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,
193 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
194 ICmpInst::Predicate Cond, Instruction &I);
195 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
197 Instruction *commonCastTransforms(CastInst &CI);
198 Instruction *commonIntCastTransforms(CastInst &CI);
199 Instruction *commonPointerCastTransforms(CastInst &CI);
200 Instruction *visitTrunc(TruncInst &CI);
201 Instruction *visitZExt(ZExtInst &CI);
202 Instruction *visitSExt(SExtInst &CI);
203 Instruction *visitFPTrunc(CastInst &CI);
204 Instruction *visitFPExt(CastInst &CI);
205 Instruction *visitFPToUI(CastInst &CI);
206 Instruction *visitFPToSI(CastInst &CI);
207 Instruction *visitUIToFP(CastInst &CI);
208 Instruction *visitSIToFP(CastInst &CI);
209 Instruction *visitPtrToInt(CastInst &CI);
210 Instruction *visitIntToPtr(CastInst &CI);
211 Instruction *visitBitCast(BitCastInst &CI);
212 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
214 Instruction *visitSelectInst(SelectInst &CI);
215 Instruction *visitCallInst(CallInst &CI);
216 Instruction *visitInvokeInst(InvokeInst &II);
217 Instruction *visitPHINode(PHINode &PN);
218 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
219 Instruction *visitAllocationInst(AllocationInst &AI);
220 Instruction *visitFreeInst(FreeInst &FI);
221 Instruction *visitLoadInst(LoadInst &LI);
222 Instruction *visitStoreInst(StoreInst &SI);
223 Instruction *visitBranchInst(BranchInst &BI);
224 Instruction *visitSwitchInst(SwitchInst &SI);
225 Instruction *visitInsertElementInst(InsertElementInst &IE);
226 Instruction *visitExtractElementInst(ExtractElementInst &EI);
227 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
229 // visitInstruction - Specify what to return for unhandled instructions...
230 Instruction *visitInstruction(Instruction &I) { return 0; }
233 Instruction *visitCallSite(CallSite CS);
234 bool transformConstExprCastCall(CallSite CS);
237 // InsertNewInstBefore - insert an instruction New before instruction Old
238 // in the program. Add the new instruction to the worklist.
240 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
241 assert(New && New->getParent() == 0 &&
242 "New instruction already inserted into a basic block!");
243 BasicBlock *BB = Old.getParent();
244 BB->getInstList().insert(&Old, New); // Insert inst
249 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
250 /// This also adds the cast to the worklist. Finally, this returns the
252 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
254 if (V->getType() == Ty) return V;
256 if (Constant *CV = dyn_cast<Constant>(V))
257 return ConstantExpr::getCast(opc, CV, Ty);
259 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
264 // ReplaceInstUsesWith - This method is to be used when an instruction is
265 // found to be dead, replacable with another preexisting expression. Here
266 // we add all uses of I to the worklist, replace all uses of I with the new
267 // value, then return I, so that the inst combiner will know that I was
270 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
271 AddUsersToWorkList(I); // Add all modified instrs to worklist
273 I.replaceAllUsesWith(V);
276 // If we are replacing the instruction with itself, this must be in a
277 // segment of unreachable code, so just clobber the instruction.
278 I.replaceAllUsesWith(UndefValue::get(I.getType()));
283 // UpdateValueUsesWith - This method is to be used when an value is
284 // found to be replacable with another preexisting expression or was
285 // updated. Here we add all uses of I to the worklist, replace all uses of
286 // I with the new value (unless the instruction was just updated), then
287 // return true, so that the inst combiner will know that I was modified.
289 bool UpdateValueUsesWith(Value *Old, Value *New) {
290 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
292 Old->replaceAllUsesWith(New);
293 if (Instruction *I = dyn_cast<Instruction>(Old))
295 if (Instruction *I = dyn_cast<Instruction>(New))
300 // EraseInstFromFunction - When dealing with an instruction that has side
301 // effects or produces a void value, we can't rely on DCE to delete the
302 // instruction. Instead, visit methods should return the value returned by
304 Instruction *EraseInstFromFunction(Instruction &I) {
305 assert(I.use_empty() && "Cannot erase instruction that is used!");
306 AddUsesToWorkList(I);
307 RemoveFromWorkList(&I);
309 return 0; // Don't do anything with FI
313 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
314 /// InsertBefore instruction. This is specialized a bit to avoid inserting
315 /// casts that are known to not do anything...
317 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
318 Value *V, const Type *DestTy,
319 Instruction *InsertBefore);
321 /// SimplifyCommutative - This performs a few simplifications for
322 /// commutative operators.
323 bool SimplifyCommutative(BinaryOperator &I);
325 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
326 /// most-complex to least-complex order.
327 bool SimplifyCompare(CmpInst &I);
329 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
330 /// on the demanded bits.
331 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
332 APInt& KnownZero, APInt& KnownOne,
335 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
336 uint64_t &UndefElts, unsigned Depth = 0);
338 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
339 // PHI node as operand #0, see if we can fold the instruction into the PHI
340 // (which is only possible if all operands to the PHI are constants).
341 Instruction *FoldOpIntoPhi(Instruction &I);
343 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
344 // operator and they all are only used by the PHI, PHI together their
345 // inputs, and do the operation once, to the result of the PHI.
346 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
347 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
350 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
351 ConstantInt *AndRHS, BinaryOperator &TheAnd);
353 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
354 bool isSub, Instruction &I);
355 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
356 bool isSigned, bool Inside, Instruction &IB);
357 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
358 Instruction *MatchBSwap(BinaryOperator &I);
359 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
361 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
364 char InstCombiner::ID = 0;
365 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
368 // getComplexity: Assign a complexity or rank value to LLVM Values...
369 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
370 static unsigned getComplexity(Value *V) {
371 if (isa<Instruction>(V)) {
372 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
376 if (isa<Argument>(V)) return 3;
377 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
380 // isOnlyUse - Return true if this instruction will be deleted if we stop using
382 static bool isOnlyUse(Value *V) {
383 return V->hasOneUse() || isa<Constant>(V);
386 // getPromotedType - Return the specified type promoted as it would be to pass
387 // though a va_arg area...
388 static const Type *getPromotedType(const Type *Ty) {
389 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
390 if (ITy->getBitWidth() < 32)
391 return Type::Int32Ty;
392 } else if (Ty == Type::FloatTy)
393 return Type::DoubleTy;
397 /// getBitCastOperand - If the specified operand is a CastInst or a constant
398 /// expression bitcast, return the operand value, otherwise return null.
399 static Value *getBitCastOperand(Value *V) {
400 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
401 return I->getOperand(0);
402 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
403 if (CE->getOpcode() == Instruction::BitCast)
404 return CE->getOperand(0);
408 /// This function is a wrapper around CastInst::isEliminableCastPair. It
409 /// simply extracts arguments and returns what that function returns.
410 static Instruction::CastOps
411 isEliminableCastPair(
412 const CastInst *CI, ///< The first cast instruction
413 unsigned opcode, ///< The opcode of the second cast instruction
414 const Type *DstTy, ///< The target type for the second cast instruction
415 TargetData *TD ///< The target data for pointer size
418 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
419 const Type *MidTy = CI->getType(); // B from above
421 // Get the opcodes of the two Cast instructions
422 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
423 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
425 return Instruction::CastOps(
426 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
427 DstTy, TD->getIntPtrType()));
430 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
431 /// in any code being generated. It does not require codegen if V is simple
432 /// enough or if the cast can be folded into other casts.
433 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
434 const Type *Ty, TargetData *TD) {
435 if (V->getType() == Ty || isa<Constant>(V)) return false;
437 // If this is another cast that can be eliminated, it isn't codegen either.
438 if (const CastInst *CI = dyn_cast<CastInst>(V))
439 if (isEliminableCastPair(CI, opcode, Ty, TD))
444 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
445 /// InsertBefore instruction. This is specialized a bit to avoid inserting
446 /// casts that are known to not do anything...
448 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
449 Value *V, const Type *DestTy,
450 Instruction *InsertBefore) {
451 if (V->getType() == DestTy) return V;
452 if (Constant *C = dyn_cast<Constant>(V))
453 return ConstantExpr::getCast(opcode, C, DestTy);
455 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
458 // SimplifyCommutative - This performs a few simplifications for commutative
461 // 1. Order operands such that they are listed from right (least complex) to
462 // left (most complex). This puts constants before unary operators before
465 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
466 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
468 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
469 bool Changed = false;
470 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
471 Changed = !I.swapOperands();
473 if (!I.isAssociative()) return Changed;
474 Instruction::BinaryOps Opcode = I.getOpcode();
475 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
476 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
477 if (isa<Constant>(I.getOperand(1))) {
478 Constant *Folded = ConstantExpr::get(I.getOpcode(),
479 cast<Constant>(I.getOperand(1)),
480 cast<Constant>(Op->getOperand(1)));
481 I.setOperand(0, Op->getOperand(0));
482 I.setOperand(1, Folded);
484 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
485 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
486 isOnlyUse(Op) && isOnlyUse(Op1)) {
487 Constant *C1 = cast<Constant>(Op->getOperand(1));
488 Constant *C2 = cast<Constant>(Op1->getOperand(1));
490 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
491 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
492 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
496 I.setOperand(0, New);
497 I.setOperand(1, Folded);
504 /// SimplifyCompare - For a CmpInst this function just orders the operands
505 /// so that theyare listed from right (least complex) to left (most complex).
506 /// This puts constants before unary operators before binary operators.
507 bool InstCombiner::SimplifyCompare(CmpInst &I) {
508 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
511 // Compare instructions are not associative so there's nothing else we can do.
515 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
516 // if the LHS is a constant zero (which is the 'negate' form).
518 static inline Value *dyn_castNegVal(Value *V) {
519 if (BinaryOperator::isNeg(V))
520 return BinaryOperator::getNegArgument(V);
522 // Constants can be considered to be negated values if they can be folded.
523 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
524 return ConstantExpr::getNeg(C);
528 static inline Value *dyn_castNotVal(Value *V) {
529 if (BinaryOperator::isNot(V))
530 return BinaryOperator::getNotArgument(V);
532 // Constants can be considered to be not'ed values...
533 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
534 return ConstantInt::get(~C->getValue());
538 // dyn_castFoldableMul - If this value is a multiply that can be folded into
539 // other computations (because it has a constant operand), return the
540 // non-constant operand of the multiply, and set CST to point to the multiplier.
541 // Otherwise, return null.
543 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
544 if (V->hasOneUse() && V->getType()->isInteger())
545 if (Instruction *I = dyn_cast<Instruction>(V)) {
546 if (I->getOpcode() == Instruction::Mul)
547 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
548 return I->getOperand(0);
549 if (I->getOpcode() == Instruction::Shl)
550 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
551 // The multiplier is really 1 << CST.
552 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
553 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
554 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
555 return I->getOperand(0);
561 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
562 /// expression, return it.
563 static User *dyn_castGetElementPtr(Value *V) {
564 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
565 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
566 if (CE->getOpcode() == Instruction::GetElementPtr)
567 return cast<User>(V);
571 /// AddOne - Add one to a ConstantInt
572 static ConstantInt *AddOne(ConstantInt *C) {
573 APInt Val(C->getValue());
574 return ConstantInt::get(++Val);
576 /// SubOne - Subtract one from a ConstantInt
577 static ConstantInt *SubOne(ConstantInt *C) {
578 APInt Val(C->getValue());
579 return ConstantInt::get(--Val);
581 /// Add - Add two ConstantInts together
582 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
583 return ConstantInt::get(C1->getValue() + C2->getValue());
585 /// And - Bitwise AND two ConstantInts together
586 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
587 return ConstantInt::get(C1->getValue() & C2->getValue());
589 /// Subtract - Subtract one ConstantInt from another
590 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
591 return ConstantInt::get(C1->getValue() - C2->getValue());
593 /// Multiply - Multiply two ConstantInts together
594 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
595 return ConstantInt::get(C1->getValue() * C2->getValue());
598 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
599 /// known to be either zero or one and return them in the KnownZero/KnownOne
600 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
602 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
603 /// we cannot optimize based on the assumption that it is zero without changing
604 /// it to be an explicit zero. If we don't change it to zero, other code could
605 /// optimized based on the contradictory assumption that it is non-zero.
606 /// Because instcombine aggressively folds operations with undef args anyway,
607 /// this won't lose us code quality.
608 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
609 APInt& KnownOne, unsigned Depth = 0) {
610 assert(V && "No Value?");
611 assert(Depth <= 6 && "Limit Search Depth");
612 uint32_t BitWidth = Mask.getBitWidth();
613 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
614 KnownZero.getBitWidth() == BitWidth &&
615 KnownOne.getBitWidth() == BitWidth &&
616 "V, Mask, KnownOne and KnownZero should have same BitWidth");
617 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
618 // We know all of the bits for a constant!
619 KnownOne = CI->getValue() & Mask;
620 KnownZero = ~KnownOne & Mask;
624 if (Depth == 6 || Mask == 0)
625 return; // Limit search depth.
627 Instruction *I = dyn_cast<Instruction>(V);
630 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
631 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
633 switch (I->getOpcode()) {
634 case Instruction::And: {
635 // If either the LHS or the RHS are Zero, the result is zero.
636 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
637 APInt Mask2(Mask & ~KnownZero);
638 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
639 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
640 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
642 // Output known-1 bits are only known if set in both the LHS & RHS.
643 KnownOne &= KnownOne2;
644 // Output known-0 are known to be clear if zero in either the LHS | RHS.
645 KnownZero |= KnownZero2;
648 case Instruction::Or: {
649 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
650 APInt Mask2(Mask & ~KnownOne);
651 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
652 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
653 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
655 // Output known-0 bits are only known if clear in both the LHS & RHS.
656 KnownZero &= KnownZero2;
657 // Output known-1 are known to be set if set in either the LHS | RHS.
658 KnownOne |= KnownOne2;
661 case Instruction::Xor: {
662 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
663 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
664 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
665 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
667 // Output known-0 bits are known if clear or set in both the LHS & RHS.
668 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
669 // Output known-1 are known to be set if set in only one of the LHS, RHS.
670 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
671 KnownZero = KnownZeroOut;
674 case Instruction::Select:
675 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
676 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
677 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
678 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
680 // Only known if known in both the LHS and RHS.
681 KnownOne &= KnownOne2;
682 KnownZero &= KnownZero2;
684 case Instruction::FPTrunc:
685 case Instruction::FPExt:
686 case Instruction::FPToUI:
687 case Instruction::FPToSI:
688 case Instruction::SIToFP:
689 case Instruction::PtrToInt:
690 case Instruction::UIToFP:
691 case Instruction::IntToPtr:
692 return; // Can't work with floating point or pointers
693 case Instruction::Trunc: {
694 // All these have integer operands
695 uint32_t SrcBitWidth =
696 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
698 MaskIn.zext(SrcBitWidth);
699 KnownZero.zext(SrcBitWidth);
700 KnownOne.zext(SrcBitWidth);
701 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
702 KnownZero.trunc(BitWidth);
703 KnownOne.trunc(BitWidth);
706 case Instruction::BitCast: {
707 const Type *SrcTy = I->getOperand(0)->getType();
708 if (SrcTy->isInteger()) {
709 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
714 case Instruction::ZExt: {
715 // Compute the bits in the result that are not present in the input.
716 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
717 uint32_t SrcBitWidth = SrcTy->getBitWidth();
720 MaskIn.trunc(SrcBitWidth);
721 KnownZero.trunc(SrcBitWidth);
722 KnownOne.trunc(SrcBitWidth);
723 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
724 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
725 // The top bits are known to be zero.
726 KnownZero.zext(BitWidth);
727 KnownOne.zext(BitWidth);
728 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
731 case Instruction::SExt: {
732 // Compute the bits in the result that are not present in the input.
733 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
734 uint32_t SrcBitWidth = SrcTy->getBitWidth();
737 MaskIn.trunc(SrcBitWidth);
738 KnownZero.trunc(SrcBitWidth);
739 KnownOne.trunc(SrcBitWidth);
740 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
741 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
742 KnownZero.zext(BitWidth);
743 KnownOne.zext(BitWidth);
745 // If the sign bit of the input is known set or clear, then we know the
746 // top bits of the result.
747 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
748 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
749 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
750 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
753 case Instruction::Shl:
754 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
755 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
756 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
757 APInt Mask2(Mask.lshr(ShiftAmt));
758 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
759 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
760 KnownZero <<= ShiftAmt;
761 KnownOne <<= ShiftAmt;
762 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
766 case Instruction::LShr:
767 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
768 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
769 // Compute the new bits that are at the top now.
770 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
772 // Unsigned shift right.
773 APInt Mask2(Mask.shl(ShiftAmt));
774 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
775 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
776 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
777 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
778 // high bits known zero.
779 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
783 case Instruction::AShr:
784 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
785 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
786 // Compute the new bits that are at the top now.
787 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
789 // Signed shift right.
790 APInt Mask2(Mask.shl(ShiftAmt));
791 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
792 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
793 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
794 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
796 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
797 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
798 KnownZero |= HighBits;
799 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
800 KnownOne |= HighBits;
807 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
808 /// this predicate to simplify operations downstream. Mask is known to be zero
809 /// for bits that V cannot have.
810 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
811 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
812 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
813 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
814 return (KnownZero & Mask) == Mask;
817 /// ShrinkDemandedConstant - Check to see if the specified operand of the
818 /// specified instruction is a constant integer. If so, check to see if there
819 /// are any bits set in the constant that are not demanded. If so, shrink the
820 /// constant and return true.
821 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
823 assert(I && "No instruction?");
824 assert(OpNo < I->getNumOperands() && "Operand index too large");
826 // If the operand is not a constant integer, nothing to do.
827 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
828 if (!OpC) return false;
830 // If there are no bits set that aren't demanded, nothing to do.
831 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
832 if ((~Demanded & OpC->getValue()) == 0)
835 // This instruction is producing bits that are not demanded. Shrink the RHS.
836 Demanded &= OpC->getValue();
837 I->setOperand(OpNo, ConstantInt::get(Demanded));
841 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
842 // set of known zero and one bits, compute the maximum and minimum values that
843 // could have the specified known zero and known one bits, returning them in
845 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
846 const APInt& KnownZero,
847 const APInt& KnownOne,
848 APInt& Min, APInt& Max) {
849 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
850 assert(KnownZero.getBitWidth() == BitWidth &&
851 KnownOne.getBitWidth() == BitWidth &&
852 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
853 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
854 APInt UnknownBits = ~(KnownZero|KnownOne);
856 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
857 // bit if it is unknown.
859 Max = KnownOne|UnknownBits;
861 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
863 Max.clear(BitWidth-1);
867 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
868 // a set of known zero and one bits, compute the maximum and minimum values that
869 // could have the specified known zero and known one bits, returning them in
871 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
872 const APInt& KnownZero,
873 const APInt& KnownOne,
876 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
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 (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1346 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1348 // Signed shift right.
1349 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1350 if (SimplifyDemandedBits(I->getOperand(0),
1352 RHSKnownZero, RHSKnownOne, Depth+1))
1354 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1355 "Bits known to be one AND zero?");
1356 // Compute the new bits that are at the top now.
1357 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1358 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1359 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1361 // Handle the sign bits.
1362 APInt SignBit(APInt::getSignBit(BitWidth));
1363 // Adjust to where it is now in the mask.
1364 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1366 // If the input sign bit is known to be zero, or if none of the top bits
1367 // are demanded, turn this into an unsigned shift right.
1368 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1369 (HighBits & ~DemandedMask) == HighBits) {
1370 // Perform the logical shift right.
1371 Value *NewVal = BinaryOperator::createLShr(
1372 I->getOperand(0), SA, I->getName());
1373 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1374 return UpdateValueUsesWith(I, NewVal);
1375 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1376 RHSKnownOne |= HighBits;
1382 // If the client is only demanding bits that we know, return the known
1384 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1385 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1390 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1391 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1392 /// actually used by the caller. This method analyzes which elements of the
1393 /// operand are undef and returns that information in UndefElts.
1395 /// If the information about demanded elements can be used to simplify the
1396 /// operation, the operation is simplified, then the resultant value is
1397 /// returned. This returns null if no change was made.
1398 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1399 uint64_t &UndefElts,
1401 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1402 assert(VWidth <= 64 && "Vector too wide to analyze!");
1403 uint64_t EltMask = ~0ULL >> (64-VWidth);
1404 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1405 "Invalid DemandedElts!");
1407 if (isa<UndefValue>(V)) {
1408 // If the entire vector is undefined, just return this info.
1409 UndefElts = EltMask;
1411 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1412 UndefElts = EltMask;
1413 return UndefValue::get(V->getType());
1417 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1418 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1419 Constant *Undef = UndefValue::get(EltTy);
1421 std::vector<Constant*> Elts;
1422 for (unsigned i = 0; i != VWidth; ++i)
1423 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1424 Elts.push_back(Undef);
1425 UndefElts |= (1ULL << i);
1426 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1427 Elts.push_back(Undef);
1428 UndefElts |= (1ULL << i);
1429 } else { // Otherwise, defined.
1430 Elts.push_back(CP->getOperand(i));
1433 // If we changed the constant, return it.
1434 Constant *NewCP = ConstantVector::get(Elts);
1435 return NewCP != CP ? NewCP : 0;
1436 } else if (isa<ConstantAggregateZero>(V)) {
1437 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1439 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1440 Constant *Zero = Constant::getNullValue(EltTy);
1441 Constant *Undef = UndefValue::get(EltTy);
1442 std::vector<Constant*> Elts;
1443 for (unsigned i = 0; i != VWidth; ++i)
1444 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1445 UndefElts = DemandedElts ^ EltMask;
1446 return ConstantVector::get(Elts);
1449 if (!V->hasOneUse()) { // Other users may use these bits.
1450 if (Depth != 0) { // Not at the root.
1451 // TODO: Just compute the UndefElts information recursively.
1455 } else if (Depth == 10) { // Limit search depth.
1459 Instruction *I = dyn_cast<Instruction>(V);
1460 if (!I) return false; // Only analyze instructions.
1462 bool MadeChange = false;
1463 uint64_t UndefElts2;
1465 switch (I->getOpcode()) {
1468 case Instruction::InsertElement: {
1469 // If this is a variable index, we don't know which element it overwrites.
1470 // demand exactly the same input as we produce.
1471 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1473 // Note that we can't propagate undef elt info, because we don't know
1474 // which elt is getting updated.
1475 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1476 UndefElts2, Depth+1);
1477 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1481 // If this is inserting an element that isn't demanded, remove this
1483 unsigned IdxNo = Idx->getZExtValue();
1484 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1485 return AddSoonDeadInstToWorklist(*I, 0);
1487 // Otherwise, the element inserted overwrites whatever was there, so the
1488 // input demanded set is simpler than the output set.
1489 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1490 DemandedElts & ~(1ULL << IdxNo),
1491 UndefElts, Depth+1);
1492 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1494 // The inserted element is defined.
1495 UndefElts |= 1ULL << IdxNo;
1498 case Instruction::BitCast: {
1499 // Packed->packed casts only.
1500 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1502 unsigned InVWidth = VTy->getNumElements();
1503 uint64_t InputDemandedElts = 0;
1506 if (VWidth == InVWidth) {
1507 // If we are converting from <4x i32> -> <4 x f32>, we demand the same
1508 // elements as are demanded of us.
1510 InputDemandedElts = DemandedElts;
1511 } else if (VWidth > InVWidth) {
1515 // If there are more elements in the result than there are in the source,
1516 // then an input element is live if any of the corresponding output
1517 // elements are live.
1518 Ratio = VWidth/InVWidth;
1519 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1520 if (DemandedElts & (1ULL << OutIdx))
1521 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1527 // If there are more elements in the source than there are in the result,
1528 // then an input element is live if the corresponding output element is
1530 Ratio = InVWidth/VWidth;
1531 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1532 if (DemandedElts & (1ULL << InIdx/Ratio))
1533 InputDemandedElts |= 1ULL << InIdx;
1536 // div/rem demand all inputs, because they don't want divide by zero.
1537 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1538 UndefElts2, Depth+1);
1540 I->setOperand(0, TmpV);
1544 UndefElts = UndefElts2;
1545 if (VWidth > InVWidth) {
1546 assert(0 && "Unimp");
1547 // If there are more elements in the result than there are in the source,
1548 // then an output element is undef if the corresponding input element is
1550 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1551 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1552 UndefElts |= 1ULL << OutIdx;
1553 } else if (VWidth < InVWidth) {
1554 assert(0 && "Unimp");
1555 // If there are more elements in the source than there are in the result,
1556 // then a result element is undef if all of the corresponding input
1557 // elements are undef.
1558 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1559 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1560 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1561 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1565 case Instruction::And:
1566 case Instruction::Or:
1567 case Instruction::Xor:
1568 case Instruction::Add:
1569 case Instruction::Sub:
1570 case Instruction::Mul:
1571 // div/rem demand all inputs, because they don't want divide by zero.
1572 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1573 UndefElts, Depth+1);
1574 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1575 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1576 UndefElts2, Depth+1);
1577 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1579 // Output elements are undefined if both are undefined. Consider things
1580 // like undef&0. The result is known zero, not undef.
1581 UndefElts &= UndefElts2;
1584 case Instruction::Call: {
1585 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1587 switch (II->getIntrinsicID()) {
1590 // Binary vector operations that work column-wise. A dest element is a
1591 // function of the corresponding input elements from the two inputs.
1592 case Intrinsic::x86_sse_sub_ss:
1593 case Intrinsic::x86_sse_mul_ss:
1594 case Intrinsic::x86_sse_min_ss:
1595 case Intrinsic::x86_sse_max_ss:
1596 case Intrinsic::x86_sse2_sub_sd:
1597 case Intrinsic::x86_sse2_mul_sd:
1598 case Intrinsic::x86_sse2_min_sd:
1599 case Intrinsic::x86_sse2_max_sd:
1600 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1601 UndefElts, Depth+1);
1602 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1603 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1604 UndefElts2, Depth+1);
1605 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1607 // If only the low elt is demanded and this is a scalarizable intrinsic,
1608 // scalarize it now.
1609 if (DemandedElts == 1) {
1610 switch (II->getIntrinsicID()) {
1612 case Intrinsic::x86_sse_sub_ss:
1613 case Intrinsic::x86_sse_mul_ss:
1614 case Intrinsic::x86_sse2_sub_sd:
1615 case Intrinsic::x86_sse2_mul_sd:
1616 // TODO: Lower MIN/MAX/ABS/etc
1617 Value *LHS = II->getOperand(1);
1618 Value *RHS = II->getOperand(2);
1619 // Extract the element as scalars.
1620 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1621 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1623 switch (II->getIntrinsicID()) {
1624 default: assert(0 && "Case stmts out of sync!");
1625 case Intrinsic::x86_sse_sub_ss:
1626 case Intrinsic::x86_sse2_sub_sd:
1627 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1628 II->getName()), *II);
1630 case Intrinsic::x86_sse_mul_ss:
1631 case Intrinsic::x86_sse2_mul_sd:
1632 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1633 II->getName()), *II);
1638 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1640 InsertNewInstBefore(New, *II);
1641 AddSoonDeadInstToWorklist(*II, 0);
1646 // Output elements are undefined if both are undefined. Consider things
1647 // like undef&0. The result is known zero, not undef.
1648 UndefElts &= UndefElts2;
1654 return MadeChange ? I : 0;
1657 /// @returns true if the specified compare instruction is
1658 /// true when both operands are equal...
1659 /// @brief Determine if the ICmpInst returns true if both operands are equal
1660 static bool isTrueWhenEqual(ICmpInst &ICI) {
1661 ICmpInst::Predicate pred = ICI.getPredicate();
1662 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1663 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1664 pred == ICmpInst::ICMP_SLE;
1667 /// AssociativeOpt - Perform an optimization on an associative operator. This
1668 /// function is designed to check a chain of associative operators for a
1669 /// potential to apply a certain optimization. Since the optimization may be
1670 /// applicable if the expression was reassociated, this checks the chain, then
1671 /// reassociates the expression as necessary to expose the optimization
1672 /// opportunity. This makes use of a special Functor, which must define
1673 /// 'shouldApply' and 'apply' methods.
1675 template<typename Functor>
1676 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1677 unsigned Opcode = Root.getOpcode();
1678 Value *LHS = Root.getOperand(0);
1680 // Quick check, see if the immediate LHS matches...
1681 if (F.shouldApply(LHS))
1682 return F.apply(Root);
1684 // Otherwise, if the LHS is not of the same opcode as the root, return.
1685 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1686 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1687 // Should we apply this transform to the RHS?
1688 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1690 // If not to the RHS, check to see if we should apply to the LHS...
1691 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1692 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1696 // If the functor wants to apply the optimization to the RHS of LHSI,
1697 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1699 BasicBlock *BB = Root.getParent();
1701 // Now all of the instructions are in the current basic block, go ahead
1702 // and perform the reassociation.
1703 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1705 // First move the selected RHS to the LHS of the root...
1706 Root.setOperand(0, LHSI->getOperand(1));
1708 // Make what used to be the LHS of the root be the user of the root...
1709 Value *ExtraOperand = TmpLHSI->getOperand(1);
1710 if (&Root == TmpLHSI) {
1711 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1714 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1715 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1716 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1717 BasicBlock::iterator ARI = &Root; ++ARI;
1718 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1721 // Now propagate the ExtraOperand down the chain of instructions until we
1723 while (TmpLHSI != LHSI) {
1724 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1725 // Move the instruction to immediately before the chain we are
1726 // constructing to avoid breaking dominance properties.
1727 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1728 BB->getInstList().insert(ARI, NextLHSI);
1731 Value *NextOp = NextLHSI->getOperand(1);
1732 NextLHSI->setOperand(1, ExtraOperand);
1734 ExtraOperand = NextOp;
1737 // Now that the instructions are reassociated, have the functor perform
1738 // the transformation...
1739 return F.apply(Root);
1742 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1748 // AddRHS - Implements: X + X --> X << 1
1751 AddRHS(Value *rhs) : RHS(rhs) {}
1752 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1753 Instruction *apply(BinaryOperator &Add) const {
1754 return BinaryOperator::createShl(Add.getOperand(0),
1755 ConstantInt::get(Add.getType(), 1));
1759 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1761 struct AddMaskingAnd {
1763 AddMaskingAnd(Constant *c) : C2(c) {}
1764 bool shouldApply(Value *LHS) const {
1766 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1767 ConstantExpr::getAnd(C1, C2)->isNullValue();
1769 Instruction *apply(BinaryOperator &Add) const {
1770 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1774 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1776 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1777 if (Constant *SOC = dyn_cast<Constant>(SO))
1778 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1780 return IC->InsertNewInstBefore(CastInst::create(
1781 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1784 // Figure out if the constant is the left or the right argument.
1785 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1786 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1788 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1790 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1791 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1794 Value *Op0 = SO, *Op1 = ConstOperand;
1796 std::swap(Op0, Op1);
1798 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1799 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1800 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1801 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1802 SO->getName()+".cmp");
1804 assert(0 && "Unknown binary instruction type!");
1807 return IC->InsertNewInstBefore(New, I);
1810 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1811 // constant as the other operand, try to fold the binary operator into the
1812 // select arguments. This also works for Cast instructions, which obviously do
1813 // not have a second operand.
1814 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1816 // Don't modify shared select instructions
1817 if (!SI->hasOneUse()) return 0;
1818 Value *TV = SI->getOperand(1);
1819 Value *FV = SI->getOperand(2);
1821 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1822 // Bool selects with constant operands can be folded to logical ops.
1823 if (SI->getType() == Type::Int1Ty) return 0;
1825 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1826 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1828 return new SelectInst(SI->getCondition(), SelectTrueVal,
1835 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1836 /// node as operand #0, see if we can fold the instruction into the PHI (which
1837 /// is only possible if all operands to the PHI are constants).
1838 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1839 PHINode *PN = cast<PHINode>(I.getOperand(0));
1840 unsigned NumPHIValues = PN->getNumIncomingValues();
1841 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1843 // Check to see if all of the operands of the PHI are constants. If there is
1844 // one non-constant value, remember the BB it is. If there is more than one
1845 // or if *it* is a PHI, bail out.
1846 BasicBlock *NonConstBB = 0;
1847 for (unsigned i = 0; i != NumPHIValues; ++i)
1848 if (!isa<Constant>(PN->getIncomingValue(i))) {
1849 if (NonConstBB) return 0; // More than one non-const value.
1850 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1851 NonConstBB = PN->getIncomingBlock(i);
1853 // If the incoming non-constant value is in I's block, we have an infinite
1855 if (NonConstBB == I.getParent())
1859 // If there is exactly one non-constant value, we can insert a copy of the
1860 // operation in that block. However, if this is a critical edge, we would be
1861 // inserting the computation one some other paths (e.g. inside a loop). Only
1862 // do this if the pred block is unconditionally branching into the phi block.
1864 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1865 if (!BI || !BI->isUnconditional()) return 0;
1868 // Okay, we can do the transformation: create the new PHI node.
1869 PHINode *NewPN = new PHINode(I.getType(), "");
1870 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1871 InsertNewInstBefore(NewPN, *PN);
1872 NewPN->takeName(PN);
1874 // Next, add all of the operands to the PHI.
1875 if (I.getNumOperands() == 2) {
1876 Constant *C = cast<Constant>(I.getOperand(1));
1877 for (unsigned i = 0; i != NumPHIValues; ++i) {
1879 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1880 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1881 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1883 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1885 assert(PN->getIncomingBlock(i) == NonConstBB);
1886 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1887 InV = BinaryOperator::create(BO->getOpcode(),
1888 PN->getIncomingValue(i), C, "phitmp",
1889 NonConstBB->getTerminator());
1890 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1891 InV = CmpInst::create(CI->getOpcode(),
1893 PN->getIncomingValue(i), C, "phitmp",
1894 NonConstBB->getTerminator());
1896 assert(0 && "Unknown binop!");
1898 AddToWorkList(cast<Instruction>(InV));
1900 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1903 CastInst *CI = cast<CastInst>(&I);
1904 const Type *RetTy = CI->getType();
1905 for (unsigned i = 0; i != NumPHIValues; ++i) {
1907 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1908 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1910 assert(PN->getIncomingBlock(i) == NonConstBB);
1911 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1912 I.getType(), "phitmp",
1913 NonConstBB->getTerminator());
1914 AddToWorkList(cast<Instruction>(InV));
1916 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1919 return ReplaceInstUsesWith(I, NewPN);
1922 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1923 bool Changed = SimplifyCommutative(I);
1924 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1926 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1927 // X + undef -> undef
1928 if (isa<UndefValue>(RHS))
1929 return ReplaceInstUsesWith(I, RHS);
1932 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1933 if (RHSC->isNullValue())
1934 return ReplaceInstUsesWith(I, LHS);
1935 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1936 if (CFP->isExactlyValue(-0.0))
1937 return ReplaceInstUsesWith(I, LHS);
1940 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1941 // X + (signbit) --> X ^ signbit
1942 const APInt& Val = CI->getValue();
1943 uint32_t BitWidth = Val.getBitWidth();
1944 if (Val == APInt::getSignBit(BitWidth))
1945 return BinaryOperator::createXor(LHS, RHS);
1947 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1948 // (X & 254)+1 -> (X&254)|1
1949 if (!isa<VectorType>(I.getType())) {
1950 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1951 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1952 KnownZero, KnownOne))
1957 if (isa<PHINode>(LHS))
1958 if (Instruction *NV = FoldOpIntoPhi(I))
1961 ConstantInt *XorRHS = 0;
1963 if (isa<ConstantInt>(RHSC) &&
1964 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1965 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1966 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1968 uint32_t Size = TySizeBits / 2;
1969 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1970 APInt CFF80Val(-C0080Val);
1972 if (TySizeBits > Size) {
1973 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1974 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1975 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1976 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1977 // This is a sign extend if the top bits are known zero.
1978 if (!MaskedValueIsZero(XorLHS,
1979 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1980 Size = 0; // Not a sign ext, but can't be any others either.
1985 C0080Val = APIntOps::lshr(C0080Val, Size);
1986 CFF80Val = APIntOps::ashr(CFF80Val, Size);
1987 } while (Size >= 1);
1989 // FIXME: This shouldn't be necessary. When the backends can handle types
1990 // with funny bit widths then this whole cascade of if statements should
1991 // be removed. It is just here to get the size of the "middle" type back
1992 // up to something that the back ends can handle.
1993 const Type *MiddleType = 0;
1996 case 32: MiddleType = Type::Int32Ty; break;
1997 case 16: MiddleType = Type::Int16Ty; break;
1998 case 8: MiddleType = Type::Int8Ty; break;
2001 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2002 InsertNewInstBefore(NewTrunc, I);
2003 return new SExtInst(NewTrunc, I.getType(), I.getName());
2009 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2010 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2012 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2013 if (RHSI->getOpcode() == Instruction::Sub)
2014 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2015 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2017 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2018 if (LHSI->getOpcode() == Instruction::Sub)
2019 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2020 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2025 if (Value *V = dyn_castNegVal(LHS))
2026 return BinaryOperator::createSub(RHS, V);
2029 if (!isa<Constant>(RHS))
2030 if (Value *V = dyn_castNegVal(RHS))
2031 return BinaryOperator::createSub(LHS, V);
2035 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2036 if (X == RHS) // X*C + X --> X * (C+1)
2037 return BinaryOperator::createMul(RHS, AddOne(C2));
2039 // X*C1 + X*C2 --> X * (C1+C2)
2041 if (X == dyn_castFoldableMul(RHS, C1))
2042 return BinaryOperator::createMul(X, Add(C1, C2));
2045 // X + X*C --> X * (C+1)
2046 if (dyn_castFoldableMul(RHS, C2) == LHS)
2047 return BinaryOperator::createMul(LHS, AddOne(C2));
2049 // X + ~X --> -1 since ~X = -X-1
2050 if (dyn_castNotVal(LHS) == RHS ||
2051 dyn_castNotVal(RHS) == LHS)
2052 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
2055 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2056 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2057 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2060 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2062 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2063 return BinaryOperator::createSub(SubOne(CRHS), X);
2065 // (X & FF00) + xx00 -> (X+xx00) & FF00
2066 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2067 Constant *Anded = And(CRHS, C2);
2068 if (Anded == CRHS) {
2069 // See if all bits from the first bit set in the Add RHS up are included
2070 // in the mask. First, get the rightmost bit.
2071 const APInt& AddRHSV = CRHS->getValue();
2073 // Form a mask of all bits from the lowest bit added through the top.
2074 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2076 // See if the and mask includes all of these bits.
2077 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2079 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2080 // Okay, the xform is safe. Insert the new add pronto.
2081 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2082 LHS->getName()), I);
2083 return BinaryOperator::createAnd(NewAdd, C2);
2088 // Try to fold constant add into select arguments.
2089 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2090 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2094 // add (cast *A to intptrtype) B ->
2095 // cast (GEP (cast *A to sbyte*) B) ->
2098 CastInst *CI = dyn_cast<CastInst>(LHS);
2101 CI = dyn_cast<CastInst>(RHS);
2104 if (CI && CI->getType()->isSized() &&
2105 (CI->getType()->getPrimitiveSizeInBits() ==
2106 TD->getIntPtrType()->getPrimitiveSizeInBits())
2107 && isa<PointerType>(CI->getOperand(0)->getType())) {
2108 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2109 PointerType::get(Type::Int8Ty), I);
2110 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2111 return new PtrToIntInst(I2, CI->getType());
2115 return Changed ? &I : 0;
2118 // isSignBit - Return true if the value represented by the constant only has the
2119 // highest order bit set.
2120 static bool isSignBit(ConstantInt *CI) {
2121 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2122 return CI->getValue() == APInt::getSignBit(NumBits);
2125 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2126 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2128 if (Op0 == Op1) // sub X, X -> 0
2129 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2131 // If this is a 'B = x-(-A)', change to B = x+A...
2132 if (Value *V = dyn_castNegVal(Op1))
2133 return BinaryOperator::createAdd(Op0, V);
2135 if (isa<UndefValue>(Op0))
2136 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2137 if (isa<UndefValue>(Op1))
2138 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2140 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2141 // Replace (-1 - A) with (~A)...
2142 if (C->isAllOnesValue())
2143 return BinaryOperator::createNot(Op1);
2145 // C - ~X == X + (1+C)
2147 if (match(Op1, m_Not(m_Value(X))))
2148 return BinaryOperator::createAdd(X, AddOne(C));
2150 // -(X >>u 31) -> (X >>s 31)
2151 // -(X >>s 31) -> (X >>u 31)
2153 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2154 if (SI->getOpcode() == Instruction::LShr) {
2155 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2156 // Check to see if we are shifting out everything but the sign bit.
2157 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2158 SI->getType()->getPrimitiveSizeInBits()-1) {
2159 // Ok, the transformation is safe. Insert AShr.
2160 return BinaryOperator::create(Instruction::AShr,
2161 SI->getOperand(0), CU, SI->getName());
2165 else if (SI->getOpcode() == Instruction::AShr) {
2166 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2167 // Check to see if we are shifting out everything but the sign bit.
2168 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2169 SI->getType()->getPrimitiveSizeInBits()-1) {
2170 // Ok, the transformation is safe. Insert LShr.
2171 return BinaryOperator::createLShr(
2172 SI->getOperand(0), CU, SI->getName());
2178 // Try to fold constant sub into select arguments.
2179 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2180 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2183 if (isa<PHINode>(Op0))
2184 if (Instruction *NV = FoldOpIntoPhi(I))
2188 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2189 if (Op1I->getOpcode() == Instruction::Add &&
2190 !Op0->getType()->isFPOrFPVector()) {
2191 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2192 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2193 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2194 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2195 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2196 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2197 // C1-(X+C2) --> (C1-C2)-X
2198 return BinaryOperator::createSub(Subtract(CI1, CI2),
2199 Op1I->getOperand(0));
2203 if (Op1I->hasOneUse()) {
2204 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2205 // is not used by anyone else...
2207 if (Op1I->getOpcode() == Instruction::Sub &&
2208 !Op1I->getType()->isFPOrFPVector()) {
2209 // Swap the two operands of the subexpr...
2210 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2211 Op1I->setOperand(0, IIOp1);
2212 Op1I->setOperand(1, IIOp0);
2214 // Create the new top level add instruction...
2215 return BinaryOperator::createAdd(Op0, Op1);
2218 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2220 if (Op1I->getOpcode() == Instruction::And &&
2221 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2222 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2225 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2226 return BinaryOperator::createAnd(Op0, NewNot);
2229 // 0 - (X sdiv C) -> (X sdiv -C)
2230 if (Op1I->getOpcode() == Instruction::SDiv)
2231 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2233 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2234 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2235 ConstantExpr::getNeg(DivRHS));
2237 // X - X*C --> X * (1-C)
2238 ConstantInt *C2 = 0;
2239 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2240 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2241 return BinaryOperator::createMul(Op0, CP1);
2246 if (!Op0->getType()->isFPOrFPVector())
2247 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2248 if (Op0I->getOpcode() == Instruction::Add) {
2249 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2250 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2251 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2252 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2253 } else if (Op0I->getOpcode() == Instruction::Sub) {
2254 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2255 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2259 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2260 if (X == Op1) // X*C - X --> X * (C-1)
2261 return BinaryOperator::createMul(Op1, SubOne(C1));
2263 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2264 if (X == dyn_castFoldableMul(Op1, C2))
2265 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2270 /// isSignBitCheck - Given an exploded icmp instruction, return true if it
2271 /// really just returns true if the most significant (sign) bit is set.
2272 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS) {
2274 case ICmpInst::ICMP_SLT:
2275 // True if LHS s< RHS and RHS == 0
2276 return RHS->isZero();
2277 case ICmpInst::ICMP_SLE:
2278 // True if LHS s<= RHS and RHS == -1
2279 return RHS->isAllOnesValue();
2280 case ICmpInst::ICMP_UGE:
2281 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2282 return RHS->getValue() ==
2283 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2284 case ICmpInst::ICMP_UGT:
2285 // True if LHS u> RHS and RHS == high-bit-mask - 1
2286 return RHS->getValue() ==
2287 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2293 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2294 bool Changed = SimplifyCommutative(I);
2295 Value *Op0 = I.getOperand(0);
2297 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2298 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2300 // Simplify mul instructions with a constant RHS...
2301 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2302 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2304 // ((X << C1)*C2) == (X * (C2 << C1))
2305 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2306 if (SI->getOpcode() == Instruction::Shl)
2307 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2308 return BinaryOperator::createMul(SI->getOperand(0),
2309 ConstantExpr::getShl(CI, ShOp));
2312 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2313 if (CI->equalsInt(1)) // X * 1 == X
2314 return ReplaceInstUsesWith(I, Op0);
2315 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2316 return BinaryOperator::createNeg(Op0, I.getName());
2318 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2319 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2320 return BinaryOperator::createShl(Op0,
2321 ConstantInt::get(Op0->getType(), Val.logBase2()));
2323 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2324 if (Op1F->isNullValue())
2325 return ReplaceInstUsesWith(I, Op1);
2327 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2328 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2329 if (Op1F->getValue() == 1.0)
2330 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2333 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2334 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2335 isa<ConstantInt>(Op0I->getOperand(1))) {
2336 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2337 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2339 InsertNewInstBefore(Add, I);
2340 Value *C1C2 = ConstantExpr::getMul(Op1,
2341 cast<Constant>(Op0I->getOperand(1)));
2342 return BinaryOperator::createAdd(Add, C1C2);
2346 // Try to fold constant mul into select arguments.
2347 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2348 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2351 if (isa<PHINode>(Op0))
2352 if (Instruction *NV = FoldOpIntoPhi(I))
2356 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2357 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2358 return BinaryOperator::createMul(Op0v, Op1v);
2360 // If one of the operands of the multiply is a cast from a boolean value, then
2361 // we know the bool is either zero or one, so this is a 'masking' multiply.
2362 // See if we can simplify things based on how the boolean was originally
2364 CastInst *BoolCast = 0;
2365 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2366 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2369 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2370 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2373 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2374 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2375 const Type *SCOpTy = SCIOp0->getType();
2377 // If the icmp is true iff the sign bit of X is set, then convert this
2378 // multiply into a shift/and combination.
2379 if (isa<ConstantInt>(SCIOp1) &&
2380 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1))) {
2381 // Shift the X value right to turn it into "all signbits".
2382 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2383 SCOpTy->getPrimitiveSizeInBits()-1);
2385 InsertNewInstBefore(
2386 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2387 BoolCast->getOperand(0)->getName()+
2390 // If the multiply type is not the same as the source type, sign extend
2391 // or truncate to the multiply type.
2392 if (I.getType() != V->getType()) {
2393 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2394 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2395 Instruction::CastOps opcode =
2396 (SrcBits == DstBits ? Instruction::BitCast :
2397 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2398 V = InsertCastBefore(opcode, V, I.getType(), I);
2401 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2402 return BinaryOperator::createAnd(V, OtherOp);
2407 return Changed ? &I : 0;
2410 /// This function implements the transforms on div instructions that work
2411 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2412 /// used by the visitors to those instructions.
2413 /// @brief Transforms common to all three div instructions
2414 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2415 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2418 if (isa<UndefValue>(Op0))
2419 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2421 // X / undef -> undef
2422 if (isa<UndefValue>(Op1))
2423 return ReplaceInstUsesWith(I, Op1);
2425 // Handle cases involving: div X, (select Cond, Y, Z)
2426 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2427 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2428 // same basic block, then we replace the select with Y, and the condition
2429 // of the select with false (if the cond value is in the same BB). If the
2430 // select has uses other than the div, this allows them to be simplified
2431 // also. Note that div X, Y is just as good as div X, 0 (undef)
2432 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2433 if (ST->isNullValue()) {
2434 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2435 if (CondI && CondI->getParent() == I.getParent())
2436 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2437 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2438 I.setOperand(1, SI->getOperand(2));
2440 UpdateValueUsesWith(SI, SI->getOperand(2));
2444 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2445 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2446 if (ST->isNullValue()) {
2447 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2448 if (CondI && CondI->getParent() == I.getParent())
2449 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2450 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2451 I.setOperand(1, SI->getOperand(1));
2453 UpdateValueUsesWith(SI, SI->getOperand(1));
2461 /// This function implements the transforms common to both integer division
2462 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2463 /// division instructions.
2464 /// @brief Common integer divide transforms
2465 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2466 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2468 if (Instruction *Common = commonDivTransforms(I))
2471 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2473 if (RHS->equalsInt(1))
2474 return ReplaceInstUsesWith(I, Op0);
2476 // (X / C1) / C2 -> X / (C1*C2)
2477 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2478 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2479 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2480 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2481 Multiply(RHS, LHSRHS));
2484 if (!RHS->isZero()) { // avoid X udiv 0
2485 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2486 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2488 if (isa<PHINode>(Op0))
2489 if (Instruction *NV = FoldOpIntoPhi(I))
2494 // 0 / X == 0, we don't need to preserve faults!
2495 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2496 if (LHS->equalsInt(0))
2497 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2502 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2503 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2505 // Handle the integer div common cases
2506 if (Instruction *Common = commonIDivTransforms(I))
2509 // X udiv C^2 -> X >> C
2510 // Check to see if this is an unsigned division with an exact power of 2,
2511 // if so, convert to a right shift.
2512 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2513 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2514 return BinaryOperator::createLShr(Op0,
2515 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2518 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2519 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2520 if (RHSI->getOpcode() == Instruction::Shl &&
2521 isa<ConstantInt>(RHSI->getOperand(0))) {
2522 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2523 if (C1.isPowerOf2()) {
2524 Value *N = RHSI->getOperand(1);
2525 const Type *NTy = N->getType();
2526 if (uint32_t C2 = C1.logBase2()) {
2527 Constant *C2V = ConstantInt::get(NTy, C2);
2528 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2530 return BinaryOperator::createLShr(Op0, N);
2535 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2536 // where C1&C2 are powers of two.
2537 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2538 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2539 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2540 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2541 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2542 // Compute the shift amounts
2543 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2544 // Construct the "on true" case of the select
2545 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2546 Instruction *TSI = BinaryOperator::createLShr(
2547 Op0, TC, SI->getName()+".t");
2548 TSI = InsertNewInstBefore(TSI, I);
2550 // Construct the "on false" case of the select
2551 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2552 Instruction *FSI = BinaryOperator::createLShr(
2553 Op0, FC, SI->getName()+".f");
2554 FSI = InsertNewInstBefore(FSI, I);
2556 // construct the select instruction and return it.
2557 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2563 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2564 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2566 // Handle the integer div common cases
2567 if (Instruction *Common = commonIDivTransforms(I))
2570 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2572 if (RHS->isAllOnesValue())
2573 return BinaryOperator::createNeg(Op0);
2576 if (Value *LHSNeg = dyn_castNegVal(Op0))
2577 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2580 // If the sign bits of both operands are zero (i.e. we can prove they are
2581 // unsigned inputs), turn this into a udiv.
2582 if (I.getType()->isInteger()) {
2583 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2584 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2585 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2592 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2593 return commonDivTransforms(I);
2596 /// GetFactor - If we can prove that the specified value is at least a multiple
2597 /// of some factor, return that factor.
2598 static Constant *GetFactor(Value *V) {
2599 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2602 // Unless we can be tricky, we know this is a multiple of 1.
2603 Constant *Result = ConstantInt::get(V->getType(), 1);
2605 Instruction *I = dyn_cast<Instruction>(V);
2606 if (!I) return Result;
2608 if (I->getOpcode() == Instruction::Mul) {
2609 // Handle multiplies by a constant, etc.
2610 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2611 GetFactor(I->getOperand(1)));
2612 } else if (I->getOpcode() == Instruction::Shl) {
2613 // (X<<C) -> X * (1 << C)
2614 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2615 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2616 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2618 } else if (I->getOpcode() == Instruction::And) {
2619 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2620 // X & 0xFFF0 is known to be a multiple of 16.
2621 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2622 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2623 return ConstantExpr::getShl(Result,
2624 ConstantInt::get(Result->getType(), Zeros));
2626 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2627 // Only handle int->int casts.
2628 if (!CI->isIntegerCast())
2630 Value *Op = CI->getOperand(0);
2631 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2636 /// This function implements the transforms on rem instructions that work
2637 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2638 /// is used by the visitors to those instructions.
2639 /// @brief Transforms common to all three rem instructions
2640 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2641 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2643 // 0 % X == 0, we don't need to preserve faults!
2644 if (Constant *LHS = dyn_cast<Constant>(Op0))
2645 if (LHS->isNullValue())
2646 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2648 if (isa<UndefValue>(Op0)) // undef % X -> 0
2649 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2650 if (isa<UndefValue>(Op1))
2651 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2653 // Handle cases involving: rem X, (select Cond, Y, Z)
2654 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2655 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2656 // the same basic block, then we replace the select with Y, and the
2657 // condition of the select with false (if the cond value is in the same
2658 // BB). If the select has uses other than the div, this allows them to be
2660 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2661 if (ST->isNullValue()) {
2662 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2663 if (CondI && CondI->getParent() == I.getParent())
2664 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2665 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2666 I.setOperand(1, SI->getOperand(2));
2668 UpdateValueUsesWith(SI, SI->getOperand(2));
2671 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2672 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2673 if (ST->isNullValue()) {
2674 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2675 if (CondI && CondI->getParent() == I.getParent())
2676 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2677 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2678 I.setOperand(1, SI->getOperand(1));
2680 UpdateValueUsesWith(SI, SI->getOperand(1));
2688 /// This function implements the transforms common to both integer remainder
2689 /// instructions (urem and srem). It is called by the visitors to those integer
2690 /// remainder instructions.
2691 /// @brief Common integer remainder transforms
2692 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2693 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2695 if (Instruction *common = commonRemTransforms(I))
2698 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2699 // X % 0 == undef, we don't need to preserve faults!
2700 if (RHS->equalsInt(0))
2701 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2703 if (RHS->equalsInt(1)) // X % 1 == 0
2704 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2706 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2707 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2708 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2710 } else if (isa<PHINode>(Op0I)) {
2711 if (Instruction *NV = FoldOpIntoPhi(I))
2714 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2715 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2716 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2723 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2724 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2726 if (Instruction *common = commonIRemTransforms(I))
2729 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2730 // X urem C^2 -> X and C
2731 // Check to see if this is an unsigned remainder with an exact power of 2,
2732 // if so, convert to a bitwise and.
2733 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2734 if (C->getValue().isPowerOf2())
2735 return BinaryOperator::createAnd(Op0, SubOne(C));
2738 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2739 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2740 if (RHSI->getOpcode() == Instruction::Shl &&
2741 isa<ConstantInt>(RHSI->getOperand(0))) {
2742 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2743 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2744 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2746 return BinaryOperator::createAnd(Op0, Add);
2751 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2752 // where C1&C2 are powers of two.
2753 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2754 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2755 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2756 // STO == 0 and SFO == 0 handled above.
2757 if ((STO->getValue().isPowerOf2()) &&
2758 (SFO->getValue().isPowerOf2())) {
2759 Value *TrueAnd = InsertNewInstBefore(
2760 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2761 Value *FalseAnd = InsertNewInstBefore(
2762 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2763 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2771 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2772 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2774 if (Instruction *common = commonIRemTransforms(I))
2777 if (Value *RHSNeg = dyn_castNegVal(Op1))
2778 if (!isa<ConstantInt>(RHSNeg) ||
2779 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2781 AddUsesToWorkList(I);
2782 I.setOperand(1, RHSNeg);
2786 // If the top bits of both operands are zero (i.e. we can prove they are
2787 // unsigned inputs), turn this into a urem.
2788 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2789 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2790 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2791 return BinaryOperator::createURem(Op0, Op1, I.getName());
2797 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2798 return commonRemTransforms(I);
2801 // isMaxValueMinusOne - return true if this is Max-1
2802 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2803 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2805 // Calculate 0111111111..11111
2806 APInt Val(APInt::getSignedMaxValue(TypeBits));
2807 return C->getValue() == Val-1;
2809 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2812 // isMinValuePlusOne - return true if this is Min+1
2813 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2815 // Calculate 1111111111000000000000
2816 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2817 APInt Val(APInt::getSignedMinValue(TypeBits));
2818 return C->getValue() == Val+1;
2820 return C->getValue() == 1; // unsigned
2823 // isOneBitSet - Return true if there is exactly one bit set in the specified
2825 static bool isOneBitSet(const ConstantInt *CI) {
2826 return CI->getValue().isPowerOf2();
2829 // isHighOnes - Return true if the constant is of the form 1+0+.
2830 // This is the same as lowones(~X).
2831 static bool isHighOnes(const ConstantInt *CI) {
2832 return (~CI->getValue() + 1).isPowerOf2();
2835 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2836 /// are carefully arranged to allow folding of expressions such as:
2838 /// (A < B) | (A > B) --> (A != B)
2840 /// Note that this is only valid if the first and second predicates have the
2841 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2843 /// Three bits are used to represent the condition, as follows:
2848 /// <=> Value Definition
2849 /// 000 0 Always false
2856 /// 111 7 Always true
2858 static unsigned getICmpCode(const ICmpInst *ICI) {
2859 switch (ICI->getPredicate()) {
2861 case ICmpInst::ICMP_UGT: return 1; // 001
2862 case ICmpInst::ICMP_SGT: return 1; // 001
2863 case ICmpInst::ICMP_EQ: return 2; // 010
2864 case ICmpInst::ICMP_UGE: return 3; // 011
2865 case ICmpInst::ICMP_SGE: return 3; // 011
2866 case ICmpInst::ICMP_ULT: return 4; // 100
2867 case ICmpInst::ICMP_SLT: return 4; // 100
2868 case ICmpInst::ICMP_NE: return 5; // 101
2869 case ICmpInst::ICMP_ULE: return 6; // 110
2870 case ICmpInst::ICMP_SLE: return 6; // 110
2873 assert(0 && "Invalid ICmp predicate!");
2878 /// getICmpValue - This is the complement of getICmpCode, which turns an
2879 /// opcode and two operands into either a constant true or false, or a brand
2880 /// new /// ICmp instruction. The sign is passed in to determine which kind
2881 /// of predicate to use in new icmp instructions.
2882 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2884 default: assert(0 && "Illegal ICmp code!");
2885 case 0: return ConstantInt::getFalse();
2888 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2890 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2891 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2894 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2896 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2899 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2901 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2902 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2905 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2907 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2908 case 7: return ConstantInt::getTrue();
2912 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2913 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2914 (ICmpInst::isSignedPredicate(p1) &&
2915 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2916 (ICmpInst::isSignedPredicate(p2) &&
2917 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2921 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2922 struct FoldICmpLogical {
2925 ICmpInst::Predicate pred;
2926 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2927 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2928 pred(ICI->getPredicate()) {}
2929 bool shouldApply(Value *V) const {
2930 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2931 if (PredicatesFoldable(pred, ICI->getPredicate()))
2932 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2933 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2936 Instruction *apply(Instruction &Log) const {
2937 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2938 if (ICI->getOperand(0) != LHS) {
2939 assert(ICI->getOperand(1) == LHS);
2940 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2943 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2944 unsigned LHSCode = getICmpCode(ICI);
2945 unsigned RHSCode = getICmpCode(RHSICI);
2947 switch (Log.getOpcode()) {
2948 case Instruction::And: Code = LHSCode & RHSCode; break;
2949 case Instruction::Or: Code = LHSCode | RHSCode; break;
2950 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2951 default: assert(0 && "Illegal logical opcode!"); return 0;
2954 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2955 ICmpInst::isSignedPredicate(ICI->getPredicate());
2957 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2958 if (Instruction *I = dyn_cast<Instruction>(RV))
2960 // Otherwise, it's a constant boolean value...
2961 return IC.ReplaceInstUsesWith(Log, RV);
2964 } // end anonymous namespace
2966 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2967 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2968 // guaranteed to be a binary operator.
2969 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2971 ConstantInt *AndRHS,
2972 BinaryOperator &TheAnd) {
2973 Value *X = Op->getOperand(0);
2974 Constant *Together = 0;
2976 Together = And(AndRHS, OpRHS);
2978 switch (Op->getOpcode()) {
2979 case Instruction::Xor:
2980 if (Op->hasOneUse()) {
2981 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2982 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
2983 InsertNewInstBefore(And, TheAnd);
2985 return BinaryOperator::createXor(And, Together);
2988 case Instruction::Or:
2989 if (Together == AndRHS) // (X | C) & C --> C
2990 return ReplaceInstUsesWith(TheAnd, AndRHS);
2992 if (Op->hasOneUse() && Together != OpRHS) {
2993 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2994 Instruction *Or = BinaryOperator::createOr(X, Together);
2995 InsertNewInstBefore(Or, TheAnd);
2997 return BinaryOperator::createAnd(Or, AndRHS);
3000 case Instruction::Add:
3001 if (Op->hasOneUse()) {
3002 // Adding a one to a single bit bit-field should be turned into an XOR
3003 // of the bit. First thing to check is to see if this AND is with a
3004 // single bit constant.
3005 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3007 // If there is only one bit set...
3008 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3009 // Ok, at this point, we know that we are masking the result of the
3010 // ADD down to exactly one bit. If the constant we are adding has
3011 // no bits set below this bit, then we can eliminate the ADD.
3012 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3014 // Check to see if any bits below the one bit set in AndRHSV are set.
3015 if ((AddRHS & (AndRHSV-1)) == 0) {
3016 // If not, the only thing that can effect the output of the AND is
3017 // the bit specified by AndRHSV. If that bit is set, the effect of
3018 // the XOR is to toggle the bit. If it is clear, then the ADD has
3020 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3021 TheAnd.setOperand(0, X);
3024 // Pull the XOR out of the AND.
3025 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3026 InsertNewInstBefore(NewAnd, TheAnd);
3027 NewAnd->takeName(Op);
3028 return BinaryOperator::createXor(NewAnd, AndRHS);
3035 case Instruction::Shl: {
3036 // We know that the AND will not produce any of the bits shifted in, so if
3037 // the anded constant includes them, clear them now!
3039 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3040 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3041 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3042 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3044 if (CI->getValue() == ShlMask) {
3045 // Masking out bits that the shift already masks
3046 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3047 } else if (CI != AndRHS) { // Reducing bits set in and.
3048 TheAnd.setOperand(1, CI);
3053 case Instruction::LShr:
3055 // We know that the AND will not produce any of the bits shifted in, so if
3056 // the anded constant includes them, clear them now! This only applies to
3057 // unsigned shifts, because a signed shr may bring in set bits!
3059 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3060 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3061 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3062 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3064 if (CI->getValue() == ShrMask) {
3065 // Masking out bits that the shift already masks.
3066 return ReplaceInstUsesWith(TheAnd, Op);
3067 } else if (CI != AndRHS) {
3068 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3073 case Instruction::AShr:
3075 // See if this is shifting in some sign extension, then masking it out
3077 if (Op->hasOneUse()) {
3078 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3079 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3080 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3081 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3082 if (C == AndRHS) { // Masking out bits shifted in.
3083 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3084 // Make the argument unsigned.
3085 Value *ShVal = Op->getOperand(0);
3086 ShVal = InsertNewInstBefore(
3087 BinaryOperator::createLShr(ShVal, OpRHS,
3088 Op->getName()), TheAnd);
3089 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3098 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3099 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3100 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3101 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3102 /// insert new instructions.
3103 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3104 bool isSigned, bool Inside,
3106 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3107 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3108 "Lo is not <= Hi in range emission code!");
3111 if (Lo == Hi) // Trivially false.
3112 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3114 // V >= Min && V < Hi --> V < Hi
3115 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3116 ICmpInst::Predicate pred = (isSigned ?
3117 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3118 return new ICmpInst(pred, V, Hi);
3121 // Emit V-Lo <u Hi-Lo
3122 Constant *NegLo = ConstantExpr::getNeg(Lo);
3123 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3124 InsertNewInstBefore(Add, IB);
3125 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3126 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3129 if (Lo == Hi) // Trivially true.
3130 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3132 // V < Min || V >= Hi -> V > Hi-1
3133 Hi = SubOne(cast<ConstantInt>(Hi));
3134 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3135 ICmpInst::Predicate pred = (isSigned ?
3136 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3137 return new ICmpInst(pred, V, Hi);
3140 // Emit V-Lo >u Hi-1-Lo
3141 // Note that Hi has already had one subtracted from it, above.
3142 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3143 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3144 InsertNewInstBefore(Add, IB);
3145 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3146 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3149 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3150 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3151 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3152 // not, since all 1s are not contiguous.
3153 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3154 const APInt& V = Val->getValue();
3155 uint32_t BitWidth = Val->getType()->getBitWidth();
3156 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3158 // look for the first zero bit after the run of ones
3159 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3160 // look for the first non-zero bit
3161 ME = V.getActiveBits();
3165 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3166 /// where isSub determines whether the operator is a sub. If we can fold one of
3167 /// the following xforms:
3169 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3170 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3171 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3173 /// return (A +/- B).
3175 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3176 ConstantInt *Mask, bool isSub,
3178 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3179 if (!LHSI || LHSI->getNumOperands() != 2 ||
3180 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3182 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3184 switch (LHSI->getOpcode()) {
3186 case Instruction::And:
3187 if (And(N, Mask) == Mask) {
3188 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3189 if ((Mask->getValue().countLeadingZeros() +
3190 Mask->getValue().countPopulation()) ==
3191 Mask->getValue().getBitWidth())
3194 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3195 // part, we don't need any explicit masks to take them out of A. If that
3196 // is all N is, ignore it.
3197 uint32_t MB = 0, ME = 0;
3198 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3199 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3200 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3201 if (MaskedValueIsZero(RHS, Mask))
3206 case Instruction::Or:
3207 case Instruction::Xor:
3208 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3209 if ((Mask->getValue().countLeadingZeros() +
3210 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3211 && And(N, Mask)->isZero())
3218 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3220 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3221 return InsertNewInstBefore(New, I);
3224 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3225 bool Changed = SimplifyCommutative(I);
3226 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3228 if (isa<UndefValue>(Op1)) // X & undef -> 0
3229 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3233 return ReplaceInstUsesWith(I, Op1);
3235 // See if we can simplify any instructions used by the instruction whose sole
3236 // purpose is to compute bits we don't care about.
3237 if (!isa<VectorType>(I.getType())) {
3238 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3239 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3240 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3241 KnownZero, KnownOne))
3244 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3245 if (CP->isAllOnesValue())
3246 return ReplaceInstUsesWith(I, I.getOperand(0));
3250 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3251 const APInt& AndRHSMask = AndRHS->getValue();
3252 APInt NotAndRHS(~AndRHSMask);
3254 // Optimize a variety of ((val OP C1) & C2) combinations...
3255 if (isa<BinaryOperator>(Op0)) {
3256 Instruction *Op0I = cast<Instruction>(Op0);
3257 Value *Op0LHS = Op0I->getOperand(0);
3258 Value *Op0RHS = Op0I->getOperand(1);
3259 switch (Op0I->getOpcode()) {
3260 case Instruction::Xor:
3261 case Instruction::Or:
3262 // If the mask is only needed on one incoming arm, push it up.
3263 if (Op0I->hasOneUse()) {
3264 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3265 // Not masking anything out for the LHS, move to RHS.
3266 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3267 Op0RHS->getName()+".masked");
3268 InsertNewInstBefore(NewRHS, I);
3269 return BinaryOperator::create(
3270 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3272 if (!isa<Constant>(Op0RHS) &&
3273 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3274 // Not masking anything out for the RHS, move to LHS.
3275 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3276 Op0LHS->getName()+".masked");
3277 InsertNewInstBefore(NewLHS, I);
3278 return BinaryOperator::create(
3279 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3284 case Instruction::Add:
3285 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3286 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3287 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3288 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3289 return BinaryOperator::createAnd(V, AndRHS);
3290 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3291 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3294 case Instruction::Sub:
3295 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3296 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3297 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3298 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3299 return BinaryOperator::createAnd(V, AndRHS);
3303 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3304 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3306 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3307 // If this is an integer truncation or change from signed-to-unsigned, and
3308 // if the source is an and/or with immediate, transform it. This
3309 // frequently occurs for bitfield accesses.
3310 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3311 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3312 CastOp->getNumOperands() == 2)
3313 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3314 if (CastOp->getOpcode() == Instruction::And) {
3315 // Change: and (cast (and X, C1) to T), C2
3316 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3317 // This will fold the two constants together, which may allow
3318 // other simplifications.
3319 Instruction *NewCast = CastInst::createTruncOrBitCast(
3320 CastOp->getOperand(0), I.getType(),
3321 CastOp->getName()+".shrunk");
3322 NewCast = InsertNewInstBefore(NewCast, I);
3323 // trunc_or_bitcast(C1)&C2
3324 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3325 C3 = ConstantExpr::getAnd(C3, AndRHS);
3326 return BinaryOperator::createAnd(NewCast, C3);
3327 } else if (CastOp->getOpcode() == Instruction::Or) {
3328 // Change: and (cast (or X, C1) to T), C2
3329 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3330 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3331 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3332 return ReplaceInstUsesWith(I, AndRHS);
3337 // Try to fold constant and into select arguments.
3338 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3339 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3341 if (isa<PHINode>(Op0))
3342 if (Instruction *NV = FoldOpIntoPhi(I))
3346 Value *Op0NotVal = dyn_castNotVal(Op0);
3347 Value *Op1NotVal = dyn_castNotVal(Op1);
3349 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3350 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3352 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3353 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3354 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3355 I.getName()+".demorgan");
3356 InsertNewInstBefore(Or, I);
3357 return BinaryOperator::createNot(Or);
3361 Value *A = 0, *B = 0;
3362 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3363 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3364 return ReplaceInstUsesWith(I, Op1);
3365 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3366 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3367 return ReplaceInstUsesWith(I, Op0);
3369 if (Op0->hasOneUse() &&
3370 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3371 if (A == Op1) { // (A^B)&A -> A&(A^B)
3372 I.swapOperands(); // Simplify below
3373 std::swap(Op0, Op1);
3374 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3375 cast<BinaryOperator>(Op0)->swapOperands();
3376 I.swapOperands(); // Simplify below
3377 std::swap(Op0, Op1);
3380 if (Op1->hasOneUse() &&
3381 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3382 if (B == Op0) { // B&(A^B) -> B&(B^A)
3383 cast<BinaryOperator>(Op1)->swapOperands();
3386 if (A == Op0) { // A&(A^B) -> A & ~B
3387 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3388 InsertNewInstBefore(NotB, I);
3389 return BinaryOperator::createAnd(A, NotB);
3394 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3395 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3396 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3399 Value *LHSVal, *RHSVal;
3400 ConstantInt *LHSCst, *RHSCst;
3401 ICmpInst::Predicate LHSCC, RHSCC;
3402 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3403 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3404 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3405 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3406 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3407 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3408 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3409 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3410 // Ensure that the larger constant is on the RHS.
3411 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3412 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3413 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3414 ICmpInst *LHS = cast<ICmpInst>(Op0);
3415 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3416 std::swap(LHS, RHS);
3417 std::swap(LHSCst, RHSCst);
3418 std::swap(LHSCC, RHSCC);
3421 // At this point, we know we have have two icmp instructions
3422 // comparing a value against two constants and and'ing the result
3423 // together. Because of the above check, we know that we only have
3424 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3425 // (from the FoldICmpLogical check above), that the two constants
3426 // are not equal and that the larger constant is on the RHS
3427 assert(LHSCst != RHSCst && "Compares not folded above?");
3430 default: assert(0 && "Unknown integer condition code!");
3431 case ICmpInst::ICMP_EQ:
3433 default: assert(0 && "Unknown integer condition code!");
3434 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3435 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3436 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3437 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3438 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3439 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3440 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3441 return ReplaceInstUsesWith(I, LHS);
3443 case ICmpInst::ICMP_NE:
3445 default: assert(0 && "Unknown integer condition code!");
3446 case ICmpInst::ICMP_ULT:
3447 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3448 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3449 break; // (X != 13 & X u< 15) -> no change
3450 case ICmpInst::ICMP_SLT:
3451 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3452 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3453 break; // (X != 13 & X s< 15) -> no change
3454 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3455 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3456 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3457 return ReplaceInstUsesWith(I, RHS);
3458 case ICmpInst::ICMP_NE:
3459 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3460 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3461 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3462 LHSVal->getName()+".off");
3463 InsertNewInstBefore(Add, I);
3464 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3465 ConstantInt::get(Add->getType(), 1));
3467 break; // (X != 13 & X != 15) -> no change
3470 case ICmpInst::ICMP_ULT:
3472 default: assert(0 && "Unknown integer condition code!");
3473 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3474 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3475 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3476 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3478 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3479 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3480 return ReplaceInstUsesWith(I, LHS);
3481 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3485 case ICmpInst::ICMP_SLT:
3487 default: assert(0 && "Unknown integer condition code!");
3488 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3489 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3490 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3491 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3493 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3494 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3495 return ReplaceInstUsesWith(I, LHS);
3496 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3500 case ICmpInst::ICMP_UGT:
3502 default: assert(0 && "Unknown integer condition code!");
3503 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3504 return ReplaceInstUsesWith(I, LHS);
3505 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3506 return ReplaceInstUsesWith(I, RHS);
3507 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3509 case ICmpInst::ICMP_NE:
3510 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3511 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3512 break; // (X u> 13 & X != 15) -> no change
3513 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3514 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3516 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3520 case ICmpInst::ICMP_SGT:
3522 default: assert(0 && "Unknown integer condition code!");
3523 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3524 return ReplaceInstUsesWith(I, LHS);
3525 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3526 return ReplaceInstUsesWith(I, RHS);
3527 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3529 case ICmpInst::ICMP_NE:
3530 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3531 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3532 break; // (X s> 13 & X != 15) -> no change
3533 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3534 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3536 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3544 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3545 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3546 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3547 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3548 const Type *SrcTy = Op0C->getOperand(0)->getType();
3549 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3550 // Only do this if the casts both really cause code to be generated.
3551 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3553 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3555 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3556 Op1C->getOperand(0),
3558 InsertNewInstBefore(NewOp, I);
3559 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3563 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3564 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3565 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3566 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3567 SI0->getOperand(1) == SI1->getOperand(1) &&
3568 (SI0->hasOneUse() || SI1->hasOneUse())) {
3569 Instruction *NewOp =
3570 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3572 SI0->getName()), I);
3573 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3574 SI1->getOperand(1));
3578 return Changed ? &I : 0;
3581 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3582 /// in the result. If it does, and if the specified byte hasn't been filled in
3583 /// yet, fill it in and return false.
3584 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3585 Instruction *I = dyn_cast<Instruction>(V);
3586 if (I == 0) return true;
3588 // If this is an or instruction, it is an inner node of the bswap.
3589 if (I->getOpcode() == Instruction::Or)
3590 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3591 CollectBSwapParts(I->getOperand(1), ByteValues);
3593 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3594 // If this is a shift by a constant int, and it is "24", then its operand
3595 // defines a byte. We only handle unsigned types here.
3596 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3597 // Not shifting the entire input by N-1 bytes?
3598 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3599 8*(ByteValues.size()-1))
3603 if (I->getOpcode() == Instruction::Shl) {
3604 // X << 24 defines the top byte with the lowest of the input bytes.
3605 DestNo = ByteValues.size()-1;
3607 // X >>u 24 defines the low byte with the highest of the input bytes.
3611 // If the destination byte value is already defined, the values are or'd
3612 // together, which isn't a bswap (unless it's an or of the same bits).
3613 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3615 ByteValues[DestNo] = I->getOperand(0);
3619 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3621 Value *Shift = 0, *ShiftLHS = 0;
3622 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3623 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3624 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3626 Instruction *SI = cast<Instruction>(Shift);
3628 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3629 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3630 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3633 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3635 if (AndAmt->getValue().getActiveBits() > 64)
3637 uint64_t AndAmtVal = AndAmt->getZExtValue();
3638 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3639 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3641 // Unknown mask for bswap.
3642 if (DestByte == ByteValues.size()) return true;
3644 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3646 if (SI->getOpcode() == Instruction::Shl)
3647 SrcByte = DestByte - ShiftBytes;
3649 SrcByte = DestByte + ShiftBytes;
3651 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3652 if (SrcByte != ByteValues.size()-DestByte-1)
3655 // If the destination byte value is already defined, the values are or'd
3656 // together, which isn't a bswap (unless it's an or of the same bits).
3657 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3659 ByteValues[DestByte] = SI->getOperand(0);
3663 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3664 /// If so, insert the new bswap intrinsic and return it.
3665 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3666 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3667 if (!ITy || ITy->getBitWidth() % 16)
3668 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3670 /// ByteValues - For each byte of the result, we keep track of which value
3671 /// defines each byte.
3672 SmallVector<Value*, 8> ByteValues;
3673 ByteValues.resize(ITy->getBitWidth()/8);
3675 // Try to find all the pieces corresponding to the bswap.
3676 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3677 CollectBSwapParts(I.getOperand(1), ByteValues))
3680 // Check to see if all of the bytes come from the same value.
3681 Value *V = ByteValues[0];
3682 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3684 // Check to make sure that all of the bytes come from the same value.
3685 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3686 if (ByteValues[i] != V)
3688 const Type *Tys[] = { ITy, ITy };
3689 Module *M = I.getParent()->getParent()->getParent();
3690 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 2);
3691 return new CallInst(F, V);
3695 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3696 bool Changed = SimplifyCommutative(I);
3697 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3699 if (isa<UndefValue>(Op1)) // X | undef -> -1
3700 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
3704 return ReplaceInstUsesWith(I, Op0);
3706 // See if we can simplify any instructions used by the instruction whose sole
3707 // purpose is to compute bits we don't care about.
3708 if (!isa<VectorType>(I.getType())) {
3709 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3710 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3711 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3712 KnownZero, KnownOne))
3717 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3718 ConstantInt *C1 = 0; Value *X = 0;
3719 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3720 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3721 Instruction *Or = BinaryOperator::createOr(X, RHS);
3722 InsertNewInstBefore(Or, I);
3724 return BinaryOperator::createAnd(Or,
3725 ConstantInt::get(RHS->getValue() | C1->getValue()));
3728 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3729 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3730 Instruction *Or = BinaryOperator::createOr(X, RHS);
3731 InsertNewInstBefore(Or, I);
3733 return BinaryOperator::createXor(Or,
3734 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3737 // Try to fold constant and into select arguments.
3738 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3739 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3741 if (isa<PHINode>(Op0))
3742 if (Instruction *NV = FoldOpIntoPhi(I))
3746 Value *A = 0, *B = 0;
3747 ConstantInt *C1 = 0, *C2 = 0;
3749 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3750 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3751 return ReplaceInstUsesWith(I, Op1);
3752 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3753 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3754 return ReplaceInstUsesWith(I, Op0);
3756 // (A | B) | C and A | (B | C) -> bswap if possible.
3757 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3758 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3759 match(Op1, m_Or(m_Value(), m_Value())) ||
3760 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3761 match(Op1, m_Shift(m_Value(), m_Value())))) {
3762 if (Instruction *BSwap = MatchBSwap(I))
3766 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3767 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3768 MaskedValueIsZero(Op1, C1->getValue())) {
3769 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3770 InsertNewInstBefore(NOr, I);
3772 return BinaryOperator::createXor(NOr, C1);
3775 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3776 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3777 MaskedValueIsZero(Op0, C1->getValue())) {
3778 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3779 InsertNewInstBefore(NOr, I);
3781 return BinaryOperator::createXor(NOr, C1);
3786 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3787 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3788 Value *V1 = 0, *V2 = 0, *V3 = 0;
3789 C1 = dyn_cast<ConstantInt>(C);
3790 C2 = dyn_cast<ConstantInt>(D);
3791 if (C1 && C2) { // (A & C1)|(B & C2)
3792 // If we have: ((V + N) & C1) | (V & C2)
3793 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3794 // replace with V+N.
3795 if (C1->getValue() == ~C2->getValue()) {
3796 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3797 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3798 // Add commutes, try both ways.
3799 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3800 return ReplaceInstUsesWith(I, A);
3801 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3802 return ReplaceInstUsesWith(I, A);
3804 // Or commutes, try both ways.
3805 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3806 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3807 // Add commutes, try both ways.
3808 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3809 return ReplaceInstUsesWith(I, B);
3810 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3811 return ReplaceInstUsesWith(I, B);
3814 V1 = 0; V2 = 0; V3 = 0;
3817 // Check to see if we have any common things being and'ed. If so, find the
3818 // terms for V1 & (V2|V3).
3819 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3820 if (A == B) // (A & C)|(A & D) == A & (C|D)
3821 V1 = A, V2 = C, V3 = D;
3822 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3823 V1 = A, V2 = B, V3 = C;
3824 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3825 V1 = C, V2 = A, V3 = D;
3826 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3827 V1 = C, V2 = A, V3 = B;
3831 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3832 return BinaryOperator::createAnd(V1, Or);
3835 // (V1 & V3)|(V2 & ~V3) -> ((V1 ^ V2) & V3) ^ V2
3836 if (isOnlyUse(Op0) && isOnlyUse(Op1)) {
3837 // Try all combination of terms to find V3 and ~V3.
3838 if (A->hasOneUse() && match(A, m_Not(m_Value(V3)))) {
3844 if (B->hasOneUse() && match(B, m_Not(m_Value(V3)))) {
3850 if (C->hasOneUse() && match(C, m_Not(m_Value(V3)))) {
3856 if (D->hasOneUse() && match(D, m_Not(m_Value(V3)))) {
3863 A = InsertNewInstBefore(BinaryOperator::createXor(V1, V2, "tmp"), I);
3864 A = InsertNewInstBefore(BinaryOperator::createAnd(A, V3, "tmp"), I);
3865 return BinaryOperator::createXor(A, V2);
3871 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3872 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3873 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3874 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3875 SI0->getOperand(1) == SI1->getOperand(1) &&
3876 (SI0->hasOneUse() || SI1->hasOneUse())) {
3877 Instruction *NewOp =
3878 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3880 SI0->getName()), I);
3881 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3882 SI1->getOperand(1));
3886 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3887 if (A == Op1) // ~A | A == -1
3888 return ReplaceInstUsesWith(I,
3889 ConstantInt::getAllOnesValue(I.getType()));
3893 // Note, A is still live here!
3894 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3896 return ReplaceInstUsesWith(I,
3897 ConstantInt::getAllOnesValue(I.getType()));
3899 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3900 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3901 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3902 I.getName()+".demorgan"), I);
3903 return BinaryOperator::createNot(And);
3907 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3908 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3909 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3912 Value *LHSVal, *RHSVal;
3913 ConstantInt *LHSCst, *RHSCst;
3914 ICmpInst::Predicate LHSCC, RHSCC;
3915 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3916 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3917 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3918 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3919 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3920 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3921 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3922 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3923 // We can't fold (ugt x, C) | (sgt x, C2).
3924 PredicatesFoldable(LHSCC, RHSCC)) {
3925 // Ensure that the larger constant is on the RHS.
3926 ICmpInst *LHS = cast<ICmpInst>(Op0);
3928 if (ICmpInst::isSignedPredicate(LHSCC))
3929 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3931 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3934 std::swap(LHS, RHS);
3935 std::swap(LHSCst, RHSCst);
3936 std::swap(LHSCC, RHSCC);
3939 // At this point, we know we have have two icmp instructions
3940 // comparing a value against two constants and or'ing the result
3941 // together. Because of the above check, we know that we only have
3942 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3943 // FoldICmpLogical check above), that the two constants are not
3945 assert(LHSCst != RHSCst && "Compares not folded above?");
3948 default: assert(0 && "Unknown integer condition code!");
3949 case ICmpInst::ICMP_EQ:
3951 default: assert(0 && "Unknown integer condition code!");
3952 case ICmpInst::ICMP_EQ:
3953 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3954 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3955 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3956 LHSVal->getName()+".off");
3957 InsertNewInstBefore(Add, I);
3958 AddCST = Subtract(AddOne(RHSCst), LHSCst);
3959 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3961 break; // (X == 13 | X == 15) -> no change
3962 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3963 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3965 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3966 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3967 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3968 return ReplaceInstUsesWith(I, RHS);
3971 case ICmpInst::ICMP_NE:
3973 default: assert(0 && "Unknown integer condition code!");
3974 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3975 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3976 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3977 return ReplaceInstUsesWith(I, LHS);
3978 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3979 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3980 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3981 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3984 case ICmpInst::ICMP_ULT:
3986 default: assert(0 && "Unknown integer condition code!");
3987 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3989 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3990 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
3992 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
3994 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
3995 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
3996 return ReplaceInstUsesWith(I, RHS);
3997 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4001 case ICmpInst::ICMP_SLT:
4003 default: assert(0 && "Unknown integer condition code!");
4004 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4006 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4007 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4009 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4011 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4012 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4013 return ReplaceInstUsesWith(I, RHS);
4014 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4018 case ICmpInst::ICMP_UGT:
4020 default: assert(0 && "Unknown integer condition code!");
4021 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4022 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4023 return ReplaceInstUsesWith(I, LHS);
4024 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4026 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4027 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4028 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4029 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4033 case ICmpInst::ICMP_SGT:
4035 default: assert(0 && "Unknown integer condition code!");
4036 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4037 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4038 return ReplaceInstUsesWith(I, LHS);
4039 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4041 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4042 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4043 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4044 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4052 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4053 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4054 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4055 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4056 const Type *SrcTy = Op0C->getOperand(0)->getType();
4057 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4058 // Only do this if the casts both really cause code to be generated.
4059 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4061 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4063 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4064 Op1C->getOperand(0),
4066 InsertNewInstBefore(NewOp, I);
4067 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4072 return Changed ? &I : 0;
4075 // XorSelf - Implements: X ^ X --> 0
4078 XorSelf(Value *rhs) : RHS(rhs) {}
4079 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4080 Instruction *apply(BinaryOperator &Xor) const {
4086 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4087 bool Changed = SimplifyCommutative(I);
4088 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4090 if (isa<UndefValue>(Op1))
4091 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4093 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4094 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4095 assert(Result == &I && "AssociativeOpt didn't work?");
4096 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4099 // See if we can simplify any instructions used by the instruction whose sole
4100 // purpose is to compute bits we don't care about.
4101 if (!isa<VectorType>(I.getType())) {
4102 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4103 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4104 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4105 KnownZero, KnownOne))
4109 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4110 // xor (icmp A, B), true = not (icmp A, B) = !icmp A, B
4111 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4112 if (RHS == ConstantInt::getTrue() && ICI->hasOneUse())
4113 return new ICmpInst(ICI->getInversePredicate(),
4114 ICI->getOperand(0), ICI->getOperand(1));
4116 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4117 // ~(c-X) == X-c-1 == X+(-c-1)
4118 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4119 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4120 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4121 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4122 ConstantInt::get(I.getType(), 1));
4123 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4126 // ~(~X & Y) --> (X | ~Y)
4127 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
4128 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4129 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4131 BinaryOperator::createNot(Op0I->getOperand(1),
4132 Op0I->getOperand(1)->getName()+".not");
4133 InsertNewInstBefore(NotY, I);
4134 return BinaryOperator::createOr(Op0NotVal, NotY);
4138 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4139 if (Op0I->getOpcode() == Instruction::Add) {
4140 // ~(X-c) --> (-c-1)-X
4141 if (RHS->isAllOnesValue()) {
4142 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4143 return BinaryOperator::createSub(
4144 ConstantExpr::getSub(NegOp0CI,
4145 ConstantInt::get(I.getType(), 1)),
4146 Op0I->getOperand(0));
4147 } else if (RHS->getValue().isSignBit()) {
4148 // (X + C) ^ signbit -> (X + C + signbit)
4149 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4150 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4153 } else if (Op0I->getOpcode() == Instruction::Or) {
4154 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4155 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4156 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4157 // Anything in both C1 and C2 is known to be zero, remove it from
4159 Constant *CommonBits = And(Op0CI, RHS);
4160 NewRHS = ConstantExpr::getAnd(NewRHS,
4161 ConstantExpr::getNot(CommonBits));
4162 AddToWorkList(Op0I);
4163 I.setOperand(0, Op0I->getOperand(0));
4164 I.setOperand(1, NewRHS);
4170 // Try to fold constant and into select arguments.
4171 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4172 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4174 if (isa<PHINode>(Op0))
4175 if (Instruction *NV = FoldOpIntoPhi(I))
4179 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4181 return ReplaceInstUsesWith(I,
4182 ConstantInt::getAllOnesValue(I.getType()));
4184 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4186 return ReplaceInstUsesWith(I, ConstantInt::getAllOnesValue(I.getType()));
4189 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4192 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4193 if (A == Op0) { // B^(B|A) == (A|B)^B
4194 Op1I->swapOperands();
4196 std::swap(Op0, Op1);
4197 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4198 I.swapOperands(); // Simplified below.
4199 std::swap(Op0, Op1);
4201 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4202 if (Op0 == A) // A^(A^B) == B
4203 return ReplaceInstUsesWith(I, B);
4204 else if (Op0 == B) // A^(B^A) == B
4205 return ReplaceInstUsesWith(I, A);
4206 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4207 if (A == Op0) { // A^(A&B) -> A^(B&A)
4208 Op1I->swapOperands();
4211 if (B == Op0) { // A^(B&A) -> (B&A)^A
4212 I.swapOperands(); // Simplified below.
4213 std::swap(Op0, Op1);
4218 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4221 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4222 if (A == Op1) // (B|A)^B == (A|B)^B
4224 if (B == Op1) { // (A|B)^B == A & ~B
4226 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4227 return BinaryOperator::createAnd(A, NotB);
4229 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4230 if (Op1 == A) // (A^B)^A == B
4231 return ReplaceInstUsesWith(I, B);
4232 else if (Op1 == B) // (B^A)^A == B
4233 return ReplaceInstUsesWith(I, A);
4234 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4235 if (A == Op1) // (A&B)^A -> (B&A)^A
4237 if (B == Op1 && // (B&A)^A == ~B & A
4238 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4240 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4241 return BinaryOperator::createAnd(N, Op1);
4246 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4247 if (Op0I && Op1I && Op0I->isShift() &&
4248 Op0I->getOpcode() == Op1I->getOpcode() &&
4249 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4250 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4251 Instruction *NewOp =
4252 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4253 Op1I->getOperand(0),
4254 Op0I->getName()), I);
4255 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4256 Op1I->getOperand(1));
4260 Value *A, *B, *C, *D;
4261 // (A & B)^(A | B) -> A ^ B
4262 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4263 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4264 if ((A == C && B == D) || (A == D && B == C))
4265 return BinaryOperator::createXor(A, B);
4267 // (A | B)^(A & B) -> A ^ B
4268 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4269 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4270 if ((A == C && B == D) || (A == D && B == C))
4271 return BinaryOperator::createXor(A, B);
4275 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4276 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4277 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4278 // (X & Y)^(X & Y) -> (Y^Z) & X
4279 Value *X = 0, *Y = 0, *Z = 0;
4281 X = A, Y = B, Z = D;
4283 X = A, Y = B, Z = C;
4285 X = B, Y = A, Z = D;
4287 X = B, Y = A, Z = C;
4290 Instruction *NewOp =
4291 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4292 return BinaryOperator::createAnd(NewOp, X);
4297 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4298 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4299 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4302 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4303 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4304 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4305 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4306 const Type *SrcTy = Op0C->getOperand(0)->getType();
4307 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4308 // Only do this if the casts both really cause code to be generated.
4309 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4311 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4313 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4314 Op1C->getOperand(0),
4316 InsertNewInstBefore(NewOp, I);
4317 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4321 return Changed ? &I : 0;
4324 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4325 /// overflowed for this type.
4326 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4327 ConstantInt *In2, bool IsSigned = false) {
4328 Result = cast<ConstantInt>(Add(In1, In2));
4331 if (In2->getValue().isNegative())
4332 return Result->getValue().sgt(In1->getValue());
4334 return Result->getValue().slt(In1->getValue());
4336 return Result->getValue().ult(In1->getValue());
4339 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4340 /// code necessary to compute the offset from the base pointer (without adding
4341 /// in the base pointer). Return the result as a signed integer of intptr size.
4342 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4343 TargetData &TD = IC.getTargetData();
4344 gep_type_iterator GTI = gep_type_begin(GEP);
4345 const Type *IntPtrTy = TD.getIntPtrType();
4346 Value *Result = Constant::getNullValue(IntPtrTy);
4348 // Build a mask for high order bits.
4349 unsigned IntPtrWidth = TD.getPointerSize()*8;
4350 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4352 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4353 Value *Op = GEP->getOperand(i);
4354 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4355 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4356 if (OpC->isZero()) continue;
4358 // Handle a struct index, which adds its field offset to the pointer.
4359 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4360 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4362 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4363 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4365 Result = IC.InsertNewInstBefore(
4366 BinaryOperator::createAdd(Result,
4367 ConstantInt::get(IntPtrTy, Size),
4368 GEP->getName()+".offs"), I);
4372 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4373 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4374 Scale = ConstantExpr::getMul(OC, Scale);
4375 if (Constant *RC = dyn_cast<Constant>(Result))
4376 Result = ConstantExpr::getAdd(RC, Scale);
4378 // Emit an add instruction.
4379 Result = IC.InsertNewInstBefore(
4380 BinaryOperator::createAdd(Result, Scale,
4381 GEP->getName()+".offs"), I);
4385 // Convert to correct type.
4386 if (Op->getType() != IntPtrTy) {
4387 if (Constant *OpC = dyn_cast<Constant>(Op))
4388 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4390 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4391 Op->getName()+".c"), I);
4394 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4395 if (Constant *OpC = dyn_cast<Constant>(Op))
4396 Op = ConstantExpr::getMul(OpC, Scale);
4397 else // We'll let instcombine(mul) convert this to a shl if possible.
4398 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4399 GEP->getName()+".idx"), I);
4402 // Emit an add instruction.
4403 if (isa<Constant>(Op) && isa<Constant>(Result))
4404 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4405 cast<Constant>(Result));
4407 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4408 GEP->getName()+".offs"), I);
4413 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4414 /// else. At this point we know that the GEP is on the LHS of the comparison.
4415 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4416 ICmpInst::Predicate Cond,
4418 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4420 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4421 if (isa<PointerType>(CI->getOperand(0)->getType()))
4422 RHS = CI->getOperand(0);
4424 Value *PtrBase = GEPLHS->getOperand(0);
4425 if (PtrBase == RHS) {
4426 // As an optimization, we don't actually have to compute the actual value of
4427 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4428 // each index is zero or not.
4429 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4430 Instruction *InVal = 0;
4431 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4432 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4434 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4435 if (isa<UndefValue>(C)) // undef index -> undef.
4436 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4437 if (C->isNullValue())
4439 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4440 EmitIt = false; // This is indexing into a zero sized array?
4441 } else if (isa<ConstantInt>(C))
4442 return ReplaceInstUsesWith(I, // No comparison is needed here.
4443 ConstantInt::get(Type::Int1Ty,
4444 Cond == ICmpInst::ICMP_NE));
4449 new ICmpInst(Cond, GEPLHS->getOperand(i),
4450 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4454 InVal = InsertNewInstBefore(InVal, I);
4455 InsertNewInstBefore(Comp, I);
4456 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4457 InVal = BinaryOperator::createOr(InVal, Comp);
4458 else // True if all are equal
4459 InVal = BinaryOperator::createAnd(InVal, Comp);
4467 // No comparison is needed here, all indexes = 0
4468 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4469 Cond == ICmpInst::ICMP_EQ));
4472 // Only lower this if the icmp is the only user of the GEP or if we expect
4473 // the result to fold to a constant!
4474 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4475 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4476 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4477 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4478 Constant::getNullValue(Offset->getType()));
4480 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4481 // If the base pointers are different, but the indices are the same, just
4482 // compare the base pointer.
4483 if (PtrBase != GEPRHS->getOperand(0)) {
4484 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4485 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4486 GEPRHS->getOperand(0)->getType();
4488 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4489 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4490 IndicesTheSame = false;
4494 // If all indices are the same, just compare the base pointers.
4496 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4497 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4499 // Otherwise, the base pointers are different and the indices are
4500 // different, bail out.
4504 // If one of the GEPs has all zero indices, recurse.
4505 bool AllZeros = true;
4506 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4507 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4508 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4513 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4514 ICmpInst::getSwappedPredicate(Cond), I);
4516 // If the other GEP has all zero indices, recurse.
4518 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4519 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4520 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4525 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4527 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4528 // If the GEPs only differ by one index, compare it.
4529 unsigned NumDifferences = 0; // Keep track of # differences.
4530 unsigned DiffOperand = 0; // The operand that differs.
4531 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4532 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4533 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4534 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4535 // Irreconcilable differences.
4539 if (NumDifferences++) break;
4544 if (NumDifferences == 0) // SAME GEP?
4545 return ReplaceInstUsesWith(I, // No comparison is needed here.
4546 ConstantInt::get(Type::Int1Ty,
4547 Cond == ICmpInst::ICMP_EQ));
4548 else if (NumDifferences == 1) {
4549 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4550 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4551 // Make sure we do a signed comparison here.
4552 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4556 // Only lower this if the icmp is the only user of the GEP or if we expect
4557 // the result to fold to a constant!
4558 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4559 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4560 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4561 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4562 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4563 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4569 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4570 bool Changed = SimplifyCompare(I);
4571 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4573 // Fold trivial predicates.
4574 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4575 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4576 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4577 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4579 // Simplify 'fcmp pred X, X'
4581 switch (I.getPredicate()) {
4582 default: assert(0 && "Unknown predicate!");
4583 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4584 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4585 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4586 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4587 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4588 case FCmpInst::FCMP_OLT: // True if ordered and less than
4589 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4590 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4592 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4593 case FCmpInst::FCMP_ULT: // True if unordered or less than
4594 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4595 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4596 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4597 I.setPredicate(FCmpInst::FCMP_UNO);
4598 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4601 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4602 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4603 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4604 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4605 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4606 I.setPredicate(FCmpInst::FCMP_ORD);
4607 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4612 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4613 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4615 // Handle fcmp with constant RHS
4616 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4617 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4618 switch (LHSI->getOpcode()) {
4619 case Instruction::PHI:
4620 if (Instruction *NV = FoldOpIntoPhi(I))
4623 case Instruction::Select:
4624 // If either operand of the select is a constant, we can fold the
4625 // comparison into the select arms, which will cause one to be
4626 // constant folded and the select turned into a bitwise or.
4627 Value *Op1 = 0, *Op2 = 0;
4628 if (LHSI->hasOneUse()) {
4629 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4630 // Fold the known value into the constant operand.
4631 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4632 // Insert a new FCmp of the other select operand.
4633 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4634 LHSI->getOperand(2), RHSC,
4636 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4637 // Fold the known value into the constant operand.
4638 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4639 // Insert a new FCmp of the other select operand.
4640 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4641 LHSI->getOperand(1), RHSC,
4647 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4652 return Changed ? &I : 0;
4655 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4656 bool Changed = SimplifyCompare(I);
4657 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4658 const Type *Ty = Op0->getType();
4662 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4663 isTrueWhenEqual(I)));
4665 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4666 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4668 // icmp of GlobalValues can never equal each other as long as they aren't
4669 // external weak linkage type.
4670 if (GlobalValue *GV0 = dyn_cast<GlobalValue>(Op0))
4671 if (GlobalValue *GV1 = dyn_cast<GlobalValue>(Op1))
4672 if (!GV0->hasExternalWeakLinkage() || !GV1->hasExternalWeakLinkage())
4673 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4674 !isTrueWhenEqual(I)));
4676 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4677 // addresses never equal each other! We already know that Op0 != Op1.
4678 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4679 isa<ConstantPointerNull>(Op0)) &&
4680 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4681 isa<ConstantPointerNull>(Op1)))
4682 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4683 !isTrueWhenEqual(I)));
4685 // icmp's with boolean values can always be turned into bitwise operations
4686 if (Ty == Type::Int1Ty) {
4687 switch (I.getPredicate()) {
4688 default: assert(0 && "Invalid icmp instruction!");
4689 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4690 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4691 InsertNewInstBefore(Xor, I);
4692 return BinaryOperator::createNot(Xor);
4694 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4695 return BinaryOperator::createXor(Op0, Op1);
4697 case ICmpInst::ICMP_UGT:
4698 case ICmpInst::ICMP_SGT:
4699 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4701 case ICmpInst::ICMP_ULT:
4702 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4703 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4704 InsertNewInstBefore(Not, I);
4705 return BinaryOperator::createAnd(Not, Op1);
4707 case ICmpInst::ICMP_UGE:
4708 case ICmpInst::ICMP_SGE:
4709 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4711 case ICmpInst::ICMP_ULE:
4712 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4713 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4714 InsertNewInstBefore(Not, I);
4715 return BinaryOperator::createOr(Not, Op1);
4720 // See if we are doing a comparison between a constant and an instruction that
4721 // can be folded into the comparison.
4722 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4723 switch (I.getPredicate()) {
4725 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4726 if (CI->isMinValue(false))
4727 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4728 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4729 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4730 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4731 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4732 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4733 if (CI->isMinValue(true))
4734 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4735 ConstantInt::getAllOnesValue(Op0->getType()));
4739 case ICmpInst::ICMP_SLT:
4740 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4741 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4742 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4743 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4744 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4745 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4748 case ICmpInst::ICMP_UGT:
4749 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4750 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4751 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4752 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4753 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4754 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4756 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4757 if (CI->isMaxValue(true))
4758 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4759 ConstantInt::getNullValue(Op0->getType()));
4762 case ICmpInst::ICMP_SGT:
4763 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4764 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4765 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4766 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4767 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4768 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4771 case ICmpInst::ICMP_ULE:
4772 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4773 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4774 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4775 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4776 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4777 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4780 case ICmpInst::ICMP_SLE:
4781 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4782 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4783 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4784 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4785 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4786 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4789 case ICmpInst::ICMP_UGE:
4790 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4791 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4792 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4793 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4794 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4795 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4798 case ICmpInst::ICMP_SGE:
4799 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4800 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4801 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4802 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4803 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4804 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4808 // If we still have a icmp le or icmp ge instruction, turn it into the
4809 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4810 // already been handled above, this requires little checking.
4812 switch (I.getPredicate()) {
4814 case ICmpInst::ICMP_ULE:
4815 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4816 case ICmpInst::ICMP_SLE:
4817 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4818 case ICmpInst::ICMP_UGE:
4819 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4820 case ICmpInst::ICMP_SGE:
4821 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4824 // See if we can fold the comparison based on bits known to be zero or one
4826 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4827 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4828 if (SimplifyDemandedBits(Op0, APInt::getAllOnesValue(BitWidth),
4829 KnownZero, KnownOne, 0))
4832 // Given the known and unknown bits, compute a range that the LHS could be
4834 if ((KnownOne | KnownZero) != 0) {
4835 // Compute the Min, Max and RHS values based on the known bits. For the
4836 // EQ and NE we use unsigned values.
4837 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4838 const APInt& RHSVal = CI->getValue();
4839 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4840 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4843 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4846 switch (I.getPredicate()) { // LE/GE have been folded already.
4847 default: assert(0 && "Unknown icmp opcode!");
4848 case ICmpInst::ICMP_EQ:
4849 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4850 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4852 case ICmpInst::ICMP_NE:
4853 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4854 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4856 case ICmpInst::ICMP_ULT:
4857 if (Max.ult(RHSVal))
4858 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4859 if (Min.uge(RHSVal))
4860 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4862 case ICmpInst::ICMP_UGT:
4863 if (Min.ugt(RHSVal))
4864 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4865 if (Max.ule(RHSVal))
4866 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4868 case ICmpInst::ICMP_SLT:
4869 if (Max.slt(RHSVal))
4870 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4871 if (Min.sgt(RHSVal))
4872 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4874 case ICmpInst::ICMP_SGT:
4875 if (Min.sgt(RHSVal))
4876 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4877 if (Max.sle(RHSVal))
4878 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4883 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4884 // instruction, see if that instruction also has constants so that the
4885 // instruction can be folded into the icmp
4886 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4887 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4891 // Handle icmp with constant (but not simple integer constant) RHS
4892 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4893 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4894 switch (LHSI->getOpcode()) {
4895 case Instruction::GetElementPtr:
4896 if (RHSC->isNullValue()) {
4897 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4898 bool isAllZeros = true;
4899 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4900 if (!isa<Constant>(LHSI->getOperand(i)) ||
4901 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4906 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
4907 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4911 case Instruction::PHI:
4912 if (Instruction *NV = FoldOpIntoPhi(I))
4915 case Instruction::Select: {
4916 // If either operand of the select is a constant, we can fold the
4917 // comparison into the select arms, which will cause one to be
4918 // constant folded and the select turned into a bitwise or.
4919 Value *Op1 = 0, *Op2 = 0;
4920 if (LHSI->hasOneUse()) {
4921 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4922 // Fold the known value into the constant operand.
4923 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4924 // Insert a new ICmp of the other select operand.
4925 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4926 LHSI->getOperand(2), RHSC,
4928 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4929 // Fold the known value into the constant operand.
4930 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4931 // Insert a new ICmp of the other select operand.
4932 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4933 LHSI->getOperand(1), RHSC,
4939 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4942 case Instruction::Malloc:
4943 // If we have (malloc != null), and if the malloc has a single use, we
4944 // can assume it is successful and remove the malloc.
4945 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
4946 AddToWorkList(LHSI);
4947 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4948 !isTrueWhenEqual(I)));
4954 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4955 if (User *GEP = dyn_castGetElementPtr(Op0))
4956 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
4958 if (User *GEP = dyn_castGetElementPtr(Op1))
4959 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
4960 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
4963 // Test to see if the operands of the icmp are casted versions of other
4964 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4966 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4967 if (isa<PointerType>(Op0->getType()) &&
4968 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4969 // We keep moving the cast from the left operand over to the right
4970 // operand, where it can often be eliminated completely.
4971 Op0 = CI->getOperand(0);
4973 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
4974 // so eliminate it as well.
4975 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
4976 Op1 = CI2->getOperand(0);
4978 // If Op1 is a constant, we can fold the cast into the constant.
4979 if (Op0->getType() != Op1->getType())
4980 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4981 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4983 // Otherwise, cast the RHS right before the icmp
4984 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
4986 return new ICmpInst(I.getPredicate(), Op0, Op1);
4990 if (isa<CastInst>(Op0)) {
4991 // Handle the special case of: icmp (cast bool to X), <cst>
4992 // This comes up when you have code like
4995 // For generality, we handle any zero-extension of any operand comparison
4996 // with a constant or another cast from the same type.
4997 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
4998 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5002 if (I.isEquality()) {
5003 Value *A, *B, *C, *D;
5004 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5005 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5006 Value *OtherVal = A == Op1 ? B : A;
5007 return new ICmpInst(I.getPredicate(), OtherVal,
5008 Constant::getNullValue(A->getType()));
5011 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5012 // A^c1 == C^c2 --> A == C^(c1^c2)
5013 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5014 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5015 if (Op1->hasOneUse()) {
5016 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5017 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5018 return new ICmpInst(I.getPredicate(), A,
5019 InsertNewInstBefore(Xor, I));
5022 // A^B == A^D -> B == D
5023 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5024 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5025 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5026 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5030 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5031 (A == Op0 || B == Op0)) {
5032 // A == (A^B) -> B == 0
5033 Value *OtherVal = A == Op0 ? B : A;
5034 return new ICmpInst(I.getPredicate(), OtherVal,
5035 Constant::getNullValue(A->getType()));
5037 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5038 // (A-B) == A -> B == 0
5039 return new ICmpInst(I.getPredicate(), B,
5040 Constant::getNullValue(B->getType()));
5042 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5043 // A == (A-B) -> B == 0
5044 return new ICmpInst(I.getPredicate(), B,
5045 Constant::getNullValue(B->getType()));
5048 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5049 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5050 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5051 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5052 Value *X = 0, *Y = 0, *Z = 0;
5055 X = B; Y = D; Z = A;
5056 } else if (A == D) {
5057 X = B; Y = C; Z = A;
5058 } else if (B == C) {
5059 X = A; Y = D; Z = B;
5060 } else if (B == D) {
5061 X = A; Y = C; Z = B;
5064 if (X) { // Build (X^Y) & Z
5065 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5066 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5067 I.setOperand(0, Op1);
5068 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5073 return Changed ? &I : 0;
5076 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5078 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5081 const APInt &RHSV = RHS->getValue();
5083 switch (LHSI->getOpcode()) {
5084 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5085 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5086 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5088 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5089 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5090 Value *CompareVal = LHSI->getOperand(0);
5092 // If the sign bit of the XorCST is not set, there is no change to
5093 // the operation, just stop using the Xor.
5094 if (!XorCST->getValue().isNegative()) {
5095 ICI.setOperand(0, CompareVal);
5096 AddToWorkList(LHSI);
5100 // Was the old condition true if the operand is positive?
5101 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5103 // If so, the new one isn't.
5104 isTrueIfPositive ^= true;
5106 if (isTrueIfPositive)
5107 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5109 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5113 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5114 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5115 LHSI->getOperand(0)->hasOneUse()) {
5116 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5118 // If the LHS is an AND of a truncating cast, we can widen the
5119 // and/compare to be the input width without changing the value
5120 // produced, eliminating a cast.
5121 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5122 // We can do this transformation if either the AND constant does not
5123 // have its sign bit set or if it is an equality comparison.
5124 // Extending a relational comparison when we're checking the sign
5125 // bit would not work.
5126 if (Cast->hasOneUse() &&
5127 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5128 RHSV.isPositive())) {
5130 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5131 APInt NewCST = AndCST->getValue();
5132 NewCST.zext(BitWidth);
5134 NewCI.zext(BitWidth);
5135 Instruction *NewAnd =
5136 BinaryOperator::createAnd(Cast->getOperand(0),
5137 ConstantInt::get(NewCST),LHSI->getName());
5138 InsertNewInstBefore(NewAnd, ICI);
5139 return new ICmpInst(ICI.getPredicate(), NewAnd,
5140 ConstantInt::get(NewCI));
5144 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5145 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5146 // happens a LOT in code produced by the C front-end, for bitfield
5148 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5149 if (Shift && !Shift->isShift())
5153 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5154 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5155 const Type *AndTy = AndCST->getType(); // Type of the and.
5157 // We can fold this as long as we can't shift unknown bits
5158 // into the mask. This can only happen with signed shift
5159 // rights, as they sign-extend.
5161 bool CanFold = Shift->isLogicalShift();
5163 // To test for the bad case of the signed shr, see if any
5164 // of the bits shifted in could be tested after the mask.
5165 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5166 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5168 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5169 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5170 AndCST->getValue()) == 0)
5176 if (Shift->getOpcode() == Instruction::Shl)
5177 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5179 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5181 // Check to see if we are shifting out any of the bits being
5183 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5184 // If we shifted bits out, the fold is not going to work out.
5185 // As a special case, check to see if this means that the
5186 // result is always true or false now.
5187 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5188 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5189 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5190 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5192 ICI.setOperand(1, NewCst);
5193 Constant *NewAndCST;
5194 if (Shift->getOpcode() == Instruction::Shl)
5195 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5197 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5198 LHSI->setOperand(1, NewAndCST);
5199 LHSI->setOperand(0, Shift->getOperand(0));
5200 AddToWorkList(Shift); // Shift is dead.
5201 AddUsesToWorkList(ICI);
5207 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5208 // preferable because it allows the C<<Y expression to be hoisted out
5209 // of a loop if Y is invariant and X is not.
5210 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5211 ICI.isEquality() && !Shift->isArithmeticShift() &&
5212 isa<Instruction>(Shift->getOperand(0))) {
5215 if (Shift->getOpcode() == Instruction::LShr) {
5216 NS = BinaryOperator::createShl(AndCST,
5217 Shift->getOperand(1), "tmp");
5219 // Insert a logical shift.
5220 NS = BinaryOperator::createLShr(AndCST,
5221 Shift->getOperand(1), "tmp");
5223 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5225 // Compute X & (C << Y).
5226 Instruction *NewAnd =
5227 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5228 InsertNewInstBefore(NewAnd, ICI);
5230 ICI.setOperand(0, NewAnd);
5236 case Instruction::Shl: // (icmp pred (shl X, ShAmt), CI)
5237 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5238 if (ICI.isEquality()) {
5239 uint32_t TypeBits = RHSV.getBitWidth();
5241 // Check that the shift amount is in range. If not, don't perform
5242 // undefined shifts. When the shift is visited it will be
5244 if (ShAmt->uge(TypeBits))
5247 // If we are comparing against bits always shifted out, the
5248 // comparison cannot succeed.
5250 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5251 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5252 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5253 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5254 return ReplaceInstUsesWith(ICI, Cst);
5257 if (LHSI->hasOneUse()) {
5258 // Otherwise strength reduce the shift into an and.
5259 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5261 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5264 BinaryOperator::createAnd(LHSI->getOperand(0),
5265 Mask, LHSI->getName()+".mask");
5266 Value *And = InsertNewInstBefore(AndI, ICI);
5267 return new ICmpInst(ICI.getPredicate(), And,
5268 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5274 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5275 case Instruction::AShr:
5276 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5277 if (ICI.isEquality()) {
5278 // Check that the shift amount is in range. If not, don't perform
5279 // undefined shifts. When the shift is visited it will be
5281 uint32_t TypeBits = RHSV.getBitWidth();
5282 if (ShAmt->uge(TypeBits))
5284 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5286 // If we are comparing against bits always shifted out, the
5287 // comparison cannot succeed.
5288 APInt Comp = RHSV << ShAmtVal;
5289 if (LHSI->getOpcode() == Instruction::LShr)
5290 Comp = Comp.lshr(ShAmtVal);
5292 Comp = Comp.ashr(ShAmtVal);
5294 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5295 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5296 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5297 return ReplaceInstUsesWith(ICI, Cst);
5300 if (LHSI->hasOneUse() || RHSV == 0) {
5301 // Otherwise strength reduce the shift into an and.
5302 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5303 Constant *Mask = ConstantInt::get(Val);
5306 BinaryOperator::createAnd(LHSI->getOperand(0),
5307 Mask, LHSI->getName()+".mask");
5308 Value *And = InsertNewInstBefore(AndI, ICI);
5309 return new ICmpInst(ICI.getPredicate(), And,
5310 ConstantExpr::getShl(RHS, ShAmt));
5316 case Instruction::SDiv:
5317 case Instruction::UDiv:
5318 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5319 // Fold this div into the comparison, producing a range check.
5320 // Determine, based on the divide type, what the range is being
5321 // checked. If there is an overflow on the low or high side, remember
5322 // it, otherwise compute the range [low, hi) bounding the new value.
5323 // See: InsertRangeTest above for the kinds of replacements possible.
5324 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5325 // FIXME: If the operand types don't match the type of the divide
5326 // then don't attempt this transform. The code below doesn't have the
5327 // logic to deal with a signed divide and an unsigned compare (and
5328 // vice versa). This is because (x /s C1) <s C2 produces different
5329 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5330 // (x /u C1) <u C2. Simply casting the operands and result won't
5331 // work. :( The if statement below tests that condition and bails
5333 bool DivIsSigned = LHSI->getOpcode() == Instruction::SDiv;
5334 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5336 if (DivRHS->isZero())
5337 break; // Don't hack on div by zero
5339 // Initialize the variables that will indicate the nature of the
5341 bool LoOverflow = false, HiOverflow = false;
5342 ConstantInt *LoBound = 0, *HiBound = 0;
5344 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5345 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5346 // C2 (CI). By solving for X we can turn this into a range check
5347 // instead of computing a divide.
5348 ConstantInt *Prod = Multiply(RHS, DivRHS);
5350 // Determine if the product overflows by seeing if the product is
5351 // not equal to the divide. Make sure we do the same kind of divide
5352 // as in the LHS instruction that we're folding.
5353 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5354 ConstantExpr::getUDiv(Prod, DivRHS)) != RHS;
5356 // Get the ICmp opcode
5357 ICmpInst::Predicate predicate = ICI.getPredicate();
5359 if (!DivIsSigned) { // udiv
5361 LoOverflow = ProdOV;
5362 HiOverflow = ProdOV ||
5363 AddWithOverflow(HiBound, LoBound, DivRHS, false);
5364 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5365 if (RHSV == 0) { // (X / pos) op 0
5367 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5369 } else if (RHSV.isPositive()) { // (X / pos) op pos
5371 LoOverflow = ProdOV;
5372 HiOverflow = ProdOV ||
5373 AddWithOverflow(HiBound, Prod, DivRHS, true);
5374 } else { // (X / pos) op neg
5375 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5376 LoOverflow = AddWithOverflow(LoBound, Prod,
5377 cast<ConstantInt>(DivRHSH), true);
5378 HiBound = AddOne(Prod);
5379 HiOverflow = ProdOV;
5381 } else { // Divisor is < 0.
5382 if (RHSV == 0) { // (X / neg) op 0
5383 LoBound = AddOne(DivRHS);
5384 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5385 if (HiBound == DivRHS)
5386 LoBound = 0; // - INTMIN = INTMIN
5387 } else if (RHSV.isPositive()) { // (X / neg) op pos
5388 HiOverflow = LoOverflow = ProdOV;
5390 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS),
5392 HiBound = AddOne(Prod);
5393 } else { // (X / neg) op neg
5395 LoOverflow = HiOverflow = ProdOV;
5396 HiBound = Subtract(Prod, DivRHS);
5399 // Dividing by a negate swaps the condition.
5400 predicate = ICmpInst::getSwappedPredicate(predicate);
5404 Value *X = LHSI->getOperand(0);
5405 switch (predicate) {
5406 default: assert(0 && "Unhandled icmp opcode!");
5407 case ICmpInst::ICMP_EQ:
5408 if (LoOverflow && HiOverflow)
5409 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5410 else if (HiOverflow)
5411 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5412 ICmpInst::ICMP_UGE, X, LoBound);
5413 else if (LoOverflow)
5414 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5415 ICmpInst::ICMP_ULT, X, HiBound);
5417 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5419 case ICmpInst::ICMP_NE:
5420 if (LoOverflow && HiOverflow)
5421 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5422 else if (HiOverflow)
5423 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5424 ICmpInst::ICMP_ULT, X, LoBound);
5425 else if (LoOverflow)
5426 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5427 ICmpInst::ICMP_UGE, X, HiBound);
5429 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned,
5431 case ICmpInst::ICMP_ULT:
5432 case ICmpInst::ICMP_SLT:
5434 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5435 return new ICmpInst(predicate, X, LoBound);
5436 case ICmpInst::ICMP_UGT:
5437 case ICmpInst::ICMP_SGT:
5439 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5440 if (predicate == ICmpInst::ICMP_UGT)
5441 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5443 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5450 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5451 if (ICI.isEquality()) {
5452 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5454 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5455 // the second operand is a constant, simplify a bit.
5456 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5457 switch (BO->getOpcode()) {
5458 case Instruction::SRem:
5459 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5460 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5461 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5462 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5463 Instruction *NewRem =
5464 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5466 InsertNewInstBefore(NewRem, ICI);
5467 return new ICmpInst(ICI.getPredicate(), NewRem,
5468 Constant::getNullValue(BO->getType()));
5472 case Instruction::Add:
5473 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5474 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5475 if (BO->hasOneUse())
5476 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5477 Subtract(RHS, BOp1C));
5478 } else if (RHSV == 0) {
5479 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5480 // efficiently invertible, or if the add has just this one use.
5481 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5483 if (Value *NegVal = dyn_castNegVal(BOp1))
5484 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5485 else if (Value *NegVal = dyn_castNegVal(BOp0))
5486 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5487 else if (BO->hasOneUse()) {
5488 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5489 InsertNewInstBefore(Neg, ICI);
5491 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5495 case Instruction::Xor:
5496 // For the xor case, we can xor two constants together, eliminating
5497 // the explicit xor.
5498 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5499 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5500 ConstantExpr::getXor(RHS, BOC));
5503 case Instruction::Sub:
5504 // Replace (([sub|xor] A, B) != 0) with (A != B)
5506 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5510 case Instruction::Or:
5511 // If bits are being or'd in that are not present in the constant we
5512 // are comparing against, then the comparison could never succeed!
5513 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5514 Constant *NotCI = ConstantExpr::getNot(RHS);
5515 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5516 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5521 case Instruction::And:
5522 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5523 // If bits are being compared against that are and'd out, then the
5524 // comparison can never succeed!
5525 if ((RHSV & ~BOC->getValue()) != 0)
5526 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5529 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5530 if (RHS == BOC && RHSV.isPowerOf2())
5531 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5532 ICmpInst::ICMP_NE, LHSI,
5533 Constant::getNullValue(RHS->getType()));
5535 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5536 if (isSignBit(BOC)) {
5537 Value *X = BO->getOperand(0);
5538 Constant *Zero = Constant::getNullValue(X->getType());
5539 ICmpInst::Predicate pred = isICMP_NE ?
5540 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5541 return new ICmpInst(pred, X, Zero);
5544 // ((X & ~7) == 0) --> X < 8
5545 if (RHSV == 0 && isHighOnes(BOC)) {
5546 Value *X = BO->getOperand(0);
5547 Constant *NegX = ConstantExpr::getNeg(BOC);
5548 ICmpInst::Predicate pred = isICMP_NE ?
5549 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5550 return new ICmpInst(pred, X, NegX);
5555 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5556 // Handle icmp {eq|ne} <intrinsic>, intcst.
5557 if (II->getIntrinsicID() == Intrinsic::bswap) {
5559 ICI.setOperand(0, II->getOperand(1));
5560 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5564 } else { // Not a ICMP_EQ/ICMP_NE
5565 // If the LHS is a cast from an integral value of the same size,
5566 // then since we know the RHS is a constant, try to simlify.
5567 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5568 Value *CastOp = Cast->getOperand(0);
5569 const Type *SrcTy = CastOp->getType();
5570 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5571 if (SrcTy->isInteger() &&
5572 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5573 // If this is an unsigned comparison, try to make the comparison use
5574 // smaller constant values.
5575 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5576 // X u< 128 => X s> -1
5577 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5578 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5579 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5580 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5581 // X u> 127 => X s< 0
5582 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5583 Constant::getNullValue(SrcTy));
5591 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5592 /// We only handle extending casts so far.
5594 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5595 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5596 Value *LHSCIOp = LHSCI->getOperand(0);
5597 const Type *SrcTy = LHSCIOp->getType();
5598 const Type *DestTy = LHSCI->getType();
5601 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5602 // integer type is the same size as the pointer type.
5603 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5604 getTargetData().getPointerSizeInBits() ==
5605 cast<IntegerType>(DestTy)->getBitWidth()) {
5607 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5608 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5609 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5610 RHSOp = RHSC->getOperand(0);
5611 // If the pointer types don't match, insert a bitcast.
5612 if (LHSCIOp->getType() != RHSOp->getType())
5613 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5614 LHSCIOp->getType(), ICI);
5618 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5621 // The code below only handles extension cast instructions, so far.
5623 if (LHSCI->getOpcode() != Instruction::ZExt &&
5624 LHSCI->getOpcode() != Instruction::SExt)
5627 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5628 bool isSignedCmp = ICI.isSignedPredicate();
5630 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5631 // Not an extension from the same type?
5632 RHSCIOp = CI->getOperand(0);
5633 if (RHSCIOp->getType() != LHSCIOp->getType())
5636 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5637 // and the other is a zext), then we can't handle this.
5638 if (CI->getOpcode() != LHSCI->getOpcode())
5641 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5642 // then we can't handle this.
5643 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5646 // Okay, just insert a compare of the reduced operands now!
5647 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5650 // If we aren't dealing with a constant on the RHS, exit early
5651 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5655 // Compute the constant that would happen if we truncated to SrcTy then
5656 // reextended to DestTy.
5657 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5658 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5660 // If the re-extended constant didn't change...
5662 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5663 // For example, we might have:
5664 // %A = sext short %X to uint
5665 // %B = icmp ugt uint %A, 1330
5666 // It is incorrect to transform this into
5667 // %B = icmp ugt short %X, 1330
5668 // because %A may have negative value.
5670 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5671 // OR operation is EQ/NE.
5672 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5673 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5678 // The re-extended constant changed so the constant cannot be represented
5679 // in the shorter type. Consequently, we cannot emit a simple comparison.
5681 // First, handle some easy cases. We know the result cannot be equal at this
5682 // point so handle the ICI.isEquality() cases
5683 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5684 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5685 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5686 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5688 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5689 // should have been folded away previously and not enter in here.
5692 // We're performing a signed comparison.
5693 if (cast<ConstantInt>(CI)->getValue().isNegative())
5694 Result = ConstantInt::getFalse(); // X < (small) --> false
5696 Result = ConstantInt::getTrue(); // X < (large) --> true
5698 // We're performing an unsigned comparison.
5700 // We're performing an unsigned comp with a sign extended value.
5701 // This is true if the input is >= 0. [aka >s -1]
5702 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5703 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5704 NegOne, ICI.getName()), ICI);
5706 // Unsigned extend & unsigned compare -> always true.
5707 Result = ConstantInt::getTrue();
5711 // Finally, return the value computed.
5712 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5713 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5714 return ReplaceInstUsesWith(ICI, Result);
5716 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5717 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5718 "ICmp should be folded!");
5719 if (Constant *CI = dyn_cast<Constant>(Result))
5720 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5722 return BinaryOperator::createNot(Result);
5726 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5727 return commonShiftTransforms(I);
5730 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5731 return commonShiftTransforms(I);
5734 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5735 return commonShiftTransforms(I);
5738 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5739 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5740 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5742 // shl X, 0 == X and shr X, 0 == X
5743 // shl 0, X == 0 and shr 0, X == 0
5744 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5745 Op0 == Constant::getNullValue(Op0->getType()))
5746 return ReplaceInstUsesWith(I, Op0);
5748 if (isa<UndefValue>(Op0)) {
5749 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5750 return ReplaceInstUsesWith(I, Op0);
5751 else // undef << X -> 0, undef >>u X -> 0
5752 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5754 if (isa<UndefValue>(Op1)) {
5755 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5756 return ReplaceInstUsesWith(I, Op0);
5757 else // X << undef, X >>u undef -> 0
5758 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5761 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5762 if (I.getOpcode() == Instruction::AShr)
5763 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5764 if (CSI->isAllOnesValue())
5765 return ReplaceInstUsesWith(I, CSI);
5767 // Try to fold constant and into select arguments.
5768 if (isa<Constant>(Op0))
5769 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5770 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5773 // See if we can turn a signed shr into an unsigned shr.
5774 if (I.isArithmeticShift()) {
5775 if (MaskedValueIsZero(Op0,
5776 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5777 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5781 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5782 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5787 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5788 BinaryOperator &I) {
5789 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5791 // See if we can simplify any instructions used by the instruction whose sole
5792 // purpose is to compute bits we don't care about.
5793 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5794 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5795 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5796 KnownZero, KnownOne))
5799 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5800 // of a signed value.
5802 if (Op1->uge(TypeBits)) {
5803 if (I.getOpcode() != Instruction::AShr)
5804 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5806 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5811 // ((X*C1) << C2) == (X * (C1 << C2))
5812 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5813 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5814 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5815 return BinaryOperator::createMul(BO->getOperand(0),
5816 ConstantExpr::getShl(BOOp, Op1));
5818 // Try to fold constant and into select arguments.
5819 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5820 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5822 if (isa<PHINode>(Op0))
5823 if (Instruction *NV = FoldOpIntoPhi(I))
5826 if (Op0->hasOneUse()) {
5827 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5828 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5831 switch (Op0BO->getOpcode()) {
5833 case Instruction::Add:
5834 case Instruction::And:
5835 case Instruction::Or:
5836 case Instruction::Xor: {
5837 // These operators commute.
5838 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5839 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5840 match(Op0BO->getOperand(1),
5841 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5842 Instruction *YS = BinaryOperator::createShl(
5843 Op0BO->getOperand(0), Op1,
5845 InsertNewInstBefore(YS, I); // (Y << C)
5847 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5848 Op0BO->getOperand(1)->getName());
5849 InsertNewInstBefore(X, I); // (X + (Y << C))
5850 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5851 return BinaryOperator::createAnd(X, ConstantInt::get(
5852 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5855 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5856 Value *Op0BOOp1 = Op0BO->getOperand(1);
5857 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5859 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5860 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5862 Instruction *YS = BinaryOperator::createShl(
5863 Op0BO->getOperand(0), Op1,
5865 InsertNewInstBefore(YS, I); // (Y << C)
5867 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5868 V1->getName()+".mask");
5869 InsertNewInstBefore(XM, I); // X & (CC << C)
5871 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5876 case Instruction::Sub: {
5877 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5878 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5879 match(Op0BO->getOperand(0),
5880 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5881 Instruction *YS = BinaryOperator::createShl(
5882 Op0BO->getOperand(1), Op1,
5884 InsertNewInstBefore(YS, I); // (Y << C)
5886 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5887 Op0BO->getOperand(0)->getName());
5888 InsertNewInstBefore(X, I); // (X + (Y << C))
5889 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5890 return BinaryOperator::createAnd(X, ConstantInt::get(
5891 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5894 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5895 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5896 match(Op0BO->getOperand(0),
5897 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5898 m_ConstantInt(CC))) && V2 == Op1 &&
5899 cast<BinaryOperator>(Op0BO->getOperand(0))
5900 ->getOperand(0)->hasOneUse()) {
5901 Instruction *YS = BinaryOperator::createShl(
5902 Op0BO->getOperand(1), Op1,
5904 InsertNewInstBefore(YS, I); // (Y << C)
5906 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5907 V1->getName()+".mask");
5908 InsertNewInstBefore(XM, I); // X & (CC << C)
5910 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5918 // If the operand is an bitwise operator with a constant RHS, and the
5919 // shift is the only use, we can pull it out of the shift.
5920 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5921 bool isValid = true; // Valid only for And, Or, Xor
5922 bool highBitSet = false; // Transform if high bit of constant set?
5924 switch (Op0BO->getOpcode()) {
5925 default: isValid = false; break; // Do not perform transform!
5926 case Instruction::Add:
5927 isValid = isLeftShift;
5929 case Instruction::Or:
5930 case Instruction::Xor:
5933 case Instruction::And:
5938 // If this is a signed shift right, and the high bit is modified
5939 // by the logical operation, do not perform the transformation.
5940 // The highBitSet boolean indicates the value of the high bit of
5941 // the constant which would cause it to be modified for this
5944 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
5945 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
5949 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5951 Instruction *NewShift =
5952 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
5953 InsertNewInstBefore(NewShift, I);
5954 NewShift->takeName(Op0BO);
5956 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5963 // Find out if this is a shift of a shift by a constant.
5964 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
5965 if (ShiftOp && !ShiftOp->isShift())
5968 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5969 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5970 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
5971 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
5972 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
5973 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
5974 Value *X = ShiftOp->getOperand(0);
5976 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5977 if (AmtSum > TypeBits)
5980 const IntegerType *Ty = cast<IntegerType>(I.getType());
5982 // Check for (X << c1) << c2 and (X >> c1) >> c2
5983 if (I.getOpcode() == ShiftOp->getOpcode()) {
5984 return BinaryOperator::create(I.getOpcode(), X,
5985 ConstantInt::get(Ty, AmtSum));
5986 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
5987 I.getOpcode() == Instruction::AShr) {
5988 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
5989 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
5990 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
5991 I.getOpcode() == Instruction::LShr) {
5992 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
5993 Instruction *Shift =
5994 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
5995 InsertNewInstBefore(Shift, I);
5997 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
5998 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6001 // Okay, if we get here, one shift must be left, and the other shift must be
6002 // right. See if the amounts are equal.
6003 if (ShiftAmt1 == ShiftAmt2) {
6004 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6005 if (I.getOpcode() == Instruction::Shl) {
6006 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6007 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6009 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6010 if (I.getOpcode() == Instruction::LShr) {
6011 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6012 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6014 // We can simplify ((X << C) >>s C) into a trunc + sext.
6015 // NOTE: we could do this for any C, but that would make 'unusual' integer
6016 // types. For now, just stick to ones well-supported by the code
6018 const Type *SExtType = 0;
6019 switch (Ty->getBitWidth() - ShiftAmt1) {
6026 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6031 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6032 InsertNewInstBefore(NewTrunc, I);
6033 return new SExtInst(NewTrunc, Ty);
6035 // Otherwise, we can't handle it yet.
6036 } else if (ShiftAmt1 < ShiftAmt2) {
6037 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6039 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6040 if (I.getOpcode() == Instruction::Shl) {
6041 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6042 ShiftOp->getOpcode() == Instruction::AShr);
6043 Instruction *Shift =
6044 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6045 InsertNewInstBefore(Shift, I);
6047 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6048 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6051 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6052 if (I.getOpcode() == Instruction::LShr) {
6053 assert(ShiftOp->getOpcode() == Instruction::Shl);
6054 Instruction *Shift =
6055 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6056 InsertNewInstBefore(Shift, I);
6058 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6059 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6062 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6064 assert(ShiftAmt2 < ShiftAmt1);
6065 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6067 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6068 if (I.getOpcode() == Instruction::Shl) {
6069 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6070 ShiftOp->getOpcode() == Instruction::AShr);
6071 Instruction *Shift =
6072 BinaryOperator::create(ShiftOp->getOpcode(), X,
6073 ConstantInt::get(Ty, ShiftDiff));
6074 InsertNewInstBefore(Shift, I);
6076 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6077 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6080 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6081 if (I.getOpcode() == Instruction::LShr) {
6082 assert(ShiftOp->getOpcode() == Instruction::Shl);
6083 Instruction *Shift =
6084 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6085 InsertNewInstBefore(Shift, I);
6087 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6088 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6091 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6098 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6099 /// expression. If so, decompose it, returning some value X, such that Val is
6102 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6104 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6105 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6106 Offset = CI->getZExtValue();
6108 return ConstantInt::get(Type::Int32Ty, 0);
6109 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
6110 if (I->getNumOperands() == 2) {
6111 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6112 if (I->getOpcode() == Instruction::Shl) {
6113 // This is a value scaled by '1 << the shift amt'.
6114 Scale = 1U << CUI->getZExtValue();
6116 return I->getOperand(0);
6117 } else if (I->getOpcode() == Instruction::Mul) {
6118 // This value is scaled by 'CUI'.
6119 Scale = CUI->getZExtValue();
6121 return I->getOperand(0);
6122 } else if (I->getOpcode() == Instruction::Add) {
6123 // We have X+C. Check to see if we really have (X*C2)+C1,
6124 // where C1 is divisible by C2.
6127 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6128 Offset += CUI->getZExtValue();
6129 if (SubScale > 1 && (Offset % SubScale == 0)) {
6138 // Otherwise, we can't look past this.
6145 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6146 /// try to eliminate the cast by moving the type information into the alloc.
6147 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6148 AllocationInst &AI) {
6149 const PointerType *PTy = cast<PointerType>(CI.getType());
6151 // Remove any uses of AI that are dead.
6152 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6154 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6155 Instruction *User = cast<Instruction>(*UI++);
6156 if (isInstructionTriviallyDead(User)) {
6157 while (UI != E && *UI == User)
6158 ++UI; // If this instruction uses AI more than once, don't break UI.
6161 DOUT << "IC: DCE: " << *User;
6162 EraseInstFromFunction(*User);
6166 // Get the type really allocated and the type casted to.
6167 const Type *AllocElTy = AI.getAllocatedType();
6168 const Type *CastElTy = PTy->getElementType();
6169 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6171 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6172 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6173 if (CastElTyAlign < AllocElTyAlign) return 0;
6175 // If the allocation has multiple uses, only promote it if we are strictly
6176 // increasing the alignment of the resultant allocation. If we keep it the
6177 // same, we open the door to infinite loops of various kinds.
6178 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6180 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
6181 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
6182 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6184 // See if we can satisfy the modulus by pulling a scale out of the array
6186 unsigned ArraySizeScale;
6188 Value *NumElements = // See if the array size is a decomposable linear expr.
6189 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6191 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6193 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6194 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6196 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6201 // If the allocation size is constant, form a constant mul expression
6202 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6203 if (isa<ConstantInt>(NumElements))
6204 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6205 // otherwise multiply the amount and the number of elements
6206 else if (Scale != 1) {
6207 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6208 Amt = InsertNewInstBefore(Tmp, AI);
6212 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6213 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6214 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6215 Amt = InsertNewInstBefore(Tmp, AI);
6218 AllocationInst *New;
6219 if (isa<MallocInst>(AI))
6220 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6222 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6223 InsertNewInstBefore(New, AI);
6226 // If the allocation has multiple uses, insert a cast and change all things
6227 // that used it to use the new cast. This will also hack on CI, but it will
6229 if (!AI.hasOneUse()) {
6230 AddUsesToWorkList(AI);
6231 // New is the allocation instruction, pointer typed. AI is the original
6232 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6233 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6234 InsertNewInstBefore(NewCast, AI);
6235 AI.replaceAllUsesWith(NewCast);
6237 return ReplaceInstUsesWith(CI, New);
6240 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6241 /// and return it as type Ty without inserting any new casts and without
6242 /// changing the computed value. This is used by code that tries to decide
6243 /// whether promoting or shrinking integer operations to wider or smaller types
6244 /// will allow us to eliminate a truncate or extend.
6246 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6247 /// extension operation if Ty is larger.
6248 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6249 int &NumCastsRemoved) {
6250 // We can always evaluate constants in another type.
6251 if (isa<ConstantInt>(V))
6254 Instruction *I = dyn_cast<Instruction>(V);
6255 if (!I) return false;
6257 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6259 switch (I->getOpcode()) {
6260 case Instruction::Add:
6261 case Instruction::Sub:
6262 case Instruction::And:
6263 case Instruction::Or:
6264 case Instruction::Xor:
6265 if (!I->hasOneUse()) return false;
6266 // These operators can all arbitrarily be extended or truncated.
6267 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
6268 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
6270 case Instruction::Shl:
6271 if (!I->hasOneUse()) return false;
6272 // If we are truncating the result of this SHL, and if it's a shift of a
6273 // constant amount, we can always perform a SHL in a smaller type.
6274 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6275 uint32_t BitWidth = Ty->getBitWidth();
6276 if (BitWidth < OrigTy->getBitWidth() &&
6277 CI->getLimitedValue(BitWidth) < BitWidth)
6278 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6281 case Instruction::LShr:
6282 if (!I->hasOneUse()) return false;
6283 // If this is a truncate of a logical shr, we can truncate it to a smaller
6284 // lshr iff we know that the bits we would otherwise be shifting in are
6286 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6287 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6288 uint32_t BitWidth = Ty->getBitWidth();
6289 if (BitWidth < OrigBitWidth &&
6290 MaskedValueIsZero(I->getOperand(0),
6291 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6292 CI->getLimitedValue(BitWidth) < BitWidth) {
6293 return CanEvaluateInDifferentType(I->getOperand(0), Ty,NumCastsRemoved);
6297 case Instruction::Trunc:
6298 case Instruction::ZExt:
6299 case Instruction::SExt:
6300 // If this is a cast from the destination type, we can trivially eliminate
6301 // it, and this will remove a cast overall.
6302 if (I->getOperand(0)->getType() == Ty) {
6303 // If the first operand is itself a cast, and is eliminable, do not count
6304 // this as an eliminable cast. We would prefer to eliminate those two
6306 if (isa<CastInst>(I->getOperand(0)))
6314 // TODO: Can handle more cases here.
6321 /// EvaluateInDifferentType - Given an expression that
6322 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6323 /// evaluate the expression.
6324 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6326 if (Constant *C = dyn_cast<Constant>(V))
6327 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6329 // Otherwise, it must be an instruction.
6330 Instruction *I = cast<Instruction>(V);
6331 Instruction *Res = 0;
6332 switch (I->getOpcode()) {
6333 case Instruction::Add:
6334 case Instruction::Sub:
6335 case Instruction::And:
6336 case Instruction::Or:
6337 case Instruction::Xor:
6338 case Instruction::AShr:
6339 case Instruction::LShr:
6340 case Instruction::Shl: {
6341 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6342 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6343 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6344 LHS, RHS, I->getName());
6347 case Instruction::Trunc:
6348 case Instruction::ZExt:
6349 case Instruction::SExt:
6350 case Instruction::BitCast:
6351 // If the source type of the cast is the type we're trying for then we can
6352 // just return the source. There's no need to insert it because its not new.
6353 if (I->getOperand(0)->getType() == Ty)
6354 return I->getOperand(0);
6356 // Some other kind of cast, which shouldn't happen, so just ..
6359 // TODO: Can handle more cases here.
6360 assert(0 && "Unreachable!");
6364 return InsertNewInstBefore(Res, *I);
6367 /// @brief Implement the transforms common to all CastInst visitors.
6368 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6369 Value *Src = CI.getOperand(0);
6371 // Casting undef to anything results in undef so might as just replace it and
6372 // get rid of the cast.
6373 if (isa<UndefValue>(Src)) // cast undef -> undef
6374 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
6376 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6377 // eliminate it now.
6378 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6379 if (Instruction::CastOps opc =
6380 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6381 // The first cast (CSrc) is eliminable so we need to fix up or replace
6382 // the second cast (CI). CSrc will then have a good chance of being dead.
6383 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6387 // If we are casting a select then fold the cast into the select
6388 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6389 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6392 // If we are casting a PHI then fold the cast into the PHI
6393 if (isa<PHINode>(Src))
6394 if (Instruction *NV = FoldOpIntoPhi(CI))
6400 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6401 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6402 Value *Src = CI.getOperand(0);
6404 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6405 // If casting the result of a getelementptr instruction with no offset, turn
6406 // this into a cast of the original pointer!
6407 if (GEP->hasAllZeroIndices()) {
6408 // Changing the cast operand is usually not a good idea but it is safe
6409 // here because the pointer operand is being replaced with another
6410 // pointer operand so the opcode doesn't need to change.
6412 CI.setOperand(0, GEP->getOperand(0));
6416 // If the GEP has a single use, and the base pointer is a bitcast, and the
6417 // GEP computes a constant offset, see if we can convert these three
6418 // instructions into fewer. This typically happens with unions and other
6419 // non-type-safe code.
6420 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6421 if (GEP->hasAllConstantIndices()) {
6422 // We are guaranteed to get a constant from EmitGEPOffset.
6423 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6424 int64_t Offset = OffsetV->getSExtValue();
6426 // Get the base pointer input of the bitcast, and the type it points to.
6427 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6428 const Type *GEPIdxTy =
6429 cast<PointerType>(OrigBase->getType())->getElementType();
6430 if (GEPIdxTy->isSized()) {
6431 SmallVector<Value*, 8> NewIndices;
6433 // Start with the index over the outer type. Note that the type size
6434 // might be zero (even if the offset isn't zero) if the indexed type
6435 // is something like [0 x {int, int}]
6436 const Type *IntPtrTy = TD->getIntPtrType();
6437 int64_t FirstIdx = 0;
6438 if (int64_t TySize = TD->getTypeSize(GEPIdxTy)) {
6439 FirstIdx = Offset/TySize;
6442 // Handle silly modulus not returning values values [0..TySize).
6446 assert(Offset >= 0);
6448 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6451 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6453 // Index into the types. If we fail, set OrigBase to null.
6455 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6456 const StructLayout *SL = TD->getStructLayout(STy);
6457 if (Offset < (int64_t)SL->getSizeInBytes()) {
6458 unsigned Elt = SL->getElementContainingOffset(Offset);
6459 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6461 Offset -= SL->getElementOffset(Elt);
6462 GEPIdxTy = STy->getElementType(Elt);
6464 // Otherwise, we can't index into this, bail out.
6468 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6469 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6470 if (uint64_t EltSize = TD->getTypeSize(STy->getElementType())) {
6471 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6474 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6476 GEPIdxTy = STy->getElementType();
6478 // Otherwise, we can't index into this, bail out.
6484 // If we were able to index down into an element, create the GEP
6485 // and bitcast the result. This eliminates one bitcast, potentially
6487 Instruction *NGEP = new GetElementPtrInst(OrigBase, &NewIndices[0],
6488 NewIndices.size(), "");
6489 InsertNewInstBefore(NGEP, CI);
6490 NGEP->takeName(GEP);
6492 if (isa<BitCastInst>(CI))
6493 return new BitCastInst(NGEP, CI.getType());
6494 assert(isa<PtrToIntInst>(CI));
6495 return new PtrToIntInst(NGEP, CI.getType());
6502 return commonCastTransforms(CI);
6507 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6508 /// integer types. This function implements the common transforms for all those
6510 /// @brief Implement the transforms common to CastInst with integer operands
6511 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6512 if (Instruction *Result = commonCastTransforms(CI))
6515 Value *Src = CI.getOperand(0);
6516 const Type *SrcTy = Src->getType();
6517 const Type *DestTy = CI.getType();
6518 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6519 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6521 // See if we can simplify any instructions used by the LHS whose sole
6522 // purpose is to compute bits we don't care about.
6523 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6524 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6525 KnownZero, KnownOne))
6528 // If the source isn't an instruction or has more than one use then we
6529 // can't do anything more.
6530 Instruction *SrcI = dyn_cast<Instruction>(Src);
6531 if (!SrcI || !Src->hasOneUse())
6534 // Attempt to propagate the cast into the instruction for int->int casts.
6535 int NumCastsRemoved = 0;
6536 if (!isa<BitCastInst>(CI) &&
6537 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6539 // If this cast is a truncate, evaluting in a different type always
6540 // eliminates the cast, so it is always a win. If this is a noop-cast
6541 // this just removes a noop cast which isn't pointful, but simplifies
6542 // the code. If this is a zero-extension, we need to do an AND to
6543 // maintain the clear top-part of the computation, so we require that
6544 // the input have eliminated at least one cast. If this is a sign
6545 // extension, we insert two new casts (to do the extension) so we
6546 // require that two casts have been eliminated.
6548 switch (CI.getOpcode()) {
6550 // All the others use floating point so we shouldn't actually
6551 // get here because of the check above.
6552 assert(0 && "Unknown cast type");
6553 case Instruction::Trunc:
6556 case Instruction::ZExt:
6557 DoXForm = NumCastsRemoved >= 1;
6559 case Instruction::SExt:
6560 DoXForm = NumCastsRemoved >= 2;
6562 case Instruction::BitCast:
6568 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6569 CI.getOpcode() == Instruction::SExt);
6570 assert(Res->getType() == DestTy);
6571 switch (CI.getOpcode()) {
6572 default: assert(0 && "Unknown cast type!");
6573 case Instruction::Trunc:
6574 case Instruction::BitCast:
6575 // Just replace this cast with the result.
6576 return ReplaceInstUsesWith(CI, Res);
6577 case Instruction::ZExt: {
6578 // We need to emit an AND to clear the high bits.
6579 assert(SrcBitSize < DestBitSize && "Not a zext?");
6580 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6582 return BinaryOperator::createAnd(Res, C);
6584 case Instruction::SExt:
6585 // We need to emit a cast to truncate, then a cast to sext.
6586 return CastInst::create(Instruction::SExt,
6587 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6593 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6594 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6596 switch (SrcI->getOpcode()) {
6597 case Instruction::Add:
6598 case Instruction::Mul:
6599 case Instruction::And:
6600 case Instruction::Or:
6601 case Instruction::Xor:
6602 // If we are discarding information, rewrite.
6603 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6604 // Don't insert two casts if they cannot be eliminated. We allow
6605 // two casts to be inserted if the sizes are the same. This could
6606 // only be converting signedness, which is a noop.
6607 if (DestBitSize == SrcBitSize ||
6608 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6609 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6610 Instruction::CastOps opcode = CI.getOpcode();
6611 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6612 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6613 return BinaryOperator::create(
6614 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6618 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6619 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6620 SrcI->getOpcode() == Instruction::Xor &&
6621 Op1 == ConstantInt::getTrue() &&
6622 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6623 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6624 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6627 case Instruction::SDiv:
6628 case Instruction::UDiv:
6629 case Instruction::SRem:
6630 case Instruction::URem:
6631 // If we are just changing the sign, rewrite.
6632 if (DestBitSize == SrcBitSize) {
6633 // Don't insert two casts if they cannot be eliminated. We allow
6634 // two casts to be inserted if the sizes are the same. This could
6635 // only be converting signedness, which is a noop.
6636 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6637 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6638 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6640 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6642 return BinaryOperator::create(
6643 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6648 case Instruction::Shl:
6649 // Allow changing the sign of the source operand. Do not allow
6650 // changing the size of the shift, UNLESS the shift amount is a
6651 // constant. We must not change variable sized shifts to a smaller
6652 // size, because it is undefined to shift more bits out than exist
6654 if (DestBitSize == SrcBitSize ||
6655 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6656 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6657 Instruction::BitCast : Instruction::Trunc);
6658 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6659 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6660 return BinaryOperator::createShl(Op0c, Op1c);
6663 case Instruction::AShr:
6664 // If this is a signed shr, and if all bits shifted in are about to be
6665 // truncated off, turn it into an unsigned shr to allow greater
6667 if (DestBitSize < SrcBitSize &&
6668 isa<ConstantInt>(Op1)) {
6669 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6670 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6671 // Insert the new logical shift right.
6672 return BinaryOperator::createLShr(Op0, Op1);
6680 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6681 if (Instruction *Result = commonIntCastTransforms(CI))
6684 Value *Src = CI.getOperand(0);
6685 const Type *Ty = CI.getType();
6686 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6687 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6689 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6690 switch (SrcI->getOpcode()) {
6692 case Instruction::LShr:
6693 // We can shrink lshr to something smaller if we know the bits shifted in
6694 // are already zeros.
6695 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6696 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6698 // Get a mask for the bits shifting in.
6699 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6700 Value* SrcIOp0 = SrcI->getOperand(0);
6701 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6702 if (ShAmt >= DestBitWidth) // All zeros.
6703 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6705 // Okay, we can shrink this. Truncate the input, then return a new
6707 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6708 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6710 return BinaryOperator::createLShr(V1, V2);
6712 } else { // This is a variable shr.
6714 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6715 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6716 // loop-invariant and CSE'd.
6717 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6718 Value *One = ConstantInt::get(SrcI->getType(), 1);
6720 Value *V = InsertNewInstBefore(
6721 BinaryOperator::createShl(One, SrcI->getOperand(1),
6723 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6724 SrcI->getOperand(0),
6726 Value *Zero = Constant::getNullValue(V->getType());
6727 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6737 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6738 // If one of the common conversion will work ..
6739 if (Instruction *Result = commonIntCastTransforms(CI))
6742 Value *Src = CI.getOperand(0);
6744 // If this is a cast of a cast
6745 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6746 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6747 // types and if the sizes are just right we can convert this into a logical
6748 // 'and' which will be much cheaper than the pair of casts.
6749 if (isa<TruncInst>(CSrc)) {
6750 // Get the sizes of the types involved
6751 Value *A = CSrc->getOperand(0);
6752 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6753 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6754 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6755 // If we're actually extending zero bits and the trunc is a no-op
6756 if (MidSize < DstSize && SrcSize == DstSize) {
6757 // Replace both of the casts with an And of the type mask.
6758 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6759 Constant *AndConst = ConstantInt::get(AndValue);
6761 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6762 // Unfortunately, if the type changed, we need to cast it back.
6763 if (And->getType() != CI.getType()) {
6764 And->setName(CSrc->getName()+".mask");
6765 InsertNewInstBefore(And, CI);
6766 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6773 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6774 // If we are just checking for a icmp eq of a single bit and zext'ing it
6775 // to an integer, then shift the bit to the appropriate place and then
6776 // cast to integer to avoid the comparison.
6777 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6778 const APInt &Op1CV = Op1C->getValue();
6780 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
6781 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
6782 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6783 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6784 Value *In = ICI->getOperand(0);
6785 Value *Sh = ConstantInt::get(In->getType(),
6786 In->getType()->getPrimitiveSizeInBits()-1);
6787 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
6788 In->getName()+".lobit"),
6790 if (In->getType() != CI.getType())
6791 In = CastInst::createIntegerCast(In, CI.getType(),
6792 false/*ZExt*/, "tmp", &CI);
6794 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
6795 Constant *One = ConstantInt::get(In->getType(), 1);
6796 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
6797 In->getName()+".not"),
6801 return ReplaceInstUsesWith(CI, In);
6806 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
6807 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6808 // zext (X == 1) to i32 --> X iff X has only the low bit set.
6809 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
6810 // zext (X != 0) to i32 --> X iff X has only the low bit set.
6811 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
6812 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
6813 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6814 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
6815 // This only works for EQ and NE
6816 ICI->isEquality()) {
6817 // If Op1C some other power of two, convert:
6818 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6819 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6820 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6821 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
6823 APInt KnownZeroMask(~KnownZero);
6824 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6825 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
6826 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6827 // (X&4) == 2 --> false
6828 // (X&4) != 2 --> true
6829 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6830 Res = ConstantExpr::getZExt(Res, CI.getType());
6831 return ReplaceInstUsesWith(CI, Res);
6834 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6835 Value *In = ICI->getOperand(0);
6837 // Perform a logical shr by shiftamt.
6838 // Insert the shift to put the result in the low bit.
6839 In = InsertNewInstBefore(
6840 BinaryOperator::createLShr(In,
6841 ConstantInt::get(In->getType(), ShiftAmt),
6842 In->getName()+".lobit"), CI);
6845 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6846 Constant *One = ConstantInt::get(In->getType(), 1);
6847 In = BinaryOperator::createXor(In, One, "tmp");
6848 InsertNewInstBefore(cast<Instruction>(In), CI);
6851 if (CI.getType() == In->getType())
6852 return ReplaceInstUsesWith(CI, In);
6854 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6862 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
6863 if (Instruction *I = commonIntCastTransforms(CI))
6866 Value *Src = CI.getOperand(0);
6868 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
6869 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
6870 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6871 // If we are just checking for a icmp eq of a single bit and zext'ing it
6872 // to an integer, then shift the bit to the appropriate place and then
6873 // cast to integer to avoid the comparison.
6874 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6875 const APInt &Op1CV = Op1C->getValue();
6877 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
6878 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
6879 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6880 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6881 Value *In = ICI->getOperand(0);
6882 Value *Sh = ConstantInt::get(In->getType(),
6883 In->getType()->getPrimitiveSizeInBits()-1);
6884 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
6885 In->getName()+".lobit"),
6887 if (In->getType() != CI.getType())
6888 In = CastInst::createIntegerCast(In, CI.getType(),
6889 true/*SExt*/, "tmp", &CI);
6891 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
6892 In = InsertNewInstBefore(BinaryOperator::createNot(In,
6893 In->getName()+".not"), CI);
6895 return ReplaceInstUsesWith(CI, In);
6903 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6904 return commonCastTransforms(CI);
6907 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6908 return commonCastTransforms(CI);
6911 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6912 return commonCastTransforms(CI);
6915 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6916 return commonCastTransforms(CI);
6919 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6920 return commonCastTransforms(CI);
6923 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6924 return commonCastTransforms(CI);
6927 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6928 return commonPointerCastTransforms(CI);
6931 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6932 return commonCastTransforms(CI);
6935 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
6936 // If the operands are integer typed then apply the integer transforms,
6937 // otherwise just apply the common ones.
6938 Value *Src = CI.getOperand(0);
6939 const Type *SrcTy = Src->getType();
6940 const Type *DestTy = CI.getType();
6942 if (SrcTy->isInteger() && DestTy->isInteger()) {
6943 if (Instruction *Result = commonIntCastTransforms(CI))
6945 } else if (isa<PointerType>(SrcTy)) {
6946 if (Instruction *I = commonPointerCastTransforms(CI))
6949 if (Instruction *Result = commonCastTransforms(CI))
6954 // Get rid of casts from one type to the same type. These are useless and can
6955 // be replaced by the operand.
6956 if (DestTy == Src->getType())
6957 return ReplaceInstUsesWith(CI, Src);
6959 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6960 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
6961 const Type *DstElTy = DstPTy->getElementType();
6962 const Type *SrcElTy = SrcPTy->getElementType();
6964 // If we are casting a malloc or alloca to a pointer to a type of the same
6965 // size, rewrite the allocation instruction to allocate the "right" type.
6966 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
6967 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
6970 // If the source and destination are pointers, and this cast is equivalent
6971 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
6972 // This can enhance SROA and other transforms that want type-safe pointers.
6973 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
6974 unsigned NumZeros = 0;
6975 while (SrcElTy != DstElTy &&
6976 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6977 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6978 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6982 // If we found a path from the src to dest, create the getelementptr now.
6983 if (SrcElTy == DstElTy) {
6984 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
6985 return new GetElementPtrInst(Src, &Idxs[0], Idxs.size());
6989 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6990 if (SVI->hasOneUse()) {
6991 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6992 // a bitconvert to a vector with the same # elts.
6993 if (isa<VectorType>(DestTy) &&
6994 cast<VectorType>(DestTy)->getNumElements() ==
6995 SVI->getType()->getNumElements()) {
6997 // If either of the operands is a cast from CI.getType(), then
6998 // evaluating the shuffle in the casted destination's type will allow
6999 // us to eliminate at least one cast.
7000 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7001 Tmp->getOperand(0)->getType() == DestTy) ||
7002 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7003 Tmp->getOperand(0)->getType() == DestTy)) {
7004 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7005 SVI->getOperand(0), DestTy, &CI);
7006 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7007 SVI->getOperand(1), DestTy, &CI);
7008 // Return a new shuffle vector. Use the same element ID's, as we
7009 // know the vector types match #elts.
7010 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7018 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7020 /// %D = select %cond, %C, %A
7022 /// %C = select %cond, %B, 0
7025 /// Assuming that the specified instruction is an operand to the select, return
7026 /// a bitmask indicating which operands of this instruction are foldable if they
7027 /// equal the other incoming value of the select.
7029 static unsigned GetSelectFoldableOperands(Instruction *I) {
7030 switch (I->getOpcode()) {
7031 case Instruction::Add:
7032 case Instruction::Mul:
7033 case Instruction::And:
7034 case Instruction::Or:
7035 case Instruction::Xor:
7036 return 3; // Can fold through either operand.
7037 case Instruction::Sub: // Can only fold on the amount subtracted.
7038 case Instruction::Shl: // Can only fold on the shift amount.
7039 case Instruction::LShr:
7040 case Instruction::AShr:
7043 return 0; // Cannot fold
7047 /// GetSelectFoldableConstant - For the same transformation as the previous
7048 /// function, return the identity constant that goes into the select.
7049 static Constant *GetSelectFoldableConstant(Instruction *I) {
7050 switch (I->getOpcode()) {
7051 default: assert(0 && "This cannot happen!"); abort();
7052 case Instruction::Add:
7053 case Instruction::Sub:
7054 case Instruction::Or:
7055 case Instruction::Xor:
7056 case Instruction::Shl:
7057 case Instruction::LShr:
7058 case Instruction::AShr:
7059 return Constant::getNullValue(I->getType());
7060 case Instruction::And:
7061 return ConstantInt::getAllOnesValue(I->getType());
7062 case Instruction::Mul:
7063 return ConstantInt::get(I->getType(), 1);
7067 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7068 /// have the same opcode and only one use each. Try to simplify this.
7069 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7071 if (TI->getNumOperands() == 1) {
7072 // If this is a non-volatile load or a cast from the same type,
7075 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7078 return 0; // unknown unary op.
7081 // Fold this by inserting a select from the input values.
7082 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7083 FI->getOperand(0), SI.getName()+".v");
7084 InsertNewInstBefore(NewSI, SI);
7085 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7089 // Only handle binary operators here.
7090 if (!isa<BinaryOperator>(TI))
7093 // Figure out if the operations have any operands in common.
7094 Value *MatchOp, *OtherOpT, *OtherOpF;
7096 if (TI->getOperand(0) == FI->getOperand(0)) {
7097 MatchOp = TI->getOperand(0);
7098 OtherOpT = TI->getOperand(1);
7099 OtherOpF = FI->getOperand(1);
7100 MatchIsOpZero = true;
7101 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7102 MatchOp = TI->getOperand(1);
7103 OtherOpT = TI->getOperand(0);
7104 OtherOpF = FI->getOperand(0);
7105 MatchIsOpZero = false;
7106 } else if (!TI->isCommutative()) {
7108 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7109 MatchOp = TI->getOperand(0);
7110 OtherOpT = TI->getOperand(1);
7111 OtherOpF = FI->getOperand(0);
7112 MatchIsOpZero = true;
7113 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7114 MatchOp = TI->getOperand(1);
7115 OtherOpT = TI->getOperand(0);
7116 OtherOpF = FI->getOperand(1);
7117 MatchIsOpZero = true;
7122 // If we reach here, they do have operations in common.
7123 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7124 OtherOpF, SI.getName()+".v");
7125 InsertNewInstBefore(NewSI, SI);
7127 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7129 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7131 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7133 assert(0 && "Shouldn't get here");
7137 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7138 Value *CondVal = SI.getCondition();
7139 Value *TrueVal = SI.getTrueValue();
7140 Value *FalseVal = SI.getFalseValue();
7142 // select true, X, Y -> X
7143 // select false, X, Y -> Y
7144 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7145 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7147 // select C, X, X -> X
7148 if (TrueVal == FalseVal)
7149 return ReplaceInstUsesWith(SI, TrueVal);
7151 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7152 return ReplaceInstUsesWith(SI, FalseVal);
7153 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7154 return ReplaceInstUsesWith(SI, TrueVal);
7155 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7156 if (isa<Constant>(TrueVal))
7157 return ReplaceInstUsesWith(SI, TrueVal);
7159 return ReplaceInstUsesWith(SI, FalseVal);
7162 if (SI.getType() == Type::Int1Ty) {
7163 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7164 if (C->getZExtValue()) {
7165 // Change: A = select B, true, C --> A = or B, C
7166 return BinaryOperator::createOr(CondVal, FalseVal);
7168 // Change: A = select B, false, C --> A = and !B, C
7170 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7171 "not."+CondVal->getName()), SI);
7172 return BinaryOperator::createAnd(NotCond, FalseVal);
7174 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7175 if (C->getZExtValue() == false) {
7176 // Change: A = select B, C, false --> A = and B, C
7177 return BinaryOperator::createAnd(CondVal, TrueVal);
7179 // Change: A = select B, C, true --> A = or !B, C
7181 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7182 "not."+CondVal->getName()), SI);
7183 return BinaryOperator::createOr(NotCond, TrueVal);
7188 // Selecting between two integer constants?
7189 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7190 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7191 // select C, 1, 0 -> zext C to int
7192 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7193 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7194 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7195 // select C, 0, 1 -> zext !C to int
7197 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7198 "not."+CondVal->getName()), SI);
7199 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7202 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7204 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7206 // (x <s 0) ? -1 : 0 -> ashr x, 31
7207 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7208 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7209 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7210 // The comparison constant and the result are not neccessarily the
7211 // same width. Make an all-ones value by inserting a AShr.
7212 Value *X = IC->getOperand(0);
7213 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7214 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7215 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7217 InsertNewInstBefore(SRA, SI);
7219 // Finally, convert to the type of the select RHS. We figure out
7220 // if this requires a SExt, Trunc or BitCast based on the sizes.
7221 Instruction::CastOps opc = Instruction::BitCast;
7222 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7223 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7224 if (SRASize < SISize)
7225 opc = Instruction::SExt;
7226 else if (SRASize > SISize)
7227 opc = Instruction::Trunc;
7228 return CastInst::create(opc, SRA, SI.getType());
7233 // If one of the constants is zero (we know they can't both be) and we
7234 // have an icmp instruction with zero, and we have an 'and' with the
7235 // non-constant value, eliminate this whole mess. This corresponds to
7236 // cases like this: ((X & 27) ? 27 : 0)
7237 if (TrueValC->isZero() || FalseValC->isZero())
7238 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7239 cast<Constant>(IC->getOperand(1))->isNullValue())
7240 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7241 if (ICA->getOpcode() == Instruction::And &&
7242 isa<ConstantInt>(ICA->getOperand(1)) &&
7243 (ICA->getOperand(1) == TrueValC ||
7244 ICA->getOperand(1) == FalseValC) &&
7245 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7246 // Okay, now we know that everything is set up, we just don't
7247 // know whether we have a icmp_ne or icmp_eq and whether the
7248 // true or false val is the zero.
7249 bool ShouldNotVal = !TrueValC->isZero();
7250 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7253 V = InsertNewInstBefore(BinaryOperator::create(
7254 Instruction::Xor, V, ICA->getOperand(1)), SI);
7255 return ReplaceInstUsesWith(SI, V);
7260 // See if we are selecting two values based on a comparison of the two values.
7261 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7262 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7263 // Transform (X == Y) ? X : Y -> Y
7264 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7265 return ReplaceInstUsesWith(SI, FalseVal);
7266 // Transform (X != Y) ? X : Y -> X
7267 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7268 return ReplaceInstUsesWith(SI, TrueVal);
7269 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7271 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7272 // Transform (X == Y) ? Y : X -> X
7273 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7274 return ReplaceInstUsesWith(SI, FalseVal);
7275 // Transform (X != Y) ? Y : X -> Y
7276 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7277 return ReplaceInstUsesWith(SI, TrueVal);
7278 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7282 // See if we are selecting two values based on a comparison of the two values.
7283 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7284 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7285 // Transform (X == Y) ? X : Y -> Y
7286 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7287 return ReplaceInstUsesWith(SI, FalseVal);
7288 // Transform (X != Y) ? X : Y -> X
7289 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7290 return ReplaceInstUsesWith(SI, TrueVal);
7291 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7293 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7294 // Transform (X == Y) ? Y : X -> X
7295 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7296 return ReplaceInstUsesWith(SI, FalseVal);
7297 // Transform (X != Y) ? Y : X -> Y
7298 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7299 return ReplaceInstUsesWith(SI, TrueVal);
7300 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7304 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7305 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7306 if (TI->hasOneUse() && FI->hasOneUse()) {
7307 Instruction *AddOp = 0, *SubOp = 0;
7309 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7310 if (TI->getOpcode() == FI->getOpcode())
7311 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7314 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7315 // even legal for FP.
7316 if (TI->getOpcode() == Instruction::Sub &&
7317 FI->getOpcode() == Instruction::Add) {
7318 AddOp = FI; SubOp = TI;
7319 } else if (FI->getOpcode() == Instruction::Sub &&
7320 TI->getOpcode() == Instruction::Add) {
7321 AddOp = TI; SubOp = FI;
7325 Value *OtherAddOp = 0;
7326 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7327 OtherAddOp = AddOp->getOperand(1);
7328 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7329 OtherAddOp = AddOp->getOperand(0);
7333 // So at this point we know we have (Y -> OtherAddOp):
7334 // select C, (add X, Y), (sub X, Z)
7335 Value *NegVal; // Compute -Z
7336 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7337 NegVal = ConstantExpr::getNeg(C);
7339 NegVal = InsertNewInstBefore(
7340 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7343 Value *NewTrueOp = OtherAddOp;
7344 Value *NewFalseOp = NegVal;
7346 std::swap(NewTrueOp, NewFalseOp);
7347 Instruction *NewSel =
7348 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7350 NewSel = InsertNewInstBefore(NewSel, SI);
7351 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7356 // See if we can fold the select into one of our operands.
7357 if (SI.getType()->isInteger()) {
7358 // See the comment above GetSelectFoldableOperands for a description of the
7359 // transformation we are doing here.
7360 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7361 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7362 !isa<Constant>(FalseVal))
7363 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7364 unsigned OpToFold = 0;
7365 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7367 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7372 Constant *C = GetSelectFoldableConstant(TVI);
7373 Instruction *NewSel =
7374 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7375 InsertNewInstBefore(NewSel, SI);
7376 NewSel->takeName(TVI);
7377 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7378 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7380 assert(0 && "Unknown instruction!!");
7385 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7386 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7387 !isa<Constant>(TrueVal))
7388 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7389 unsigned OpToFold = 0;
7390 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7392 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7397 Constant *C = GetSelectFoldableConstant(FVI);
7398 Instruction *NewSel =
7399 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7400 InsertNewInstBefore(NewSel, SI);
7401 NewSel->takeName(FVI);
7402 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7403 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7405 assert(0 && "Unknown instruction!!");
7410 if (BinaryOperator::isNot(CondVal)) {
7411 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7412 SI.setOperand(1, FalseVal);
7413 SI.setOperand(2, TrueVal);
7420 /// GetKnownAlignment - If the specified pointer has an alignment that we can
7421 /// determine, return it, otherwise return 0.
7422 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
7423 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7424 unsigned Align = GV->getAlignment();
7425 if (Align == 0 && TD)
7426 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7428 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7429 unsigned Align = AI->getAlignment();
7430 if (Align == 0 && TD) {
7431 if (isa<AllocaInst>(AI))
7432 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7433 else if (isa<MallocInst>(AI)) {
7434 // Malloc returns maximally aligned memory.
7435 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7438 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7441 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7445 } else if (isa<BitCastInst>(V) ||
7446 (isa<ConstantExpr>(V) &&
7447 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7448 User *CI = cast<User>(V);
7449 if (isa<PointerType>(CI->getOperand(0)->getType()))
7450 return GetKnownAlignment(CI->getOperand(0), TD);
7452 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7453 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
7454 if (BaseAlignment == 0) return 0;
7456 // If all indexes are zero, it is just the alignment of the base pointer.
7457 bool AllZeroOperands = true;
7458 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7459 if (!isa<Constant>(GEPI->getOperand(i)) ||
7460 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7461 AllZeroOperands = false;
7464 if (AllZeroOperands)
7465 return BaseAlignment;
7467 // Otherwise, if the base alignment is >= the alignment we expect for the
7468 // base pointer type, then we know that the resultant pointer is aligned at
7469 // least as much as its type requires.
7472 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7473 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7474 if (TD->getABITypeAlignment(PtrTy->getElementType())
7476 const Type *GEPTy = GEPI->getType();
7477 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7478 return TD->getABITypeAlignment(GEPPtrTy->getElementType());
7486 /// visitCallInst - CallInst simplification. This mostly only handles folding
7487 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7488 /// the heavy lifting.
7490 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7491 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7492 if (!II) return visitCallSite(&CI);
7494 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7496 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7497 bool Changed = false;
7499 // memmove/cpy/set of zero bytes is a noop.
7500 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7501 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7503 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7504 if (CI->getZExtValue() == 1) {
7505 // Replace the instruction with just byte operations. We would
7506 // transform other cases to loads/stores, but we don't know if
7507 // alignment is sufficient.
7511 // If we have a memmove and the source operation is a constant global,
7512 // then the source and dest pointers can't alias, so we can change this
7513 // into a call to memcpy.
7514 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7515 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7516 if (GVSrc->isConstant()) {
7517 Module *M = CI.getParent()->getParent()->getParent();
7519 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7521 Name = "llvm.memcpy.i32";
7523 Name = "llvm.memcpy.i64";
7524 Constant *MemCpy = M->getOrInsertFunction(Name,
7525 CI.getCalledFunction()->getFunctionType());
7526 CI.setOperand(0, MemCpy);
7531 // If we can determine a pointer alignment that is bigger than currently
7532 // set, update the alignment.
7533 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7534 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
7535 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
7536 unsigned Align = std::min(Alignment1, Alignment2);
7537 if (MI->getAlignment()->getZExtValue() < Align) {
7538 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7541 } else if (isa<MemSetInst>(MI)) {
7542 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
7543 if (MI->getAlignment()->getZExtValue() < Alignment) {
7544 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7549 if (Changed) return II;
7551 switch (II->getIntrinsicID()) {
7553 case Intrinsic::ppc_altivec_lvx:
7554 case Intrinsic::ppc_altivec_lvxl:
7555 case Intrinsic::x86_sse_loadu_ps:
7556 case Intrinsic::x86_sse2_loadu_pd:
7557 case Intrinsic::x86_sse2_loadu_dq:
7558 // Turn PPC lvx -> load if the pointer is known aligned.
7559 // Turn X86 loadups -> load if the pointer is known aligned.
7560 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7561 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7562 PointerType::get(II->getType()), CI);
7563 return new LoadInst(Ptr);
7566 case Intrinsic::ppc_altivec_stvx:
7567 case Intrinsic::ppc_altivec_stvxl:
7568 // Turn stvx -> store if the pointer is known aligned.
7569 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
7570 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7571 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7573 return new StoreInst(II->getOperand(1), Ptr);
7576 case Intrinsic::x86_sse_storeu_ps:
7577 case Intrinsic::x86_sse2_storeu_pd:
7578 case Intrinsic::x86_sse2_storeu_dq:
7579 case Intrinsic::x86_sse2_storel_dq:
7580 // Turn X86 storeu -> store if the pointer is known aligned.
7581 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
7582 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7583 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7585 return new StoreInst(II->getOperand(2), Ptr);
7589 case Intrinsic::x86_sse_cvttss2si: {
7590 // These intrinsics only demands the 0th element of its input vector. If
7591 // we can simplify the input based on that, do so now.
7593 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7595 II->setOperand(1, V);
7601 case Intrinsic::ppc_altivec_vperm:
7602 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7603 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7604 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7606 // Check that all of the elements are integer constants or undefs.
7607 bool AllEltsOk = true;
7608 for (unsigned i = 0; i != 16; ++i) {
7609 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7610 !isa<UndefValue>(Mask->getOperand(i))) {
7617 // Cast the input vectors to byte vectors.
7618 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7619 II->getOperand(1), Mask->getType(), CI);
7620 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7621 II->getOperand(2), Mask->getType(), CI);
7622 Value *Result = UndefValue::get(Op0->getType());
7624 // Only extract each element once.
7625 Value *ExtractedElts[32];
7626 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7628 for (unsigned i = 0; i != 16; ++i) {
7629 if (isa<UndefValue>(Mask->getOperand(i)))
7631 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7632 Idx &= 31; // Match the hardware behavior.
7634 if (ExtractedElts[Idx] == 0) {
7636 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7637 InsertNewInstBefore(Elt, CI);
7638 ExtractedElts[Idx] = Elt;
7641 // Insert this value into the result vector.
7642 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7643 InsertNewInstBefore(cast<Instruction>(Result), CI);
7645 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7650 case Intrinsic::stackrestore: {
7651 // If the save is right next to the restore, remove the restore. This can
7652 // happen when variable allocas are DCE'd.
7653 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7654 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7655 BasicBlock::iterator BI = SS;
7657 return EraseInstFromFunction(CI);
7661 // If the stack restore is in a return/unwind block and if there are no
7662 // allocas or calls between the restore and the return, nuke the restore.
7663 TerminatorInst *TI = II->getParent()->getTerminator();
7664 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7665 BasicBlock::iterator BI = II;
7666 bool CannotRemove = false;
7667 for (++BI; &*BI != TI; ++BI) {
7668 if (isa<AllocaInst>(BI) ||
7669 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7670 CannotRemove = true;
7675 return EraseInstFromFunction(CI);
7682 return visitCallSite(II);
7685 // InvokeInst simplification
7687 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7688 return visitCallSite(&II);
7691 // visitCallSite - Improvements for call and invoke instructions.
7693 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7694 bool Changed = false;
7696 // If the callee is a constexpr cast of a function, attempt to move the cast
7697 // to the arguments of the call/invoke.
7698 if (transformConstExprCastCall(CS)) return 0;
7700 Value *Callee = CS.getCalledValue();
7702 if (Function *CalleeF = dyn_cast<Function>(Callee))
7703 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7704 Instruction *OldCall = CS.getInstruction();
7705 // If the call and callee calling conventions don't match, this call must
7706 // be unreachable, as the call is undefined.
7707 new StoreInst(ConstantInt::getTrue(),
7708 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7709 if (!OldCall->use_empty())
7710 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7711 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7712 return EraseInstFromFunction(*OldCall);
7716 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7717 // This instruction is not reachable, just remove it. We insert a store to
7718 // undef so that we know that this code is not reachable, despite the fact
7719 // that we can't modify the CFG here.
7720 new StoreInst(ConstantInt::getTrue(),
7721 UndefValue::get(PointerType::get(Type::Int1Ty)),
7722 CS.getInstruction());
7724 if (!CS.getInstruction()->use_empty())
7725 CS.getInstruction()->
7726 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7728 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7729 // Don't break the CFG, insert a dummy cond branch.
7730 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7731 ConstantInt::getTrue(), II);
7733 return EraseInstFromFunction(*CS.getInstruction());
7736 const PointerType *PTy = cast<PointerType>(Callee->getType());
7737 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7738 if (FTy->isVarArg()) {
7739 // See if we can optimize any arguments passed through the varargs area of
7741 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7742 E = CS.arg_end(); I != E; ++I)
7743 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7744 // If this cast does not effect the value passed through the varargs
7745 // area, we can eliminate the use of the cast.
7746 Value *Op = CI->getOperand(0);
7747 if (CI->isLosslessCast()) {
7754 return Changed ? CS.getInstruction() : 0;
7757 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7758 // attempt to move the cast to the arguments of the call/invoke.
7760 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7761 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7762 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7763 if (CE->getOpcode() != Instruction::BitCast ||
7764 !isa<Function>(CE->getOperand(0)))
7766 Function *Callee = cast<Function>(CE->getOperand(0));
7767 Instruction *Caller = CS.getInstruction();
7769 // Okay, this is a cast from a function to a different type. Unless doing so
7770 // would cause a type conversion of one of our arguments, change this call to
7771 // be a direct call with arguments casted to the appropriate types.
7773 const FunctionType *FT = Callee->getFunctionType();
7774 const Type *OldRetTy = Caller->getType();
7776 const FunctionType *ActualFT =
7777 cast<FunctionType>(cast<PointerType>(CE->getType())->getElementType());
7779 // If the parameter attributes don't match up, don't do the xform. We don't
7780 // want to lose an sret attribute or something.
7781 if (FT->getParamAttrs() != ActualFT->getParamAttrs())
7784 // Check to see if we are changing the return type...
7785 if (OldRetTy != FT->getReturnType()) {
7786 if (Callee->isDeclaration() && !Caller->use_empty() &&
7787 // Conversion is ok if changing from pointer to int of same size.
7788 !(isa<PointerType>(FT->getReturnType()) &&
7789 TD->getIntPtrType() == OldRetTy))
7790 return false; // Cannot transform this return value.
7792 // If the callsite is an invoke instruction, and the return value is used by
7793 // a PHI node in a successor, we cannot change the return type of the call
7794 // because there is no place to put the cast instruction (without breaking
7795 // the critical edge). Bail out in this case.
7796 if (!Caller->use_empty())
7797 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
7798 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
7800 if (PHINode *PN = dyn_cast<PHINode>(*UI))
7801 if (PN->getParent() == II->getNormalDest() ||
7802 PN->getParent() == II->getUnwindDest())
7806 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
7807 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
7809 CallSite::arg_iterator AI = CS.arg_begin();
7810 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
7811 const Type *ParamTy = FT->getParamType(i);
7812 const Type *ActTy = (*AI)->getType();
7813 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
7814 //Some conversions are safe even if we do not have a body.
7815 //Either we can cast directly, or we can upconvert the argument
7816 bool isConvertible = ActTy == ParamTy ||
7817 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
7818 (ParamTy->isInteger() && ActTy->isInteger() &&
7819 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
7820 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
7821 && c->getValue().isStrictlyPositive());
7822 if (Callee->isDeclaration() && !isConvertible) return false;
7824 // Most other conversions can be done if we have a body, even if these
7825 // lose information, e.g. int->short.
7826 // Some conversions cannot be done at all, e.g. float to pointer.
7827 // Logic here parallels CastInst::getCastOpcode (the design there
7828 // requires legality checks like this be done before calling it).
7829 if (ParamTy->isInteger()) {
7830 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7831 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7834 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
7835 !isa<PointerType>(ActTy))
7837 } else if (ParamTy->isFloatingPoint()) {
7838 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7839 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
7842 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
7844 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
7845 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
7846 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
7849 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
7851 } else if (isa<PointerType>(ParamTy)) {
7852 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
7859 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
7860 Callee->isDeclaration())
7861 return false; // Do not delete arguments unless we have a function body...
7863 // Okay, we decided that this is a safe thing to do: go ahead and start
7864 // inserting cast instructions as necessary...
7865 std::vector<Value*> Args;
7866 Args.reserve(NumActualArgs);
7868 AI = CS.arg_begin();
7869 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7870 const Type *ParamTy = FT->getParamType(i);
7871 if ((*AI)->getType() == ParamTy) {
7872 Args.push_back(*AI);
7874 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7875 false, ParamTy, false);
7876 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7877 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7881 // If the function takes more arguments than the call was taking, add them
7883 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7884 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7886 // If we are removing arguments to the function, emit an obnoxious warning...
7887 if (FT->getNumParams() < NumActualArgs)
7888 if (!FT->isVarArg()) {
7889 cerr << "WARNING: While resolving call to function '"
7890 << Callee->getName() << "' arguments were dropped!\n";
7892 // Add all of the arguments in their promoted form to the arg list...
7893 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7894 const Type *PTy = getPromotedType((*AI)->getType());
7895 if (PTy != (*AI)->getType()) {
7896 // Must promote to pass through va_arg area!
7897 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
7899 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7900 InsertNewInstBefore(Cast, *Caller);
7901 Args.push_back(Cast);
7903 Args.push_back(*AI);
7908 if (FT->getReturnType() == Type::VoidTy)
7909 Caller->setName(""); // Void type should not have a name.
7912 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7913 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7914 &Args[0], Args.size(), Caller->getName(), Caller);
7915 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7917 NC = new CallInst(Callee, &Args[0], Args.size(), Caller->getName(), Caller);
7918 if (cast<CallInst>(Caller)->isTailCall())
7919 cast<CallInst>(NC)->setTailCall();
7920 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7923 // Insert a cast of the return type as necessary.
7925 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7926 if (NV->getType() != Type::VoidTy) {
7927 const Type *CallerTy = Caller->getType();
7928 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
7930 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7932 // If this is an invoke instruction, we should insert it after the first
7933 // non-phi, instruction in the normal successor block.
7934 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7935 BasicBlock::iterator I = II->getNormalDest()->begin();
7936 while (isa<PHINode>(I)) ++I;
7937 InsertNewInstBefore(NC, *I);
7939 // Otherwise, it's a call, just insert cast right after the call instr
7940 InsertNewInstBefore(NC, *Caller);
7942 AddUsersToWorkList(*Caller);
7944 NV = UndefValue::get(Caller->getType());
7948 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7949 Caller->replaceAllUsesWith(NV);
7950 Caller->eraseFromParent();
7951 RemoveFromWorkList(Caller);
7955 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7956 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7957 /// and a single binop.
7958 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7959 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7960 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
7961 isa<CmpInst>(FirstInst));
7962 unsigned Opc = FirstInst->getOpcode();
7963 Value *LHSVal = FirstInst->getOperand(0);
7964 Value *RHSVal = FirstInst->getOperand(1);
7966 const Type *LHSType = LHSVal->getType();
7967 const Type *RHSType = RHSVal->getType();
7969 // Scan to see if all operands are the same opcode, all have one use, and all
7970 // kill their operands (i.e. the operands have one use).
7971 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7972 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7973 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7974 // Verify type of the LHS matches so we don't fold cmp's of different
7975 // types or GEP's with different index types.
7976 I->getOperand(0)->getType() != LHSType ||
7977 I->getOperand(1)->getType() != RHSType)
7980 // If they are CmpInst instructions, check their predicates
7981 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
7982 if (cast<CmpInst>(I)->getPredicate() !=
7983 cast<CmpInst>(FirstInst)->getPredicate())
7986 // Keep track of which operand needs a phi node.
7987 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7988 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7991 // Otherwise, this is safe to transform, determine if it is profitable.
7993 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7994 // Indexes are often folded into load/store instructions, so we don't want to
7995 // hide them behind a phi.
7996 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7999 Value *InLHS = FirstInst->getOperand(0);
8000 Value *InRHS = FirstInst->getOperand(1);
8001 PHINode *NewLHS = 0, *NewRHS = 0;
8003 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8004 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8005 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8006 InsertNewInstBefore(NewLHS, PN);
8011 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8012 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8013 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8014 InsertNewInstBefore(NewRHS, PN);
8018 // Add all operands to the new PHIs.
8019 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8021 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8022 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8025 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8026 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8030 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8031 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8032 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8033 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8036 assert(isa<GetElementPtrInst>(FirstInst));
8037 return new GetElementPtrInst(LHSVal, RHSVal);
8041 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8042 /// of the block that defines it. This means that it must be obvious the value
8043 /// of the load is not changed from the point of the load to the end of the
8046 /// Finally, it is safe, but not profitable, to sink a load targetting a
8047 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8049 static bool isSafeToSinkLoad(LoadInst *L) {
8050 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8052 for (++BBI; BBI != E; ++BBI)
8053 if (BBI->mayWriteToMemory())
8056 // Check for non-address taken alloca. If not address-taken already, it isn't
8057 // profitable to do this xform.
8058 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8059 bool isAddressTaken = false;
8060 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8062 if (isa<LoadInst>(UI)) continue;
8063 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8064 // If storing TO the alloca, then the address isn't taken.
8065 if (SI->getOperand(1) == AI) continue;
8067 isAddressTaken = true;
8071 if (!isAddressTaken)
8079 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8080 // operator and they all are only used by the PHI, PHI together their
8081 // inputs, and do the operation once, to the result of the PHI.
8082 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8083 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8085 // Scan the instruction, looking for input operations that can be folded away.
8086 // If all input operands to the phi are the same instruction (e.g. a cast from
8087 // the same type or "+42") we can pull the operation through the PHI, reducing
8088 // code size and simplifying code.
8089 Constant *ConstantOp = 0;
8090 const Type *CastSrcTy = 0;
8091 bool isVolatile = false;
8092 if (isa<CastInst>(FirstInst)) {
8093 CastSrcTy = FirstInst->getOperand(0)->getType();
8094 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8095 // Can fold binop, compare or shift here if the RHS is a constant,
8096 // otherwise call FoldPHIArgBinOpIntoPHI.
8097 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8098 if (ConstantOp == 0)
8099 return FoldPHIArgBinOpIntoPHI(PN);
8100 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8101 isVolatile = LI->isVolatile();
8102 // We can't sink the load if the loaded value could be modified between the
8103 // load and the PHI.
8104 if (LI->getParent() != PN.getIncomingBlock(0) ||
8105 !isSafeToSinkLoad(LI))
8107 } else if (isa<GetElementPtrInst>(FirstInst)) {
8108 if (FirstInst->getNumOperands() == 2)
8109 return FoldPHIArgBinOpIntoPHI(PN);
8110 // Can't handle general GEPs yet.
8113 return 0; // Cannot fold this operation.
8116 // Check to see if all arguments are the same operation.
8117 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8118 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8119 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8120 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8123 if (I->getOperand(0)->getType() != CastSrcTy)
8124 return 0; // Cast operation must match.
8125 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8126 // We can't sink the load if the loaded value could be modified between
8127 // the load and the PHI.
8128 if (LI->isVolatile() != isVolatile ||
8129 LI->getParent() != PN.getIncomingBlock(i) ||
8130 !isSafeToSinkLoad(LI))
8132 } else if (I->getOperand(1) != ConstantOp) {
8137 // Okay, they are all the same operation. Create a new PHI node of the
8138 // correct type, and PHI together all of the LHS's of the instructions.
8139 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8140 PN.getName()+".in");
8141 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8143 Value *InVal = FirstInst->getOperand(0);
8144 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8146 // Add all operands to the new PHI.
8147 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8148 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8149 if (NewInVal != InVal)
8151 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8156 // The new PHI unions all of the same values together. This is really
8157 // common, so we handle it intelligently here for compile-time speed.
8161 InsertNewInstBefore(NewPN, PN);
8165 // Insert and return the new operation.
8166 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8167 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8168 else if (isa<LoadInst>(FirstInst))
8169 return new LoadInst(PhiVal, "", isVolatile);
8170 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8171 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8172 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8173 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8174 PhiVal, ConstantOp);
8176 assert(0 && "Unknown operation");
8180 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8182 static bool DeadPHICycle(PHINode *PN,
8183 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8184 if (PN->use_empty()) return true;
8185 if (!PN->hasOneUse()) return false;
8187 // Remember this node, and if we find the cycle, return.
8188 if (!PotentiallyDeadPHIs.insert(PN))
8191 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8192 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8197 // PHINode simplification
8199 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8200 // If LCSSA is around, don't mess with Phi nodes
8201 if (MustPreserveLCSSA) return 0;
8203 if (Value *V = PN.hasConstantValue())
8204 return ReplaceInstUsesWith(PN, V);
8206 // If all PHI operands are the same operation, pull them through the PHI,
8207 // reducing code size.
8208 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8209 PN.getIncomingValue(0)->hasOneUse())
8210 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8213 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8214 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8215 // PHI)... break the cycle.
8216 if (PN.hasOneUse()) {
8217 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8218 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8219 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8220 PotentiallyDeadPHIs.insert(&PN);
8221 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8222 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8225 // If this phi has a single use, and if that use just computes a value for
8226 // the next iteration of a loop, delete the phi. This occurs with unused
8227 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8228 // common case here is good because the only other things that catch this
8229 // are induction variable analysis (sometimes) and ADCE, which is only run
8231 if (PHIUser->hasOneUse() &&
8232 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8233 PHIUser->use_back() == &PN) {
8234 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8241 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8242 Instruction *InsertPoint,
8244 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8245 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8246 // We must cast correctly to the pointer type. Ensure that we
8247 // sign extend the integer value if it is smaller as this is
8248 // used for address computation.
8249 Instruction::CastOps opcode =
8250 (VTySize < PtrSize ? Instruction::SExt :
8251 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8252 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8256 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8257 Value *PtrOp = GEP.getOperand(0);
8258 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8259 // If so, eliminate the noop.
8260 if (GEP.getNumOperands() == 1)
8261 return ReplaceInstUsesWith(GEP, PtrOp);
8263 if (isa<UndefValue>(GEP.getOperand(0)))
8264 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8266 bool HasZeroPointerIndex = false;
8267 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8268 HasZeroPointerIndex = C->isNullValue();
8270 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8271 return ReplaceInstUsesWith(GEP, PtrOp);
8273 // Eliminate unneeded casts for indices.
8274 bool MadeChange = false;
8276 gep_type_iterator GTI = gep_type_begin(GEP);
8277 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8278 if (isa<SequentialType>(*GTI)) {
8279 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8280 if (CI->getOpcode() == Instruction::ZExt ||
8281 CI->getOpcode() == Instruction::SExt) {
8282 const Type *SrcTy = CI->getOperand(0)->getType();
8283 // We can eliminate a cast from i32 to i64 iff the target
8284 // is a 32-bit pointer target.
8285 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8287 GEP.setOperand(i, CI->getOperand(0));
8291 // If we are using a wider index than needed for this platform, shrink it
8292 // to what we need. If the incoming value needs a cast instruction,
8293 // insert it. This explicit cast can make subsequent optimizations more
8295 Value *Op = GEP.getOperand(i);
8296 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
8297 if (Constant *C = dyn_cast<Constant>(Op)) {
8298 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8301 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8303 GEP.setOperand(i, Op);
8308 if (MadeChange) return &GEP;
8310 // If this GEP instruction doesn't move the pointer, and if the input operand
8311 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8312 // real input to the dest type.
8313 if (GEP.hasAllZeroIndices() && isa<BitCastInst>(GEP.getOperand(0)))
8314 return new BitCastInst(cast<BitCastInst>(GEP.getOperand(0))->getOperand(0),
8317 // Combine Indices - If the source pointer to this getelementptr instruction
8318 // is a getelementptr instruction, combine the indices of the two
8319 // getelementptr instructions into a single instruction.
8321 SmallVector<Value*, 8> SrcGEPOperands;
8322 if (User *Src = dyn_castGetElementPtr(PtrOp))
8323 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8325 if (!SrcGEPOperands.empty()) {
8326 // Note that if our source is a gep chain itself that we wait for that
8327 // chain to be resolved before we perform this transformation. This
8328 // avoids us creating a TON of code in some cases.
8330 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8331 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8332 return 0; // Wait until our source is folded to completion.
8334 SmallVector<Value*, 8> Indices;
8336 // Find out whether the last index in the source GEP is a sequential idx.
8337 bool EndsWithSequential = false;
8338 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8339 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8340 EndsWithSequential = !isa<StructType>(*I);
8342 // Can we combine the two pointer arithmetics offsets?
8343 if (EndsWithSequential) {
8344 // Replace: gep (gep %P, long B), long A, ...
8345 // With: T = long A+B; gep %P, T, ...
8347 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8348 if (SO1 == Constant::getNullValue(SO1->getType())) {
8350 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8353 // If they aren't the same type, convert both to an integer of the
8354 // target's pointer size.
8355 if (SO1->getType() != GO1->getType()) {
8356 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8357 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8358 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8359 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8361 unsigned PS = TD->getPointerSize();
8362 if (TD->getTypeSize(SO1->getType()) == PS) {
8363 // Convert GO1 to SO1's type.
8364 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8366 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8367 // Convert SO1 to GO1's type.
8368 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8370 const Type *PT = TD->getIntPtrType();
8371 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8372 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8376 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8377 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8379 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8380 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8384 // Recycle the GEP we already have if possible.
8385 if (SrcGEPOperands.size() == 2) {
8386 GEP.setOperand(0, SrcGEPOperands[0]);
8387 GEP.setOperand(1, Sum);
8390 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8391 SrcGEPOperands.end()-1);
8392 Indices.push_back(Sum);
8393 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8395 } else if (isa<Constant>(*GEP.idx_begin()) &&
8396 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8397 SrcGEPOperands.size() != 1) {
8398 // Otherwise we can do the fold if the first index of the GEP is a zero
8399 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8400 SrcGEPOperands.end());
8401 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8404 if (!Indices.empty())
8405 return new GetElementPtrInst(SrcGEPOperands[0], &Indices[0],
8406 Indices.size(), GEP.getName());
8408 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8409 // GEP of global variable. If all of the indices for this GEP are
8410 // constants, we can promote this to a constexpr instead of an instruction.
8412 // Scan for nonconstants...
8413 SmallVector<Constant*, 8> Indices;
8414 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8415 for (; I != E && isa<Constant>(*I); ++I)
8416 Indices.push_back(cast<Constant>(*I));
8418 if (I == E) { // If they are all constants...
8419 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8420 &Indices[0],Indices.size());
8422 // Replace all uses of the GEP with the new constexpr...
8423 return ReplaceInstUsesWith(GEP, CE);
8425 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8426 if (!isa<PointerType>(X->getType())) {
8427 // Not interesting. Source pointer must be a cast from pointer.
8428 } else if (HasZeroPointerIndex) {
8429 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8430 // into : GEP [10 x ubyte]* X, long 0, ...
8432 // This occurs when the program declares an array extern like "int X[];"
8434 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8435 const PointerType *XTy = cast<PointerType>(X->getType());
8436 if (const ArrayType *XATy =
8437 dyn_cast<ArrayType>(XTy->getElementType()))
8438 if (const ArrayType *CATy =
8439 dyn_cast<ArrayType>(CPTy->getElementType()))
8440 if (CATy->getElementType() == XATy->getElementType()) {
8441 // At this point, we know that the cast source type is a pointer
8442 // to an array of the same type as the destination pointer
8443 // array. Because the array type is never stepped over (there
8444 // is a leading zero) we can fold the cast into this GEP.
8445 GEP.setOperand(0, X);
8448 } else if (GEP.getNumOperands() == 2) {
8449 // Transform things like:
8450 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8451 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8452 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8453 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8454 if (isa<ArrayType>(SrcElTy) &&
8455 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8456 TD->getTypeSize(ResElTy)) {
8457 Value *V = InsertNewInstBefore(
8458 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8459 GEP.getOperand(1), GEP.getName()), GEP);
8460 // V and GEP are both pointer types --> BitCast
8461 return new BitCastInst(V, GEP.getType());
8464 // Transform things like:
8465 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8466 // (where tmp = 8*tmp2) into:
8467 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8469 if (isa<ArrayType>(SrcElTy) &&
8470 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8471 uint64_t ArrayEltSize =
8472 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8474 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8475 // allow either a mul, shift, or constant here.
8477 ConstantInt *Scale = 0;
8478 if (ArrayEltSize == 1) {
8479 NewIdx = GEP.getOperand(1);
8480 Scale = ConstantInt::get(NewIdx->getType(), 1);
8481 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8482 NewIdx = ConstantInt::get(CI->getType(), 1);
8484 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8485 if (Inst->getOpcode() == Instruction::Shl &&
8486 isa<ConstantInt>(Inst->getOperand(1))) {
8487 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8488 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8489 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8490 NewIdx = Inst->getOperand(0);
8491 } else if (Inst->getOpcode() == Instruction::Mul &&
8492 isa<ConstantInt>(Inst->getOperand(1))) {
8493 Scale = cast<ConstantInt>(Inst->getOperand(1));
8494 NewIdx = Inst->getOperand(0);
8498 // If the index will be to exactly the right offset with the scale taken
8499 // out, perform the transformation.
8500 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8501 if (isa<ConstantInt>(Scale))
8502 Scale = ConstantInt::get(Scale->getType(),
8503 Scale->getZExtValue() / ArrayEltSize);
8504 if (Scale->getZExtValue() != 1) {
8505 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8507 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8508 NewIdx = InsertNewInstBefore(Sc, GEP);
8511 // Insert the new GEP instruction.
8512 Instruction *NewGEP =
8513 new GetElementPtrInst(X, Constant::getNullValue(Type::Int32Ty),
8514 NewIdx, GEP.getName());
8515 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8516 // The NewGEP must be pointer typed, so must the old one -> BitCast
8517 return new BitCastInst(NewGEP, GEP.getType());
8526 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8527 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8528 if (AI.isArrayAllocation()) // Check C != 1
8529 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8531 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8532 AllocationInst *New = 0;
8534 // Create and insert the replacement instruction...
8535 if (isa<MallocInst>(AI))
8536 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8538 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8539 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8542 InsertNewInstBefore(New, AI);
8544 // Scan to the end of the allocation instructions, to skip over a block of
8545 // allocas if possible...
8547 BasicBlock::iterator It = New;
8548 while (isa<AllocationInst>(*It)) ++It;
8550 // Now that I is pointing to the first non-allocation-inst in the block,
8551 // insert our getelementptr instruction...
8553 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8554 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
8555 New->getName()+".sub", It);
8557 // Now make everything use the getelementptr instead of the original
8559 return ReplaceInstUsesWith(AI, V);
8560 } else if (isa<UndefValue>(AI.getArraySize())) {
8561 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8564 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8565 // Note that we only do this for alloca's, because malloc should allocate and
8566 // return a unique pointer, even for a zero byte allocation.
8567 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8568 TD->getTypeSize(AI.getAllocatedType()) == 0)
8569 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8574 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8575 Value *Op = FI.getOperand(0);
8577 // free undef -> unreachable.
8578 if (isa<UndefValue>(Op)) {
8579 // Insert a new store to null because we cannot modify the CFG here.
8580 new StoreInst(ConstantInt::getTrue(),
8581 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8582 return EraseInstFromFunction(FI);
8585 // If we have 'free null' delete the instruction. This can happen in stl code
8586 // when lots of inlining happens.
8587 if (isa<ConstantPointerNull>(Op))
8588 return EraseInstFromFunction(FI);
8590 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8591 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
8592 FI.setOperand(0, CI->getOperand(0));
8596 // Change free (gep X, 0,0,0,0) into free(X)
8597 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
8598 if (GEPI->hasAllZeroIndices()) {
8599 AddToWorkList(GEPI);
8600 FI.setOperand(0, GEPI->getOperand(0));
8605 // Change free(malloc) into nothing, if the malloc has a single use.
8606 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
8607 if (MI->hasOneUse()) {
8608 EraseInstFromFunction(FI);
8609 return EraseInstFromFunction(*MI);
8616 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8617 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8618 User *CI = cast<User>(LI.getOperand(0));
8619 Value *CastOp = CI->getOperand(0);
8621 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8622 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8623 const Type *SrcPTy = SrcTy->getElementType();
8625 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8626 isa<VectorType>(DestPTy)) {
8627 // If the source is an array, the code below will not succeed. Check to
8628 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8630 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8631 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8632 if (ASrcTy->getNumElements() != 0) {
8634 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8635 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8636 SrcTy = cast<PointerType>(CastOp->getType());
8637 SrcPTy = SrcTy->getElementType();
8640 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8641 isa<VectorType>(SrcPTy)) &&
8642 // Do not allow turning this into a load of an integer, which is then
8643 // casted to a pointer, this pessimizes pointer analysis a lot.
8644 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8645 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8646 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8648 // Okay, we are casting from one integer or pointer type to another of
8649 // the same size. Instead of casting the pointer before the load, cast
8650 // the result of the loaded value.
8651 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8653 LI.isVolatile()),LI);
8654 // Now cast the result of the load.
8655 return new BitCastInst(NewLoad, LI.getType());
8662 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8663 /// from this value cannot trap. If it is not obviously safe to load from the
8664 /// specified pointer, we do a quick local scan of the basic block containing
8665 /// ScanFrom, to determine if the address is already accessed.
8666 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8667 // If it is an alloca or global variable, it is always safe to load from.
8668 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8670 // Otherwise, be a little bit agressive by scanning the local block where we
8671 // want to check to see if the pointer is already being loaded or stored
8672 // from/to. If so, the previous load or store would have already trapped,
8673 // so there is no harm doing an extra load (also, CSE will later eliminate
8674 // the load entirely).
8675 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8680 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8681 if (LI->getOperand(0) == V) return true;
8682 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8683 if (SI->getOperand(1) == V) return true;
8689 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8690 Value *Op = LI.getOperand(0);
8692 // load (cast X) --> cast (load X) iff safe
8693 if (isa<CastInst>(Op))
8694 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8697 // None of the following transforms are legal for volatile loads.
8698 if (LI.isVolatile()) return 0;
8700 if (&LI.getParent()->front() != &LI) {
8701 BasicBlock::iterator BBI = &LI; --BBI;
8702 // If the instruction immediately before this is a store to the same
8703 // address, do a simple form of store->load forwarding.
8704 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8705 if (SI->getOperand(1) == LI.getOperand(0))
8706 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8707 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8708 if (LIB->getOperand(0) == LI.getOperand(0))
8709 return ReplaceInstUsesWith(LI, LIB);
8712 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8713 if (isa<ConstantPointerNull>(GEPI->getOperand(0))) {
8714 // Insert a new store to null instruction before the load to indicate
8715 // that this code is not reachable. We do this instead of inserting
8716 // an unreachable instruction directly because we cannot modify the
8718 new StoreInst(UndefValue::get(LI.getType()),
8719 Constant::getNullValue(Op->getType()), &LI);
8720 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8723 if (Constant *C = dyn_cast<Constant>(Op)) {
8724 // load null/undef -> undef
8725 if ((C->isNullValue() || isa<UndefValue>(C))) {
8726 // Insert a new store to null instruction before the load to indicate that
8727 // this code is not reachable. We do this instead of inserting an
8728 // unreachable instruction directly because we cannot modify the CFG.
8729 new StoreInst(UndefValue::get(LI.getType()),
8730 Constant::getNullValue(Op->getType()), &LI);
8731 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8734 // Instcombine load (constant global) into the value loaded.
8735 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
8736 if (GV->isConstant() && !GV->isDeclaration())
8737 return ReplaceInstUsesWith(LI, GV->getInitializer());
8739 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
8740 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
8741 if (CE->getOpcode() == Instruction::GetElementPtr) {
8742 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
8743 if (GV->isConstant() && !GV->isDeclaration())
8745 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
8746 return ReplaceInstUsesWith(LI, V);
8747 if (CE->getOperand(0)->isNullValue()) {
8748 // Insert a new store to null instruction before the load to indicate
8749 // that this code is not reachable. We do this instead of inserting
8750 // an unreachable instruction directly because we cannot modify the
8752 new StoreInst(UndefValue::get(LI.getType()),
8753 Constant::getNullValue(Op->getType()), &LI);
8754 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8757 } else if (CE->isCast()) {
8758 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8763 if (Op->hasOneUse()) {
8764 // Change select and PHI nodes to select values instead of addresses: this
8765 // helps alias analysis out a lot, allows many others simplifications, and
8766 // exposes redundancy in the code.
8768 // Note that we cannot do the transformation unless we know that the
8769 // introduced loads cannot trap! Something like this is valid as long as
8770 // the condition is always false: load (select bool %C, int* null, int* %G),
8771 // but it would not be valid if we transformed it to load from null
8774 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
8775 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
8776 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
8777 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
8778 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
8779 SI->getOperand(1)->getName()+".val"), LI);
8780 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
8781 SI->getOperand(2)->getName()+".val"), LI);
8782 return new SelectInst(SI->getCondition(), V1, V2);
8785 // load (select (cond, null, P)) -> load P
8786 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
8787 if (C->isNullValue()) {
8788 LI.setOperand(0, SI->getOperand(2));
8792 // load (select (cond, P, null)) -> load P
8793 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
8794 if (C->isNullValue()) {
8795 LI.setOperand(0, SI->getOperand(1));
8803 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
8805 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
8806 User *CI = cast<User>(SI.getOperand(1));
8807 Value *CastOp = CI->getOperand(0);
8809 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8810 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8811 const Type *SrcPTy = SrcTy->getElementType();
8813 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
8814 // If the source is an array, the code below will not succeed. Check to
8815 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8817 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8818 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8819 if (ASrcTy->getNumElements() != 0) {
8821 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8822 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8823 SrcTy = cast<PointerType>(CastOp->getType());
8824 SrcPTy = SrcTy->getElementType();
8827 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
8828 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8829 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8831 // Okay, we are casting from one integer or pointer type to another of
8832 // the same size. Instead of casting the pointer before
8833 // the store, cast the value to be stored.
8835 Value *SIOp0 = SI.getOperand(0);
8836 Instruction::CastOps opcode = Instruction::BitCast;
8837 const Type* CastSrcTy = SIOp0->getType();
8838 const Type* CastDstTy = SrcPTy;
8839 if (isa<PointerType>(CastDstTy)) {
8840 if (CastSrcTy->isInteger())
8841 opcode = Instruction::IntToPtr;
8842 } else if (isa<IntegerType>(CastDstTy)) {
8843 if (isa<PointerType>(SIOp0->getType()))
8844 opcode = Instruction::PtrToInt;
8846 if (Constant *C = dyn_cast<Constant>(SIOp0))
8847 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
8849 NewCast = IC.InsertNewInstBefore(
8850 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
8852 return new StoreInst(NewCast, CastOp);
8859 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
8860 Value *Val = SI.getOperand(0);
8861 Value *Ptr = SI.getOperand(1);
8863 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
8864 EraseInstFromFunction(SI);
8869 // If the RHS is an alloca with a single use, zapify the store, making the
8871 if (Ptr->hasOneUse()) {
8872 if (isa<AllocaInst>(Ptr)) {
8873 EraseInstFromFunction(SI);
8878 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
8879 if (isa<AllocaInst>(GEP->getOperand(0)) &&
8880 GEP->getOperand(0)->hasOneUse()) {
8881 EraseInstFromFunction(SI);
8887 // Do really simple DSE, to catch cases where there are several consequtive
8888 // stores to the same location, separated by a few arithmetic operations. This
8889 // situation often occurs with bitfield accesses.
8890 BasicBlock::iterator BBI = &SI;
8891 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
8895 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
8896 // Prev store isn't volatile, and stores to the same location?
8897 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
8900 EraseInstFromFunction(*PrevSI);
8906 // If this is a load, we have to stop. However, if the loaded value is from
8907 // the pointer we're loading and is producing the pointer we're storing,
8908 // then *this* store is dead (X = load P; store X -> P).
8909 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8910 if (LI == Val && LI->getOperand(0) == Ptr) {
8911 EraseInstFromFunction(SI);
8915 // Otherwise, this is a load from some other location. Stores before it
8920 // Don't skip over loads or things that can modify memory.
8921 if (BBI->mayWriteToMemory())
8926 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
8928 // store X, null -> turns into 'unreachable' in SimplifyCFG
8929 if (isa<ConstantPointerNull>(Ptr)) {
8930 if (!isa<UndefValue>(Val)) {
8931 SI.setOperand(0, UndefValue::get(Val->getType()));
8932 if (Instruction *U = dyn_cast<Instruction>(Val))
8933 AddToWorkList(U); // Dropped a use.
8936 return 0; // Do not modify these!
8939 // store undef, Ptr -> noop
8940 if (isa<UndefValue>(Val)) {
8941 EraseInstFromFunction(SI);
8946 // If the pointer destination is a cast, see if we can fold the cast into the
8948 if (isa<CastInst>(Ptr))
8949 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8951 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8953 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8957 // If this store is the last instruction in the basic block, and if the block
8958 // ends with an unconditional branch, try to move it to the successor block.
8960 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8961 if (BI->isUnconditional())
8962 if (SimplifyStoreAtEndOfBlock(SI))
8963 return 0; // xform done!
8968 /// SimplifyStoreAtEndOfBlock - Turn things like:
8969 /// if () { *P = v1; } else { *P = v2 }
8970 /// into a phi node with a store in the successor.
8972 /// Simplify things like:
8973 /// *P = v1; if () { *P = v2; }
8974 /// into a phi node with a store in the successor.
8976 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
8977 BasicBlock *StoreBB = SI.getParent();
8979 // Check to see if the successor block has exactly two incoming edges. If
8980 // so, see if the other predecessor contains a store to the same location.
8981 // if so, insert a PHI node (if needed) and move the stores down.
8982 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
8984 // Determine whether Dest has exactly two predecessors and, if so, compute
8985 // the other predecessor.
8986 pred_iterator PI = pred_begin(DestBB);
8987 BasicBlock *OtherBB = 0;
8991 if (PI == pred_end(DestBB))
8994 if (*PI != StoreBB) {
8999 if (++PI != pred_end(DestBB))
9003 // Verify that the other block ends in a branch and is not otherwise empty.
9004 BasicBlock::iterator BBI = OtherBB->getTerminator();
9005 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9006 if (!OtherBr || BBI == OtherBB->begin())
9009 // If the other block ends in an unconditional branch, check for the 'if then
9010 // else' case. there is an instruction before the branch.
9011 StoreInst *OtherStore = 0;
9012 if (OtherBr->isUnconditional()) {
9013 // If this isn't a store, or isn't a store to the same location, bail out.
9015 OtherStore = dyn_cast<StoreInst>(BBI);
9016 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9019 // Otherwise, the other block ended with a conditional branch. If one of the
9020 // destinations is StoreBB, then we have the if/then case.
9021 if (OtherBr->getSuccessor(0) != StoreBB &&
9022 OtherBr->getSuccessor(1) != StoreBB)
9025 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9026 // if/then triangle. See if there is a store to the same ptr as SI that
9027 // lives in OtherBB.
9029 // Check to see if we find the matching store.
9030 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9031 if (OtherStore->getOperand(1) != SI.getOperand(1))
9035 // If we find something that may be using the stored value, or if we run
9036 // out of instructions, we can't do the xform.
9037 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9038 BBI == OtherBB->begin())
9042 // In order to eliminate the store in OtherBr, we have to
9043 // make sure nothing reads the stored value in StoreBB.
9044 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9045 // FIXME: This should really be AA driven.
9046 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9051 // Insert a PHI node now if we need it.
9052 Value *MergedVal = OtherStore->getOperand(0);
9053 if (MergedVal != SI.getOperand(0)) {
9054 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9055 PN->reserveOperandSpace(2);
9056 PN->addIncoming(SI.getOperand(0), SI.getParent());
9057 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9058 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9061 // Advance to a place where it is safe to insert the new store and
9063 BBI = DestBB->begin();
9064 while (isa<PHINode>(BBI)) ++BBI;
9065 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9066 OtherStore->isVolatile()), *BBI);
9068 // Nuke the old stores.
9069 EraseInstFromFunction(SI);
9070 EraseInstFromFunction(*OtherStore);
9076 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9077 // Change br (not X), label True, label False to: br X, label False, True
9079 BasicBlock *TrueDest;
9080 BasicBlock *FalseDest;
9081 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9082 !isa<Constant>(X)) {
9083 // Swap Destinations and condition...
9085 BI.setSuccessor(0, FalseDest);
9086 BI.setSuccessor(1, TrueDest);
9090 // Cannonicalize fcmp_one -> fcmp_oeq
9091 FCmpInst::Predicate FPred; Value *Y;
9092 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9093 TrueDest, FalseDest)))
9094 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9095 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9096 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9097 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9098 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9099 NewSCC->takeName(I);
9100 // Swap Destinations and condition...
9101 BI.setCondition(NewSCC);
9102 BI.setSuccessor(0, FalseDest);
9103 BI.setSuccessor(1, TrueDest);
9104 RemoveFromWorkList(I);
9105 I->eraseFromParent();
9106 AddToWorkList(NewSCC);
9110 // Cannonicalize icmp_ne -> icmp_eq
9111 ICmpInst::Predicate IPred;
9112 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9113 TrueDest, FalseDest)))
9114 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9115 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9116 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9117 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9118 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9119 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9120 NewSCC->takeName(I);
9121 // Swap Destinations and condition...
9122 BI.setCondition(NewSCC);
9123 BI.setSuccessor(0, FalseDest);
9124 BI.setSuccessor(1, TrueDest);
9125 RemoveFromWorkList(I);
9126 I->eraseFromParent();;
9127 AddToWorkList(NewSCC);
9134 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9135 Value *Cond = SI.getCondition();
9136 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9137 if (I->getOpcode() == Instruction::Add)
9138 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9139 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9140 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9141 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9143 SI.setOperand(0, I->getOperand(0));
9151 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9152 /// is to leave as a vector operation.
9153 static bool CheapToScalarize(Value *V, bool isConstant) {
9154 if (isa<ConstantAggregateZero>(V))
9156 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9157 if (isConstant) return true;
9158 // If all elts are the same, we can extract.
9159 Constant *Op0 = C->getOperand(0);
9160 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9161 if (C->getOperand(i) != Op0)
9165 Instruction *I = dyn_cast<Instruction>(V);
9166 if (!I) return false;
9168 // Insert element gets simplified to the inserted element or is deleted if
9169 // this is constant idx extract element and its a constant idx insertelt.
9170 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9171 isa<ConstantInt>(I->getOperand(2)))
9173 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9175 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9176 if (BO->hasOneUse() &&
9177 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9178 CheapToScalarize(BO->getOperand(1), isConstant)))
9180 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9181 if (CI->hasOneUse() &&
9182 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9183 CheapToScalarize(CI->getOperand(1), isConstant)))
9189 /// Read and decode a shufflevector mask.
9191 /// It turns undef elements into values that are larger than the number of
9192 /// elements in the input.
9193 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9194 unsigned NElts = SVI->getType()->getNumElements();
9195 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9196 return std::vector<unsigned>(NElts, 0);
9197 if (isa<UndefValue>(SVI->getOperand(2)))
9198 return std::vector<unsigned>(NElts, 2*NElts);
9200 std::vector<unsigned> Result;
9201 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9202 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9203 if (isa<UndefValue>(CP->getOperand(i)))
9204 Result.push_back(NElts*2); // undef -> 8
9206 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9210 /// FindScalarElement - Given a vector and an element number, see if the scalar
9211 /// value is already around as a register, for example if it were inserted then
9212 /// extracted from the vector.
9213 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9214 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9215 const VectorType *PTy = cast<VectorType>(V->getType());
9216 unsigned Width = PTy->getNumElements();
9217 if (EltNo >= Width) // Out of range access.
9218 return UndefValue::get(PTy->getElementType());
9220 if (isa<UndefValue>(V))
9221 return UndefValue::get(PTy->getElementType());
9222 else if (isa<ConstantAggregateZero>(V))
9223 return Constant::getNullValue(PTy->getElementType());
9224 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9225 return CP->getOperand(EltNo);
9226 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9227 // If this is an insert to a variable element, we don't know what it is.
9228 if (!isa<ConstantInt>(III->getOperand(2)))
9230 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9232 // If this is an insert to the element we are looking for, return the
9235 return III->getOperand(1);
9237 // Otherwise, the insertelement doesn't modify the value, recurse on its
9239 return FindScalarElement(III->getOperand(0), EltNo);
9240 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9241 unsigned InEl = getShuffleMask(SVI)[EltNo];
9243 return FindScalarElement(SVI->getOperand(0), InEl);
9244 else if (InEl < Width*2)
9245 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9247 return UndefValue::get(PTy->getElementType());
9250 // Otherwise, we don't know.
9254 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9256 // If packed val is undef, replace extract with scalar undef.
9257 if (isa<UndefValue>(EI.getOperand(0)))
9258 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9260 // If packed val is constant 0, replace extract with scalar 0.
9261 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9262 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9264 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9265 // If packed val is constant with uniform operands, replace EI
9266 // with that operand
9267 Constant *op0 = C->getOperand(0);
9268 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9269 if (C->getOperand(i) != op0) {
9274 return ReplaceInstUsesWith(EI, op0);
9277 // If extracting a specified index from the vector, see if we can recursively
9278 // find a previously computed scalar that was inserted into the vector.
9279 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9280 unsigned IndexVal = IdxC->getZExtValue();
9281 unsigned VectorWidth =
9282 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9284 // If this is extracting an invalid index, turn this into undef, to avoid
9285 // crashing the code below.
9286 if (IndexVal >= VectorWidth)
9287 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9289 // This instruction only demands the single element from the input vector.
9290 // If the input vector has a single use, simplify it based on this use
9292 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9294 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9297 EI.setOperand(0, V);
9302 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9303 return ReplaceInstUsesWith(EI, Elt);
9305 // If the this extractelement is directly using a bitcast from a vector of
9306 // the same number of elements, see if we can find the source element from
9307 // it. In this case, we will end up needing to bitcast the scalars.
9308 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9309 if (const VectorType *VT =
9310 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9311 if (VT->getNumElements() == VectorWidth)
9312 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9313 return new BitCastInst(Elt, EI.getType());
9317 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9318 if (I->hasOneUse()) {
9319 // Push extractelement into predecessor operation if legal and
9320 // profitable to do so
9321 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9322 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9323 if (CheapToScalarize(BO, isConstantElt)) {
9324 ExtractElementInst *newEI0 =
9325 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9326 EI.getName()+".lhs");
9327 ExtractElementInst *newEI1 =
9328 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9329 EI.getName()+".rhs");
9330 InsertNewInstBefore(newEI0, EI);
9331 InsertNewInstBefore(newEI1, EI);
9332 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9334 } else if (isa<LoadInst>(I)) {
9335 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9336 PointerType::get(EI.getType()), EI);
9337 GetElementPtrInst *GEP =
9338 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9339 InsertNewInstBefore(GEP, EI);
9340 return new LoadInst(GEP);
9343 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9344 // Extracting the inserted element?
9345 if (IE->getOperand(2) == EI.getOperand(1))
9346 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9347 // If the inserted and extracted elements are constants, they must not
9348 // be the same value, extract from the pre-inserted value instead.
9349 if (isa<Constant>(IE->getOperand(2)) &&
9350 isa<Constant>(EI.getOperand(1))) {
9351 AddUsesToWorkList(EI);
9352 EI.setOperand(0, IE->getOperand(0));
9355 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9356 // If this is extracting an element from a shufflevector, figure out where
9357 // it came from and extract from the appropriate input element instead.
9358 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9359 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9361 if (SrcIdx < SVI->getType()->getNumElements())
9362 Src = SVI->getOperand(0);
9363 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9364 SrcIdx -= SVI->getType()->getNumElements();
9365 Src = SVI->getOperand(1);
9367 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9369 return new ExtractElementInst(Src, SrcIdx);
9376 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9377 /// elements from either LHS or RHS, return the shuffle mask and true.
9378 /// Otherwise, return false.
9379 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9380 std::vector<Constant*> &Mask) {
9381 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9382 "Invalid CollectSingleShuffleElements");
9383 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9385 if (isa<UndefValue>(V)) {
9386 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9388 } else if (V == LHS) {
9389 for (unsigned i = 0; i != NumElts; ++i)
9390 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9392 } else if (V == RHS) {
9393 for (unsigned i = 0; i != NumElts; ++i)
9394 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9396 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9397 // If this is an insert of an extract from some other vector, include it.
9398 Value *VecOp = IEI->getOperand(0);
9399 Value *ScalarOp = IEI->getOperand(1);
9400 Value *IdxOp = IEI->getOperand(2);
9402 if (!isa<ConstantInt>(IdxOp))
9404 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9406 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9407 // Okay, we can handle this if the vector we are insertinting into is
9409 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9410 // If so, update the mask to reflect the inserted undef.
9411 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9414 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9415 if (isa<ConstantInt>(EI->getOperand(1)) &&
9416 EI->getOperand(0)->getType() == V->getType()) {
9417 unsigned ExtractedIdx =
9418 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9420 // This must be extracting from either LHS or RHS.
9421 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9422 // Okay, we can handle this if the vector we are insertinting into is
9424 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9425 // If so, update the mask to reflect the inserted value.
9426 if (EI->getOperand(0) == LHS) {
9427 Mask[InsertedIdx & (NumElts-1)] =
9428 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9430 assert(EI->getOperand(0) == RHS);
9431 Mask[InsertedIdx & (NumElts-1)] =
9432 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9441 // TODO: Handle shufflevector here!
9446 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9447 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9448 /// that computes V and the LHS value of the shuffle.
9449 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9451 assert(isa<VectorType>(V->getType()) &&
9452 (RHS == 0 || V->getType() == RHS->getType()) &&
9453 "Invalid shuffle!");
9454 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9456 if (isa<UndefValue>(V)) {
9457 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9459 } else if (isa<ConstantAggregateZero>(V)) {
9460 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9462 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9463 // If this is an insert of an extract from some other vector, include it.
9464 Value *VecOp = IEI->getOperand(0);
9465 Value *ScalarOp = IEI->getOperand(1);
9466 Value *IdxOp = IEI->getOperand(2);
9468 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9469 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9470 EI->getOperand(0)->getType() == V->getType()) {
9471 unsigned ExtractedIdx =
9472 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9473 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9475 // Either the extracted from or inserted into vector must be RHSVec,
9476 // otherwise we'd end up with a shuffle of three inputs.
9477 if (EI->getOperand(0) == RHS || RHS == 0) {
9478 RHS = EI->getOperand(0);
9479 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9480 Mask[InsertedIdx & (NumElts-1)] =
9481 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9486 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9487 // Everything but the extracted element is replaced with the RHS.
9488 for (unsigned i = 0; i != NumElts; ++i) {
9489 if (i != InsertedIdx)
9490 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9495 // If this insertelement is a chain that comes from exactly these two
9496 // vectors, return the vector and the effective shuffle.
9497 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9498 return EI->getOperand(0);
9503 // TODO: Handle shufflevector here!
9505 // Otherwise, can't do anything fancy. Return an identity vector.
9506 for (unsigned i = 0; i != NumElts; ++i)
9507 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9511 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9512 Value *VecOp = IE.getOperand(0);
9513 Value *ScalarOp = IE.getOperand(1);
9514 Value *IdxOp = IE.getOperand(2);
9516 // Inserting an undef or into an undefined place, remove this.
9517 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
9518 ReplaceInstUsesWith(IE, VecOp);
9520 // If the inserted element was extracted from some other vector, and if the
9521 // indexes are constant, try to turn this into a shufflevector operation.
9522 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9523 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9524 EI->getOperand(0)->getType() == IE.getType()) {
9525 unsigned NumVectorElts = IE.getType()->getNumElements();
9526 unsigned ExtractedIdx =
9527 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9528 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9530 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9531 return ReplaceInstUsesWith(IE, VecOp);
9533 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9534 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9536 // If we are extracting a value from a vector, then inserting it right
9537 // back into the same place, just use the input vector.
9538 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9539 return ReplaceInstUsesWith(IE, VecOp);
9541 // We could theoretically do this for ANY input. However, doing so could
9542 // turn chains of insertelement instructions into a chain of shufflevector
9543 // instructions, and right now we do not merge shufflevectors. As such,
9544 // only do this in a situation where it is clear that there is benefit.
9545 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9546 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9547 // the values of VecOp, except then one read from EIOp0.
9548 // Build a new shuffle mask.
9549 std::vector<Constant*> Mask;
9550 if (isa<UndefValue>(VecOp))
9551 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9553 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9554 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9557 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9558 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9559 ConstantVector::get(Mask));
9562 // If this insertelement isn't used by some other insertelement, turn it
9563 // (and any insertelements it points to), into one big shuffle.
9564 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9565 std::vector<Constant*> Mask;
9567 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9568 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9569 // We now have a shuffle of LHS, RHS, Mask.
9570 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9579 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9580 Value *LHS = SVI.getOperand(0);
9581 Value *RHS = SVI.getOperand(1);
9582 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9584 bool MadeChange = false;
9586 // Undefined shuffle mask -> undefined value.
9587 if (isa<UndefValue>(SVI.getOperand(2)))
9588 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9590 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9591 // the undef, change them to undefs.
9592 if (isa<UndefValue>(SVI.getOperand(1))) {
9593 // Scan to see if there are any references to the RHS. If so, replace them
9594 // with undef element refs and set MadeChange to true.
9595 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9596 if (Mask[i] >= e && Mask[i] != 2*e) {
9603 // Remap any references to RHS to use LHS.
9604 std::vector<Constant*> Elts;
9605 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9607 Elts.push_back(UndefValue::get(Type::Int32Ty));
9609 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9611 SVI.setOperand(2, ConstantVector::get(Elts));
9615 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9616 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9617 if (LHS == RHS || isa<UndefValue>(LHS)) {
9618 if (isa<UndefValue>(LHS) && LHS == RHS) {
9619 // shuffle(undef,undef,mask) -> undef.
9620 return ReplaceInstUsesWith(SVI, LHS);
9623 // Remap any references to RHS to use LHS.
9624 std::vector<Constant*> Elts;
9625 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9627 Elts.push_back(UndefValue::get(Type::Int32Ty));
9629 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9630 (Mask[i] < e && isa<UndefValue>(LHS)))
9631 Mask[i] = 2*e; // Turn into undef.
9633 Mask[i] &= (e-1); // Force to LHS.
9634 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9637 SVI.setOperand(0, SVI.getOperand(1));
9638 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9639 SVI.setOperand(2, ConstantVector::get(Elts));
9640 LHS = SVI.getOperand(0);
9641 RHS = SVI.getOperand(1);
9645 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9646 bool isLHSID = true, isRHSID = true;
9648 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9649 if (Mask[i] >= e*2) continue; // Ignore undef values.
9650 // Is this an identity shuffle of the LHS value?
9651 isLHSID &= (Mask[i] == i);
9653 // Is this an identity shuffle of the RHS value?
9654 isRHSID &= (Mask[i]-e == i);
9657 // Eliminate identity shuffles.
9658 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9659 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9661 // If the LHS is a shufflevector itself, see if we can combine it with this
9662 // one without producing an unusual shuffle. Here we are really conservative:
9663 // we are absolutely afraid of producing a shuffle mask not in the input
9664 // program, because the code gen may not be smart enough to turn a merged
9665 // shuffle into two specific shuffles: it may produce worse code. As such,
9666 // we only merge two shuffles if the result is one of the two input shuffle
9667 // masks. In this case, merging the shuffles just removes one instruction,
9668 // which we know is safe. This is good for things like turning:
9669 // (splat(splat)) -> splat.
9670 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9671 if (isa<UndefValue>(RHS)) {
9672 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9674 std::vector<unsigned> NewMask;
9675 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9677 NewMask.push_back(2*e);
9679 NewMask.push_back(LHSMask[Mask[i]]);
9681 // If the result mask is equal to the src shuffle or this shuffle mask, do
9683 if (NewMask == LHSMask || NewMask == Mask) {
9684 std::vector<Constant*> Elts;
9685 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9686 if (NewMask[i] >= e*2) {
9687 Elts.push_back(UndefValue::get(Type::Int32Ty));
9689 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9692 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9693 LHSSVI->getOperand(1),
9694 ConstantVector::get(Elts));
9699 return MadeChange ? &SVI : 0;
9705 /// TryToSinkInstruction - Try to move the specified instruction from its
9706 /// current block into the beginning of DestBlock, which can only happen if it's
9707 /// safe to move the instruction past all of the instructions between it and the
9708 /// end of its block.
9709 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
9710 assert(I->hasOneUse() && "Invariants didn't hold!");
9712 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
9713 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
9715 // Do not sink alloca instructions out of the entry block.
9716 if (isa<AllocaInst>(I) && I->getParent() ==
9717 &DestBlock->getParent()->getEntryBlock())
9720 // We can only sink load instructions if there is nothing between the load and
9721 // the end of block that could change the value.
9722 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
9723 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
9725 if (Scan->mayWriteToMemory())
9729 BasicBlock::iterator InsertPos = DestBlock->begin();
9730 while (isa<PHINode>(InsertPos)) ++InsertPos;
9732 I->moveBefore(InsertPos);
9738 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
9739 /// all reachable code to the worklist.
9741 /// This has a couple of tricks to make the code faster and more powerful. In
9742 /// particular, we constant fold and DCE instructions as we go, to avoid adding
9743 /// them to the worklist (this significantly speeds up instcombine on code where
9744 /// many instructions are dead or constant). Additionally, if we find a branch
9745 /// whose condition is a known constant, we only visit the reachable successors.
9747 static void AddReachableCodeToWorklist(BasicBlock *BB,
9748 SmallPtrSet<BasicBlock*, 64> &Visited,
9750 const TargetData *TD) {
9751 std::vector<BasicBlock*> Worklist;
9752 Worklist.push_back(BB);
9754 while (!Worklist.empty()) {
9755 BB = Worklist.back();
9756 Worklist.pop_back();
9758 // We have now visited this block! If we've already been here, ignore it.
9759 if (!Visited.insert(BB)) continue;
9761 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
9762 Instruction *Inst = BBI++;
9764 // DCE instruction if trivially dead.
9765 if (isInstructionTriviallyDead(Inst)) {
9767 DOUT << "IC: DCE: " << *Inst;
9768 Inst->eraseFromParent();
9772 // ConstantProp instruction if trivially constant.
9773 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
9774 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
9775 Inst->replaceAllUsesWith(C);
9777 Inst->eraseFromParent();
9781 IC.AddToWorkList(Inst);
9784 // Recursively visit successors. If this is a branch or switch on a
9785 // constant, only visit the reachable successor.
9786 TerminatorInst *TI = BB->getTerminator();
9787 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
9788 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
9789 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
9790 Worklist.push_back(BI->getSuccessor(!CondVal));
9793 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
9794 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
9795 // See if this is an explicit destination.
9796 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
9797 if (SI->getCaseValue(i) == Cond) {
9798 Worklist.push_back(SI->getSuccessor(i));
9802 // Otherwise it is the default destination.
9803 Worklist.push_back(SI->getSuccessor(0));
9808 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
9809 Worklist.push_back(TI->getSuccessor(i));
9813 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
9814 bool Changed = false;
9815 TD = &getAnalysis<TargetData>();
9817 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
9818 << F.getNameStr() << "\n");
9821 // Do a depth-first traversal of the function, populate the worklist with
9822 // the reachable instructions. Ignore blocks that are not reachable. Keep
9823 // track of which blocks we visit.
9824 SmallPtrSet<BasicBlock*, 64> Visited;
9825 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
9827 // Do a quick scan over the function. If we find any blocks that are
9828 // unreachable, remove any instructions inside of them. This prevents
9829 // the instcombine code from having to deal with some bad special cases.
9830 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
9831 if (!Visited.count(BB)) {
9832 Instruction *Term = BB->getTerminator();
9833 while (Term != BB->begin()) { // Remove instrs bottom-up
9834 BasicBlock::iterator I = Term; --I;
9836 DOUT << "IC: DCE: " << *I;
9839 if (!I->use_empty())
9840 I->replaceAllUsesWith(UndefValue::get(I->getType()));
9841 I->eraseFromParent();
9846 while (!Worklist.empty()) {
9847 Instruction *I = RemoveOneFromWorkList();
9848 if (I == 0) continue; // skip null values.
9850 // Check to see if we can DCE the instruction.
9851 if (isInstructionTriviallyDead(I)) {
9852 // Add operands to the worklist.
9853 if (I->getNumOperands() < 4)
9854 AddUsesToWorkList(*I);
9857 DOUT << "IC: DCE: " << *I;
9859 I->eraseFromParent();
9860 RemoveFromWorkList(I);
9864 // Instruction isn't dead, see if we can constant propagate it.
9865 if (Constant *C = ConstantFoldInstruction(I, TD)) {
9866 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
9868 // Add operands to the worklist.
9869 AddUsesToWorkList(*I);
9870 ReplaceInstUsesWith(*I, C);
9873 I->eraseFromParent();
9874 RemoveFromWorkList(I);
9878 // See if we can trivially sink this instruction to a successor basic block.
9879 if (I->hasOneUse()) {
9880 BasicBlock *BB = I->getParent();
9881 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
9882 if (UserParent != BB) {
9883 bool UserIsSuccessor = false;
9884 // See if the user is one of our successors.
9885 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
9886 if (*SI == UserParent) {
9887 UserIsSuccessor = true;
9891 // If the user is one of our immediate successors, and if that successor
9892 // only has us as a predecessors (we'd have to split the critical edge
9893 // otherwise), we can keep going.
9894 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
9895 next(pred_begin(UserParent)) == pred_end(UserParent))
9896 // Okay, the CFG is simple enough, try to sink this instruction.
9897 Changed |= TryToSinkInstruction(I, UserParent);
9901 // Now that we have an instruction, try combining it to simplify it...
9905 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
9906 if (Instruction *Result = visit(*I)) {
9908 // Should we replace the old instruction with a new one?
9910 DOUT << "IC: Old = " << *I
9911 << " New = " << *Result;
9913 // Everything uses the new instruction now.
9914 I->replaceAllUsesWith(Result);
9916 // Push the new instruction and any users onto the worklist.
9917 AddToWorkList(Result);
9918 AddUsersToWorkList(*Result);
9920 // Move the name to the new instruction first.
9921 Result->takeName(I);
9923 // Insert the new instruction into the basic block...
9924 BasicBlock *InstParent = I->getParent();
9925 BasicBlock::iterator InsertPos = I;
9927 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
9928 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
9931 InstParent->getInstList().insert(InsertPos, Result);
9933 // Make sure that we reprocess all operands now that we reduced their
9935 AddUsesToWorkList(*I);
9937 // Instructions can end up on the worklist more than once. Make sure
9938 // we do not process an instruction that has been deleted.
9939 RemoveFromWorkList(I);
9941 // Erase the old instruction.
9942 InstParent->getInstList().erase(I);
9945 DOUT << "IC: Mod = " << OrigI
9949 // If the instruction was modified, it's possible that it is now dead.
9950 // if so, remove it.
9951 if (isInstructionTriviallyDead(I)) {
9952 // Make sure we process all operands now that we are reducing their
9954 AddUsesToWorkList(*I);
9956 // Instructions may end up in the worklist more than once. Erase all
9957 // occurrences of this instruction.
9958 RemoveFromWorkList(I);
9959 I->eraseFromParent();
9962 AddUsersToWorkList(*I);
9969 assert(WorklistMap.empty() && "Worklist empty, but map not?");
9974 bool InstCombiner::runOnFunction(Function &F) {
9975 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
9977 bool EverMadeChange = false;
9979 // Iterate while there is work to do.
9980 unsigned Iteration = 0;
9981 while (DoOneIteration(F, Iteration++))
9982 EverMadeChange = true;
9983 return EverMadeChange;
9986 FunctionPass *llvm::createInstructionCombiningPass() {
9987 return new InstCombiner();