1, NewCst);
5370 Constant *NewAndCST;
5371 if (Shift->getOpcode() == Instruction::Shl)
5372 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5374 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5375 LHSI->setOperand(1, NewAndCST);
5376 LHSI->setOperand(0, Shift->getOperand(0));
5377 AddToWorkList(Shift); // Shift is dead.
5378 AddUsesToWorkList(ICI);
5384 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5385 // preferable because it allows the C<<Y expression to be hoisted out
5386 // of a loop if Y is invariant and X is not.
5387 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5388 ICI.isEquality() && !Shift->isArithmeticShift() &&
5389 isa<Instruction>(Shift->getOperand(0))) {
5392 if (Shift->getOpcode() == Instruction::LShr) {
5393 NS = BinaryOperator::createShl(AndCST,
5394 Shift->getOperand(1), "tmp");
5396 // Insert a logical shift.
5397 NS = BinaryOperator::createLShr(AndCST,
5398 Shift->getOperand(1), "tmp");
5400 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5402 // Compute X & (C << Y).
5403 Instruction *NewAnd =
5404 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5405 InsertNewInstBefore(NewAnd, ICI);
5407 ICI.setOperand(0, NewAnd);
5413 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5414 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5417 uint32_t TypeBits = RHSV.getBitWidth();
5419 // Check that the shift amount is in range. If not, don't perform
5420 // undefined shifts. When the shift is visited it will be
5422 if (ShAmt->uge(TypeBits))
5425 if (ICI.isEquality()) {
5426 // If we are comparing against bits always shifted out, the
5427 // comparison cannot succeed.
5429 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5430 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5431 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5432 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5433 return ReplaceInstUsesWith(ICI, Cst);
5436 if (LHSI->hasOneUse()) {
5437 // Otherwise strength reduce the shift into an and.
5438 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5440 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5443 BinaryOperator::createAnd(LHSI->getOperand(0),
5444 Mask, LHSI->getName()+".mask");
5445 Value *And = InsertNewInstBefore(AndI, ICI);
5446 return new ICmpInst(ICI.getPredicate(), And,
5447 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5451 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5452 bool TrueIfSigned = false;
5453 if (LHSI->hasOneUse() &&
5454 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5455 // (X << 31) <s 0 --> (X&1) != 0
5456 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5457 (TypeBits-ShAmt->getZExtValue()-1));
5459 BinaryOperator::createAnd(LHSI->getOperand(0),
5460 Mask, LHSI->getName()+".mask");
5461 Value *And = InsertNewInstBefore(AndI, ICI);
5463 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5464 And, Constant::getNullValue(And->getType()));
5469 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5470 case Instruction::AShr: {
5471 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5474 if (ICI.isEquality()) {
5475 // Check that the shift amount is in range. If not, don't perform
5476 // undefined shifts. When the shift is visited it will be
5478 uint32_t TypeBits = RHSV.getBitWidth();
5479 if (ShAmt->uge(TypeBits))
5481 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5483 // If we are comparing against bits always shifted out, the
5484 // comparison cannot succeed.
5485 APInt Comp = RHSV << ShAmtVal;
5486 if (LHSI->getOpcode() == Instruction::LShr)
5487 Comp = Comp.lshr(ShAmtVal);
5489 Comp = Comp.ashr(ShAmtVal);
5491 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5492 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5493 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5494 return ReplaceInstUsesWith(ICI, Cst);
5497 if (LHSI->hasOneUse() || RHSV == 0) {
5498 // Otherwise strength reduce the shift into an and.
5499 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5500 Constant *Mask = ConstantInt::get(Val);
5503 BinaryOperator::createAnd(LHSI->getOperand(0),
5504 Mask, LHSI->getName()+".mask");
5505 Value *And = InsertNewInstBefore(AndI, ICI);
5506 return new ICmpInst(ICI.getPredicate(), And,
5507 ConstantExpr::getShl(RHS, ShAmt));
5513 case Instruction::SDiv:
5514 case Instruction::UDiv:
5515 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5516 // Fold this div into the comparison, producing a range check.
5517 // Determine, based on the divide type, what the range is being
5518 // checked. If there is an overflow on the low or high side, remember
5519 // it, otherwise compute the range [low, hi) bounding the new value.
5520 // See: InsertRangeTest above for the kinds of replacements possible.
5521 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5522 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5528 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5529 if (ICI.isEquality()) {
5530 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5532 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5533 // the second operand is a constant, simplify a bit.
5534 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5535 switch (BO->getOpcode()) {
5536 case Instruction::SRem:
5537 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5538 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5539 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5540 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5541 Instruction *NewRem =
5542 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5544 InsertNewInstBefore(NewRem, ICI);
5545 return new ICmpInst(ICI.getPredicate(), NewRem,
5546 Constant::getNullValue(BO->getType()));
5550 case Instruction::Add:
5551 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5552 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5553 if (BO->hasOneUse())
5554 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5555 Subtract(RHS, BOp1C));
5556 } else if (RHSV == 0) {
5557 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5558 // efficiently invertible, or if the add has just this one use.
5559 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5561 if (Value *NegVal = dyn_castNegVal(BOp1))
5562 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5563 else if (Value *NegVal = dyn_castNegVal(BOp0))
5564 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5565 else if (BO->hasOneUse()) {
5566 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5567 InsertNewInstBefore(Neg, ICI);
5569 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5573 case Instruction::Xor:
5574 // For the xor case, we can xor two constants together, eliminating
5575 // the explicit xor.
5576 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5577 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5578 ConstantExpr::getXor(RHS, BOC));
5581 case Instruction::Sub:
5582 // Replace (([sub|xor] A, B) != 0) with (A != B)
5584 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5588 case Instruction::Or:
5589 // If bits are being or'd in that are not present in the constant we
5590 // are comparing against, then the comparison could never succeed!
5591 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5592 Constant *NotCI = ConstantExpr::getNot(RHS);
5593 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5594 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5599 case Instruction::And:
5600 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5601 // If bits are being compared against that are and'd out, then the
5602 // comparison can never succeed!
5603 if ((RHSV & ~BOC->getValue()) != 0)
5604 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5607 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5608 if (RHS == BOC && RHSV.isPowerOf2())
5609 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5610 ICmpInst::ICMP_NE, LHSI,
5611 Constant::getNullValue(RHS->getType()));
5613 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5614 if (isSignBit(BOC)) {
5615 Value *X = BO->getOperand(0);
5616 Constant *Zero = Constant::getNullValue(X->getType());
5617 ICmpInst::Predicate pred = isICMP_NE ?
5618 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5619 return new ICmpInst(pred, X, Zero);
5622 // ((X & ~7) == 0) --> X < 8
5623 if (RHSV == 0 && isHighOnes(BOC)) {
5624 Value *X = BO->getOperand(0);
5625 Constant *NegX = ConstantExpr::getNeg(BOC);
5626 ICmpInst::Predicate pred = isICMP_NE ?
5627 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5628 return new ICmpInst(pred, X, NegX);
5633 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5634 // Handle icmp {eq|ne} <intrinsic>, intcst.
5635 if (II->getIntrinsicID() == Intrinsic::bswap) {
5637 ICI.setOperand(0, II->getOperand(1));
5638 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5642 } else { // Not a ICMP_EQ/ICMP_NE
5643 // If the LHS is a cast from an integral value of the same size,
5644 // then since we know the RHS is a constant, try to simlify.
5645 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5646 Value *CastOp = Cast->getOperand(0);
5647 const Type *SrcTy = CastOp->getType();
5648 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5649 if (SrcTy->isInteger() &&
5650 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5651 // If this is an unsigned comparison, try to make the comparison use
5652 // smaller constant values.
5653 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5654 // X u< 128 => X s> -1
5655 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5656 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5657 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5658 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5659 // X u> 127 => X s< 0
5660 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5661 Constant::getNullValue(SrcTy));
5669 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5670 /// We only handle extending casts so far.
5672 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5673 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5674 Value *LHSCIOp = LHSCI->getOperand(0);
5675 const Type *SrcTy = LHSCIOp->getType();
5676 const Type *DestTy = LHSCI->getType();
5679 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5680 // integer type is the same size as the pointer type.
5681 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5682 getTargetData().getPointerSizeInBits() ==
5683 cast<IntegerType>(DestTy)->getBitWidth()) {
5685 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5686 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5687 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5688 RHSOp = RHSC->getOperand(0);
5689 // If the pointer types don't match, insert a bitcast.
5690 if (LHSCIOp->getType() != RHSOp->getType())
5691 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5692 LHSCIOp->getType(), ICI);
5696 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5699 // The code below only handles extension cast instructions, so far.
5701 if (LHSCI->getOpcode() != Instruction::ZExt &&
5702 LHSCI->getOpcode() != Instruction::SExt)
5705 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5706 bool isSignedCmp = ICI.isSignedPredicate();
5708 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5709 // Not an extension from the same type?
5710 RHSCIOp = CI->getOperand(0);
5711 if (RHSCIOp->getType() != LHSCIOp->getType())
5714 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5715 // and the other is a zext), then we can't handle this.
5716 if (CI->getOpcode() != LHSCI->getOpcode())
5719 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5720 // then we can't handle this.
5721 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5724 // Okay, just insert a compare of the reduced operands now!
5725 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5728 // If we aren't dealing with a constant on the RHS, exit early
5729 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5733 // Compute the constant that would happen if we truncated to SrcTy then
5734 // reextended to DestTy.
5735 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5736 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5738 // If the re-extended constant didn't change...
5740 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5741 // For example, we might have:
5742 // %A = sext short %X to uint
5743 // %B = icmp ugt uint %A, 1330
5744 // It is incorrect to transform this into
5745 // %B = icmp ugt short %X, 1330
5746 // because %A may have negative value.
5748 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5749 // OR operation is EQ/NE.
5750 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5751 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5756 // The re-extended constant changed so the constant cannot be represented
5757 // in the shorter type. Consequently, we cannot emit a simple comparison.
5759 // First, handle some easy cases. We know the result cannot be equal at this
5760 // point so handle the ICI.isEquality() cases
5761 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5762 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5763 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5764 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5766 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5767 // should have been folded away previously and not enter in here.
5770 // We're performing a signed comparison.
5771 if (cast<ConstantInt>(CI)->getValue().isNegative())
5772 Result = ConstantInt::getFalse(); // X < (small) --> false
5774 Result = ConstantInt::getTrue(); // X < (large) --> true
5776 // We're performing an unsigned comparison.
5778 // We're performing an unsigned comp with a sign extended value.
5779 // This is true if the input is >= 0. [aka >s -1]
5780 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5781 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5782 NegOne, ICI.getName()), ICI);
5784 // Unsigned extend & unsigned compare -> always true.
5785 Result = ConstantInt::getTrue();
5789 // Finally, return the value computed.
5790 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5791 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5792 return ReplaceInstUsesWith(ICI, Result);
5794 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5795 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5796 "ICmp should be folded!");
5797 if (Constant *CI = dyn_cast<Constant>(Result))
5798 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5800 return BinaryOperator::createNot(Result);
5804 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5805 return commonShiftTransforms(I);
5808 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5809 return commonShiftTransforms(I);
5812 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5813 return commonShiftTransforms(I);
5816 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5817 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5818 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5820 // shl X, 0 == X and shr X, 0 == X
5821 // shl 0, X == 0 and shr 0, X == 0
5822 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5823 Op0 == Constant::getNullValue(Op0->getType()))
5824 return ReplaceInstUsesWith(I, Op0);
5826 if (isa<UndefValue>(Op0)) {
5827 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5828 return ReplaceInstUsesWith(I, Op0);
5829 else // undef << X -> 0, undef >>u X -> 0
5830 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5832 if (isa<UndefValue>(Op1)) {
5833 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5834 return ReplaceInstUsesWith(I, Op0);
5835 else // X << undef, X >>u undef -> 0
5836 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5839 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5840 if (I.getOpcode() == Instruction::AShr)
5841 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5842 if (CSI->isAllOnesValue())
5843 return ReplaceInstUsesWith(I, CSI);
5845 // Try to fold constant and into select arguments.
5846 if (isa<Constant>(Op0))
5847 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5848 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5851 // See if we can turn a signed shr into an unsigned shr.
5852 if (I.isArithmeticShift()) {
5853 if (MaskedValueIsZero(Op0,
5854 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5855 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5859 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5860 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5865 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5866 BinaryOperator &I) {
5867 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5869 // See if we can simplify any instructions used by the instruction whose sole
5870 // purpose is to compute bits we don't care about.
5871 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5872 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5873 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5874 KnownZero, KnownOne))
5877 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5878 // of a signed value.
5880 if (Op1->uge(TypeBits)) {
5881 if (I.getOpcode() != Instruction::AShr)
5882 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5884 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5889 // ((X*C1) << C2) == (X * (C1 << C2))
5890 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5891 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5892 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5893 return BinaryOperator::createMul(BO->getOperand(0),
5894 ConstantExpr::getShl(BOOp, Op1));
5896 // Try to fold constant and into select arguments.
5897 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5898 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5900 if (isa<PHINode>(Op0))
5901 if (Instruction *NV = FoldOpIntoPhi(I))
5904 if (Op0->hasOneUse()) {
5905 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5906 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5909 switch (Op0BO->getOpcode()) {
5911 case Instruction::Add:
5912 case Instruction::And:
5913 case Instruction::Or:
5914 case Instruction::Xor: {
5915 // These operators commute.
5916 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5917 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5918 match(Op0BO->getOperand(1),
5919 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5920 Instruction *YS = BinaryOperator::createShl(
5921 Op0BO->getOperand(0), Op1,
5923 InsertNewInstBefore(YS, I); // (Y << C)
5925 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5926 Op0BO->getOperand(1)->getName());
5927 InsertNewInstBefore(X, I); // (X + (Y << C))
5928 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5929 return BinaryOperator::createAnd(X, ConstantInt::get(
5930 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5933 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5934 Value *Op0BOOp1 = Op0BO->getOperand(1);
5935 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5937 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5938 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5940 Instruction *YS = BinaryOperator::createShl(
5941 Op0BO->getOperand(0), Op1,
5943 InsertNewInstBefore(YS, I); // (Y << C)
5945 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5946 V1->getName()+".mask");
5947 InsertNewInstBefore(XM, I); // X & (CC << C)
5949 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5954 case Instruction::Sub: {
5955 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5956 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5957 match(Op0BO->getOperand(0),
5958 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5959 Instruction *YS = BinaryOperator::createShl(
5960 Op0BO->getOperand(1), Op1,
5962 InsertNewInstBefore(YS, I); // (Y << C)
5964 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5965 Op0BO->getOperand(0)->getName());
5966 InsertNewInstBefore(X, I); // (X + (Y << C))
5967 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5968 return BinaryOperator::createAnd(X, ConstantInt::get(
5969 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5972 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5973 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5974 match(Op0BO->getOperand(0),
5975 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5976 m_ConstantInt(CC))) && V2 == Op1 &&
5977 cast<BinaryOperator>(Op0BO->getOperand(0))
5978 ->getOperand(0)->hasOneUse()) {
5979 Instruction *YS = BinaryOperator::createShl(
5980 Op0BO->getOperand(1), Op1,
5982 InsertNewInstBefore(YS, I); // (Y << C)
5984 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5985 V1->getName()+".mask");
5986 InsertNewInstBefore(XM, I); // X & (CC << C)
5988 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5996 // If the operand is an bitwise operator with a constant RHS, and the
5997 // shift is the only use, we can pull it out of the shift.
5998 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5999 bool isValid = true; // Valid only for And, Or, Xor
6000 bool highBitSet = false; // Transform if high bit of constant set?
6002 switch (Op0BO->getOpcode()) {
6003 default: isValid = false; break; // Do not perform transform!
6004 case Instruction::Add:
6005 isValid = isLeftShift;
6007 case Instruction::Or:
6008 case Instruction::Xor:
6011 case Instruction::And:
6016 // If this is a signed shift right, and the high bit is modified
6017 // by the logical operation, do not perform the transformation.
6018 // The highBitSet boolean indicates the value of the high bit of
6019 // the constant which would cause it to be modified for this
6022 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
6023 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6027 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6029 Instruction *NewShift =
6030 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6031 InsertNewInstBefore(NewShift, I);
6032 NewShift->takeName(Op0BO);
6034 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6041 // Find out if this is a shift of a shift by a constant.
6042 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6043 if (ShiftOp && !ShiftOp->isShift())
6046 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6047 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6048 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6049 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6050 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6051 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6052 Value *X = ShiftOp->getOperand(0);
6054 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6055 if (AmtSum > TypeBits)
6058 const IntegerType *Ty = cast<IntegerType>(I.getType());
6060 // Check for (X << c1) << c2 and (X >> c1) >> c2
6061 if (I.getOpcode() == ShiftOp->getOpcode()) {
6062 return BinaryOperator::create(I.getOpcode(), X,
6063 ConstantInt::get(Ty, AmtSum));
6064 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6065 I.getOpcode() == Instruction::AShr) {
6066 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6067 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6068 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6069 I.getOpcode() == Instruction::LShr) {
6070 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6071 Instruction *Shift =
6072 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6073 InsertNewInstBefore(Shift, I);
6075 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6076 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6079 // Okay, if we get here, one shift must be left, and the other shift must be
6080 // right. See if the amounts are equal.
6081 if (ShiftAmt1 == ShiftAmt2) {
6082 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6083 if (I.getOpcode() == Instruction::Shl) {
6084 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6085 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6087 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6088 if (I.getOpcode() == Instruction::LShr) {
6089 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6090 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6092 // We can simplify ((X << C) >>s C) into a trunc + sext.
6093 // NOTE: we could do this for any C, but that would make 'unusual' integer
6094 // types. For now, just stick to ones well-supported by the code
6096 const Type *SExtType = 0;
6097 switch (Ty->getBitWidth() - ShiftAmt1) {
6104 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6109 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6110 InsertNewInstBefore(NewTrunc, I);
6111 return new SExtInst(NewTrunc, Ty);
6113 // Otherwise, we can't handle it yet.
6114 } else if (ShiftAmt1 < ShiftAmt2) {
6115 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6117 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6118 if (I.getOpcode() == Instruction::Shl) {
6119 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6120 ShiftOp->getOpcode() == Instruction::AShr);
6121 Instruction *Shift =
6122 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6123 InsertNewInstBefore(Shift, I);
6125 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6126 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6129 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6130 if (I.getOpcode() == Instruction::LShr) {
6131 assert(ShiftOp->getOpcode() == Instruction::Shl);
6132 Instruction *Shift =
6133 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6134 InsertNewInstBefore(Shift, I);
6136 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6137 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6140 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6142 assert(ShiftAmt2 < ShiftAmt1);
6143 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6145 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6146 if (I.getOpcode() == Instruction::Shl) {
6147 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6148 ShiftOp->getOpcode() == Instruction::AShr);
6149 Instruction *Shift =
6150 BinaryOperator::create(ShiftOp->getOpcode(), X,
6151 ConstantInt::get(Ty, ShiftDiff));
6152 InsertNewInstBefore(Shift, I);
6154 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6155 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6158 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6159 if (I.getOpcode() == Instruction::LShr) {
6160 assert(ShiftOp->getOpcode() == Instruction::Shl);
6161 Instruction *Shift =
6162 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6163 InsertNewInstBefore(Shift, I);
6165 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6166 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6169 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6176 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6177 /// expression. If so, decompose it, returning some value X, such that Val is
6180 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6182 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6183 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6184 Offset = CI->getZExtValue();
6186 return ConstantInt::get(Type::Int32Ty, 0);
6187 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
6188 if (I->getNumOperands() == 2) {
6189 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6190 if (I->getOpcode() == Instruction::Shl) {
6191 // This is a value scaled by '1 << the shift amt'.
6192 Scale = 1U << CUI->getZExtValue();
6194 return I->getOperand(0);
6195 } else if (I->getOpcode() == Instruction::Mul) {
6196 // This value is scaled by 'CUI'.
6197 Scale = CUI->getZExtValue();
6199 return I->getOperand(0);
6200 } else if (I->getOpcode() == Instruction::Add) {
6201 // We have X+C. Check to see if we really have (X*C2)+C1,
6202 // where C1 is divisible by C2.
6205 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6206 Offset += CUI->getZExtValue();
6207 if (SubScale > 1 && (Offset % SubScale == 0)) {
6216 // Otherwise, we can't look past this.
6223 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6224 /// try to eliminate the cast by moving the type information into the alloc.
6225 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6226 AllocationInst &AI) {
6227 const PointerType *PTy = cast<PointerType>(CI.getType());
6229 // Remove any uses of AI that are dead.
6230 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6232 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6233 Instruction *User = cast<Instruction>(*UI++);
6234 if (isInstructionTriviallyDead(User)) {
6235 while (UI != E && *UI == User)
6236 ++UI; // If this instruction uses AI more than once, don't break UI.
6239 DOUT << "IC: DCE: " << *User;
6240 EraseInstFromFunction(*User);
6244 // Get the type really allocated and the type casted to.
6245 const Type *AllocElTy = AI.getAllocatedType();
6246 const Type *CastElTy = PTy->getElementType();
6247 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6249 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6250 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6251 if (CastElTyAlign < AllocElTyAlign) return 0;
6253 // If the allocation has multiple uses, only promote it if we are strictly
6254 // increasing the alignment of the resultant allocation. If we keep it the
6255 // same, we open the door to infinite loops of various kinds.
6256 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6258 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
6259 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
6260 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6262 // See if we can satisfy the modulus by pulling a scale out of the array
6264 unsigned ArraySizeScale;
6266 Value *NumElements = // See if the array size is a decomposable linear expr.
6267 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6269 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6271 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6272 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6274 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6279 // If the allocation size is constant, form a constant mul expression
6280 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6281 if (isa<ConstantInt>(NumElements))
6282 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6283 // otherwise multiply the amount and the number of elements
6284 else if (Scale != 1) {
6285 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6286 Amt = InsertNewInstBefore(Tmp, AI);
6290 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6291 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6292 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6293 Amt = InsertNewInstBefore(Tmp, AI);
6296 AllocationInst *New;
6297 if (isa<MallocInst>(AI))
6298 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6300 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6301 InsertNewInstBefore(New, AI);
6304 // If the allocation has multiple uses, insert a cast and change all things
6305 // that used it to use the new cast. This will also hack on CI, but it will
6307 if (!AI.hasOneUse()) {
6308 AddUsesToWorkList(AI);
6309 // New is the allocation instruction, pointer typed. AI is the original
6310 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6311 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6312 InsertNewInstBefore(NewCast, AI);
6313 AI.replaceAllUsesWith(NewCast);
6315 return ReplaceInstUsesWith(CI, New);
6318 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6319 /// and return it as type Ty without inserting any new casts and without
6320 /// changing the computed value. This is used by code that tries to decide
6321 /// whether promoting or shrinking integer operations to wider or smaller types
6322 /// will allow us to eliminate a truncate or extend.
6324 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6325 /// extension operation if Ty is larger.
6326 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6327 unsigned CastOpc, int &NumCastsRemoved) {
6328 // We can always evaluate constants in another type.
6329 if (isa<ConstantInt>(V))
6332 Instruction *I = dyn_cast<Instruction>(V);
6333 if (!I) return false;
6335 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6337 // If this is an extension or truncate, we can often eliminate it.
6338 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6339 // If this is a cast from the destination type, we can trivially eliminate
6340 // it, and this will remove a cast overall.
6341 if (I->getOperand(0)->getType() == Ty) {
6342 // If the first operand is itself a cast, and is eliminable, do not count
6343 // this as an eliminable cast. We would prefer to eliminate those two
6345 if (!isa<CastInst>(I->getOperand(0)))
6351 // We can't extend or shrink something that has multiple uses: doing so would
6352 // require duplicating the instruction in general, which isn't profitable.
6353 if (!I->hasOneUse()) return false;
6355 switch (I->getOpcode()) {
6356 case Instruction::Add:
6357 case Instruction::Sub:
6358 case Instruction::And:
6359 case Instruction::Or:
6360 case Instruction::Xor:
6361 // These operators can all arbitrarily be extended or truncated.
6362 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6364 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6367 case Instruction::Shl:
6368 // If we are truncating the result of this SHL, and if it's a shift of a
6369 // constant amount, we can always perform a SHL in a smaller type.
6370 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6371 uint32_t BitWidth = Ty->getBitWidth();
6372 if (BitWidth < OrigTy->getBitWidth() &&
6373 CI->getLimitedValue(BitWidth) < BitWidth)
6374 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6378 case Instruction::LShr:
6379 // If this is a truncate of a logical shr, we can truncate it to a smaller
6380 // lshr iff we know that the bits we would otherwise be shifting in are
6382 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6383 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6384 uint32_t BitWidth = Ty->getBitWidth();
6385 if (BitWidth < OrigBitWidth &&
6386 MaskedValueIsZero(I->getOperand(0),
6387 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6388 CI->getLimitedValue(BitWidth) < BitWidth) {
6389 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6394 case Instruction::ZExt:
6395 case Instruction::SExt:
6396 case Instruction::Trunc:
6397 // If this is the same kind of case as our original (e.g. zext+zext), we
6398 // can safely replace it. Note that replacing it does not reduce the number
6399 // of casts in the input.
6400 if (I->getOpcode() == CastOpc)
6404 // TODO: Can handle more cases here.
6411 /// EvaluateInDifferentType - Given an expression that
6412 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6413 /// evaluate the expression.
6414 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6416 if (Constant *C = dyn_cast<Constant>(V))
6417 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6419 // Otherwise, it must be an instruction.
6420 Instruction *I = cast<Instruction>(V);
6421 Instruction *Res = 0;
6422 switch (I->getOpcode()) {
6423 case Instruction::Add:
6424 case Instruction::Sub:
6425 case Instruction::And:
6426 case Instruction::Or:
6427 case Instruction::Xor:
6428 case Instruction::AShr:
6429 case Instruction::LShr:
6430 case Instruction::Shl: {
6431 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6432 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6433 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6434 LHS, RHS, I->getName());
6437 case Instruction::Trunc:
6438 case Instruction::ZExt:
6439 case Instruction::SExt:
6440 // If the source type of the cast is the type we're trying for then we can
6441 // just return the source. There's no need to insert it because it is not
6443 if (I->getOperand(0)->getType() == Ty)
6444 return I->getOperand(0);
6446 // Otherwise, must be the same type of case, so just reinsert a new one.
6447 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6451 // TODO: Can handle more cases here.
6452 assert(0 && "Unreachable!");
6456 return InsertNewInstBefore(Res, *I);
6459 /// @brief Implement the transforms common to all CastInst visitors.
6460 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6461 Value *Src = CI.getOperand(0);
6463 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6464 // eliminate it now.
6465 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6466 if (Instruction::CastOps opc =
6467 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6468 // The first cast (CSrc) is eliminable so we need to fix up or replace
6469 // the second cast (CI). CSrc will then have a good chance of being dead.
6470 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6474 // If we are casting a select then fold the cast into the select
6475 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6476 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6479 // If we are casting a PHI then fold the cast into the PHI
6480 if (isa<PHINode>(Src))
6481 if (Instruction *NV = FoldOpIntoPhi(CI))
6487 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6488 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6489 Value *Src = CI.getOperand(0);
6491 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6492 // If casting the result of a getelementptr instruction with no offset, turn
6493 // this into a cast of the original pointer!
6494 if (GEP->hasAllZeroIndices()) {
6495 // Changing the cast operand is usually not a good idea but it is safe
6496 // here because the pointer operand is being replaced with another
6497 // pointer operand so the opcode doesn't need to change.
6499 CI.setOperand(0, GEP->getOperand(0));
6503 // If the GEP has a single use, and the base pointer is a bitcast, and the
6504 // GEP computes a constant offset, see if we can convert these three
6505 // instructions into fewer. This typically happens with unions and other
6506 // non-type-safe code.
6507 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6508 if (GEP->hasAllConstantIndices()) {
6509 // We are guaranteed to get a constant from EmitGEPOffset.
6510 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6511 int64_t Offset = OffsetV->getSExtValue();
6513 // Get the base pointer input of the bitcast, and the type it points to.
6514 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6515 const Type *GEPIdxTy =
6516 cast<PointerType>(OrigBase->getType())->getElementType();
6517 if (GEPIdxTy->isSized()) {
6518 SmallVector<Value*, 8> NewIndices;
6520 // Start with the index over the outer type. Note that the type size
6521 // might be zero (even if the offset isn't zero) if the indexed type
6522 // is something like [0 x {int, int}]
6523 const Type *IntPtrTy = TD->getIntPtrType();
6524 int64_t FirstIdx = 0;
6525 if (int64_t TySize = TD->getTypeSize(GEPIdxTy)) {
6526 FirstIdx = Offset/TySize;
6529 // Handle silly modulus not returning values values [0..TySize).
6533 assert(Offset >= 0);
6535 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6538 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6540 // Index into the types. If we fail, set OrigBase to null.
6542 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6543 const StructLayout *SL = TD->getStructLayout(STy);
6544 if (Offset < (int64_t)SL->getSizeInBytes()) {
6545 unsigned Elt = SL->getElementContainingOffset(Offset);
6546 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6548 Offset -= SL->getElementOffset(Elt);
6549 GEPIdxTy = STy->getElementType(Elt);
6551 // Otherwise, we can't index into this, bail out.
6555 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6556 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6557 if (uint64_t EltSize = TD->getTypeSize(STy->getElementType())) {
6558 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6561 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6563 GEPIdxTy = STy->getElementType();
6565 // Otherwise, we can't index into this, bail out.
6571 // If we were able to index down into an element, create the GEP
6572 // and bitcast the result. This eliminates one bitcast, potentially
6574 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6576 NewIndices.end(), "");
6577 InsertNewInstBefore(NGEP, CI);
6578 NGEP->takeName(GEP);
6580 if (isa<BitCastInst>(CI))
6581 return new BitCastInst(NGEP, CI.getType());
6582 assert(isa<PtrToIntInst>(CI));
6583 return new PtrToIntInst(NGEP, CI.getType());
6590 return commonCastTransforms(CI);
6595 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6596 /// integer types. This function implements the common transforms for all those
6598 /// @brief Implement the transforms common to CastInst with integer operands
6599 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6600 if (Instruction *Result = commonCastTransforms(CI))
6603 Value *Src = CI.getOperand(0);
6604 const Type *SrcTy = Src->getType();
6605 const Type *DestTy = CI.getType();
6606 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6607 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6609 // See if we can simplify any instructions used by the LHS whose sole
6610 // purpose is to compute bits we don't care about.
6611 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6612 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6613 KnownZero, KnownOne))
6616 // If the source isn't an instruction or has more than one use then we
6617 // can't do anything more.
6618 Instruction *SrcI = dyn_cast<Instruction>(Src);
6619 if (!SrcI || !Src->hasOneUse())
6622 // Attempt to propagate the cast into the instruction for int->int casts.
6623 int NumCastsRemoved = 0;
6624 if (!isa<BitCastInst>(CI) &&
6625 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6626 CI.getOpcode(), NumCastsRemoved)) {
6627 // If this cast is a truncate, evaluting in a different type always
6628 // eliminates the cast, so it is always a win. If this is a zero-extension,
6629 // we need to do an AND to maintain the clear top-part of the computation,
6630 // so we require that the input have eliminated at least one cast. If this
6631 // is a sign extension, we insert two new casts (to do the extension) so we
6632 // require that two casts have been eliminated.
6634 switch (CI.getOpcode()) {
6636 // All the others use floating point so we shouldn't actually
6637 // get here because of the check above.
6638 assert(0 && "Unknown cast type");
6639 case Instruction::Trunc:
6642 case Instruction::ZExt:
6643 DoXForm = NumCastsRemoved >= 1;
6645 case Instruction::SExt:
6646 DoXForm = NumCastsRemoved >= 2;
6651 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6652 CI.getOpcode() == Instruction::SExt);
6653 assert(Res->getType() == DestTy);
6654 switch (CI.getOpcode()) {
6655 default: assert(0 && "Unknown cast type!");
6656 case Instruction::Trunc:
6657 case Instruction::BitCast:
6658 // Just replace this cast with the result.
6659 return ReplaceInstUsesWith(CI, Res);
6660 case Instruction::ZExt: {
6661 // We need to emit an AND to clear the high bits.
6662 assert(SrcBitSize < DestBitSize && "Not a zext?");
6663 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6665 return BinaryOperator::createAnd(Res, C);
6667 case Instruction::SExt:
6668 // We need to emit a cast to truncate, then a cast to sext.
6669 return CastInst::create(Instruction::SExt,
6670 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6676 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6677 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6679 switch (SrcI->getOpcode()) {
6680 case Instruction::Add:
6681 case Instruction::Mul:
6682 case Instruction::And:
6683 case Instruction::Or:
6684 case Instruction::Xor:
6685 // If we are discarding information, rewrite.
6686 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6687 // Don't insert two casts if they cannot be eliminated. We allow
6688 // two casts to be inserted if the sizes are the same. This could
6689 // only be converting signedness, which is a noop.
6690 if (DestBitSize == SrcBitSize ||
6691 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6692 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6693 Instruction::CastOps opcode = CI.getOpcode();
6694 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6695 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6696 return BinaryOperator::create(
6697 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6701 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6702 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6703 SrcI->getOpcode() == Instruction::Xor &&
6704 Op1 == ConstantInt::getTrue() &&
6705 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6706 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6707 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6710 case Instruction::SDiv:
6711 case Instruction::UDiv:
6712 case Instruction::SRem:
6713 case Instruction::URem:
6714 // If we are just changing the sign, rewrite.
6715 if (DestBitSize == SrcBitSize) {
6716 // Don't insert two casts if they cannot be eliminated. We allow
6717 // two casts to be inserted if the sizes are the same. This could
6718 // only be converting signedness, which is a noop.
6719 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6720 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6721 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6723 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6725 return BinaryOperator::create(
6726 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6731 case Instruction::Shl:
6732 // Allow changing the sign of the source operand. Do not allow
6733 // changing the size of the shift, UNLESS the shift amount is a
6734 // constant. We must not change variable sized shifts to a smaller
6735 // size, because it is undefined to shift more bits out than exist
6737 if (DestBitSize == SrcBitSize ||
6738 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6739 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6740 Instruction::BitCast : Instruction::Trunc);
6741 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6742 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6743 return BinaryOperator::createShl(Op0c, Op1c);
6746 case Instruction::AShr:
6747 // If this is a signed shr, and if all bits shifted in are about to be
6748 // truncated off, turn it into an unsigned shr to allow greater
6750 if (DestBitSize < SrcBitSize &&
6751 isa<ConstantInt>(Op1)) {
6752 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6753 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6754 // Insert the new logical shift right.
6755 return BinaryOperator::createLShr(Op0, Op1);
6763 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6764 if (Instruction *Result = commonIntCastTransforms(CI))
6767 Value *Src = CI.getOperand(0);
6768 const Type *Ty = CI.getType();
6769 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6770 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6772 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6773 switch (SrcI->getOpcode()) {
6775 case Instruction::LShr:
6776 // We can shrink lshr to something smaller if we know the bits shifted in
6777 // are already zeros.
6778 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6779 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6781 // Get a mask for the bits shifting in.
6782 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6783 Value* SrcIOp0 = SrcI->getOperand(0);
6784 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6785 if (ShAmt >= DestBitWidth) // All zeros.
6786 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6788 // Okay, we can shrink this. Truncate the input, then return a new
6790 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6791 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6793 return BinaryOperator::createLShr(V1, V2);
6795 } else { // This is a variable shr.
6797 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6798 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6799 // loop-invariant and CSE'd.
6800 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6801 Value *One = ConstantInt::get(SrcI->getType(), 1);
6803 Value *V = InsertNewInstBefore(
6804 BinaryOperator::createShl(One, SrcI->getOperand(1),
6806 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6807 SrcI->getOperand(0),
6809 Value *Zero = Constant::getNullValue(V->getType());
6810 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6820 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6821 // If one of the common conversion will work ..
6822 if (Instruction *Result = commonIntCastTransforms(CI))
6825 Value *Src = CI.getOperand(0);
6827 // If this is a cast of a cast
6828 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6829 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6830 // types and if the sizes are just right we can convert this into a logical
6831 // 'and' which will be much cheaper than the pair of casts.
6832 if (isa<TruncInst>(CSrc)) {
6833 // Get the sizes of the types involved
6834 Value *A = CSrc->getOperand(0);
6835 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6836 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6837 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6838 // If we're actually extending zero bits and the trunc is a no-op
6839 if (MidSize < DstSize && SrcSize == DstSize) {
6840 // Replace both of the casts with an And of the type mask.
6841 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6842 Constant *AndConst = ConstantInt::get(AndValue);
6844 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6845 // Unfortunately, if the type changed, we need to cast it back.
6846 if (And->getType() != CI.getType()) {
6847 And->setName(CSrc->getName()+".mask");
6848 InsertNewInstBefore(And, CI);
6849 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6856 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6857 // If we are just checking for a icmp eq of a single bit and zext'ing it
6858 // to an integer, then shift the bit to the appropriate place and then
6859 // cast to integer to avoid the comparison.
6860 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6861 const APInt &Op1CV = Op1C->getValue();
6863 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
6864 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
6865 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6866 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6867 Value *In = ICI->getOperand(0);
6868 Value *Sh = ConstantInt::get(In->getType(),
6869 In->getType()->getPrimitiveSizeInBits()-1);
6870 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
6871 In->getName()+".lobit"),
6873 if (In->getType() != CI.getType())
6874 In = CastInst::createIntegerCast(In, CI.getType(),
6875 false/*ZExt*/, "tmp", &CI);
6877 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
6878 Constant *One = ConstantInt::get(In->getType(), 1);
6879 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
6880 In->getName()+".not"),
6884 return ReplaceInstUsesWith(CI, In);
6889 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
6890 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6891 // zext (X == 1) to i32 --> X iff X has only the low bit set.
6892 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
6893 // zext (X != 0) to i32 --> X iff X has only the low bit set.
6894 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
6895 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
6896 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6897 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
6898 // This only works for EQ and NE
6899 ICI->isEquality()) {
6900 // If Op1C some other power of two, convert:
6901 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6902 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6903 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6904 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
6906 APInt KnownZeroMask(~KnownZero);
6907 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6908 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
6909 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6910 // (X&4) == 2 --> false
6911 // (X&4) != 2 --> true
6912 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6913 Res = ConstantExpr::getZExt(Res, CI.getType());
6914 return ReplaceInstUsesWith(CI, Res);
6917 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6918 Value *In = ICI->getOperand(0);
6920 // Perform a logical shr by shiftamt.
6921 // Insert the shift to put the result in the low bit.
6922 In = InsertNewInstBefore(
6923 BinaryOperator::createLShr(In,
6924 ConstantInt::get(In->getType(), ShiftAmt),
6925 In->getName()+".lobit"), CI);
6928 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6929 Constant *One = ConstantInt::get(In->getType(), 1);
6930 In = BinaryOperator::createXor(In, One, "tmp");
6931 InsertNewInstBefore(cast<Instruction>(In), CI);
6934 if (CI.getType() == In->getType())
6935 return ReplaceInstUsesWith(CI, In);
6937 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6945 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
6946 if (Instruction *I = commonIntCastTransforms(CI))
6949 Value *Src = CI.getOperand(0);
6951 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
6952 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
6953 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6954 // If we are just checking for a icmp eq of a single bit and zext'ing it
6955 // to an integer, then shift the bit to the appropriate place and then
6956 // cast to integer to avoid the comparison.
6957 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6958 const APInt &Op1CV = Op1C->getValue();
6960 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
6961 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
6962 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6963 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6964 Value *In = ICI->getOperand(0);
6965 Value *Sh = ConstantInt::get(In->getType(),
6966 In->getType()->getPrimitiveSizeInBits()-1);
6967 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
6968 In->getName()+".lobit"),
6970 if (In->getType() != CI.getType())
6971 In = CastInst::createIntegerCast(In, CI.getType(),
6972 true/*SExt*/, "tmp", &CI);
6974 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
6975 In = InsertNewInstBefore(BinaryOperator::createNot(In,
6976 In->getName()+".not"), CI);
6978 return ReplaceInstUsesWith(CI, In);
6986 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6987 return commonCastTransforms(CI);
6990 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6991 return commonCastTransforms(CI);
6994 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6995 return commonCastTransforms(CI);
6998 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6999 return commonCastTransforms(CI);
7002 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
7003 return commonCastTransforms(CI);
7006 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7007 return commonCastTransforms(CI);
7010 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7011 return commonPointerCastTransforms(CI);
7014 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7015 return commonCastTransforms(CI);
7018 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7019 // If the operands are integer typed then apply the integer transforms,
7020 // otherwise just apply the common ones.
7021 Value *Src = CI.getOperand(0);
7022 const Type *SrcTy = Src->getType();
7023 const Type *DestTy = CI.getType();
7025 if (SrcTy->isInteger() && DestTy->isInteger()) {
7026 if (Instruction *Result = commonIntCastTransforms(CI))
7028 } else if (isa<PointerType>(SrcTy)) {
7029 if (Instruction *I = commonPointerCastTransforms(CI))
7032 if (Instruction *Result = commonCastTransforms(CI))
7037 // Get rid of casts from one type to the same type. These are useless and can
7038 // be replaced by the operand.
7039 if (DestTy == Src->getType())
7040 return ReplaceInstUsesWith(CI, Src);
7042 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7043 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7044 const Type *DstElTy = DstPTy->getElementType();
7045 const Type *SrcElTy = SrcPTy->getElementType();
7047 // If we are casting a malloc or alloca to a pointer to a type of the same
7048 // size, rewrite the allocation instruction to allocate the "right" type.
7049 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7050 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7053 // If the source and destination are pointers, and this cast is equivalent
7054 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7055 // This can enhance SROA and other transforms that want type-safe pointers.
7056 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7057 unsigned NumZeros = 0;
7058 while (SrcElTy != DstElTy &&
7059 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7060 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7061 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7065 // If we found a path from the src to dest, create the getelementptr now.
7066 if (SrcElTy == DstElTy) {
7067 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7068 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7069 ((Instruction*) NULL));
7073 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7074 if (SVI->hasOneUse()) {
7075 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7076 // a bitconvert to a vector with the same # elts.
7077 if (isa<VectorType>(DestTy) &&
7078 cast<VectorType>(DestTy)->getNumElements() ==
7079 SVI->getType()->getNumElements()) {
7081 // If either of the operands is a cast from CI.getType(), then
7082 // evaluating the shuffle in the casted destination's type will allow
7083 // us to eliminate at least one cast.
7084 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7085 Tmp->getOperand(0)->getType() == DestTy) ||
7086 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7087 Tmp->getOperand(0)->getType() == DestTy)) {
7088 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7089 SVI->getOperand(0), DestTy, &CI);
7090 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7091 SVI->getOperand(1), DestTy, &CI);
7092 // Return a new shuffle vector. Use the same element ID's, as we
7093 // know the vector types match #elts.
7094 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7102 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7104 /// %D = select %cond, %C, %A
7106 /// %C = select %cond, %B, 0
7109 /// Assuming that the specified instruction is an operand to the select, return
7110 /// a bitmask indicating which operands of this instruction are foldable if they
7111 /// equal the other incoming value of the select.
7113 static unsigned GetSelectFoldableOperands(Instruction *I) {
7114 switch (I->getOpcode()) {
7115 case Instruction::Add:
7116 case Instruction::Mul:
7117 case Instruction::And:
7118 case Instruction::Or:
7119 case Instruction::Xor:
7120 return 3; // Can fold through either operand.
7121 case Instruction::Sub: // Can only fold on the amount subtracted.
7122 case Instruction::Shl: // Can only fold on the shift amount.
7123 case Instruction::LShr:
7124 case Instruction::AShr:
7127 return 0; // Cannot fold
7131 /// GetSelectFoldableConstant - For the same transformation as the previous
7132 /// function, return the identity constant that goes into the select.
7133 static Constant *GetSelectFoldableConstant(Instruction *I) {
7134 switch (I->getOpcode()) {
7135 default: assert(0 && "This cannot happen!"); abort();
7136 case Instruction::Add:
7137 case Instruction::Sub:
7138 case Instruction::Or:
7139 case Instruction::Xor:
7140 case Instruction::Shl:
7141 case Instruction::LShr:
7142 case Instruction::AShr:
7143 return Constant::getNullValue(I->getType());
7144 case Instruction::And:
7145 return Constant::getAllOnesValue(I->getType());
7146 case Instruction::Mul:
7147 return ConstantInt::get(I->getType(), 1);
7151 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7152 /// have the same opcode and only one use each. Try to simplify this.
7153 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7155 if (TI->getNumOperands() == 1) {
7156 // If this is a non-volatile load or a cast from the same type,
7159 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7162 return 0; // unknown unary op.
7165 // Fold this by inserting a select from the input values.
7166 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7167 FI->getOperand(0), SI.getName()+".v");
7168 InsertNewInstBefore(NewSI, SI);
7169 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7173 // Only handle binary operators here.
7174 if (!isa<BinaryOperator>(TI))
7177 // Figure out if the operations have any operands in common.
7178 Value *MatchOp, *OtherOpT, *OtherOpF;
7180 if (TI->getOperand(0) == FI->getOperand(0)) {
7181 MatchOp = TI->getOperand(0);
7182 OtherOpT = TI->getOperand(1);
7183 OtherOpF = FI->getOperand(1);
7184 MatchIsOpZero = true;
7185 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7186 MatchOp = TI->getOperand(1);
7187 OtherOpT = TI->getOperand(0);
7188 OtherOpF = FI->getOperand(0);
7189 MatchIsOpZero = false;
7190 } else if (!TI->isCommutative()) {
7192 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7193 MatchOp = TI->getOperand(0);
7194 OtherOpT = TI->getOperand(1);
7195 OtherOpF = FI->getOperand(0);
7196 MatchIsOpZero = true;
7197 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7198 MatchOp = TI->getOperand(1);
7199 OtherOpT = TI->getOperand(0);
7200 OtherOpF = FI->getOperand(1);
7201 MatchIsOpZero = true;
7206 // If we reach here, they do have operations in common.
7207 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7208 OtherOpF, SI.getName()+".v");
7209 InsertNewInstBefore(NewSI, SI);
7211 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7213 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7215 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7217 assert(0 && "Shouldn't get here");
7221 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7222 Value *CondVal = SI.getCondition();
7223 Value *TrueVal = SI.getTrueValue();
7224 Value *FalseVal = SI.getFalseValue();
7226 // select true, X, Y -> X
7227 // select false, X, Y -> Y
7228 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7229 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7231 // select C, X, X -> X
7232 if (TrueVal == FalseVal)
7233 return ReplaceInstUsesWith(SI, TrueVal);
7235 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7236 return ReplaceInstUsesWith(SI, FalseVal);
7237 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7238 return ReplaceInstUsesWith(SI, TrueVal);
7239 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7240 if (isa<Constant>(TrueVal))
7241 return ReplaceInstUsesWith(SI, TrueVal);
7243 return ReplaceInstUsesWith(SI, FalseVal);
7246 if (SI.getType() == Type::Int1Ty) {
7247 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7248 if (C->getZExtValue()) {
7249 // Change: A = select B, true, C --> A = or B, C
7250 return BinaryOperator::createOr(CondVal, FalseVal);
7252 // Change: A = select B, false, C --> A = and !B, C
7254 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7255 "not."+CondVal->getName()), SI);
7256 return BinaryOperator::createAnd(NotCond, FalseVal);
7258 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7259 if (C->getZExtValue() == false) {
7260 // Change: A = select B, C, false --> A = and B, C
7261 return BinaryOperator::createAnd(CondVal, TrueVal);
7263 // Change: A = select B, C, true --> A = or !B, C
7265 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7266 "not."+CondVal->getName()), SI);
7267 return BinaryOperator::createOr(NotCond, TrueVal);
7272 // Selecting between two integer constants?
7273 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7274 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7275 // select C, 1, 0 -> zext C to int
7276 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7277 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7278 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7279 // select C, 0, 1 -> zext !C to int
7281 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7282 "not."+CondVal->getName()), SI);
7283 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7286 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7288 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7290 // (x <s 0) ? -1 : 0 -> ashr x, 31
7291 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7292 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7293 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7294 // The comparison constant and the result are not neccessarily the
7295 // same width. Make an all-ones value by inserting a AShr.
7296 Value *X = IC->getOperand(0);
7297 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7298 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7299 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7301 InsertNewInstBefore(SRA, SI);
7303 // Finally, convert to the type of the select RHS. We figure out
7304 // if this requires a SExt, Trunc or BitCast based on the sizes.
7305 Instruction::CastOps opc = Instruction::BitCast;
7306 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7307 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7308 if (SRASize < SISize)
7309 opc = Instruction::SExt;
7310 else if (SRASize > SISize)
7311 opc = Instruction::Trunc;
7312 return CastInst::create(opc, SRA, SI.getType());
7317 // If one of the constants is zero (we know they can't both be) and we
7318 // have an icmp instruction with zero, and we have an 'and' with the
7319 // non-constant value, eliminate this whole mess. This corresponds to
7320 // cases like this: ((X & 27) ? 27 : 0)
7321 if (TrueValC->isZero() || FalseValC->isZero())
7322 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7323 cast<Constant>(IC->getOperand(1))->isNullValue())
7324 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7325 if (ICA->getOpcode() == Instruction::And &&
7326 isa<ConstantInt>(ICA->getOperand(1)) &&
7327 (ICA->getOperand(1) == TrueValC ||
7328 ICA->getOperand(1) == FalseValC) &&
7329 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7330 // Okay, now we know that everything is set up, we just don't
7331 // know whether we have a icmp_ne or icmp_eq and whether the
7332 // true or false val is the zero.
7333 bool ShouldNotVal = !TrueValC->isZero();
7334 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7337 V = InsertNewInstBefore(BinaryOperator::create(
7338 Instruction::Xor, V, ICA->getOperand(1)), SI);
7339 return ReplaceInstUsesWith(SI, V);
7344 // See if we are selecting two values based on a comparison of the two values.
7345 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7346 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7347 // Transform (X == Y) ? X : Y -> Y
7348 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7349 return ReplaceInstUsesWith(SI, FalseVal);
7350 // Transform (X != Y) ? X : Y -> X
7351 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7352 return ReplaceInstUsesWith(SI, TrueVal);
7353 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7355 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7356 // Transform (X == Y) ? Y : X -> X
7357 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7358 return ReplaceInstUsesWith(SI, FalseVal);
7359 // Transform (X != Y) ? Y : X -> Y
7360 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7361 return ReplaceInstUsesWith(SI, TrueVal);
7362 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7366 // See if we are selecting two values based on a comparison of the two values.
7367 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7368 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7369 // Transform (X == Y) ? X : Y -> Y
7370 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7371 return ReplaceInstUsesWith(SI, FalseVal);
7372 // Transform (X != Y) ? X : Y -> X
7373 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7374 return ReplaceInstUsesWith(SI, TrueVal);
7375 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7377 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7378 // Transform (X == Y) ? Y : X -> X
7379 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7380 return ReplaceInstUsesWith(SI, FalseVal);
7381 // Transform (X != Y) ? Y : X -> Y
7382 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7383 return ReplaceInstUsesWith(SI, TrueVal);
7384 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7388 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7389 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7390 if (TI->hasOneUse() && FI->hasOneUse()) {
7391 Instruction *AddOp = 0, *SubOp = 0;
7393 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7394 if (TI->getOpcode() == FI->getOpcode())
7395 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7398 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7399 // even legal for FP.
7400 if (TI->getOpcode() == Instruction::Sub &&
7401 FI->getOpcode() == Instruction::Add) {
7402 AddOp = FI; SubOp = TI;
7403 } else if (FI->getOpcode() == Instruction::Sub &&
7404 TI->getOpcode() == Instruction::Add) {
7405 AddOp = TI; SubOp = FI;
7409 Value *OtherAddOp = 0;
7410 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7411 OtherAddOp = AddOp->getOperand(1);
7412 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7413 OtherAddOp = AddOp->getOperand(0);
7417 // So at this point we know we have (Y -> OtherAddOp):
7418 // select C, (add X, Y), (sub X, Z)
7419 Value *NegVal; // Compute -Z
7420 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7421 NegVal = ConstantExpr::getNeg(C);
7423 NegVal = InsertNewInstBefore(
7424 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7427 Value *NewTrueOp = OtherAddOp;
7428 Value *NewFalseOp = NegVal;
7430 std::swap(NewTrueOp, NewFalseOp);
7431 Instruction *NewSel =
7432 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7434 NewSel = InsertNewInstBefore(NewSel, SI);
7435 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7440 // See if we can fold the select into one of our operands.
7441 if (SI.getType()->isInteger()) {
7442 // See the comment above GetSelectFoldableOperands for a description of the
7443 // transformation we are doing here.
7444 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7445 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7446 !isa<Constant>(FalseVal))
7447 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7448 unsigned OpToFold = 0;
7449 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7451 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7456 Constant *C = GetSelectFoldableConstant(TVI);
7457 Instruction *NewSel =
7458 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7459 InsertNewInstBefore(NewSel, SI);
7460 NewSel->takeName(TVI);
7461 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7462 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7464 assert(0 && "Unknown instruction!!");
7469 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7470 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7471 !isa<Constant>(TrueVal))
7472 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7473 unsigned OpToFold = 0;
7474 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7476 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7481 Constant *C = GetSelectFoldableConstant(FVI);
7482 Instruction *NewSel =
7483 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7484 InsertNewInstBefore(NewSel, SI);
7485 NewSel->takeName(FVI);
7486 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7487 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7489 assert(0 && "Unknown instruction!!");
7494 if (BinaryOperator::isNot(CondVal)) {
7495 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7496 SI.setOperand(1, FalseVal);
7497 SI.setOperand(2, TrueVal);
7504 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7505 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7506 /// and it is more than the alignment of the ultimate object, see if we can
7507 /// increase the alignment of the ultimate object, making this check succeed.
7508 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7509 unsigned PrefAlign = 0) {
7510 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7511 unsigned Align = GV->getAlignment();
7512 if (Align == 0 && TD)
7513 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7515 // If there is a large requested alignment and we can, bump up the alignment
7517 if (PrefAlign > Align && GV->hasInitializer()) {
7518 GV->setAlignment(PrefAlign);
7522 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7523 unsigned Align = AI->getAlignment();
7524 if (Align == 0 && TD) {
7525 if (isa<AllocaInst>(AI))
7526 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7527 else if (isa<MallocInst>(AI)) {
7528 // Malloc returns maximally aligned memory.
7529 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7532 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7535 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7539 // If there is a requested alignment and if this is an alloca, round up. We
7540 // don't do this for malloc, because some systems can't respect the request.
7541 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7542 AI->setAlignment(PrefAlign);
7546 } else if (isa<BitCastInst>(V) ||
7547 (isa<ConstantExpr>(V) &&
7548 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7549 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7551 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7552 // If all indexes are zero, it is just the alignment of the base pointer.
7553 bool AllZeroOperands = true;
7554 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7555 if (!isa<Constant>(GEPI->getOperand(i)) ||
7556 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7557 AllZeroOperands = false;
7561 if (AllZeroOperands) {
7562 // Treat this like a bitcast.
7563 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7566 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7567 if (BaseAlignment == 0) return 0;
7569 // Otherwise, if the base alignment is >= the alignment we expect for the
7570 // base pointer type, then we know that the resultant pointer is aligned at
7571 // least as much as its type requires.
7574 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7575 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7576 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7577 if (Align <= BaseAlignment) {
7578 const Type *GEPTy = GEPI->getType();
7579 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7580 Align = std::min(Align, (unsigned)
7581 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7590 /// visitCallInst - CallInst simplification. This mostly only handles folding
7591 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7592 /// the heavy lifting.
7594 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7595 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7596 if (!II) return visitCallSite(&CI);
7598 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7600 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7601 bool Changed = false;
7603 // memmove/cpy/set of zero bytes is a noop.
7604 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7605 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7607 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7608 if (CI->getZExtValue() == 1) {
7609 // Replace the instruction with just byte operations. We would
7610 // transform other cases to loads/stores, but we don't know if
7611 // alignment is sufficient.
7615 // If we have a memmove and the source operation is a constant global,
7616 // then the source and dest pointers can't alias, so we can change this
7617 // into a call to memcpy.
7618 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7619 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7620 if (GVSrc->isConstant()) {
7621 Module *M = CI.getParent()->getParent()->getParent();
7623 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7625 Name = "llvm.memcpy.i32";
7627 Name = "llvm.memcpy.i64";
7628 Constant *MemCpy = M->getOrInsertFunction(Name,
7629 CI.getCalledFunction()->getFunctionType());
7630 CI.setOperand(0, MemCpy);
7635 // If we can determine a pointer alignment that is bigger than currently
7636 // set, update the alignment.
7637 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7638 unsigned Alignment1 = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7639 unsigned Alignment2 = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7640 unsigned Align = std::min(Alignment1, Alignment2);
7641 if (MI->getAlignment()->getZExtValue() < Align) {
7642 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7645 } else if (isa<MemSetInst>(MI)) {
7646 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
7647 if (MI->getAlignment()->getZExtValue() < Alignment) {
7648 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7653 if (Changed) return II;
7655 switch (II->getIntrinsicID()) {
7657 case Intrinsic::ppc_altivec_lvx:
7658 case Intrinsic::ppc_altivec_lvxl:
7659 case Intrinsic::x86_sse_loadu_ps:
7660 case Intrinsic::x86_sse2_loadu_pd:
7661 case Intrinsic::x86_sse2_loadu_dq:
7662 // Turn PPC lvx -> load if the pointer is known aligned.
7663 // Turn X86 loadups -> load if the pointer is known aligned.
7664 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7665 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7666 PointerType::get(II->getType()), CI);
7667 return new LoadInst(Ptr);
7670 case Intrinsic::ppc_altivec_stvx:
7671 case Intrinsic::ppc_altivec_stvxl:
7672 // Turn stvx -> store if the pointer is known aligned.
7673 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
7674 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7675 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7677 return new StoreInst(II->getOperand(1), Ptr);
7680 case Intrinsic::x86_sse_storeu_ps:
7681 case Intrinsic::x86_sse2_storeu_pd:
7682 case Intrinsic::x86_sse2_storeu_dq:
7683 case Intrinsic::x86_sse2_storel_dq:
7684 // Turn X86 storeu -> store if the pointer is known aligned.
7685 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7686 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7687 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7689 return new StoreInst(II->getOperand(2), Ptr);
7693 case Intrinsic::x86_sse_cvttss2si: {
7694 // These intrinsics only demands the 0th element of its input vector. If
7695 // we can simplify the input based on that, do so now.
7697 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7699 II->setOperand(1, V);
7705 case Intrinsic::ppc_altivec_vperm:
7706 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7707 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7708 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7710 // Check that all of the elements are integer constants or undefs.
7711 bool AllEltsOk = true;
7712 for (unsigned i = 0; i != 16; ++i) {
7713 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7714 !isa<UndefValue>(Mask->getOperand(i))) {
7721 // Cast the input vectors to byte vectors.
7722 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7723 II->getOperand(1), Mask->getType(), CI);
7724 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7725 II->getOperand(2), Mask->getType(), CI);
7726 Value *Result = UndefValue::get(Op0->getType());
7728 // Only extract each element once.
7729 Value *ExtractedElts[32];
7730 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7732 for (unsigned i = 0; i != 16; ++i) {
7733 if (isa<UndefValue>(Mask->getOperand(i)))
7735 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7736 Idx &= 31; // Match the hardware behavior.
7738 if (ExtractedElts[Idx] == 0) {
7740 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7741 InsertNewInstBefore(Elt, CI);
7742 ExtractedElts[Idx] = Elt;
7745 // Insert this value into the result vector.
7746 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7747 InsertNewInstBefore(cast<Instruction>(Result), CI);
7749 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7754 case Intrinsic::stackrestore: {
7755 // If the save is right next to the restore, remove the restore. This can
7756 // happen when variable allocas are DCE'd.
7757 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7758 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7759 BasicBlock::iterator BI = SS;
7761 return EraseInstFromFunction(CI);
7765 // If the stack restore is in a return/unwind block and if there are no
7766 // allocas or calls between the restore and the return, nuke the restore.
7767 TerminatorInst *TI = II->getParent()->getTerminator();
7768 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7769 BasicBlock::iterator BI = II;
7770 bool CannotRemove = false;
7771 for (++BI; &*BI != TI; ++BI) {
7772 if (isa<AllocaInst>(BI) ||
7773 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7774 CannotRemove = true;
7779 return EraseInstFromFunction(CI);
7786 return visitCallSite(II);
7789 // InvokeInst simplification
7791 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7792 return visitCallSite(&II);
7795 // visitCallSite - Improvements for call and invoke instructions.
7797 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7798 bool Changed = false;
7800 // If the callee is a constexpr cast of a function, attempt to move the cast
7801 // to the arguments of the call/invoke.
7802 if (transformConstExprCastCall(CS)) return 0;
7804 Value *Callee = CS.getCalledValue();
7806 if (Function *CalleeF = dyn_cast<Function>(Callee))
7807 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7808 Instruction *OldCall = CS.getInstruction();
7809 // If the call and callee calling conventions don't match, this call must
7810 // be unreachable, as the call is undefined.
7811 new StoreInst(ConstantInt::getTrue(),
7812 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7813 if (!OldCall->use_empty())
7814 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7815 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7816 return EraseInstFromFunction(*OldCall);
7820 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7821 // This instruction is not reachable, just remove it. We insert a store to
7822 // undef so that we know that this code is not reachable, despite the fact
7823 // that we can't modify the CFG here.
7824 new StoreInst(ConstantInt::getTrue(),
7825 UndefValue::get(PointerType::get(Type::Int1Ty)),
7826 CS.getInstruction());
7828 if (!CS.getInstruction()->use_empty())
7829 CS.getInstruction()->
7830 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7832 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7833 // Don't break the CFG, insert a dummy cond branch.
7834 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7835 ConstantInt::getTrue(), II);
7837 return EraseInstFromFunction(*CS.getInstruction());
7840 const PointerType *PTy = cast<PointerType>(Callee->getType());
7841 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7842 if (FTy->isVarArg()) {
7843 // See if we can optimize any arguments passed through the varargs area of
7845 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7846 E = CS.arg_end(); I != E; ++I)
7847 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7848 // If this cast does not effect the value passed through the varargs
7849 // area, we can eliminate the use of the cast.
7850 Value *Op = CI->getOperand(0);
7851 if (CI->isLosslessCast()) {
7858 return Changed ? CS.getInstruction() : 0;
7861 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7862 // attempt to move the cast to the arguments of the call/invoke.
7864 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7865 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7866 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7867 if (CE->getOpcode() != Instruction::BitCast ||
7868 !isa<Function>(CE->getOperand(0)))
7870 Function *Callee = cast<Function>(CE->getOperand(0));
7871 Instruction *Caller = CS.getInstruction();
7873 // Okay, this is a cast from a function to a different type. Unless doing so
7874 // would cause a type conversion of one of our arguments, change this call to
7875 // be a direct call with arguments casted to the appropriate types.
7877 const FunctionType *FT = Callee->getFunctionType();
7878 const Type *OldRetTy = Caller->getType();
7880 const FunctionType *ActualFT =
7881 cast<FunctionType>(cast<PointerType>(CE->getType())->getElementType());
7883 // If the parameter attributes don't match up, don't do the xform. We don't
7884 // want to lose an sret attribute or something.
7885 if (FT->getParamAttrs() != ActualFT->getParamAttrs())
7888 // Check to see if we are changing the return type...
7889 if (OldRetTy != FT->getReturnType()) {
7890 if (Callee->isDeclaration() && !Caller->use_empty() &&
7891 // Conversion is ok if changing from pointer to int of same size.
7892 !(isa<PointerType>(FT->getReturnType()) &&
7893 TD->getIntPtrType() == OldRetTy))
7894 return false; // Cannot transform this return value.
7896 // If the callsite is an invoke instruction, and the return value is used by
7897 // a PHI node in a successor, we cannot change the return type of the call
7898 // because there is no place to put the cast instruction (without breaking
7899 // the critical edge). Bail out in this case.
7900 if (!Caller->use_empty())
7901 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7902 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7904 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7905 if (PN->getParent() == II->getNormalDest() ||
7906 PN->getParent() == II->getUnwindDest())
7910 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7911 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7913 CallSite::arg_iterator AI = CS.arg_begin();
7914 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7915 const Type *ParamTy = FT->getParamType(i);
7916 const Type *ActTy = (*AI)->getType();
7917 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7918 //Some conversions are safe even if we do not have a body.
7919 //Either we can cast directly, or we can upconvert the argument
7920 bool isConvertible = ActTy == ParamTy ||
7921 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7922 (ParamTy->isInteger() && ActTy->isInteger() &&
7923 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7924 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7925 && c->getValue().isStrictlyPositive());
7926 if (Callee->isDeclaration() && !isConvertible) return false;
7928 // Most other conversions can be done if we have a body, even if these
7929 // lose information, e.g. int->short.
7930 // Some conversions cannot be done at all, e.g. float to pointer.
7931 // Logic here parallels CastInst::getCastOpcode (the design there
7932 // requires legality checks like this be done before calling it).
7933 if (ParamTy->isInteger()) {
7934 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7935 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7938 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
7939 !isa<PointerType>(ActTy))
7941 } else if (ParamTy->isFloatingPoint()) {
7942 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7943 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7946 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
7948 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
7949 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7950 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
7953 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
7955 } else if (isa<PointerType>(ParamTy)) {
7956 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
7963 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7964 Callee->isDeclaration())
7965 return false; // Do not delete arguments unless we have a function body...
7967 // Okay, we decided that this is a safe thing to do: go ahead and start
7968 // inserting cast instructions as necessary...
7969 std::vector<Value*> Args;
7970 Args.reserve(NumActualArgs);
7972 AI = CS.arg_begin();
7973 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7974 const Type *ParamTy = FT->getParamType(i);
7975 if ((*AI)->getType() == ParamTy) {
7976 Args.push_back(*AI);
7978 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7979 false, ParamTy, false);
7980 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7981 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7985 // If the function takes more arguments than the call was taking, add them
7987 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7988 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7990 // If we are removing arguments to the function, emit an obnoxious warning...
7991 if (FT->getNumParams() < NumActualArgs)
7992 if (!FT->isVarArg()) {
7993 cerr << "WARNING: While resolving call to function '"
7994 << Callee->getName() << "' arguments were dropped!\n";
7996 // Add all of the arguments in their promoted form to the arg list...
7997 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7998 const Type *PTy = getPromotedType((*AI)->getType());
7999 if (PTy != (*AI)->getType()) {
8000 // Must promote to pass through va_arg area!
8001 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8003 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8004 InsertNewInstBefore(Cast, *Caller);
8005 Args.push_back(Cast);
8007 Args.push_back(*AI);
8012 if (FT->getReturnType() == Type::VoidTy)
8013 Caller->setName(""); // Void type should not have a name.
8016 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8017 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8018 Args.begin(), Args.end(), Caller->getName(), Caller);
8019 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8021 NC = new CallInst(Callee, Args.begin(), Args.end(),
8022 Caller->getName(), Caller);
8023 if (cast<CallInst>(Caller)->isTailCall())
8024 cast<CallInst>(NC)->setTailCall();
8025 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8028 // Insert a cast of the return type as necessary.
8030 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
8031 if (NV->getType() != Type::VoidTy) {
8032 const Type *CallerTy = Caller->getType();
8033 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8035 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
8037 // If this is an invoke instruction, we should insert it after the first
8038 // non-phi, instruction in the normal successor block.
8039 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8040 BasicBlock::iterator I = II->getNormalDest()->begin();
8041 while (isa<PHINode>(I)) ++I;
8042 InsertNewInstBefore(NC, *I);
8044 // Otherwise, it's a call, just insert cast right after the call instr
8045 InsertNewInstBefore(NC, *Caller);
8047 AddUsersToWorkList(*Caller);
8049 NV = UndefValue::get(Caller->getType());
8053 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8054 Caller->replaceAllUsesWith(NV);
8055 Caller->eraseFromParent();
8056 RemoveFromWorkList(Caller);
8060 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8061 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8062 /// and a single binop.
8063 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8064 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8065 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8066 isa<CmpInst>(FirstInst));
8067 unsigned Opc = FirstInst->getOpcode();
8068 Value *LHSVal = FirstInst->getOperand(0);
8069 Value *RHSVal = FirstInst->getOperand(1);
8071 const Type *LHSType = LHSVal->getType();
8072 const Type *RHSType = RHSVal->getType();
8074 // Scan to see if all operands are the same opcode, all have one use, and all
8075 // kill their operands (i.e. the operands have one use).
8076 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8077 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8078 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8079 // Verify type of the LHS matches so we don't fold cmp's of different
8080 // types or GEP's with different index types.
8081 I->getOperand(0)->getType() != LHSType ||
8082 I->getOperand(1)->getType() != RHSType)
8085 // If they are CmpInst instructions, check their predicates
8086 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8087 if (cast<CmpInst>(I)->getPredicate() !=
8088 cast<CmpInst>(FirstInst)->getPredicate())
8091 // Keep track of which operand needs a phi node.
8092 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8093 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8096 // Otherwise, this is safe to transform, determine if it is profitable.
8098 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8099 // Indexes are often folded into load/store instructions, so we don't want to
8100 // hide them behind a phi.
8101 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8104 Value *InLHS = FirstInst->getOperand(0);
8105 Value *InRHS = FirstInst->getOperand(1);
8106 PHINode *NewLHS = 0, *NewRHS = 0;
8108 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8109 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8110 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8111 InsertNewInstBefore(NewLHS, PN);
8116 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8117 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8118 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8119 InsertNewInstBefore(NewRHS, PN);
8123 // Add all operands to the new PHIs.
8124 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8126 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8127 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8130 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8131 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8135 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8136 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8137 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8138 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8141 assert(isa<GetElementPtrInst>(FirstInst));
8142 return new GetElementPtrInst(LHSVal, RHSVal);
8146 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8147 /// of the block that defines it. This means that it must be obvious the value
8148 /// of the load is not changed from the point of the load to the end of the
8151 /// Finally, it is safe, but not profitable, to sink a load targetting a
8152 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8154 static bool isSafeToSinkLoad(LoadInst *L) {
8155 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8157 for (++BBI; BBI != E; ++BBI)
8158 if (BBI->mayWriteToMemory())
8161 // Check for non-address taken alloca. If not address-taken already, it isn't
8162 // profitable to do this xform.
8163 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8164 bool isAddressTaken = false;
8165 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8167 if (isa<LoadInst>(UI)) continue;
8168 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8169 // If storing TO the alloca, then the address isn't taken.
8170 if (SI->getOperand(1) == AI) continue;
8172 isAddressTaken = true;
8176 if (!isAddressTaken)
8184 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8185 // operator and they all are only used by the PHI, PHI together their
8186 // inputs, and do the operation once, to the result of the PHI.
8187 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8188 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8190 // Scan the instruction, looking for input operations that can be folded away.
8191 // If all input operands to the phi are the same instruction (e.g. a cast from
8192 // the same type or "+42") we can pull the operation through the PHI, reducing
8193 // code size and simplifying code.
8194 Constant *ConstantOp = 0;
8195 const Type *CastSrcTy = 0;
8196 bool isVolatile = false;
8197 if (isa<CastInst>(FirstInst)) {
8198 CastSrcTy = FirstInst->getOperand(0)->getType();
8199 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8200 // Can fold binop, compare or shift here if the RHS is a constant,
8201 // otherwise call FoldPHIArgBinOpIntoPHI.
8202 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8203 if (ConstantOp == 0)
8204 return FoldPHIArgBinOpIntoPHI(PN);
8205 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8206 isVolatile = LI->isVolatile();
8207 // We can't sink the load if the loaded value could be modified between the
8208 // load and the PHI.
8209 if (LI->getParent() != PN.getIncomingBlock(0) ||
8210 !isSafeToSinkLoad(LI))
8212 } else if (isa<GetElementPtrInst>(FirstInst)) {
8213 if (FirstInst->getNumOperands() == 2)
8214 return FoldPHIArgBinOpIntoPHI(PN);
8215 // Can't handle general GEPs yet.
8218 return 0; // Cannot fold this operation.
8221 // Check to see if all arguments are the same operation.
8222 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8223 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8224 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8225 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8228 if (I->getOperand(0)->getType() != CastSrcTy)
8229 return 0; // Cast operation must match.
8230 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8231 // We can't sink the load if the loaded value could be modified between
8232 // the load and the PHI.
8233 if (LI->isVolatile() != isVolatile ||
8234 LI->getParent() != PN.getIncomingBlock(i) ||
8235 !isSafeToSinkLoad(LI))
8237 } else if (I->getOperand(1) != ConstantOp) {
8242 // Okay, they are all the same operation. Create a new PHI node of the
8243 // correct type, and PHI together all of the LHS's of the instructions.
8244 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8245 PN.getName()+".in");
8246 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8248 Value *InVal = FirstInst->getOperand(0);
8249 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8251 // Add all operands to the new PHI.
8252 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8253 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8254 if (NewInVal != InVal)
8256 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8261 // The new PHI unions all of the same values together. This is really
8262 // common, so we handle it intelligently here for compile-time speed.
8266 InsertNewInstBefore(NewPN, PN);
8270 // Insert and return the new operation.
8271 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8272 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8273 else if (isa<LoadInst>(FirstInst))
8274 return new LoadInst(PhiVal, "", isVolatile);
8275 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8276 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8277 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8278 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8279 PhiVal, ConstantOp);
8281 assert(0 && "Unknown operation");
8285 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8287 static bool DeadPHICycle(PHINode *PN,
8288 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8289 if (PN->use_empty()) return true;
8290 if (!PN->hasOneUse()) return false;
8292 // Remember this node, and if we find the cycle, return.
8293 if (!PotentiallyDeadPHIs.insert(PN))
8296 // Don't scan crazily complex things.
8297 if (PotentiallyDeadPHIs.size() == 16)
8300 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8301 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8306 // PHINode simplification
8308 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8309 // If LCSSA is around, don't mess with Phi nodes
8310 if (MustPreserveLCSSA) return 0;
8312 if (Value *V = PN.hasConstantValue())
8313 return ReplaceInstUsesWith(PN, V);
8315 // If all PHI operands are the same operation, pull them through the PHI,
8316 // reducing code size.
8317 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8318 PN.getIncomingValue(0)->hasOneUse())
8319 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8322 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8323 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8324 // PHI)... break the cycle.
8325 if (PN.hasOneUse()) {
8326 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8327 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8328 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8329 PotentiallyDeadPHIs.insert(&PN);
8330 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8331 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8334 // If this phi has a single use, and if that use just computes a value for
8335 // the next iteration of a loop, delete the phi. This occurs with unused
8336 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8337 // common case here is good because the only other things that catch this
8338 // are induction variable analysis (sometimes) and ADCE, which is only run
8340 if (PHIUser->hasOneUse() &&
8341 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8342 PHIUser->use_back() == &PN) {
8343 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8350 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8351 Instruction *InsertPoint,
8353 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8354 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8355 // We must cast correctly to the pointer type. Ensure that we
8356 // sign extend the integer value if it is smaller as this is
8357 // used for address computation.
8358 Instruction::CastOps opcode =
8359 (VTySize < PtrSize ? Instruction::SExt :
8360 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8361 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8365 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8366 Value *PtrOp = GEP.getOperand(0);
8367 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8368 // If so, eliminate the noop.
8369 if (GEP.getNumOperands() == 1)
8370 return ReplaceInstUsesWith(GEP, PtrOp);
8372 if (isa<UndefValue>(GEP.getOperand(0)))
8373 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8375 bool HasZeroPointerIndex = false;
8376 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8377 HasZeroPointerIndex = C->isNullValue();
8379 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8380 return ReplaceInstUsesWith(GEP, PtrOp);
8382 // Eliminate unneeded casts for indices.
8383 bool MadeChange = false;
8385 gep_type_iterator GTI = gep_type_begin(GEP);
8386 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8387 if (isa<SequentialType>(*GTI)) {
8388 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8389 if (CI->getOpcode() == Instruction::ZExt ||
8390 CI->getOpcode() == Instruction::SExt) {
8391 const Type *SrcTy = CI->getOperand(0)->getType();
8392 // We can eliminate a cast from i32 to i64 iff the target
8393 // is a 32-bit pointer target.
8394 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8396 GEP.setOperand(i, CI->getOperand(0));
8400 // If we are using a wider index than needed for this platform, shrink it
8401 // to what we need. If the incoming value needs a cast instruction,
8402 // insert it. This explicit cast can make subsequent optimizations more
8404 Value *Op = GEP.getOperand(i);
8405 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
8406 if (Constant *C = dyn_cast<Constant>(Op)) {
8407 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8410 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8412 GEP.setOperand(i, Op);
8417 if (MadeChange) return &GEP;
8419 // If this GEP instruction doesn't move the pointer, and if the input operand
8420 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8421 // real input to the dest type.
8422 if (GEP.hasAllZeroIndices() && isa<BitCastInst>(GEP.getOperand(0)))
8423 return new BitCastInst(cast<BitCastInst>(GEP.getOperand(0))->getOperand(0),
8426 // Combine Indices - If the source pointer to this getelementptr instruction
8427 // is a getelementptr instruction, combine the indices of the two
8428 // getelementptr instructions into a single instruction.
8430 SmallVector<Value*, 8> SrcGEPOperands;
8431 if (User *Src = dyn_castGetElementPtr(PtrOp))
8432 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8434 if (!SrcGEPOperands.empty()) {
8435 // Note that if our source is a gep chain itself that we wait for that
8436 // chain to be resolved before we perform this transformation. This
8437 // avoids us creating a TON of code in some cases.
8439 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8440 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8441 return 0; // Wait until our source is folded to completion.
8443 SmallVector<Value*, 8> Indices;
8445 // Find out whether the last index in the source GEP is a sequential idx.
8446 bool EndsWithSequential = false;
8447 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8448 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8449 EndsWithSequential = !isa<StructType>(*I);
8451 // Can we combine the two pointer arithmetics offsets?
8452 if (EndsWithSequential) {
8453 // Replace: gep (gep %P, long B), long A, ...
8454 // With: T = long A+B; gep %P, T, ...
8456 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8457 if (SO1 == Constant::getNullValue(SO1->getType())) {
8459 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8462 // If they aren't the same type, convert both to an integer of the
8463 // target's pointer size.
8464 if (SO1->getType() != GO1->getType()) {
8465 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8466 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8467 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8468 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8470 unsigned PS = TD->getPointerSize();
8471 if (TD->getTypeSize(SO1->getType()) == PS) {
8472 // Convert GO1 to SO1's type.
8473 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8475 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8476 // Convert SO1 to GO1's type.
8477 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8479 const Type *PT = TD->getIntPtrType();
8480 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8481 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8485 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8486 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8488 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8489 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8493 // Recycle the GEP we already have if possible.
8494 if (SrcGEPOperands.size() == 2) {
8495 GEP.setOperand(0, SrcGEPOperands[0]);
8496 GEP.setOperand(1, Sum);
8499 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8500 SrcGEPOperands.end()-1);
8501 Indices.push_back(Sum);
8502 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8504 } else if (isa<Constant>(*GEP.idx_begin()) &&
8505 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8506 SrcGEPOperands.size() != 1) {
8507 // Otherwise we can do the fold if the first index of the GEP is a zero
8508 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8509 SrcGEPOperands.end());
8510 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8513 if (!Indices.empty())
8514 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
8515 Indices.end(), GEP.getName());
8517 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8518 // GEP of global variable. If all of the indices for this GEP are
8519 // constants, we can promote this to a constexpr instead of an instruction.
8521 // Scan for nonconstants...
8522 SmallVector<Constant*, 8> Indices;
8523 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8524 for (; I != E && isa<Constant>(*I); ++I)
8525 Indices.push_back(cast<Constant>(*I));
8527 if (I == E) { // If they are all constants...
8528 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8529 &Indices[0],Indices.size());
8531 // Replace all uses of the GEP with the new constexpr...
8532 return ReplaceInstUsesWith(GEP, CE);
8534 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8535 if (!isa<PointerType>(X->getType())) {
8536 // Not interesting. Source pointer must be a cast from pointer.
8537 } else if (HasZeroPointerIndex) {
8538 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8539 // into : GEP [10 x ubyte]* X, long 0, ...
8541 // This occurs when the program declares an array extern like "int X[];"
8543 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8544 const PointerType *XTy = cast<PointerType>(X->getType());
8545 if (const ArrayType *XATy =
8546 dyn_cast<ArrayType>(XTy->getElementType()))
8547 if (const ArrayType *CATy =
8548 dyn_cast<ArrayType>(CPTy->getElementType()))
8549 if (CATy->getElementType() == XATy->getElementType()) {
8550 // At this point, we know that the cast source type is a pointer
8551 // to an array of the same type as the destination pointer
8552 // array. Because the array type is never stepped over (there
8553 // is a leading zero) we can fold the cast into this GEP.
8554 GEP.setOperand(0, X);
8557 } else if (GEP.getNumOperands() == 2) {
8558 // Transform things like:
8559 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8560 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8561 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8562 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8563 if (isa<ArrayType>(SrcElTy) &&
8564 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8565 TD->getTypeSize(ResElTy)) {
8567 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8568 Idx[1] = GEP.getOperand(1);
8569 Value *V = InsertNewInstBefore(
8570 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
8571 // V and GEP are both pointer types --> BitCast
8572 return new BitCastInst(V, GEP.getType());
8575 // Transform things like:
8576 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8577 // (where tmp = 8*tmp2) into:
8578 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8580 if (isa<ArrayType>(SrcElTy) &&
8581 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8582 uint64_t ArrayEltSize =
8583 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8585 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8586 // allow either a mul, shift, or constant here.
8588 ConstantInt *Scale = 0;
8589 if (ArrayEltSize == 1) {
8590 NewIdx = GEP.getOperand(1);
8591 Scale = ConstantInt::get(NewIdx->getType(), 1);
8592 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8593 NewIdx = ConstantInt::get(CI->getType(), 1);
8595 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8596 if (Inst->getOpcode() == Instruction::Shl &&
8597 isa<ConstantInt>(Inst->getOperand(1))) {
8598 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8599 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8600 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8601 NewIdx = Inst->getOperand(0);
8602 } else if (Inst->getOpcode() == Instruction::Mul &&
8603 isa<ConstantInt>(Inst->getOperand(1))) {
8604 Scale = cast<ConstantInt>(Inst->getOperand(1));
8605 NewIdx = Inst->getOperand(0);
8609 // If the index will be to exactly the right offset with the scale taken
8610 // out, perform the transformation.
8611 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8612 if (isa<ConstantInt>(Scale))
8613 Scale = ConstantInt::get(Scale->getType(),
8614 Scale->getZExtValue() / ArrayEltSize);
8615 if (Scale->getZExtValue() != 1) {
8616 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8618 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8619 NewIdx = InsertNewInstBefore(Sc, GEP);
8622 // Insert the new GEP instruction.
8624 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8626 Instruction *NewGEP =
8627 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
8628 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8629 // The NewGEP must be pointer typed, so must the old one -> BitCast
8630 return new BitCastInst(NewGEP, GEP.getType());
8639 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8640 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8641 if (AI.isArrayAllocation()) // Check C != 1
8642 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8644 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8645 AllocationInst *New = 0;
8647 // Create and insert the replacement instruction...
8648 if (isa<MallocInst>(AI))
8649 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8651 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8652 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8655 InsertNewInstBefore(New, AI);
8657 // Scan to the end of the allocation instructions, to skip over a block of
8658 // allocas if possible...
8660 BasicBlock::iterator It = New;
8661 while (isa<AllocationInst>(*It)) ++It;
8663 // Now that I is pointing to the first non-allocation-inst in the block,
8664 // insert our getelementptr instruction...
8666 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8670 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
8671 New->getName()+".sub", It);
8673 // Now make everything use the getelementptr instead of the original
8675 return ReplaceInstUsesWith(AI, V);
8676 } else if (isa<UndefValue>(AI.getArraySize())) {
8677 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8680 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8681 // Note that we only do this for alloca's, because malloc should allocate and
8682 // return a unique pointer, even for a zero byte allocation.
8683 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8684 TD->getTypeSize(AI.getAllocatedType()) == 0)
8685 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8690 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8691 Value *Op = FI.getOperand(0);
8693 // free undef -> unreachable.
8694 if (isa<UndefValue>(Op)) {
8695 // Insert a new store to null because we cannot modify the CFG here.
8696 new StoreInst(ConstantInt::getTrue(),
8697 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8698 return EraseInstFromFunction(FI);
8701 // If we have 'free null' delete the instruction. This can happen in stl code
8702 // when lots of inlining happens.
8703 if (isa<ConstantPointerNull>(Op))
8704 return EraseInstFromFunction(FI);
8706 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8707 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
8708 FI.setOperand(0, CI->getOperand(0));
8712 // Change free (gep X, 0,0,0,0) into free(X)
8713 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
8714 if (GEPI->hasAllZeroIndices()) {
8715 AddToWorkList(GEPI);
8716 FI.setOperand(0, GEPI->getOperand(0));
8721 // Change free(malloc) into nothing, if the malloc has a single use.
8722 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
8723 if (MI->hasOneUse()) {
8724 EraseInstFromFunction(FI);
8725 return EraseInstFromFunction(*MI);
8732 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8733 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8734 User *CI = cast<User>(LI.getOperand(0));
8735 Value *CastOp = CI->getOperand(0);
8737 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8738 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8739 const Type *SrcPTy = SrcTy->getElementType();
8741 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8742 isa<VectorType>(DestPTy)) {
8743 // If the source is an array, the code below will not succeed. Check to
8744 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8746 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8747 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8748 if (ASrcTy->getNumElements() != 0) {
8750 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8751 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8752 SrcTy = cast<PointerType>(CastOp->getType());
8753 SrcPTy = SrcTy->getElementType();
8756 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8757 isa<VectorType>(SrcPTy)) &&
8758 // Do not allow turning this into a load of an integer, which is then
8759 // casted to a pointer, this pessimizes pointer analysis a lot.
8760 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8761 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8762 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8764 // Okay, we are casting from one integer or pointer type to another of
8765 // the same size. Instead of casting the pointer before the load, cast
8766 // the result of the loaded value.
8767 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8769 LI.isVolatile()),LI);
8770 // Now cast the result of the load.
8771 return new BitCastInst(NewLoad, LI.getType());
8778 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8779 /// from this value cannot trap. If it is not obviously safe to load from the
8780 /// specified pointer, we do a quick local scan of the basic block containing
8781 /// ScanFrom, to determine if the address is already accessed.
8782 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8783 // If it is an alloca or global variable, it is always safe to load from.
8784 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8786 // Otherwise, be a little bit agressive by scanning the local block where we
8787 // want to check to see if the pointer is already being loaded or stored
8788 // from/to. If so, the previous load or store would have already trapped,
8789 // so there is no harm doing an extra load (also, CSE will later eliminate
8790 // the load entirely).
8791 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8796 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8797 if (LI->getOperand(0) == V) return true;
8798 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8799 if (SI->getOperand(1) == V) return true;
8805 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
8806 /// until we find the underlying object a pointer is referring to or something
8807 /// we don't understand. Note that the returned pointer may be offset from the
8808 /// input, because we ignore GEP indices.
8809 static Value *GetUnderlyingObject(Value *Ptr) {
8811 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
8812 if (CE->getOpcode() == Instruction::BitCast ||
8813 CE->getOpcode() == Instruction::GetElementPtr)
8814 Ptr = CE->getOperand(0);
8817 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
8818 Ptr = BCI->getOperand(0);
8819 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
8820 Ptr = GEP->getOperand(0);
8827 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8828 Value *Op = LI.getOperand(0);
8830 // Attempt to improve the alignment.
8831 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
8832 if (KnownAlign > LI.getAlignment())
8833 LI.setAlignment(KnownAlign);
8835 // load (cast X) --> cast (load X) iff safe
8836 if (isa<CastInst>(Op))
8837 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8840 // None of the following transforms are legal for volatile loads.
8841 if (LI.isVolatile()) return 0;
8843 if (&LI.getParent()->front() != &LI) {
8844 BasicBlock::iterator BBI = &LI; --BBI;
8845 // If the instruction immediately before this is a store to the same
8846 // address, do a simple form of store->load forwarding.
8847 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8848 if (SI->getOperand(1) == LI.getOperand(0))
8849 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8850 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8851 if (LIB->getOperand(0) == LI.getOperand(0))
8852 return ReplaceInstUsesWith(LI, LIB);
8855 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8856 if (isa<ConstantPointerNull>(GEPI->getOperand(0))) {
8857 // Insert a new store to null instruction before the load to indicate
8858 // that this code is not reachable. We do this instead of inserting
8859 // an unreachable instruction directly because we cannot modify the
8861 new StoreInst(UndefValue::get(LI.getType()),
8862 Constant::getNullValue(Op->getType()), &LI);
8863 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8866 if (Constant *C = dyn_cast<Constant>(Op)) {
8867 // load null/undef -> undef
8868 if ((C->isNullValue() || isa<UndefValue>(C))) {
8869 // Insert a new store to null instruction before the load to indicate that
8870 // this code is not reachable. We do this instead of inserting an
8871 // unreachable instruction directly because we cannot modify the CFG.
8872 new StoreInst(UndefValue::get(LI.getType()),
8873 Constant::getNullValue(Op->getType()), &LI);
8874 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8877 // Instcombine load (constant global) into the value loaded.
8878 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8879 if (GV->isConstant() && !GV->isDeclaration())
8880 return ReplaceInstUsesWith(LI, GV->getInitializer());
8882 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8883 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8884 if (CE->getOpcode() == Instruction::GetElementPtr) {
8885 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8886 if (GV->isConstant() && !GV->isDeclaration())
8888 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8889 return ReplaceInstUsesWith(LI, V);
8890 if (CE->getOperand(0)->isNullValue()) {
8891 // Insert a new store to null instruction before the load to indicate
8892 // that this code is not reachable. We do this instead of inserting
8893 // an unreachable instruction directly because we cannot modify the
8895 new StoreInst(UndefValue::get(LI.getType()),
8896 Constant::getNullValue(Op->getType()), &LI);
8897 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8900 } else if (CE->isCast()) {
8901 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8906 // If this load comes from anywhere in a constant global, and if the global
8907 // is all undef or zero, we know what it loads.
8908 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
8909 if (GV->isConstant() && GV->hasInitializer()) {
8910 if (GV->getInitializer()->isNullValue())
8911 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
8912 else if (isa<UndefValue>(GV->getInitializer()))
8913 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8917 if (Op->hasOneUse()) {
8918 // Change select and PHI nodes to select values instead of addresses: this
8919 // helps alias analysis out a lot, allows many others simplifications, and
8920 // exposes redundancy in the code.
8922 // Note that we cannot do the transformation unless we know that the
8923 // introduced loads cannot trap! Something like this is valid as long as
8924 // the condition is always false: load (select bool %C, int* null, int* %G),
8925 // but it would not be valid if we transformed it to load from null
8928 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8929 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8930 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8931 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8932 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8933 SI->getOperand(1)->getName()+".val"), LI);
8934 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8935 SI->getOperand(2)->getName()+".val"), LI);
8936 return new SelectInst(SI->getCondition(), V1, V2);
8939 // load (select (cond, null, P)) -> load P
8940 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8941 if (C->isNullValue()) {
8942 LI.setOperand(0, SI->getOperand(2));
8946 // load (select (cond, P, null)) -> load P
8947 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8948 if (C->isNullValue()) {
8949 LI.setOperand(0, SI->getOperand(1));
8957 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8959 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8960 User *CI = cast<User>(SI.getOperand(1));
8961 Value *CastOp = CI->getOperand(0);
8963 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8964 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8965 const Type *SrcPTy = SrcTy->getElementType();
8967 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8968 // If the source is an array, the code below will not succeed. Check to
8969 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8971 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8972 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8973 if (ASrcTy->getNumElements() != 0) {
8975 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8976 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8977 SrcTy = cast<PointerType>(CastOp->getType());
8978 SrcPTy = SrcTy->getElementType();
8981 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8982 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8983 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8985 // Okay, we are casting from one integer or pointer type to another of
8986 // the same size. Instead of casting the pointer before
8987 // the store, cast the value to be stored.
8989 Value *SIOp0 = SI.getOperand(0);
8990 Instruction::CastOps opcode = Instruction::BitCast;
8991 const Type* CastSrcTy = SIOp0->getType();
8992 const Type* CastDstTy = SrcPTy;
8993 if (isa<PointerType>(CastDstTy)) {
8994 if (CastSrcTy->isInteger())
8995 opcode = Instruction::IntToPtr;
8996 } else if (isa<IntegerType>(CastDstTy)) {
8997 if (isa<PointerType>(SIOp0->getType()))
8998 opcode = Instruction::PtrToInt;
9000 if (Constant *C = dyn_cast<Constant>(SIOp0))
9001 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9003 NewCast = IC.InsertNewInstBefore(
9004 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9006 return new StoreInst(NewCast, CastOp);
9013 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9014 Value *Val = SI.getOperand(0);
9015 Value *Ptr = SI.getOperand(1);
9017 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9018 EraseInstFromFunction(SI);
9023 // If the RHS is an alloca with a single use, zapify the store, making the
9025 if (Ptr->hasOneUse()) {
9026 if (isa<AllocaInst>(Ptr)) {
9027 EraseInstFromFunction(SI);
9032 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9033 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9034 GEP->getOperand(0)->hasOneUse()) {
9035 EraseInstFromFunction(SI);
9041 // Attempt to improve the alignment.
9042 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9043 if (KnownAlign > SI.getAlignment())
9044 SI.setAlignment(KnownAlign);
9046 // Do really simple DSE, to catch cases where there are several consequtive
9047 // stores to the same location, separated by a few arithmetic operations. This
9048 // situation often occurs with bitfield accesses.
9049 BasicBlock::iterator BBI = &SI;
9050 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9054 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9055 // Prev store isn't volatile, and stores to the same location?
9056 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9059 EraseInstFromFunction(*PrevSI);
9065 // If this is a load, we have to stop. However, if the loaded value is from
9066 // the pointer we're loading and is producing the pointer we're storing,
9067 // then *this* store is dead (X = load P; store X -> P).
9068 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9069 if (LI == Val && LI->getOperand(0) == Ptr) {
9070 EraseInstFromFunction(SI);
9074 // Otherwise, this is a load from some other location. Stores before it
9079 // Don't skip over loads or things that can modify memory.
9080 if (BBI->mayWriteToMemory())
9085 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9087 // store X, null -> turns into 'unreachable' in SimplifyCFG
9088 if (isa<ConstantPointerNull>(Ptr)) {
9089 if (!isa<UndefValue>(Val)) {
9090 SI.setOperand(0, UndefValue::get(Val->getType()));
9091 if (Instruction *U = dyn_cast<Instruction>(Val))
9092 AddToWorkList(U); // Dropped a use.
9095 return 0; // Do not modify these!
9098 // store undef, Ptr -> noop
9099 if (isa<UndefValue>(Val)) {
9100 EraseInstFromFunction(SI);
9105 // If the pointer destination is a cast, see if we can fold the cast into the
9107 if (isa<CastInst>(Ptr))
9108 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9110 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9112 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9116 // If this store is the last instruction in the basic block, and if the block
9117 // ends with an unconditional branch, try to move it to the successor block.
9119 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9120 if (BI->isUnconditional())
9121 if (SimplifyStoreAtEndOfBlock(SI))
9122 return 0; // xform done!
9127 /// SimplifyStoreAtEndOfBlock - Turn things like:
9128 /// if () { *P = v1; } else { *P = v2 }
9129 /// into a phi node with a store in the successor.
9131 /// Simplify things like:
9132 /// *P = v1; if () { *P = v2; }
9133 /// into a phi node with a store in the successor.
9135 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9136 BasicBlock *StoreBB = SI.getParent();
9138 // Check to see if the successor block has exactly two incoming edges. If
9139 // so, see if the other predecessor contains a store to the same location.
9140 // if so, insert a PHI node (if needed) and move the stores down.
9141 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9143 // Determine whether Dest has exactly two predecessors and, if so, compute
9144 // the other predecessor.
9145 pred_iterator PI = pred_begin(DestBB);
9146 BasicBlock *OtherBB = 0;
9150 if (PI == pred_end(DestBB))
9153 if (*PI != StoreBB) {
9158 if (++PI != pred_end(DestBB))
9162 // Verify that the other block ends in a branch and is not otherwise empty.
9163 BasicBlock::iterator BBI = OtherBB->getTerminator();
9164 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9165 if (!OtherBr || BBI == OtherBB->begin())
9168 // If the other block ends in an unconditional branch, check for the 'if then
9169 // else' case. there is an instruction before the branch.
9170 StoreInst *OtherStore = 0;
9171 if (OtherBr->isUnconditional()) {
9172 // If this isn't a store, or isn't a store to the same location, bail out.
9174 OtherStore = dyn_cast<StoreInst>(BBI);
9175 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9178 // Otherwise, the other block ended with a conditional branch. If one of the
9179 // destinations is StoreBB, then we have the if/then case.
9180 if (OtherBr->getSuccessor(0) != StoreBB &&
9181 OtherBr->getSuccessor(1) != StoreBB)
9184 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9185 // if/then triangle. See if there is a store to the same ptr as SI that
9186 // lives in OtherBB.
9188 // Check to see if we find the matching store.
9189 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9190 if (OtherStore->getOperand(1) != SI.getOperand(1))
9194 // If we find something that may be using the stored value, or if we run
9195 // out of instructions, we can't do the xform.
9196 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9197 BBI == OtherBB->begin())
9201 // In order to eliminate the store in OtherBr, we have to
9202 // make sure nothing reads the stored value in StoreBB.
9203 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9204 // FIXME: This should really be AA driven.
9205 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9210 // Insert a PHI node now if we need it.
9211 Value *MergedVal = OtherStore->getOperand(0);
9212 if (MergedVal != SI.getOperand(0)) {
9213 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9214 PN->reserveOperandSpace(2);
9215 PN->addIncoming(SI.getOperand(0), SI.getParent());
9216 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9217 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9220 // Advance to a place where it is safe to insert the new store and
9222 BBI = DestBB->begin();
9223 while (isa<PHINode>(BBI)) ++BBI;
9224 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9225 OtherStore->isVolatile()), *BBI);
9227 // Nuke the old stores.
9228 EraseInstFromFunction(SI);
9229 EraseInstFromFunction(*OtherStore);
9235 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9236 // Change br (not X), label True, label False to: br X, label False, True
9238 BasicBlock *TrueDest;
9239 BasicBlock *FalseDest;
9240 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9241 !isa<Constant>(X)) {
9242 // Swap Destinations and condition...
9244 BI.setSuccessor(0, FalseDest);
9245 BI.setSuccessor(1, TrueDest);
9249 // Cannonicalize fcmp_one -> fcmp_oeq
9250 FCmpInst::Predicate FPred; Value *Y;
9251 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9252 TrueDest, FalseDest)))
9253 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9254 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9255 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9256 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9257 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9258 NewSCC->takeName(I);
9259 // Swap Destinations and condition...
9260 BI.setCondition(NewSCC);
9261 BI.setSuccessor(0, FalseDest);
9262 BI.setSuccessor(1, TrueDest);
9263 RemoveFromWorkList(I);
9264 I->eraseFromParent();
9265 AddToWorkList(NewSCC);
9269 // Cannonicalize icmp_ne -> icmp_eq
9270 ICmpInst::Predicate IPred;
9271 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9272 TrueDest, FalseDest)))
9273 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9274 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9275 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9276 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9277 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9278 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9279 NewSCC->takeName(I);
9280 // Swap Destinations and condition...
9281 BI.setCondition(NewSCC);
9282 BI.setSuccessor(0, FalseDest);
9283 BI.setSuccessor(1, TrueDest);
9284 RemoveFromWorkList(I);
9285 I->eraseFromParent();;
9286 AddToWorkList(NewSCC);
9293 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9294 Value *Cond = SI.getCondition();
9295 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9296 if (I->getOpcode() == Instruction::Add)
9297 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9298 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9299 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9300 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9302 SI.setOperand(0, I->getOperand(0));
9310 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9311 /// is to leave as a vector operation.
9312 static bool CheapToScalarize(Value *V, bool isConstant) {
9313 if (isa<ConstantAggregateZero>(V))
9315 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9316 if (isConstant) return true;
9317 // If all elts are the same, we can extract.
9318 Constant *Op0 = C->getOperand(0);
9319 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9320 if (C->getOperand(i) != Op0)
9324 Instruction *I = dyn_cast<Instruction>(V);
9325 if (!I) return false;
9327 // Insert element gets simplified to the inserted element or is deleted if
9328 // this is constant idx extract element and its a constant idx insertelt.
9329 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9330 isa<ConstantInt>(I->getOperand(2)))
9332 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9334 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9335 if (BO->hasOneUse() &&
9336 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9337 CheapToScalarize(BO->getOperand(1), isConstant)))
9339 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9340 if (CI->hasOneUse() &&
9341 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9342 CheapToScalarize(CI->getOperand(1), isConstant)))
9348 /// Read and decode a shufflevector mask.
9350 /// It turns undef elements into values that are larger than the number of
9351 /// elements in the input.
9352 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9353 unsigned NElts = SVI->getType()->getNumElements();
9354 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9355 return std::vector<unsigned>(NElts, 0);
9356 if (isa<UndefValue>(SVI->getOperand(2)))
9357 return std::vector<unsigned>(NElts, 2*NElts);
9359 std::vector<unsigned> Result;
9360 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9361 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9362 if (isa<UndefValue>(CP->getOperand(i)))
9363 Result.push_back(NElts*2); // undef -> 8
9365 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9369 /// FindScalarElement - Given a vector and an element number, see if the scalar
9370 /// value is already around as a register, for example if it were inserted then
9371 /// extracted from the vector.
9372 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9373 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9374 const VectorType *PTy = cast<VectorType>(V->getType());
9375 unsigned Width = PTy->getNumElements();
9376 if (EltNo >= Width) // Out of range access.
9377 return UndefValue::get(PTy->getElementType());
9379 if (isa<UndefValue>(V))
9380 return UndefValue::get(PTy->getElementType());
9381 else if (isa<ConstantAggregateZero>(V))
9382 return Constant::getNullValue(PTy->getElementType());
9383 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9384 return CP->getOperand(EltNo);
9385 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9386 // If this is an insert to a variable element, we don't know what it is.
9387 if (!isa<ConstantInt>(III->getOperand(2)))
9389 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9391 // If this is an insert to the element we are looking for, return the
9394 return III->getOperand(1);
9396 // Otherwise, the insertelement doesn't modify the value, recurse on its
9398 return FindScalarElement(III->getOperand(0), EltNo);
9399 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9400 unsigned InEl = getShuffleMask(SVI)[EltNo];
9402 return FindScalarElement(SVI->getOperand(0), InEl);
9403 else if (InEl < Width*2)
9404 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9406 return UndefValue::get(PTy->getElementType());
9409 // Otherwise, we don't know.
9413 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9415 // If vector val is undef, replace extract with scalar undef.
9416 if (isa<UndefValue>(EI.getOperand(0)))
9417 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9419 // If vector val is constant 0, replace extract with scalar 0.
9420 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9421 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9423 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9424 // If vector val is constant with uniform operands, replace EI
9425 // with that operand
9426 Constant *op0 = C->getOperand(0);
9427 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9428 if (C->getOperand(i) != op0) {
9433 return ReplaceInstUsesWith(EI, op0);
9436 // If extracting a specified index from the vector, see if we can recursively
9437 // find a previously computed scalar that was inserted into the vector.
9438 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9439 unsigned IndexVal = IdxC->getZExtValue();
9440 unsigned VectorWidth =
9441 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9443 // If this is extracting an invalid index, turn this into undef, to avoid
9444 // crashing the code below.
9445 if (IndexVal >= VectorWidth)
9446 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9448 // This instruction only demands the single element from the input vector.
9449 // If the input vector has a single use, simplify it based on this use
9451 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9453 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9456 EI.setOperand(0, V);
9461 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9462 return ReplaceInstUsesWith(EI, Elt);
9464 // If the this extractelement is directly using a bitcast from a vector of
9465 // the same number of elements, see if we can find the source element from
9466 // it. In this case, we will end up needing to bitcast the scalars.
9467 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9468 if (const VectorType *VT =
9469 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9470 if (VT->getNumElements() == VectorWidth)
9471 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9472 return new BitCastInst(Elt, EI.getType());
9476 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9477 if (I->hasOneUse()) {
9478 // Push extractelement into predecessor operation if legal and
9479 // profitable to do so
9480 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9481 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9482 if (CheapToScalarize(BO, isConstantElt)) {
9483 ExtractElementInst *newEI0 =
9484 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9485 EI.getName()+".lhs");
9486 ExtractElementInst *newEI1 =
9487 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9488 EI.getName()+".rhs");
9489 InsertNewInstBefore(newEI0, EI);
9490 InsertNewInstBefore(newEI1, EI);
9491 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9493 } else if (isa<LoadInst>(I)) {
9494 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9495 PointerType::get(EI.getType()), EI);
9496 GetElementPtrInst *GEP =
9497 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9498 InsertNewInstBefore(GEP, EI);
9499 return new LoadInst(GEP);
9502 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9503 // Extracting the inserted element?
9504 if (IE->getOperand(2) == EI.getOperand(1))
9505 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9506 // If the inserted and extracted elements are constants, they must not
9507 // be the same value, extract from the pre-inserted value instead.
9508 if (isa<Constant>(IE->getOperand(2)) &&
9509 isa<Constant>(EI.getOperand(1))) {
9510 AddUsesToWorkList(EI);
9511 EI.setOperand(0, IE->getOperand(0));
9514 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9515 // If this is extracting an element from a shufflevector, figure out where
9516 // it came from and extract from the appropriate input element instead.
9517 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9518 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9520 if (SrcIdx < SVI->getType()->getNumElements())
9521 Src = SVI->getOperand(0);
9522 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9523 SrcIdx -= SVI->getType()->getNumElements();
9524 Src = SVI->getOperand(1);
9526 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9528 return new ExtractElementInst(Src, SrcIdx);
9535 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9536 /// elements from either LHS or RHS, return the shuffle mask and true.
9537 /// Otherwise, return false.
9538 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9539 std::vector<Constant*> &Mask) {
9540 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9541 "Invalid CollectSingleShuffleElements");
9542 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9544 if (isa<UndefValue>(V)) {
9545 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9547 } else if (V == LHS) {
9548 for (unsigned i = 0; i != NumElts; ++i)
9549 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9551 } else if (V == RHS) {
9552 for (unsigned i = 0; i != NumElts; ++i)
9553 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9555 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9556 // If this is an insert of an extract from some other vector, include it.
9557 Value *VecOp = IEI->getOperand(0);
9558 Value *ScalarOp = IEI->getOperand(1);
9559 Value *IdxOp = IEI->getOperand(2);
9561 if (!isa<ConstantInt>(IdxOp))
9563 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9565 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9566 // Okay, we can handle this if the vector we are insertinting into is
9568 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9569 // If so, update the mask to reflect the inserted undef.
9570 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9573 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9574 if (isa<ConstantInt>(EI->getOperand(1)) &&
9575 EI->getOperand(0)->getType() == V->getType()) {
9576 unsigned ExtractedIdx =
9577 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9579 // This must be extracting from either LHS or RHS.
9580 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9581 // Okay, we can handle this if the vector we are insertinting into is
9583 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9584 // If so, update the mask to reflect the inserted value.
9585 if (EI->getOperand(0) == LHS) {
9586 Mask[InsertedIdx & (NumElts-1)] =
9587 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9589 assert(EI->getOperand(0) == RHS);
9590 Mask[InsertedIdx & (NumElts-1)] =
9591 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9600 // TODO: Handle shufflevector here!
9605 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9606 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9607 /// that computes V and the LHS value of the shuffle.
9608 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9610 assert(isa<VectorType>(V->getType()) &&
9611 (RHS == 0 || V->getType() == RHS->getType()) &&
9612 "Invalid shuffle!");
9613 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9615 if (isa<UndefValue>(V)) {
9616 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9618 } else if (isa<ConstantAggregateZero>(V)) {
9619 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9621 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9622 // If this is an insert of an extract from some other vector, include it.
9623 Value *VecOp = IEI->getOperand(0);
9624 Value *ScalarOp = IEI->getOperand(1);
9625 Value *IdxOp = IEI->getOperand(2);
9627 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9628 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9629 EI->getOperand(0)->getType() == V->getType()) {
9630 unsigned ExtractedIdx =
9631 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9632 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9634 // Either the extracted from or inserted into vector must be RHSVec,
9635 // otherwise we'd end up with a shuffle of three inputs.
9636 if (EI->getOperand(0) == RHS || RHS == 0) {
9637 RHS = EI->getOperand(0);
9638 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9639 Mask[InsertedIdx & (NumElts-1)] =
9640 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9645 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9646 // Everything but the extracted element is replaced with the RHS.
9647 for (unsigned i = 0; i != NumElts; ++i) {
9648 if (i != InsertedIdx)
9649 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9654 // If this insertelement is a chain that comes from exactly these two
9655 // vectors, return the vector and the effective shuffle.
9656 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9657 return EI->getOperand(0);
9662 // TODO: Handle shufflevector here!
9664 // Otherwise, can't do anything fancy. Return an identity vector.
9665 for (unsigned i = 0; i != NumElts; ++i)
9666 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9670 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9671 Value *VecOp = IE.getOperand(0);
9672 Value *ScalarOp = IE.getOperand(1);
9673 Value *IdxOp = IE.getOperand(2);
9675 // Inserting an undef or into an undefined place, remove this.
9676 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
9677 ReplaceInstUsesWith(IE, VecOp);
9679 // If the inserted element was extracted from some other vector, and if the
9680 // indexes are constant, try to turn this into a shufflevector operation.
9681 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9682 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9683 EI->getOperand(0)->getType() == IE.getType()) {
9684 unsigned NumVectorElts = IE.getType()->getNumElements();
9685 unsigned ExtractedIdx =
9686 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9687 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9689 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9690 return ReplaceInstUsesWith(IE, VecOp);
9692 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9693 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9695 // If we are extracting a value from a vector, then inserting it right
9696 // back into the same place, just use the input vector.
9697 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9698 return ReplaceInstUsesWith(IE, VecOp);
9700 // We could theoretically do this for ANY input. However, doing so could
9701 // turn chains of insertelement instructions into a chain of shufflevector
9702 // instructions, and right now we do not merge shufflevectors. As such,
9703 // only do this in a situation where it is clear that there is benefit.
9704 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9705 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9706 // the values of VecOp, except then one read from EIOp0.
9707 // Build a new shuffle mask.
9708 std::vector<Constant*> Mask;
9709 if (isa<UndefValue>(VecOp))
9710 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9712 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9713 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9716 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9717 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9718 ConstantVector::get(Mask));
9721 // If this insertelement isn't used by some other insertelement, turn it
9722 // (and any insertelements it points to), into one big shuffle.
9723 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9724 std::vector<Constant*> Mask;
9726 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9727 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9728 // We now have a shuffle of LHS, RHS, Mask.
9729 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9738 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9739 Value *LHS = SVI.getOperand(0);
9740 Value *RHS = SVI.getOperand(1);
9741 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9743 bool MadeChange = false;
9745 // Undefined shuffle mask -> undefined value.
9746 if (isa<UndefValue>(SVI.getOperand(2)))
9747 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9749 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9750 // the undef, change them to undefs.
9751 if (isa<UndefValue>(SVI.getOperand(1))) {
9752 // Scan to see if there are any references to the RHS. If so, replace them
9753 // with undef element refs and set MadeChange to true.
9754 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9755 if (Mask[i] >= e && Mask[i] != 2*e) {
9762 // Remap any references to RHS to use LHS.
9763 std::vector<Constant*> Elts;
9764 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9766 Elts.push_back(UndefValue::get(Type::Int32Ty));
9768 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9770 SVI.setOperand(2, ConstantVector::get(Elts));
9774 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9775 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9776 if (LHS == RHS || isa<UndefValue>(LHS)) {
9777 if (isa<UndefValue>(LHS) && LHS == RHS) {
9778 // shuffle(undef,undef,mask) -> undef.
9779 return ReplaceInstUsesWith(SVI, LHS);
9782 // Remap any references to RHS to use LHS.
9783 std::vector<Constant*> Elts;
9784 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9786 Elts.push_back(UndefValue::get(Type::Int32Ty));
9788 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9789 (Mask[i] < e && isa<UndefValue>(LHS)))
9790 Mask[i] = 2*e; // Turn into undef.
9792 Mask[i] &= (e-1); // Force to LHS.
9793 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9796 SVI.setOperand(0, SVI.getOperand(1));
9797 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9798 SVI.setOperand(2, ConstantVector::get(Elts));
9799 LHS = SVI.getOperand(0);
9800 RHS = SVI.getOperand(1);
9804 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9805 bool isLHSID = true, isRHSID = true;
9807 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9808 if (Mask[i] >= e*2) continue; // Ignore undef values.
9809 // Is this an identity shuffle of the LHS value?
9810 isLHSID &= (Mask[i] == i);
9812 // Is this an identity shuffle of the RHS value?
9813 isRHSID &= (Mask[i]-e == i);
9816 // Eliminate identity shuffles.
9817 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9818 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9820 // If the LHS is a shufflevector itself, see if we can combine it with this
9821 // one without producing an unusual shuffle. Here we are really conservative:
9822 // we are absolutely afraid of producing a shuffle mask not in the input
9823 // program, because the code gen may not be smart enough to turn a merged
9824 // shuffle into two specific shuffles: it may produce worse code. As such,
9825 // we only merge two shuffles if the result is one of the two input shuffle
9826 // masks. In this case, merging the shuffles just removes one instruction,
9827 // which we know is safe. This is good for things like turning:
9828 // (splat(splat)) -> splat.
9829 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9830 if (isa<UndefValue>(RHS)) {
9831 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9833 std::vector<unsigned> NewMask;
9834 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9836 NewMask.push_back(2*e);
9838 NewMask.push_back(LHSMask[Mask[i]]);
9840 // If the result mask is equal to the src shuffle or this shuffle mask, do
9842 if (NewMask == LHSMask || NewMask == Mask) {
9843 std::vector<Constant*> Elts;
9844 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9845 if (NewMask[i] >= e*2) {
9846 Elts.push_back(UndefValue::get(Type::Int32Ty));
9848 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9851 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9852 LHSSVI->getOperand(1),
9853 ConstantVector::get(Elts));
9858 return MadeChange ? &SVI : 0;
9864 /// TryToSinkInstruction - Try to move the specified instruction from its
9865 /// current block into the beginning of DestBlock, which can only happen if it's
9866 /// safe to move the instruction past all of the instructions between it and the
9867 /// end of its block.
9868 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9869 assert(I->hasOneUse() && "Invariants didn't hold!");
9871 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9872 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9874 // Do not sink alloca instructions out of the entry block.
9875 if (isa<AllocaInst>(I) && I->getParent() ==
9876 &DestBlock->getParent()->getEntryBlock())
9879 // We can only sink load instructions if there is nothing between the load and
9880 // the end of block that could change the value.
9881 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9882 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9884 if (Scan->mayWriteToMemory())
9888 BasicBlock::iterator InsertPos = DestBlock->begin();
9889 while (isa<PHINode>(InsertPos)) ++InsertPos;
9891 I->moveBefore(InsertPos);
9897 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9898 /// all reachable code to the worklist.
9900 /// This has a couple of tricks to make the code faster and more powerful. In
9901 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9902 /// them to the worklist (this significantly speeds up instcombine on code where
9903 /// many instructions are dead or constant). Additionally, if we find a branch
9904 /// whose condition is a known constant, we only visit the reachable successors.
9906 static void AddReachableCodeToWorklist(BasicBlock *BB,
9907 SmallPtrSet<BasicBlock*, 64> &Visited,
9909 const TargetData *TD) {
9910 std::vector<BasicBlock*> Worklist;
9911 Worklist.push_back(BB);
9913 while (!Worklist.empty()) {
9914 BB = Worklist.back();
9915 Worklist.pop_back();
9917 // We have now visited this block! If we've already been here, ignore it.
9918 if (!Visited.insert(BB)) continue;
9920 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9921 Instruction *Inst = BBI++;
9923 // DCE instruction if trivially dead.
9924 if (isInstructionTriviallyDead(Inst)) {
9926 DOUT << "IC: DCE: " << *Inst;
9927 Inst->eraseFromParent();
9931 // ConstantProp instruction if trivially constant.
9932 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9933 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9934 Inst->replaceAllUsesWith(C);
9936 Inst->eraseFromParent();
9940 IC.AddToWorkList(Inst);
9943 // Recursively visit successors. If this is a branch or switch on a
9944 // constant, only visit the reachable successor.
9945 TerminatorInst *TI = BB->getTerminator();
9946 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9947 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9948 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9949 Worklist.push_back(BI->getSuccessor(!CondVal));
9952 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9953 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9954 // See if this is an explicit destination.
9955 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9956 if (SI->getCaseValue(i) == Cond) {
9957 Worklist.push_back(SI->getSuccessor(i));
9961 // Otherwise it is the default destination.
9962 Worklist.push_back(SI->getSuccessor(0));
9967 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9968 Worklist.push_back(TI->getSuccessor(i));
9972 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9973 bool Changed = false;
9974 TD = &getAnalysis<TargetData>();
9976 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9977 << F.getNameStr() << "\n");
9980 // Do a depth-first traversal of the function, populate the worklist with
9981 // the reachable instructions. Ignore blocks that are not reachable. Keep
9982 // track of which blocks we visit.
9983 SmallPtrSet<BasicBlock*, 64> Visited;
9984 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9986 // Do a quick scan over the function. If we find any blocks that are
9987 // unreachable, remove any instructions inside of them. This prevents
9988 // the instcombine code from having to deal with some bad special cases.
9989 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9990 if (!Visited.count(BB)) {
9991 Instruction *Term = BB->getTerminator();
9992 while (Term != BB->begin()) { // Remove instrs bottom-up
9993 BasicBlock::iterator I = Term; --I;
9995 DOUT << "IC: DCE: " << *I;
9998 if (!I->use_empty())
9999 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10000 I->eraseFromParent();
10005 while (!Worklist.empty()) {
10006 Instruction *I = RemoveOneFromWorkList();
10007 if (I == 0) continue; // skip null values.
10009 // Check to see if we can DCE the instruction.
10010 if (isInstructionTriviallyDead(I)) {
10011 // Add operands to the worklist.
10012 if (I->getNumOperands() < 4)
10013 AddUsesToWorkList(*I);
10016 DOUT << "IC: DCE: " << *I;
10018 I->eraseFromParent();
10019 RemoveFromWorkList(I);
10023 // Instruction isn't dead, see if we can constant propagate it.
10024 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10025 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10027 // Add operands to the worklist.
10028 AddUsesToWorkList(*I);
10029 ReplaceInstUsesWith(*I, C);
10032 I->eraseFromParent();
10033 RemoveFromWorkList(I);
10037 // See if we can trivially sink this instruction to a successor basic block.
10038 if (I->hasOneUse()) {
10039 BasicBlock *BB = I->getParent();
10040 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10041 if (UserParent != BB) {
10042 bool UserIsSuccessor = false;
10043 // See if the user is one of our successors.
10044 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10045 if (*SI == UserParent) {
10046 UserIsSuccessor = true;
10050 // If the user is one of our immediate successors, and if that successor
10051 // only has us as a predecessors (we'd have to split the critical edge
10052 // otherwise), we can keep going.
10053 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10054 next(pred_begin(UserParent)) == pred_end(UserParent))
10055 // Okay, the CFG is simple enough, try to sink this instruction.
10056 Changed |= TryToSinkInstruction(I, UserParent);
10060 // Now that we have an instruction, try combining it to simplify it...
10064 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10065 if (Instruction *Result = visit(*I)) {
10067 // Should we replace the old instruction with a new one?
10069 DOUT << "IC: Old = " << *I
10070 << " New = " << *Result;
10072 // Everything uses the new instruction now.
10073 I->replaceAllUsesWith(Result);
10075 // Push the new instruction and any users onto the worklist.
10076 AddToWorkList(Result);
10077 AddUsersToWorkList(*Result);
10079 // Move the name to the new instruction first.
10080 Result->takeName(I);
10082 // Insert the new instruction into the basic block...
10083 BasicBlock *InstParent = I->getParent();
10084 BasicBlock::iterator InsertPos = I;
10086 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10087 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10090 InstParent->getInstList().insert(InsertPos, Result);
10092 // Make sure that we reprocess all operands now that we reduced their
10094 AddUsesToWorkList(*I);
10096 // Instructions can end up on the worklist more than once. Make sure
10097 // we do not process an instruction that has been deleted.
10098 RemoveFromWorkList(I);
10100 // Erase the old instruction.
10101 InstParent->getInstList().erase(I);
10104 DOUT << "IC: Mod = " << OrigI
10105 << " New = " << *I;
10108 // If the instruction was modified, it's possible that it is now dead.
10109 // if so, remove it.
10110 if (isInstructionTriviallyDead(I)) {
10111 // Make sure we process all operands now that we are reducing their
10113 AddUsesToWorkList(*I);
10115 // Instructions may end up in the worklist more than once. Erase all
10116 // occurrences of this instruction.
10117 RemoveFromWorkList(I);
10118 I->eraseFromParent();
10121 AddUsersToWorkList(*I);
10128 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10130 // Do an explicit clear, this shrinks the map if needed.
10131 WorklistMap.clear();
10136 bool InstCombiner::runOnFunction(Function &F) {
10137 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10139 bool EverMadeChange = false;
10141 // Iterate while there is work to do.
10142 unsigned Iteration = 0;
10143 while (DoOneIteration(F, Iteration++))
10144 EverMadeChange = true;
10145 return EverMadeChange;
10148 FunctionPass *llvm::createInstructionCombiningPass() {
10149 return new InstCombiner();