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/ParameterAttributes.h"
43 #include "llvm/Analysis/ConstantFolding.h"
44 #include "llvm/Target/TargetData.h"
45 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
46 #include "llvm/Transforms/Utils/Local.h"
47 #include "llvm/Support/CallSite.h"
48 #include "llvm/Support/Debug.h"
49 #include "llvm/Support/GetElementPtrTypeIterator.h"
50 #include "llvm/Support/InstVisitor.h"
51 #include "llvm/Support/MathExtras.h"
52 #include "llvm/Support/PatternMatch.h"
53 #include "llvm/Support/Compiler.h"
54 #include "llvm/ADT/DenseMap.h"
55 #include "llvm/ADT/SmallVector.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.h"
58 #include "llvm/ADT/STLExtras.h"
62 using namespace llvm::PatternMatch;
64 STATISTIC(NumCombined , "Number of insts combined");
65 STATISTIC(NumConstProp, "Number of constant folds");
66 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
67 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
68 STATISTIC(NumSunkInst , "Number of instructions sunk");
71 class VISIBILITY_HIDDEN InstCombiner
72 : public FunctionPass,
73 public InstVisitor<InstCombiner, Instruction*> {
74 // Worklist of all of the instructions that need to be simplified.
75 std::vector<Instruction*> Worklist;
76 DenseMap<Instruction*, unsigned> WorklistMap;
78 bool MustPreserveLCSSA;
80 static char ID; // Pass identification, replacement for typeid
81 InstCombiner() : FunctionPass((intptr_t)&ID) {}
83 /// AddToWorkList - Add the specified instruction to the worklist if it
84 /// isn't already in it.
85 void AddToWorkList(Instruction *I) {
86 if (WorklistMap.insert(std::make_pair(I, Worklist.size())))
87 Worklist.push_back(I);
90 // RemoveFromWorkList - remove I from the worklist if it exists.
91 void RemoveFromWorkList(Instruction *I) {
92 DenseMap<Instruction*, unsigned>::iterator It = WorklistMap.find(I);
93 if (It == WorklistMap.end()) return; // Not in worklist.
95 // Don't bother moving everything down, just null out the slot.
96 Worklist[It->second] = 0;
98 WorklistMap.erase(It);
101 Instruction *RemoveOneFromWorkList() {
102 Instruction *I = Worklist.back();
104 WorklistMap.erase(I);
109 /// AddUsersToWorkList - When an instruction is simplified, add all users of
110 /// the instruction to the work lists because they might get more simplified
113 void AddUsersToWorkList(Value &I) {
114 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
116 AddToWorkList(cast<Instruction>(*UI));
119 /// AddUsesToWorkList - When an instruction is simplified, add operands to
120 /// the work lists because they might get more simplified now.
122 void AddUsesToWorkList(Instruction &I) {
123 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
124 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
128 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
129 /// dead. Add all of its operands to the worklist, turning them into
130 /// undef's to reduce the number of uses of those instructions.
132 /// Return the specified operand before it is turned into an undef.
134 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
135 Value *R = I.getOperand(op);
137 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
138 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
140 // Set the operand to undef to drop the use.
141 I.setOperand(i, UndefValue::get(Op->getType()));
148 virtual bool runOnFunction(Function &F);
150 bool DoOneIteration(Function &F, unsigned ItNum);
152 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
153 AU.addRequired<TargetData>();
154 AU.addPreservedID(LCSSAID);
155 AU.setPreservesCFG();
158 TargetData &getTargetData() const { return *TD; }
160 // Visitation implementation - Implement instruction combining for different
161 // instruction types. The semantics are as follows:
163 // null - No change was made
164 // I - Change was made, I is still valid, I may be dead though
165 // otherwise - Change was made, replace I with returned instruction
167 Instruction *visitAdd(BinaryOperator &I);
168 Instruction *visitSub(BinaryOperator &I);
169 Instruction *visitMul(BinaryOperator &I);
170 Instruction *visitURem(BinaryOperator &I);
171 Instruction *visitSRem(BinaryOperator &I);
172 Instruction *visitFRem(BinaryOperator &I);
173 Instruction *commonRemTransforms(BinaryOperator &I);
174 Instruction *commonIRemTransforms(BinaryOperator &I);
175 Instruction *commonDivTransforms(BinaryOperator &I);
176 Instruction *commonIDivTransforms(BinaryOperator &I);
177 Instruction *visitUDiv(BinaryOperator &I);
178 Instruction *visitSDiv(BinaryOperator &I);
179 Instruction *visitFDiv(BinaryOperator &I);
180 Instruction *visitAnd(BinaryOperator &I);
181 Instruction *visitOr (BinaryOperator &I);
182 Instruction *visitXor(BinaryOperator &I);
183 Instruction *visitShl(BinaryOperator &I);
184 Instruction *visitAShr(BinaryOperator &I);
185 Instruction *visitLShr(BinaryOperator &I);
186 Instruction *commonShiftTransforms(BinaryOperator &I);
187 Instruction *visitFCmpInst(FCmpInst &I);
188 Instruction *visitICmpInst(ICmpInst &I);
189 Instruction *visitICmpInstWithCastAndCast(ICmpInst &ICI);
190 Instruction *visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
193 Instruction *FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
194 ConstantInt *DivRHS);
196 Instruction *FoldGEPICmp(User *GEPLHS, Value *RHS,
197 ICmpInst::Predicate Cond, Instruction &I);
198 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
200 Instruction *commonCastTransforms(CastInst &CI);
201 Instruction *commonIntCastTransforms(CastInst &CI);
202 Instruction *commonPointerCastTransforms(CastInst &CI);
203 Instruction *visitTrunc(TruncInst &CI);
204 Instruction *visitZExt(ZExtInst &CI);
205 Instruction *visitSExt(SExtInst &CI);
206 Instruction *visitFPTrunc(CastInst &CI);
207 Instruction *visitFPExt(CastInst &CI);
208 Instruction *visitFPToUI(CastInst &CI);
209 Instruction *visitFPToSI(CastInst &CI);
210 Instruction *visitUIToFP(CastInst &CI);
211 Instruction *visitSIToFP(CastInst &CI);
212 Instruction *visitPtrToInt(CastInst &CI);
213 Instruction *visitIntToPtr(CastInst &CI);
214 Instruction *visitBitCast(BitCastInst &CI);
215 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
217 Instruction *visitSelectInst(SelectInst &CI);
218 Instruction *visitCallInst(CallInst &CI);
219 Instruction *visitInvokeInst(InvokeInst &II);
220 Instruction *visitPHINode(PHINode &PN);
221 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
222 Instruction *visitAllocationInst(AllocationInst &AI);
223 Instruction *visitFreeInst(FreeInst &FI);
224 Instruction *visitLoadInst(LoadInst &LI);
225 Instruction *visitStoreInst(StoreInst &SI);
226 Instruction *visitBranchInst(BranchInst &BI);
227 Instruction *visitSwitchInst(SwitchInst &SI);
228 Instruction *visitInsertElementInst(InsertElementInst &IE);
229 Instruction *visitExtractElementInst(ExtractElementInst &EI);
230 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
232 // visitInstruction - Specify what to return for unhandled instructions...
233 Instruction *visitInstruction(Instruction &I) { return 0; }
236 Instruction *visitCallSite(CallSite CS);
237 bool transformConstExprCastCall(CallSite CS);
240 // InsertNewInstBefore - insert an instruction New before instruction Old
241 // in the program. Add the new instruction to the worklist.
243 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
244 assert(New && New->getParent() == 0 &&
245 "New instruction already inserted into a basic block!");
246 BasicBlock *BB = Old.getParent();
247 BB->getInstList().insert(&Old, New); // Insert inst
252 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
253 /// This also adds the cast to the worklist. Finally, this returns the
255 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
257 if (V->getType() == Ty) return V;
259 if (Constant *CV = dyn_cast<Constant>(V))
260 return ConstantExpr::getCast(opc, CV, Ty);
262 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
267 // ReplaceInstUsesWith - This method is to be used when an instruction is
268 // found to be dead, replacable with another preexisting expression. Here
269 // we add all uses of I to the worklist, replace all uses of I with the new
270 // value, then return I, so that the inst combiner will know that I was
273 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
274 AddUsersToWorkList(I); // Add all modified instrs to worklist
276 I.replaceAllUsesWith(V);
279 // If we are replacing the instruction with itself, this must be in a
280 // segment of unreachable code, so just clobber the instruction.
281 I.replaceAllUsesWith(UndefValue::get(I.getType()));
286 // UpdateValueUsesWith - This method is to be used when an value is
287 // found to be replacable with another preexisting expression or was
288 // updated. Here we add all uses of I to the worklist, replace all uses of
289 // I with the new value (unless the instruction was just updated), then
290 // return true, so that the inst combiner will know that I was modified.
292 bool UpdateValueUsesWith(Value *Old, Value *New) {
293 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
295 Old->replaceAllUsesWith(New);
296 if (Instruction *I = dyn_cast<Instruction>(Old))
298 if (Instruction *I = dyn_cast<Instruction>(New))
303 // EraseInstFromFunction - When dealing with an instruction that has side
304 // effects or produces a void value, we can't rely on DCE to delete the
305 // instruction. Instead, visit methods should return the value returned by
307 Instruction *EraseInstFromFunction(Instruction &I) {
308 assert(I.use_empty() && "Cannot erase instruction that is used!");
309 AddUsesToWorkList(I);
310 RemoveFromWorkList(&I);
312 return 0; // Don't do anything with FI
316 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
317 /// InsertBefore instruction. This is specialized a bit to avoid inserting
318 /// casts that are known to not do anything...
320 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
321 Value *V, const Type *DestTy,
322 Instruction *InsertBefore);
324 /// SimplifyCommutative - This performs a few simplifications for
325 /// commutative operators.
326 bool SimplifyCommutative(BinaryOperator &I);
328 /// SimplifyCompare - This reorders the operands of a CmpInst to get them in
329 /// most-complex to least-complex order.
330 bool SimplifyCompare(CmpInst &I);
332 /// SimplifyDemandedBits - Attempts to replace V with a simpler value based
333 /// on the demanded bits.
334 bool SimplifyDemandedBits(Value *V, APInt DemandedMask,
335 APInt& KnownZero, APInt& KnownOne,
338 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
339 uint64_t &UndefElts, unsigned Depth = 0);
341 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
342 // PHI node as operand #0, see if we can fold the instruction into the PHI
343 // (which is only possible if all operands to the PHI are constants).
344 Instruction *FoldOpIntoPhi(Instruction &I);
346 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
347 // operator and they all are only used by the PHI, PHI together their
348 // inputs, and do the operation once, to the result of the PHI.
349 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
350 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
353 Instruction *OptAndOp(Instruction *Op, ConstantInt *OpRHS,
354 ConstantInt *AndRHS, BinaryOperator &TheAnd);
356 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantInt *Mask,
357 bool isSub, Instruction &I);
358 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
359 bool isSigned, bool Inside, Instruction &IB);
360 Instruction *PromoteCastOfAllocation(BitCastInst &CI, AllocationInst &AI);
361 Instruction *MatchBSwap(BinaryOperator &I);
362 bool SimplifyStoreAtEndOfBlock(StoreInst &SI);
364 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
367 char InstCombiner::ID = 0;
368 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
371 // getComplexity: Assign a complexity or rank value to LLVM Values...
372 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
373 static unsigned getComplexity(Value *V) {
374 if (isa<Instruction>(V)) {
375 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
379 if (isa<Argument>(V)) return 3;
380 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
383 // isOnlyUse - Return true if this instruction will be deleted if we stop using
385 static bool isOnlyUse(Value *V) {
386 return V->hasOneUse() || isa<Constant>(V);
389 // getPromotedType - Return the specified type promoted as it would be to pass
390 // though a va_arg area...
391 static const Type *getPromotedType(const Type *Ty) {
392 if (const IntegerType* ITy = dyn_cast<IntegerType>(Ty)) {
393 if (ITy->getBitWidth() < 32)
394 return Type::Int32Ty;
399 /// getBitCastOperand - If the specified operand is a CastInst or a constant
400 /// expression bitcast, return the operand value, otherwise return null.
401 static Value *getBitCastOperand(Value *V) {
402 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
403 return I->getOperand(0);
404 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
405 if (CE->getOpcode() == Instruction::BitCast)
406 return CE->getOperand(0);
410 /// This function is a wrapper around CastInst::isEliminableCastPair. It
411 /// simply extracts arguments and returns what that function returns.
412 static Instruction::CastOps
413 isEliminableCastPair(
414 const CastInst *CI, ///< The first cast instruction
415 unsigned opcode, ///< The opcode of the second cast instruction
416 const Type *DstTy, ///< The target type for the second cast instruction
417 TargetData *TD ///< The target data for pointer size
420 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
421 const Type *MidTy = CI->getType(); // B from above
423 // Get the opcodes of the two Cast instructions
424 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
425 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
427 return Instruction::CastOps(
428 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
429 DstTy, TD->getIntPtrType()));
432 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
433 /// in any code being generated. It does not require codegen if V is simple
434 /// enough or if the cast can be folded into other casts.
435 static bool ValueRequiresCast(Instruction::CastOps opcode, const Value *V,
436 const Type *Ty, TargetData *TD) {
437 if (V->getType() == Ty || isa<Constant>(V)) return false;
439 // If this is another cast that can be eliminated, it isn't codegen either.
440 if (const CastInst *CI = dyn_cast<CastInst>(V))
441 if (isEliminableCastPair(CI, opcode, Ty, TD))
446 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
447 /// InsertBefore instruction. This is specialized a bit to avoid inserting
448 /// casts that are known to not do anything...
450 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
451 Value *V, const Type *DestTy,
452 Instruction *InsertBefore) {
453 if (V->getType() == DestTy) return V;
454 if (Constant *C = dyn_cast<Constant>(V))
455 return ConstantExpr::getCast(opcode, C, DestTy);
457 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
460 // SimplifyCommutative - This performs a few simplifications for commutative
463 // 1. Order operands such that they are listed from right (least complex) to
464 // left (most complex). This puts constants before unary operators before
467 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
468 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
470 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
471 bool Changed = false;
472 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
473 Changed = !I.swapOperands();
475 if (!I.isAssociative()) return Changed;
476 Instruction::BinaryOps Opcode = I.getOpcode();
477 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
478 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
479 if (isa<Constant>(I.getOperand(1))) {
480 Constant *Folded = ConstantExpr::get(I.getOpcode(),
481 cast<Constant>(I.getOperand(1)),
482 cast<Constant>(Op->getOperand(1)));
483 I.setOperand(0, Op->getOperand(0));
484 I.setOperand(1, Folded);
486 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
487 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
488 isOnlyUse(Op) && isOnlyUse(Op1)) {
489 Constant *C1 = cast<Constant>(Op->getOperand(1));
490 Constant *C2 = cast<Constant>(Op1->getOperand(1));
492 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
493 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
494 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
498 I.setOperand(0, New);
499 I.setOperand(1, Folded);
506 /// SimplifyCompare - For a CmpInst this function just orders the operands
507 /// so that theyare listed from right (least complex) to left (most complex).
508 /// This puts constants before unary operators before binary operators.
509 bool InstCombiner::SimplifyCompare(CmpInst &I) {
510 if (getComplexity(I.getOperand(0)) >= getComplexity(I.getOperand(1)))
513 // Compare instructions are not associative so there's nothing else we can do.
517 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
518 // if the LHS is a constant zero (which is the 'negate' form).
520 static inline Value *dyn_castNegVal(Value *V) {
521 if (BinaryOperator::isNeg(V))
522 return BinaryOperator::getNegArgument(V);
524 // Constants can be considered to be negated values if they can be folded.
525 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
526 return ConstantExpr::getNeg(C);
530 static inline Value *dyn_castNotVal(Value *V) {
531 if (BinaryOperator::isNot(V))
532 return BinaryOperator::getNotArgument(V);
534 // Constants can be considered to be not'ed values...
535 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
536 return ConstantInt::get(~C->getValue());
540 // dyn_castFoldableMul - If this value is a multiply that can be folded into
541 // other computations (because it has a constant operand), return the
542 // non-constant operand of the multiply, and set CST to point to the multiplier.
543 // Otherwise, return null.
545 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
546 if (V->hasOneUse() && V->getType()->isInteger())
547 if (Instruction *I = dyn_cast<Instruction>(V)) {
548 if (I->getOpcode() == Instruction::Mul)
549 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
550 return I->getOperand(0);
551 if (I->getOpcode() == Instruction::Shl)
552 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
553 // The multiplier is really 1 << CST.
554 uint32_t BitWidth = cast<IntegerType>(V->getType())->getBitWidth();
555 uint32_t CSTVal = CST->getLimitedValue(BitWidth);
556 CST = ConstantInt::get(APInt(BitWidth, 1).shl(CSTVal));
557 return I->getOperand(0);
563 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
564 /// expression, return it.
565 static User *dyn_castGetElementPtr(Value *V) {
566 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
567 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
568 if (CE->getOpcode() == Instruction::GetElementPtr)
569 return cast<User>(V);
573 /// AddOne - Add one to a ConstantInt
574 static ConstantInt *AddOne(ConstantInt *C) {
575 APInt Val(C->getValue());
576 return ConstantInt::get(++Val);
578 /// SubOne - Subtract one from a ConstantInt
579 static ConstantInt *SubOne(ConstantInt *C) {
580 APInt Val(C->getValue());
581 return ConstantInt::get(--Val);
583 /// Add - Add two ConstantInts together
584 static ConstantInt *Add(ConstantInt *C1, ConstantInt *C2) {
585 return ConstantInt::get(C1->getValue() + C2->getValue());
587 /// And - Bitwise AND two ConstantInts together
588 static ConstantInt *And(ConstantInt *C1, ConstantInt *C2) {
589 return ConstantInt::get(C1->getValue() & C2->getValue());
591 /// Subtract - Subtract one ConstantInt from another
592 static ConstantInt *Subtract(ConstantInt *C1, ConstantInt *C2) {
593 return ConstantInt::get(C1->getValue() - C2->getValue());
595 /// Multiply - Multiply two ConstantInts together
596 static ConstantInt *Multiply(ConstantInt *C1, ConstantInt *C2) {
597 return ConstantInt::get(C1->getValue() * C2->getValue());
600 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
601 /// known to be either zero or one and return them in the KnownZero/KnownOne
602 /// bit sets. This code only analyzes bits in Mask, in order to short-circuit
604 /// NOTE: we cannot consider 'undef' to be "IsZero" here. The problem is that
605 /// we cannot optimize based on the assumption that it is zero without changing
606 /// it to be an explicit zero. If we don't change it to zero, other code could
607 /// optimized based on the contradictory assumption that it is non-zero.
608 /// Because instcombine aggressively folds operations with undef args anyway,
609 /// this won't lose us code quality.
610 static void ComputeMaskedBits(Value *V, const APInt &Mask, APInt& KnownZero,
611 APInt& KnownOne, unsigned Depth = 0) {
612 assert(V && "No Value?");
613 assert(Depth <= 6 && "Limit Search Depth");
614 uint32_t BitWidth = Mask.getBitWidth();
615 assert(cast<IntegerType>(V->getType())->getBitWidth() == BitWidth &&
616 KnownZero.getBitWidth() == BitWidth &&
617 KnownOne.getBitWidth() == BitWidth &&
618 "V, Mask, KnownOne and KnownZero should have same BitWidth");
619 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
620 // We know all of the bits for a constant!
621 KnownOne = CI->getValue() & Mask;
622 KnownZero = ~KnownOne & Mask;
626 if (Depth == 6 || Mask == 0)
627 return; // Limit search depth.
629 Instruction *I = dyn_cast<Instruction>(V);
632 KnownZero.clear(); KnownOne.clear(); // Don't know anything.
633 APInt KnownZero2(KnownZero), KnownOne2(KnownOne);
635 switch (I->getOpcode()) {
636 case Instruction::And: {
637 // If either the LHS or the RHS are Zero, the result is zero.
638 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
639 APInt Mask2(Mask & ~KnownZero);
640 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
641 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
642 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
644 // Output known-1 bits are only known if set in both the LHS & RHS.
645 KnownOne &= KnownOne2;
646 // Output known-0 are known to be clear if zero in either the LHS | RHS.
647 KnownZero |= KnownZero2;
650 case Instruction::Or: {
651 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
652 APInt Mask2(Mask & ~KnownOne);
653 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero2, KnownOne2, Depth+1);
654 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
655 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
657 // Output known-0 bits are only known if clear in both the LHS & RHS.
658 KnownZero &= KnownZero2;
659 // Output known-1 are known to be set if set in either the LHS | RHS.
660 KnownOne |= KnownOne2;
663 case Instruction::Xor: {
664 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
665 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
666 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
667 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
669 // Output known-0 bits are known if clear or set in both the LHS & RHS.
670 APInt KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
671 // Output known-1 are known to be set if set in only one of the LHS, RHS.
672 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
673 KnownZero = KnownZeroOut;
676 case Instruction::Select:
677 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
678 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
679 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
680 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
682 // Only known if known in both the LHS and RHS.
683 KnownOne &= KnownOne2;
684 KnownZero &= KnownZero2;
686 case Instruction::FPTrunc:
687 case Instruction::FPExt:
688 case Instruction::FPToUI:
689 case Instruction::FPToSI:
690 case Instruction::SIToFP:
691 case Instruction::PtrToInt:
692 case Instruction::UIToFP:
693 case Instruction::IntToPtr:
694 return; // Can't work with floating point or pointers
695 case Instruction::Trunc: {
696 // All these have integer operands
697 uint32_t SrcBitWidth =
698 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
700 MaskIn.zext(SrcBitWidth);
701 KnownZero.zext(SrcBitWidth);
702 KnownOne.zext(SrcBitWidth);
703 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
704 KnownZero.trunc(BitWidth);
705 KnownOne.trunc(BitWidth);
708 case Instruction::BitCast: {
709 const Type *SrcTy = I->getOperand(0)->getType();
710 if (SrcTy->isInteger()) {
711 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
716 case Instruction::ZExt: {
717 // Compute the bits in the result that are not present in the input.
718 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
719 uint32_t SrcBitWidth = SrcTy->getBitWidth();
722 MaskIn.trunc(SrcBitWidth);
723 KnownZero.trunc(SrcBitWidth);
724 KnownOne.trunc(SrcBitWidth);
725 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
726 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
727 // The top bits are known to be zero.
728 KnownZero.zext(BitWidth);
729 KnownOne.zext(BitWidth);
730 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
733 case Instruction::SExt: {
734 // Compute the bits in the result that are not present in the input.
735 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
736 uint32_t SrcBitWidth = SrcTy->getBitWidth();
739 MaskIn.trunc(SrcBitWidth);
740 KnownZero.trunc(SrcBitWidth);
741 KnownOne.trunc(SrcBitWidth);
742 ComputeMaskedBits(I->getOperand(0), MaskIn, KnownZero, KnownOne, Depth+1);
743 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
744 KnownZero.zext(BitWidth);
745 KnownOne.zext(BitWidth);
747 // If the sign bit of the input is known set or clear, then we know the
748 // top bits of the result.
749 if (KnownZero[SrcBitWidth-1]) // Input sign bit known zero
750 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
751 else if (KnownOne[SrcBitWidth-1]) // Input sign bit known set
752 KnownOne |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
755 case Instruction::Shl:
756 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
757 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
758 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
759 APInt Mask2(Mask.lshr(ShiftAmt));
760 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero, KnownOne, Depth+1);
761 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
762 KnownZero <<= ShiftAmt;
763 KnownOne <<= ShiftAmt;
764 KnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt); // low bits known 0
768 case Instruction::LShr:
769 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
770 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
771 // Compute the new bits that are at the top now.
772 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
774 // Unsigned shift right.
775 APInt Mask2(Mask.shl(ShiftAmt));
776 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
777 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
778 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
779 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
780 // high bits known zero.
781 KnownZero |= APInt::getHighBitsSet(BitWidth, ShiftAmt);
785 case Instruction::AShr:
786 // (ashr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
787 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
788 // Compute the new bits that are at the top now.
789 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
791 // Signed shift right.
792 APInt Mask2(Mask.shl(ShiftAmt));
793 ComputeMaskedBits(I->getOperand(0), Mask2, KnownZero,KnownOne,Depth+1);
794 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
795 KnownZero = APIntOps::lshr(KnownZero, ShiftAmt);
796 KnownOne = APIntOps::lshr(KnownOne, ShiftAmt);
798 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
799 if (KnownZero[BitWidth-ShiftAmt-1]) // New bits are known zero.
800 KnownZero |= HighBits;
801 else if (KnownOne[BitWidth-ShiftAmt-1]) // New bits are known one.
802 KnownOne |= HighBits;
809 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
810 /// this predicate to simplify operations downstream. Mask is known to be zero
811 /// for bits that V cannot have.
812 static bool MaskedValueIsZero(Value *V, const APInt& Mask, unsigned Depth = 0) {
813 APInt KnownZero(Mask.getBitWidth(), 0), KnownOne(Mask.getBitWidth(), 0);
814 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
815 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
816 return (KnownZero & Mask) == Mask;
819 /// ShrinkDemandedConstant - Check to see if the specified operand of the
820 /// specified instruction is a constant integer. If so, check to see if there
821 /// are any bits set in the constant that are not demanded. If so, shrink the
822 /// constant and return true.
823 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
825 assert(I && "No instruction?");
826 assert(OpNo < I->getNumOperands() && "Operand index too large");
828 // If the operand is not a constant integer, nothing to do.
829 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
830 if (!OpC) return false;
832 // If there are no bits set that aren't demanded, nothing to do.
833 Demanded.zextOrTrunc(OpC->getValue().getBitWidth());
834 if ((~Demanded & OpC->getValue()) == 0)
837 // This instruction is producing bits that are not demanded. Shrink the RHS.
838 Demanded &= OpC->getValue();
839 I->setOperand(OpNo, ConstantInt::get(Demanded));
843 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
844 // set of known zero and one bits, compute the maximum and minimum values that
845 // could have the specified known zero and known one bits, returning them in
847 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
848 const APInt& KnownZero,
849 const APInt& KnownOne,
850 APInt& Min, APInt& Max) {
851 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
852 assert(KnownZero.getBitWidth() == BitWidth &&
853 KnownOne.getBitWidth() == BitWidth &&
854 Min.getBitWidth() == BitWidth && Max.getBitWidth() == BitWidth &&
855 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
856 APInt UnknownBits = ~(KnownZero|KnownOne);
858 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
859 // bit if it is unknown.
861 Max = KnownOne|UnknownBits;
863 if (UnknownBits[BitWidth-1]) { // Sign bit is unknown
865 Max.clear(BitWidth-1);
869 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
870 // a set of known zero and one bits, compute the maximum and minimum values that
871 // could have the specified known zero and known one bits, returning them in
873 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
874 const APInt &KnownZero,
875 const APInt &KnownOne,
876 APInt &Min, APInt &Max) {
877 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth(); BitWidth = BitWidth;
878 assert(KnownZero.getBitWidth() == BitWidth &&
879 KnownOne.getBitWidth() == BitWidth &&
880 Min.getBitWidth() == BitWidth && Max.getBitWidth() &&
881 "Ty, KnownZero, KnownOne and Min, Max must have equal bitwidth.");
882 APInt UnknownBits = ~(KnownZero|KnownOne);
884 // The minimum value is when the unknown bits are all zeros.
886 // The maximum value is when the unknown bits are all ones.
887 Max = KnownOne|UnknownBits;
890 /// SimplifyDemandedBits - This function attempts to replace V with a simpler
891 /// value based on the demanded bits. When this function is called, it is known
892 /// that only the bits set in DemandedMask of the result of V are ever used
893 /// downstream. Consequently, depending on the mask and V, it may be possible
894 /// to replace V with a constant or one of its operands. In such cases, this
895 /// function does the replacement and returns true. In all other cases, it
896 /// returns false after analyzing the expression and setting KnownOne and known
897 /// to be one in the expression. KnownZero contains all the bits that are known
898 /// to be zero in the expression. These are provided to potentially allow the
899 /// caller (which might recursively be SimplifyDemandedBits itself) to simplify
900 /// the expression. KnownOne and KnownZero always follow the invariant that
901 /// KnownOne & KnownZero == 0. That is, a bit can't be both 1 and 0. Note that
902 /// the bits in KnownOne and KnownZero may only be accurate for those bits set
903 /// in DemandedMask. Note also that the bitwidth of V, DemandedMask, KnownZero
904 /// and KnownOne must all be the same.
905 bool InstCombiner::SimplifyDemandedBits(Value *V, APInt DemandedMask,
906 APInt& KnownZero, APInt& KnownOne,
908 assert(V != 0 && "Null pointer of Value???");
909 assert(Depth <= 6 && "Limit Search Depth");
910 uint32_t BitWidth = DemandedMask.getBitWidth();
911 const IntegerType *VTy = cast<IntegerType>(V->getType());
912 assert(VTy->getBitWidth() == BitWidth &&
913 KnownZero.getBitWidth() == BitWidth &&
914 KnownOne.getBitWidth() == BitWidth &&
915 "Value *V, DemandedMask, KnownZero and KnownOne \
916 must have same BitWidth");
917 if (ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
918 // We know all of the bits for a constant!
919 KnownOne = CI->getValue() & DemandedMask;
920 KnownZero = ~KnownOne & DemandedMask;
926 if (!V->hasOneUse()) { // Other users may use these bits.
927 if (Depth != 0) { // Not at the root.
928 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
929 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
932 // If this is the root being simplified, allow it to have multiple uses,
933 // just set the DemandedMask to all bits.
934 DemandedMask = APInt::getAllOnesValue(BitWidth);
935 } else if (DemandedMask == 0) { // Not demanding any bits from V.
936 if (V != UndefValue::get(VTy))
937 return UpdateValueUsesWith(V, UndefValue::get(VTy));
939 } else if (Depth == 6) { // Limit search depth.
943 Instruction *I = dyn_cast<Instruction>(V);
944 if (!I) return false; // Only analyze instructions.
946 APInt LHSKnownZero(BitWidth, 0), LHSKnownOne(BitWidth, 0);
947 APInt &RHSKnownZero = KnownZero, &RHSKnownOne = KnownOne;
948 switch (I->getOpcode()) {
950 case Instruction::And:
951 // If either the LHS or the RHS are Zero, the result is zero.
952 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
953 RHSKnownZero, RHSKnownOne, Depth+1))
955 assert((RHSKnownZero & RHSKnownOne) == 0 &&
956 "Bits known to be one AND zero?");
958 // If something is known zero on the RHS, the bits aren't demanded on the
960 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownZero,
961 LHSKnownZero, LHSKnownOne, Depth+1))
963 assert((LHSKnownZero & LHSKnownOne) == 0 &&
964 "Bits known to be one AND zero?");
966 // If all of the demanded bits are known 1 on one side, return the other.
967 // These bits cannot contribute to the result of the 'and'.
968 if ((DemandedMask & ~LHSKnownZero & RHSKnownOne) ==
969 (DemandedMask & ~LHSKnownZero))
970 return UpdateValueUsesWith(I, I->getOperand(0));
971 if ((DemandedMask & ~RHSKnownZero & LHSKnownOne) ==
972 (DemandedMask & ~RHSKnownZero))
973 return UpdateValueUsesWith(I, I->getOperand(1));
975 // If all of the demanded bits in the inputs are known zeros, return zero.
976 if ((DemandedMask & (RHSKnownZero|LHSKnownZero)) == DemandedMask)
977 return UpdateValueUsesWith(I, Constant::getNullValue(VTy));
979 // If the RHS is a constant, see if we can simplify it.
980 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~LHSKnownZero))
981 return UpdateValueUsesWith(I, I);
983 // Output known-1 bits are only known if set in both the LHS & RHS.
984 RHSKnownOne &= LHSKnownOne;
985 // Output known-0 are known to be clear if zero in either the LHS | RHS.
986 RHSKnownZero |= LHSKnownZero;
988 case Instruction::Or:
989 // If either the LHS or the RHS are One, the result is One.
990 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
991 RHSKnownZero, RHSKnownOne, Depth+1))
993 assert((RHSKnownZero & RHSKnownOne) == 0 &&
994 "Bits known to be one AND zero?");
995 // If something is known one on the RHS, the bits aren't demanded on the
997 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~RHSKnownOne,
998 LHSKnownZero, LHSKnownOne, Depth+1))
1000 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1001 "Bits known to be one AND zero?");
1003 // If all of the demanded bits are known zero on one side, return the other.
1004 // These bits cannot contribute to the result of the 'or'.
1005 if ((DemandedMask & ~LHSKnownOne & RHSKnownZero) ==
1006 (DemandedMask & ~LHSKnownOne))
1007 return UpdateValueUsesWith(I, I->getOperand(0));
1008 if ((DemandedMask & ~RHSKnownOne & LHSKnownZero) ==
1009 (DemandedMask & ~RHSKnownOne))
1010 return UpdateValueUsesWith(I, I->getOperand(1));
1012 // If all of the potentially set bits on one side are known to be set on
1013 // the other side, just use the 'other' side.
1014 if ((DemandedMask & (~RHSKnownZero) & LHSKnownOne) ==
1015 (DemandedMask & (~RHSKnownZero)))
1016 return UpdateValueUsesWith(I, I->getOperand(0));
1017 if ((DemandedMask & (~LHSKnownZero) & RHSKnownOne) ==
1018 (DemandedMask & (~LHSKnownZero)))
1019 return UpdateValueUsesWith(I, I->getOperand(1));
1021 // If the RHS is a constant, see if we can simplify it.
1022 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1023 return UpdateValueUsesWith(I, I);
1025 // Output known-0 bits are only known if clear in both the LHS & RHS.
1026 RHSKnownZero &= LHSKnownZero;
1027 // Output known-1 are known to be set if set in either the LHS | RHS.
1028 RHSKnownOne |= LHSKnownOne;
1030 case Instruction::Xor: {
1031 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1032 RHSKnownZero, RHSKnownOne, Depth+1))
1034 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1035 "Bits known to be one AND zero?");
1036 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1037 LHSKnownZero, LHSKnownOne, Depth+1))
1039 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1040 "Bits known to be one AND zero?");
1042 // If all of the demanded bits are known zero on one side, return the other.
1043 // These bits cannot contribute to the result of the 'xor'.
1044 if ((DemandedMask & RHSKnownZero) == DemandedMask)
1045 return UpdateValueUsesWith(I, I->getOperand(0));
1046 if ((DemandedMask & LHSKnownZero) == DemandedMask)
1047 return UpdateValueUsesWith(I, I->getOperand(1));
1049 // Output known-0 bits are known if clear or set in both the LHS & RHS.
1050 APInt KnownZeroOut = (RHSKnownZero & LHSKnownZero) |
1051 (RHSKnownOne & LHSKnownOne);
1052 // Output known-1 are known to be set if set in only one of the LHS, RHS.
1053 APInt KnownOneOut = (RHSKnownZero & LHSKnownOne) |
1054 (RHSKnownOne & LHSKnownZero);
1056 // If all of the demanded bits are known to be zero on one side or the
1057 // other, turn this into an *inclusive* or.
1058 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
1059 if ((DemandedMask & ~RHSKnownZero & ~LHSKnownZero) == 0) {
1061 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1063 InsertNewInstBefore(Or, *I);
1064 return UpdateValueUsesWith(I, Or);
1067 // If all of the demanded bits on one side are known, and all of the set
1068 // bits on that side are also known to be set on the other side, turn this
1069 // into an AND, as we know the bits will be cleared.
1070 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
1071 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask) {
1073 if ((RHSKnownOne & LHSKnownOne) == RHSKnownOne) {
1074 Constant *AndC = ConstantInt::get(~RHSKnownOne & DemandedMask);
1076 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
1077 InsertNewInstBefore(And, *I);
1078 return UpdateValueUsesWith(I, And);
1082 // If the RHS is a constant, see if we can simplify it.
1083 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
1084 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1085 return UpdateValueUsesWith(I, I);
1087 RHSKnownZero = KnownZeroOut;
1088 RHSKnownOne = KnownOneOut;
1091 case Instruction::Select:
1092 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
1093 RHSKnownZero, RHSKnownOne, Depth+1))
1095 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
1096 LHSKnownZero, LHSKnownOne, Depth+1))
1098 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1099 "Bits known to be one AND zero?");
1100 assert((LHSKnownZero & LHSKnownOne) == 0 &&
1101 "Bits known to be one AND zero?");
1103 // If the operands are constants, see if we can simplify them.
1104 if (ShrinkDemandedConstant(I, 1, DemandedMask))
1105 return UpdateValueUsesWith(I, I);
1106 if (ShrinkDemandedConstant(I, 2, DemandedMask))
1107 return UpdateValueUsesWith(I, I);
1109 // Only known if known in both the LHS and RHS.
1110 RHSKnownOne &= LHSKnownOne;
1111 RHSKnownZero &= LHSKnownZero;
1113 case Instruction::Trunc: {
1115 cast<IntegerType>(I->getOperand(0)->getType())->getBitWidth();
1116 DemandedMask.zext(truncBf);
1117 RHSKnownZero.zext(truncBf);
1118 RHSKnownOne.zext(truncBf);
1119 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1120 RHSKnownZero, RHSKnownOne, Depth+1))
1122 DemandedMask.trunc(BitWidth);
1123 RHSKnownZero.trunc(BitWidth);
1124 RHSKnownOne.trunc(BitWidth);
1125 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1126 "Bits known to be one AND zero?");
1129 case Instruction::BitCast:
1130 if (!I->getOperand(0)->getType()->isInteger())
1133 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1134 RHSKnownZero, RHSKnownOne, Depth+1))
1136 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1137 "Bits known to be one AND zero?");
1139 case Instruction::ZExt: {
1140 // Compute the bits in the result that are not present in the input.
1141 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1142 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1144 DemandedMask.trunc(SrcBitWidth);
1145 RHSKnownZero.trunc(SrcBitWidth);
1146 RHSKnownOne.trunc(SrcBitWidth);
1147 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1148 RHSKnownZero, RHSKnownOne, Depth+1))
1150 DemandedMask.zext(BitWidth);
1151 RHSKnownZero.zext(BitWidth);
1152 RHSKnownOne.zext(BitWidth);
1153 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1154 "Bits known to be one AND zero?");
1155 // The top bits are known to be zero.
1156 RHSKnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth);
1159 case Instruction::SExt: {
1160 // Compute the bits in the result that are not present in the input.
1161 const IntegerType *SrcTy = cast<IntegerType>(I->getOperand(0)->getType());
1162 uint32_t SrcBitWidth = SrcTy->getBitWidth();
1164 APInt InputDemandedBits = DemandedMask &
1165 APInt::getLowBitsSet(BitWidth, SrcBitWidth);
1167 APInt NewBits(APInt::getHighBitsSet(BitWidth, BitWidth - SrcBitWidth));
1168 // If any of the sign extended bits are demanded, we know that the sign
1170 if ((NewBits & DemandedMask) != 0)
1171 InputDemandedBits.set(SrcBitWidth-1);
1173 InputDemandedBits.trunc(SrcBitWidth);
1174 RHSKnownZero.trunc(SrcBitWidth);
1175 RHSKnownOne.trunc(SrcBitWidth);
1176 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1177 RHSKnownZero, RHSKnownOne, Depth+1))
1179 InputDemandedBits.zext(BitWidth);
1180 RHSKnownZero.zext(BitWidth);
1181 RHSKnownOne.zext(BitWidth);
1182 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1183 "Bits known to be one AND zero?");
1185 // If the sign bit of the input is known set or clear, then we know the
1186 // top bits of the result.
1188 // If the input sign bit is known zero, or if the NewBits are not demanded
1189 // convert this into a zero extension.
1190 if (RHSKnownZero[SrcBitWidth-1] || (NewBits & ~DemandedMask) == NewBits)
1192 // Convert to ZExt cast
1193 CastInst *NewCast = new ZExtInst(I->getOperand(0), VTy, I->getName(), I);
1194 return UpdateValueUsesWith(I, NewCast);
1195 } else if (RHSKnownOne[SrcBitWidth-1]) { // Input sign bit known set
1196 RHSKnownOne |= NewBits;
1200 case Instruction::Add: {
1201 // Figure out what the input bits are. If the top bits of the and result
1202 // are not demanded, then the add doesn't demand them from its input
1204 uint32_t NLZ = DemandedMask.countLeadingZeros();
1206 // If there is a constant on the RHS, there are a variety of xformations
1208 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1209 // If null, this should be simplified elsewhere. Some of the xforms here
1210 // won't work if the RHS is zero.
1214 // If the top bit of the output is demanded, demand everything from the
1215 // input. Otherwise, we demand all the input bits except NLZ top bits.
1216 APInt InDemandedBits(APInt::getLowBitsSet(BitWidth, BitWidth - NLZ));
1218 // Find information about known zero/one bits in the input.
1219 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1220 LHSKnownZero, LHSKnownOne, Depth+1))
1223 // If the RHS of the add has bits set that can't affect the input, reduce
1225 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1226 return UpdateValueUsesWith(I, I);
1228 // Avoid excess work.
1229 if (LHSKnownZero == 0 && LHSKnownOne == 0)
1232 // Turn it into OR if input bits are zero.
1233 if ((LHSKnownZero & RHS->getValue()) == RHS->getValue()) {
1235 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1237 InsertNewInstBefore(Or, *I);
1238 return UpdateValueUsesWith(I, Or);
1241 // We can say something about the output known-zero and known-one bits,
1242 // depending on potential carries from the input constant and the
1243 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1244 // bits set and the RHS constant is 0x01001, then we know we have a known
1245 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1247 // To compute this, we first compute the potential carry bits. These are
1248 // the bits which may be modified. I'm not aware of a better way to do
1250 const APInt& RHSVal = RHS->getValue();
1251 APInt CarryBits((~LHSKnownZero + RHSVal) ^ (~LHSKnownZero ^ RHSVal));
1253 // Now that we know which bits have carries, compute the known-1/0 sets.
1255 // Bits are known one if they are known zero in one operand and one in the
1256 // other, and there is no input carry.
1257 RHSKnownOne = ((LHSKnownZero & RHSVal) |
1258 (LHSKnownOne & ~RHSVal)) & ~CarryBits;
1260 // Bits are known zero if they are known zero in both operands and there
1261 // is no input carry.
1262 RHSKnownZero = LHSKnownZero & ~RHSVal & ~CarryBits;
1264 // If the high-bits of this ADD are not demanded, then it does not demand
1265 // the high bits of its LHS or RHS.
1266 if (DemandedMask[BitWidth-1] == 0) {
1267 // Right fill the mask of bits for this ADD to demand the most
1268 // significant bit and all those below it.
1269 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1270 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1271 LHSKnownZero, LHSKnownOne, Depth+1))
1273 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1274 LHSKnownZero, LHSKnownOne, Depth+1))
1280 case Instruction::Sub:
1281 // If the high-bits of this SUB are not demanded, then it does not demand
1282 // the high bits of its LHS or RHS.
1283 if (DemandedMask[BitWidth-1] == 0) {
1284 // Right fill the mask of bits for this SUB to demand the most
1285 // significant bit and all those below it.
1286 uint32_t NLZ = DemandedMask.countLeadingZeros();
1287 APInt DemandedFromOps(APInt::getLowBitsSet(BitWidth, BitWidth-NLZ));
1288 if (SimplifyDemandedBits(I->getOperand(0), DemandedFromOps,
1289 LHSKnownZero, LHSKnownOne, Depth+1))
1291 if (SimplifyDemandedBits(I->getOperand(1), DemandedFromOps,
1292 LHSKnownZero, LHSKnownOne, Depth+1))
1296 case Instruction::Shl:
1297 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1298 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1299 APInt DemandedMaskIn(DemandedMask.lshr(ShiftAmt));
1300 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1301 RHSKnownZero, RHSKnownOne, Depth+1))
1303 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1304 "Bits known to be one AND zero?");
1305 RHSKnownZero <<= ShiftAmt;
1306 RHSKnownOne <<= ShiftAmt;
1307 // low bits known zero.
1309 RHSKnownZero |= APInt::getLowBitsSet(BitWidth, ShiftAmt);
1312 case Instruction::LShr:
1313 // For a logical shift right
1314 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1315 uint64_t ShiftAmt = SA->getLimitedValue(BitWidth);
1317 // Unsigned shift right.
1318 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1319 if (SimplifyDemandedBits(I->getOperand(0), DemandedMaskIn,
1320 RHSKnownZero, RHSKnownOne, Depth+1))
1322 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1323 "Bits known to be one AND zero?");
1324 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1325 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1327 // Compute the new bits that are at the top now.
1328 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1329 RHSKnownZero |= HighBits; // high bits known zero.
1333 case Instruction::AShr:
1334 // If this is an arithmetic shift right and only the low-bit is set, we can
1335 // always convert this into a logical shr, even if the shift amount is
1336 // variable. The low bit of the shift cannot be an input sign bit unless
1337 // the shift amount is >= the size of the datatype, which is undefined.
1338 if (DemandedMask == 1) {
1339 // Perform the logical shift right.
1340 Value *NewVal = BinaryOperator::createLShr(
1341 I->getOperand(0), I->getOperand(1), I->getName());
1342 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1343 return UpdateValueUsesWith(I, NewVal);
1346 // If the sign bit is the only bit demanded by this ashr, then there is no
1347 // need to do it, the shift doesn't change the high bit.
1348 if (DemandedMask.isSignBit())
1349 return UpdateValueUsesWith(I, I->getOperand(0));
1351 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1352 uint32_t ShiftAmt = SA->getLimitedValue(BitWidth);
1354 // Signed shift right.
1355 APInt DemandedMaskIn(DemandedMask.shl(ShiftAmt));
1356 // If any of the "high bits" are demanded, we should set the sign bit as
1358 if (DemandedMask.countLeadingZeros() <= ShiftAmt)
1359 DemandedMaskIn.set(BitWidth-1);
1360 if (SimplifyDemandedBits(I->getOperand(0),
1362 RHSKnownZero, RHSKnownOne, Depth+1))
1364 assert((RHSKnownZero & RHSKnownOne) == 0 &&
1365 "Bits known to be one AND zero?");
1366 // Compute the new bits that are at the top now.
1367 APInt HighBits(APInt::getHighBitsSet(BitWidth, ShiftAmt));
1368 RHSKnownZero = APIntOps::lshr(RHSKnownZero, ShiftAmt);
1369 RHSKnownOne = APIntOps::lshr(RHSKnownOne, ShiftAmt);
1371 // Handle the sign bits.
1372 APInt SignBit(APInt::getSignBit(BitWidth));
1373 // Adjust to where it is now in the mask.
1374 SignBit = APIntOps::lshr(SignBit, ShiftAmt);
1376 // If the input sign bit is known to be zero, or if none of the top bits
1377 // are demanded, turn this into an unsigned shift right.
1378 if (RHSKnownZero[BitWidth-ShiftAmt-1] ||
1379 (HighBits & ~DemandedMask) == HighBits) {
1380 // Perform the logical shift right.
1381 Value *NewVal = BinaryOperator::createLShr(
1382 I->getOperand(0), SA, I->getName());
1383 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1384 return UpdateValueUsesWith(I, NewVal);
1385 } else if ((RHSKnownOne & SignBit) != 0) { // New bits are known one.
1386 RHSKnownOne |= HighBits;
1392 // If the client is only demanding bits that we know, return the known
1394 if ((DemandedMask & (RHSKnownZero|RHSKnownOne)) == DemandedMask)
1395 return UpdateValueUsesWith(I, ConstantInt::get(RHSKnownOne));
1400 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1401 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1402 /// actually used by the caller. This method analyzes which elements of the
1403 /// operand are undef and returns that information in UndefElts.
1405 /// If the information about demanded elements can be used to simplify the
1406 /// operation, the operation is simplified, then the resultant value is
1407 /// returned. This returns null if no change was made.
1408 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1409 uint64_t &UndefElts,
1411 unsigned VWidth = cast<VectorType>(V->getType())->getNumElements();
1412 assert(VWidth <= 64 && "Vector too wide to analyze!");
1413 uint64_t EltMask = ~0ULL >> (64-VWidth);
1414 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1415 "Invalid DemandedElts!");
1417 if (isa<UndefValue>(V)) {
1418 // If the entire vector is undefined, just return this info.
1419 UndefElts = EltMask;
1421 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1422 UndefElts = EltMask;
1423 return UndefValue::get(V->getType());
1427 if (ConstantVector *CP = dyn_cast<ConstantVector>(V)) {
1428 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1429 Constant *Undef = UndefValue::get(EltTy);
1431 std::vector<Constant*> Elts;
1432 for (unsigned i = 0; i != VWidth; ++i)
1433 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1434 Elts.push_back(Undef);
1435 UndefElts |= (1ULL << i);
1436 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1437 Elts.push_back(Undef);
1438 UndefElts |= (1ULL << i);
1439 } else { // Otherwise, defined.
1440 Elts.push_back(CP->getOperand(i));
1443 // If we changed the constant, return it.
1444 Constant *NewCP = ConstantVector::get(Elts);
1445 return NewCP != CP ? NewCP : 0;
1446 } else if (isa<ConstantAggregateZero>(V)) {
1447 // Simplify the CAZ to a ConstantVector where the non-demanded elements are
1449 const Type *EltTy = cast<VectorType>(V->getType())->getElementType();
1450 Constant *Zero = Constant::getNullValue(EltTy);
1451 Constant *Undef = UndefValue::get(EltTy);
1452 std::vector<Constant*> Elts;
1453 for (unsigned i = 0; i != VWidth; ++i)
1454 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1455 UndefElts = DemandedElts ^ EltMask;
1456 return ConstantVector::get(Elts);
1459 if (!V->hasOneUse()) { // Other users may use these bits.
1460 if (Depth != 0) { // Not at the root.
1461 // TODO: Just compute the UndefElts information recursively.
1465 } else if (Depth == 10) { // Limit search depth.
1469 Instruction *I = dyn_cast<Instruction>(V);
1470 if (!I) return false; // Only analyze instructions.
1472 bool MadeChange = false;
1473 uint64_t UndefElts2;
1475 switch (I->getOpcode()) {
1478 case Instruction::InsertElement: {
1479 // If this is a variable index, we don't know which element it overwrites.
1480 // demand exactly the same input as we produce.
1481 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1483 // Note that we can't propagate undef elt info, because we don't know
1484 // which elt is getting updated.
1485 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1486 UndefElts2, Depth+1);
1487 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1491 // If this is inserting an element that isn't demanded, remove this
1493 unsigned IdxNo = Idx->getZExtValue();
1494 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1495 return AddSoonDeadInstToWorklist(*I, 0);
1497 // Otherwise, the element inserted overwrites whatever was there, so the
1498 // input demanded set is simpler than the output set.
1499 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1500 DemandedElts & ~(1ULL << IdxNo),
1501 UndefElts, Depth+1);
1502 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1504 // The inserted element is defined.
1505 UndefElts |= 1ULL << IdxNo;
1508 case Instruction::BitCast: {
1509 // Vector->vector casts only.
1510 const VectorType *VTy = dyn_cast<VectorType>(I->getOperand(0)->getType());
1512 unsigned InVWidth = VTy->getNumElements();
1513 uint64_t InputDemandedElts = 0;
1516 if (VWidth == InVWidth) {
1517 // If we are converting from <4 x i32> -> <4 x f32>, we demand the same
1518 // elements as are demanded of us.
1520 InputDemandedElts = DemandedElts;
1521 } else if (VWidth > InVWidth) {
1525 // If there are more elements in the result than there are in the source,
1526 // then an input element is live if any of the corresponding output
1527 // elements are live.
1528 Ratio = VWidth/InVWidth;
1529 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx) {
1530 if (DemandedElts & (1ULL << OutIdx))
1531 InputDemandedElts |= 1ULL << (OutIdx/Ratio);
1537 // If there are more elements in the source than there are in the result,
1538 // then an input element is live if the corresponding output element is
1540 Ratio = InVWidth/VWidth;
1541 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1542 if (DemandedElts & (1ULL << InIdx/Ratio))
1543 InputDemandedElts |= 1ULL << InIdx;
1546 // div/rem demand all inputs, because they don't want divide by zero.
1547 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), InputDemandedElts,
1548 UndefElts2, Depth+1);
1550 I->setOperand(0, TmpV);
1554 UndefElts = UndefElts2;
1555 if (VWidth > InVWidth) {
1556 assert(0 && "Unimp");
1557 // If there are more elements in the result than there are in the source,
1558 // then an output element is undef if the corresponding input element is
1560 for (unsigned OutIdx = 0; OutIdx != VWidth; ++OutIdx)
1561 if (UndefElts2 & (1ULL << (OutIdx/Ratio)))
1562 UndefElts |= 1ULL << OutIdx;
1563 } else if (VWidth < InVWidth) {
1564 assert(0 && "Unimp");
1565 // If there are more elements in the source than there are in the result,
1566 // then a result element is undef if all of the corresponding input
1567 // elements are undef.
1568 UndefElts = ~0ULL >> (64-VWidth); // Start out all undef.
1569 for (unsigned InIdx = 0; InIdx != InVWidth; ++InIdx)
1570 if ((UndefElts2 & (1ULL << InIdx)) == 0) // Not undef?
1571 UndefElts &= ~(1ULL << (InIdx/Ratio)); // Clear undef bit.
1575 case Instruction::And:
1576 case Instruction::Or:
1577 case Instruction::Xor:
1578 case Instruction::Add:
1579 case Instruction::Sub:
1580 case Instruction::Mul:
1581 // div/rem demand all inputs, because they don't want divide by zero.
1582 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1583 UndefElts, Depth+1);
1584 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1585 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1586 UndefElts2, Depth+1);
1587 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1589 // Output elements are undefined if both are undefined. Consider things
1590 // like undef&0. The result is known zero, not undef.
1591 UndefElts &= UndefElts2;
1594 case Instruction::Call: {
1595 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1597 switch (II->getIntrinsicID()) {
1600 // Binary vector operations that work column-wise. A dest element is a
1601 // function of the corresponding input elements from the two inputs.
1602 case Intrinsic::x86_sse_sub_ss:
1603 case Intrinsic::x86_sse_mul_ss:
1604 case Intrinsic::x86_sse_min_ss:
1605 case Intrinsic::x86_sse_max_ss:
1606 case Intrinsic::x86_sse2_sub_sd:
1607 case Intrinsic::x86_sse2_mul_sd:
1608 case Intrinsic::x86_sse2_min_sd:
1609 case Intrinsic::x86_sse2_max_sd:
1610 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1611 UndefElts, Depth+1);
1612 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1613 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1614 UndefElts2, Depth+1);
1615 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1617 // If only the low elt is demanded and this is a scalarizable intrinsic,
1618 // scalarize it now.
1619 if (DemandedElts == 1) {
1620 switch (II->getIntrinsicID()) {
1622 case Intrinsic::x86_sse_sub_ss:
1623 case Intrinsic::x86_sse_mul_ss:
1624 case Intrinsic::x86_sse2_sub_sd:
1625 case Intrinsic::x86_sse2_mul_sd:
1626 // TODO: Lower MIN/MAX/ABS/etc
1627 Value *LHS = II->getOperand(1);
1628 Value *RHS = II->getOperand(2);
1629 // Extract the element as scalars.
1630 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1631 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1633 switch (II->getIntrinsicID()) {
1634 default: assert(0 && "Case stmts out of sync!");
1635 case Intrinsic::x86_sse_sub_ss:
1636 case Intrinsic::x86_sse2_sub_sd:
1637 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1638 II->getName()), *II);
1640 case Intrinsic::x86_sse_mul_ss:
1641 case Intrinsic::x86_sse2_mul_sd:
1642 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1643 II->getName()), *II);
1648 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1650 InsertNewInstBefore(New, *II);
1651 AddSoonDeadInstToWorklist(*II, 0);
1656 // Output elements are undefined if both are undefined. Consider things
1657 // like undef&0. The result is known zero, not undef.
1658 UndefElts &= UndefElts2;
1664 return MadeChange ? I : 0;
1667 /// @returns true if the specified compare predicate is
1668 /// true when both operands are equal...
1669 /// @brief Determine if the icmp Predicate is true when both operands are equal
1670 static bool isTrueWhenEqual(ICmpInst::Predicate pred) {
1671 return pred == ICmpInst::ICMP_EQ || pred == ICmpInst::ICMP_UGE ||
1672 pred == ICmpInst::ICMP_SGE || pred == ICmpInst::ICMP_ULE ||
1673 pred == ICmpInst::ICMP_SLE;
1676 /// @returns true if the specified compare instruction is
1677 /// true when both operands are equal...
1678 /// @brief Determine if the ICmpInst returns true when both operands are equal
1679 static bool isTrueWhenEqual(ICmpInst &ICI) {
1680 return isTrueWhenEqual(ICI.getPredicate());
1683 /// AssociativeOpt - Perform an optimization on an associative operator. This
1684 /// function is designed to check a chain of associative operators for a
1685 /// potential to apply a certain optimization. Since the optimization may be
1686 /// applicable if the expression was reassociated, this checks the chain, then
1687 /// reassociates the expression as necessary to expose the optimization
1688 /// opportunity. This makes use of a special Functor, which must define
1689 /// 'shouldApply' and 'apply' methods.
1691 template<typename Functor>
1692 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1693 unsigned Opcode = Root.getOpcode();
1694 Value *LHS = Root.getOperand(0);
1696 // Quick check, see if the immediate LHS matches...
1697 if (F.shouldApply(LHS))
1698 return F.apply(Root);
1700 // Otherwise, if the LHS is not of the same opcode as the root, return.
1701 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1702 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1703 // Should we apply this transform to the RHS?
1704 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1706 // If not to the RHS, check to see if we should apply to the LHS...
1707 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1708 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1712 // If the functor wants to apply the optimization to the RHS of LHSI,
1713 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1715 BasicBlock *BB = Root.getParent();
1717 // Now all of the instructions are in the current basic block, go ahead
1718 // and perform the reassociation.
1719 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1721 // First move the selected RHS to the LHS of the root...
1722 Root.setOperand(0, LHSI->getOperand(1));
1724 // Make what used to be the LHS of the root be the user of the root...
1725 Value *ExtraOperand = TmpLHSI->getOperand(1);
1726 if (&Root == TmpLHSI) {
1727 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1730 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1731 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1732 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1733 BasicBlock::iterator ARI = &Root; ++ARI;
1734 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1737 // Now propagate the ExtraOperand down the chain of instructions until we
1739 while (TmpLHSI != LHSI) {
1740 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1741 // Move the instruction to immediately before the chain we are
1742 // constructing to avoid breaking dominance properties.
1743 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1744 BB->getInstList().insert(ARI, NextLHSI);
1747 Value *NextOp = NextLHSI->getOperand(1);
1748 NextLHSI->setOperand(1, ExtraOperand);
1750 ExtraOperand = NextOp;
1753 // Now that the instructions are reassociated, have the functor perform
1754 // the transformation...
1755 return F.apply(Root);
1758 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1764 // AddRHS - Implements: X + X --> X << 1
1767 AddRHS(Value *rhs) : RHS(rhs) {}
1768 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1769 Instruction *apply(BinaryOperator &Add) const {
1770 return BinaryOperator::createShl(Add.getOperand(0),
1771 ConstantInt::get(Add.getType(), 1));
1775 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1777 struct AddMaskingAnd {
1779 AddMaskingAnd(Constant *c) : C2(c) {}
1780 bool shouldApply(Value *LHS) const {
1782 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1783 ConstantExpr::getAnd(C1, C2)->isNullValue();
1785 Instruction *apply(BinaryOperator &Add) const {
1786 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1790 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1792 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1793 if (Constant *SOC = dyn_cast<Constant>(SO))
1794 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1796 return IC->InsertNewInstBefore(CastInst::create(
1797 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1800 // Figure out if the constant is the left or the right argument.
1801 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1802 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1804 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1806 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1807 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1810 Value *Op0 = SO, *Op1 = ConstOperand;
1812 std::swap(Op0, Op1);
1814 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1815 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1816 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1817 New = CmpInst::create(CI->getOpcode(), CI->getPredicate(), Op0, Op1,
1818 SO->getName()+".cmp");
1820 assert(0 && "Unknown binary instruction type!");
1823 return IC->InsertNewInstBefore(New, I);
1826 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1827 // constant as the other operand, try to fold the binary operator into the
1828 // select arguments. This also works for Cast instructions, which obviously do
1829 // not have a second operand.
1830 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1832 // Don't modify shared select instructions
1833 if (!SI->hasOneUse()) return 0;
1834 Value *TV = SI->getOperand(1);
1835 Value *FV = SI->getOperand(2);
1837 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1838 // Bool selects with constant operands can be folded to logical ops.
1839 if (SI->getType() == Type::Int1Ty) return 0;
1841 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1842 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1844 return new SelectInst(SI->getCondition(), SelectTrueVal,
1851 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1852 /// node as operand #0, see if we can fold the instruction into the PHI (which
1853 /// is only possible if all operands to the PHI are constants).
1854 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1855 PHINode *PN = cast<PHINode>(I.getOperand(0));
1856 unsigned NumPHIValues = PN->getNumIncomingValues();
1857 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1859 // Check to see if all of the operands of the PHI are constants. If there is
1860 // one non-constant value, remember the BB it is. If there is more than one
1861 // or if *it* is a PHI, bail out.
1862 BasicBlock *NonConstBB = 0;
1863 for (unsigned i = 0; i != NumPHIValues; ++i)
1864 if (!isa<Constant>(PN->getIncomingValue(i))) {
1865 if (NonConstBB) return 0; // More than one non-const value.
1866 if (isa<PHINode>(PN->getIncomingValue(i))) return 0; // Itself a phi.
1867 NonConstBB = PN->getIncomingBlock(i);
1869 // If the incoming non-constant value is in I's block, we have an infinite
1871 if (NonConstBB == I.getParent())
1875 // If there is exactly one non-constant value, we can insert a copy of the
1876 // operation in that block. However, if this is a critical edge, we would be
1877 // inserting the computation one some other paths (e.g. inside a loop). Only
1878 // do this if the pred block is unconditionally branching into the phi block.
1880 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1881 if (!BI || !BI->isUnconditional()) return 0;
1884 // Okay, we can do the transformation: create the new PHI node.
1885 PHINode *NewPN = new PHINode(I.getType(), "");
1886 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1887 InsertNewInstBefore(NewPN, *PN);
1888 NewPN->takeName(PN);
1890 // Next, add all of the operands to the PHI.
1891 if (I.getNumOperands() == 2) {
1892 Constant *C = cast<Constant>(I.getOperand(1));
1893 for (unsigned i = 0; i != NumPHIValues; ++i) {
1895 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1896 if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1897 InV = ConstantExpr::getCompare(CI->getPredicate(), InC, C);
1899 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1901 assert(PN->getIncomingBlock(i) == NonConstBB);
1902 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1903 InV = BinaryOperator::create(BO->getOpcode(),
1904 PN->getIncomingValue(i), C, "phitmp",
1905 NonConstBB->getTerminator());
1906 else if (CmpInst *CI = dyn_cast<CmpInst>(&I))
1907 InV = CmpInst::create(CI->getOpcode(),
1909 PN->getIncomingValue(i), C, "phitmp",
1910 NonConstBB->getTerminator());
1912 assert(0 && "Unknown binop!");
1914 AddToWorkList(cast<Instruction>(InV));
1916 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1919 CastInst *CI = cast<CastInst>(&I);
1920 const Type *RetTy = CI->getType();
1921 for (unsigned i = 0; i != NumPHIValues; ++i) {
1923 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1924 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1926 assert(PN->getIncomingBlock(i) == NonConstBB);
1927 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1928 I.getType(), "phitmp",
1929 NonConstBB->getTerminator());
1930 AddToWorkList(cast<Instruction>(InV));
1932 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1935 return ReplaceInstUsesWith(I, NewPN);
1938 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1939 bool Changed = SimplifyCommutative(I);
1940 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1942 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1943 // X + undef -> undef
1944 if (isa<UndefValue>(RHS))
1945 return ReplaceInstUsesWith(I, RHS);
1948 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1949 if (RHSC->isNullValue())
1950 return ReplaceInstUsesWith(I, LHS);
1951 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1952 if (CFP->isExactlyValue(-0.0))
1953 return ReplaceInstUsesWith(I, LHS);
1956 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1957 // X + (signbit) --> X ^ signbit
1958 const APInt& Val = CI->getValue();
1959 uint32_t BitWidth = Val.getBitWidth();
1960 if (Val == APInt::getSignBit(BitWidth))
1961 return BinaryOperator::createXor(LHS, RHS);
1963 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1964 // (X & 254)+1 -> (X&254)|1
1965 if (!isa<VectorType>(I.getType())) {
1966 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
1967 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
1968 KnownZero, KnownOne))
1973 if (isa<PHINode>(LHS))
1974 if (Instruction *NV = FoldOpIntoPhi(I))
1977 ConstantInt *XorRHS = 0;
1979 if (isa<ConstantInt>(RHSC) &&
1980 match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1981 uint32_t TySizeBits = I.getType()->getPrimitiveSizeInBits();
1982 const APInt& RHSVal = cast<ConstantInt>(RHSC)->getValue();
1984 uint32_t Size = TySizeBits / 2;
1985 APInt C0080Val(APInt(TySizeBits, 1ULL).shl(Size - 1));
1986 APInt CFF80Val(-C0080Val);
1988 if (TySizeBits > Size) {
1989 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1990 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1991 if ((RHSVal == CFF80Val && XorRHS->getValue() == C0080Val) ||
1992 (RHSVal == C0080Val && XorRHS->getValue() == CFF80Val)) {
1993 // This is a sign extend if the top bits are known zero.
1994 if (!MaskedValueIsZero(XorLHS,
1995 APInt::getHighBitsSet(TySizeBits, TySizeBits - Size)))
1996 Size = 0; // Not a sign ext, but can't be any others either.
2001 C0080Val = APIntOps::lshr(C0080Val, Size);
2002 CFF80Val = APIntOps::ashr(CFF80Val, Size);
2003 } while (Size >= 1);
2005 // FIXME: This shouldn't be necessary. When the backends can handle types
2006 // with funny bit widths then this whole cascade of if statements should
2007 // be removed. It is just here to get the size of the "middle" type back
2008 // up to something that the back ends can handle.
2009 const Type *MiddleType = 0;
2012 case 32: MiddleType = Type::Int32Ty; break;
2013 case 16: MiddleType = Type::Int16Ty; break;
2014 case 8: MiddleType = Type::Int8Ty; break;
2017 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
2018 InsertNewInstBefore(NewTrunc, I);
2019 return new SExtInst(NewTrunc, I.getType(), I.getName());
2025 if (I.getType()->isInteger() && I.getType() != Type::Int1Ty) {
2026 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
2028 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
2029 if (RHSI->getOpcode() == Instruction::Sub)
2030 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
2031 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
2033 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
2034 if (LHSI->getOpcode() == Instruction::Sub)
2035 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
2036 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
2041 if (Value *V = dyn_castNegVal(LHS))
2042 return BinaryOperator::createSub(RHS, V);
2045 if (!isa<Constant>(RHS))
2046 if (Value *V = dyn_castNegVal(RHS))
2047 return BinaryOperator::createSub(LHS, V);
2051 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
2052 if (X == RHS) // X*C + X --> X * (C+1)
2053 return BinaryOperator::createMul(RHS, AddOne(C2));
2055 // X*C1 + X*C2 --> X * (C1+C2)
2057 if (X == dyn_castFoldableMul(RHS, C1))
2058 return BinaryOperator::createMul(X, Add(C1, C2));
2061 // X + X*C --> X * (C+1)
2062 if (dyn_castFoldableMul(RHS, C2) == LHS)
2063 return BinaryOperator::createMul(LHS, AddOne(C2));
2065 // X + ~X --> -1 since ~X = -X-1
2066 if (dyn_castNotVal(LHS) == RHS || dyn_castNotVal(RHS) == LHS)
2067 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
2070 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
2071 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
2072 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2)))
2075 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
2077 if (match(LHS, m_Not(m_Value(X)))) // ~X + C --> (C-1) - X
2078 return BinaryOperator::createSub(SubOne(CRHS), X);
2080 // (X & FF00) + xx00 -> (X+xx00) & FF00
2081 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
2082 Constant *Anded = And(CRHS, C2);
2083 if (Anded == CRHS) {
2084 // See if all bits from the first bit set in the Add RHS up are included
2085 // in the mask. First, get the rightmost bit.
2086 const APInt& AddRHSV = CRHS->getValue();
2088 // Form a mask of all bits from the lowest bit added through the top.
2089 APInt AddRHSHighBits(~((AddRHSV & -AddRHSV)-1));
2091 // See if the and mask includes all of these bits.
2092 APInt AddRHSHighBitsAnd(AddRHSHighBits & C2->getValue());
2094 if (AddRHSHighBits == AddRHSHighBitsAnd) {
2095 // Okay, the xform is safe. Insert the new add pronto.
2096 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
2097 LHS->getName()), I);
2098 return BinaryOperator::createAnd(NewAdd, C2);
2103 // Try to fold constant add into select arguments.
2104 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
2105 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2109 // add (cast *A to intptrtype) B ->
2110 // cast (GEP (cast *A to sbyte*) B) ->
2113 CastInst *CI = dyn_cast<CastInst>(LHS);
2116 CI = dyn_cast<CastInst>(RHS);
2119 if (CI && CI->getType()->isSized() &&
2120 (CI->getType()->getPrimitiveSizeInBits() ==
2121 TD->getIntPtrType()->getPrimitiveSizeInBits())
2122 && isa<PointerType>(CI->getOperand(0)->getType())) {
2123 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
2124 PointerType::get(Type::Int8Ty), I);
2125 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
2126 return new PtrToIntInst(I2, CI->getType());
2130 return Changed ? &I : 0;
2133 // isSignBit - Return true if the value represented by the constant only has the
2134 // highest order bit set.
2135 static bool isSignBit(ConstantInt *CI) {
2136 uint32_t NumBits = CI->getType()->getPrimitiveSizeInBits();
2137 return CI->getValue() == APInt::getSignBit(NumBits);
2140 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
2141 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2143 if (Op0 == Op1) // sub X, X -> 0
2144 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2146 // If this is a 'B = x-(-A)', change to B = x+A...
2147 if (Value *V = dyn_castNegVal(Op1))
2148 return BinaryOperator::createAdd(Op0, V);
2150 if (isa<UndefValue>(Op0))
2151 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
2152 if (isa<UndefValue>(Op1))
2153 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
2155 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
2156 // Replace (-1 - A) with (~A)...
2157 if (C->isAllOnesValue())
2158 return BinaryOperator::createNot(Op1);
2160 // C - ~X == X + (1+C)
2162 if (match(Op1, m_Not(m_Value(X))))
2163 return BinaryOperator::createAdd(X, AddOne(C));
2165 // -(X >>u 31) -> (X >>s 31)
2166 // -(X >>s 31) -> (X >>u 31)
2168 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op1))
2169 if (SI->getOpcode() == Instruction::LShr) {
2170 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2171 // Check to see if we are shifting out everything but the sign bit.
2172 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2173 SI->getType()->getPrimitiveSizeInBits()-1) {
2174 // Ok, the transformation is safe. Insert AShr.
2175 return BinaryOperator::create(Instruction::AShr,
2176 SI->getOperand(0), CU, SI->getName());
2180 else if (SI->getOpcode() == Instruction::AShr) {
2181 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
2182 // Check to see if we are shifting out everything but the sign bit.
2183 if (CU->getLimitedValue(SI->getType()->getPrimitiveSizeInBits()) ==
2184 SI->getType()->getPrimitiveSizeInBits()-1) {
2185 // Ok, the transformation is safe. Insert LShr.
2186 return BinaryOperator::createLShr(
2187 SI->getOperand(0), CU, SI->getName());
2193 // Try to fold constant sub into select arguments.
2194 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2195 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2198 if (isa<PHINode>(Op0))
2199 if (Instruction *NV = FoldOpIntoPhi(I))
2203 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
2204 if (Op1I->getOpcode() == Instruction::Add &&
2205 !Op0->getType()->isFPOrFPVector()) {
2206 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
2207 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2208 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2209 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2210 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2211 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2212 // C1-(X+C2) --> (C1-C2)-X
2213 return BinaryOperator::createSub(Subtract(CI1, CI2),
2214 Op1I->getOperand(0));
2218 if (Op1I->hasOneUse()) {
2219 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2220 // is not used by anyone else...
2222 if (Op1I->getOpcode() == Instruction::Sub &&
2223 !Op1I->getType()->isFPOrFPVector()) {
2224 // Swap the two operands of the subexpr...
2225 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2226 Op1I->setOperand(0, IIOp1);
2227 Op1I->setOperand(1, IIOp0);
2229 // Create the new top level add instruction...
2230 return BinaryOperator::createAdd(Op0, Op1);
2233 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2235 if (Op1I->getOpcode() == Instruction::And &&
2236 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2237 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2240 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2241 return BinaryOperator::createAnd(Op0, NewNot);
2244 // 0 - (X sdiv C) -> (X sdiv -C)
2245 if (Op1I->getOpcode() == Instruction::SDiv)
2246 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2248 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2249 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2250 ConstantExpr::getNeg(DivRHS));
2252 // X - X*C --> X * (1-C)
2253 ConstantInt *C2 = 0;
2254 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2255 Constant *CP1 = Subtract(ConstantInt::get(I.getType(), 1), C2);
2256 return BinaryOperator::createMul(Op0, CP1);
2261 if (!Op0->getType()->isFPOrFPVector())
2262 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2263 if (Op0I->getOpcode() == Instruction::Add) {
2264 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2265 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2266 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2267 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2268 } else if (Op0I->getOpcode() == Instruction::Sub) {
2269 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2270 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2274 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2275 if (X == Op1) // X*C - X --> X * (C-1)
2276 return BinaryOperator::createMul(Op1, SubOne(C1));
2278 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2279 if (X == dyn_castFoldableMul(Op1, C2))
2280 return BinaryOperator::createMul(Op1, Subtract(C1, C2));
2285 /// isSignBitCheck - Given an exploded icmp instruction, return true if the
2286 /// comparison only checks the sign bit. If it only checks the sign bit, set
2287 /// TrueIfSigned if the result of the comparison is true when the input value is
2289 static bool isSignBitCheck(ICmpInst::Predicate pred, ConstantInt *RHS,
2290 bool &TrueIfSigned) {
2292 case ICmpInst::ICMP_SLT: // True if LHS s< 0
2293 TrueIfSigned = true;
2294 return RHS->isZero();
2295 case ICmpInst::ICMP_SLE: // True if LHS s<= RHS and RHS == -1
2296 TrueIfSigned = true;
2297 return RHS->isAllOnesValue();
2298 case ICmpInst::ICMP_SGT: // True if LHS s> -1
2299 TrueIfSigned = false;
2300 return RHS->isAllOnesValue();
2301 case ICmpInst::ICMP_UGT:
2302 // True if LHS u> RHS and RHS == high-bit-mask - 1
2303 TrueIfSigned = true;
2304 return RHS->getValue() ==
2305 APInt::getSignedMaxValue(RHS->getType()->getPrimitiveSizeInBits());
2306 case ICmpInst::ICMP_UGE:
2307 // True if LHS u>= RHS and RHS == high-bit-mask (2^7, 2^15, 2^31, etc)
2308 TrueIfSigned = true;
2309 return RHS->getValue() ==
2310 APInt::getSignBit(RHS->getType()->getPrimitiveSizeInBits());
2316 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2317 bool Changed = SimplifyCommutative(I);
2318 Value *Op0 = I.getOperand(0);
2320 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2321 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2323 // Simplify mul instructions with a constant RHS...
2324 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2325 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2327 // ((X << C1)*C2) == (X * (C2 << C1))
2328 if (BinaryOperator *SI = dyn_cast<BinaryOperator>(Op0))
2329 if (SI->getOpcode() == Instruction::Shl)
2330 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2331 return BinaryOperator::createMul(SI->getOperand(0),
2332 ConstantExpr::getShl(CI, ShOp));
2335 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2336 if (CI->equalsInt(1)) // X * 1 == X
2337 return ReplaceInstUsesWith(I, Op0);
2338 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2339 return BinaryOperator::createNeg(Op0, I.getName());
2341 const APInt& Val = cast<ConstantInt>(CI)->getValue();
2342 if (Val.isPowerOf2()) { // Replace X*(2^C) with X << C
2343 return BinaryOperator::createShl(Op0,
2344 ConstantInt::get(Op0->getType(), Val.logBase2()));
2346 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2347 if (Op1F->isNullValue())
2348 return ReplaceInstUsesWith(I, Op1);
2350 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2351 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2352 if (Op1F->isExactlyValue(1.0))
2353 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2356 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2357 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2358 isa<ConstantInt>(Op0I->getOperand(1))) {
2359 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2360 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2362 InsertNewInstBefore(Add, I);
2363 Value *C1C2 = ConstantExpr::getMul(Op1,
2364 cast<Constant>(Op0I->getOperand(1)));
2365 return BinaryOperator::createAdd(Add, C1C2);
2369 // Try to fold constant mul into select arguments.
2370 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2371 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2374 if (isa<PHINode>(Op0))
2375 if (Instruction *NV = FoldOpIntoPhi(I))
2379 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2380 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2381 return BinaryOperator::createMul(Op0v, Op1v);
2383 // If one of the operands of the multiply is a cast from a boolean value, then
2384 // we know the bool is either zero or one, so this is a 'masking' multiply.
2385 // See if we can simplify things based on how the boolean was originally
2387 CastInst *BoolCast = 0;
2388 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2389 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2392 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2393 if (CI->getOperand(0)->getType() == Type::Int1Ty)
2396 if (ICmpInst *SCI = dyn_cast<ICmpInst>(BoolCast->getOperand(0))) {
2397 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2398 const Type *SCOpTy = SCIOp0->getType();
2401 // If the icmp is true iff the sign bit of X is set, then convert this
2402 // multiply into a shift/and combination.
2403 if (isa<ConstantInt>(SCIOp1) &&
2404 isSignBitCheck(SCI->getPredicate(), cast<ConstantInt>(SCIOp1), TIS) &&
2406 // Shift the X value right to turn it into "all signbits".
2407 Constant *Amt = ConstantInt::get(SCIOp0->getType(),
2408 SCOpTy->getPrimitiveSizeInBits()-1);
2410 InsertNewInstBefore(
2411 BinaryOperator::create(Instruction::AShr, SCIOp0, Amt,
2412 BoolCast->getOperand(0)->getName()+
2415 // If the multiply type is not the same as the source type, sign extend
2416 // or truncate to the multiply type.
2417 if (I.getType() != V->getType()) {
2418 uint32_t SrcBits = V->getType()->getPrimitiveSizeInBits();
2419 uint32_t DstBits = I.getType()->getPrimitiveSizeInBits();
2420 Instruction::CastOps opcode =
2421 (SrcBits == DstBits ? Instruction::BitCast :
2422 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2423 V = InsertCastBefore(opcode, V, I.getType(), I);
2426 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2427 return BinaryOperator::createAnd(V, OtherOp);
2432 return Changed ? &I : 0;
2435 /// This function implements the transforms on div instructions that work
2436 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2437 /// used by the visitors to those instructions.
2438 /// @brief Transforms common to all three div instructions
2439 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2440 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2443 if (isa<UndefValue>(Op0))
2444 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2446 // X / undef -> undef
2447 if (isa<UndefValue>(Op1))
2448 return ReplaceInstUsesWith(I, Op1);
2450 // Handle cases involving: div X, (select Cond, Y, Z)
2451 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2452 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2453 // same basic block, then we replace the select with Y, and the condition
2454 // of the select with false (if the cond value is in the same BB). If the
2455 // select has uses other than the div, this allows them to be simplified
2456 // also. Note that div X, Y is just as good as div X, 0 (undef)
2457 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2458 if (ST->isNullValue()) {
2459 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2460 if (CondI && CondI->getParent() == I.getParent())
2461 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2462 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2463 I.setOperand(1, SI->getOperand(2));
2465 UpdateValueUsesWith(SI, SI->getOperand(2));
2469 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2470 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2471 if (ST->isNullValue()) {
2472 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2473 if (CondI && CondI->getParent() == I.getParent())
2474 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2475 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2476 I.setOperand(1, SI->getOperand(1));
2478 UpdateValueUsesWith(SI, SI->getOperand(1));
2486 /// This function implements the transforms common to both integer division
2487 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2488 /// division instructions.
2489 /// @brief Common integer divide transforms
2490 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2491 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2493 if (Instruction *Common = commonDivTransforms(I))
2496 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2498 if (RHS->equalsInt(1))
2499 return ReplaceInstUsesWith(I, Op0);
2501 // (X / C1) / C2 -> X / (C1*C2)
2502 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2503 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2504 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2505 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2506 Multiply(RHS, LHSRHS));
2509 if (!RHS->isZero()) { // avoid X udiv 0
2510 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2511 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2513 if (isa<PHINode>(Op0))
2514 if (Instruction *NV = FoldOpIntoPhi(I))
2519 // 0 / X == 0, we don't need to preserve faults!
2520 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2521 if (LHS->equalsInt(0))
2522 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2527 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2528 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2530 // Handle the integer div common cases
2531 if (Instruction *Common = commonIDivTransforms(I))
2534 // X udiv C^2 -> X >> C
2535 // Check to see if this is an unsigned division with an exact power of 2,
2536 // if so, convert to a right shift.
2537 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2538 if (C->getValue().isPowerOf2()) // 0 not included in isPowerOf2
2539 return BinaryOperator::createLShr(Op0,
2540 ConstantInt::get(Op0->getType(), C->getValue().logBase2()));
2543 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2544 if (BinaryOperator *RHSI = dyn_cast<BinaryOperator>(I.getOperand(1))) {
2545 if (RHSI->getOpcode() == Instruction::Shl &&
2546 isa<ConstantInt>(RHSI->getOperand(0))) {
2547 const APInt& C1 = cast<ConstantInt>(RHSI->getOperand(0))->getValue();
2548 if (C1.isPowerOf2()) {
2549 Value *N = RHSI->getOperand(1);
2550 const Type *NTy = N->getType();
2551 if (uint32_t C2 = C1.logBase2()) {
2552 Constant *C2V = ConstantInt::get(NTy, C2);
2553 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2555 return BinaryOperator::createLShr(Op0, N);
2560 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2561 // where C1&C2 are powers of two.
2562 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
2563 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2564 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2565 const APInt &TVA = STO->getValue(), &FVA = SFO->getValue();
2566 if (TVA.isPowerOf2() && FVA.isPowerOf2()) {
2567 // Compute the shift amounts
2568 uint32_t TSA = TVA.logBase2(), FSA = FVA.logBase2();
2569 // Construct the "on true" case of the select
2570 Constant *TC = ConstantInt::get(Op0->getType(), TSA);
2571 Instruction *TSI = BinaryOperator::createLShr(
2572 Op0, TC, SI->getName()+".t");
2573 TSI = InsertNewInstBefore(TSI, I);
2575 // Construct the "on false" case of the select
2576 Constant *FC = ConstantInt::get(Op0->getType(), FSA);
2577 Instruction *FSI = BinaryOperator::createLShr(
2578 Op0, FC, SI->getName()+".f");
2579 FSI = InsertNewInstBefore(FSI, I);
2581 // construct the select instruction and return it.
2582 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2588 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2589 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2591 // Handle the integer div common cases
2592 if (Instruction *Common = commonIDivTransforms(I))
2595 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2597 if (RHS->isAllOnesValue())
2598 return BinaryOperator::createNeg(Op0);
2601 if (Value *LHSNeg = dyn_castNegVal(Op0))
2602 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2605 // If the sign bits of both operands are zero (i.e. we can prove they are
2606 // unsigned inputs), turn this into a udiv.
2607 if (I.getType()->isInteger()) {
2608 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2609 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2610 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2617 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2618 return commonDivTransforms(I);
2621 /// GetFactor - If we can prove that the specified value is at least a multiple
2622 /// of some factor, return that factor.
2623 static Constant *GetFactor(Value *V) {
2624 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2627 // Unless we can be tricky, we know this is a multiple of 1.
2628 Constant *Result = ConstantInt::get(V->getType(), 1);
2630 Instruction *I = dyn_cast<Instruction>(V);
2631 if (!I) return Result;
2633 if (I->getOpcode() == Instruction::Mul) {
2634 // Handle multiplies by a constant, etc.
2635 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2636 GetFactor(I->getOperand(1)));
2637 } else if (I->getOpcode() == Instruction::Shl) {
2638 // (X<<C) -> X * (1 << C)
2639 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2640 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2641 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2643 } else if (I->getOpcode() == Instruction::And) {
2644 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2645 // X & 0xFFF0 is known to be a multiple of 16.
2646 uint32_t Zeros = RHS->getValue().countTrailingZeros();
2647 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2648 return ConstantExpr::getShl(Result,
2649 ConstantInt::get(Result->getType(), Zeros));
2651 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2652 // Only handle int->int casts.
2653 if (!CI->isIntegerCast())
2655 Value *Op = CI->getOperand(0);
2656 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2661 /// This function implements the transforms on rem instructions that work
2662 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2663 /// is used by the visitors to those instructions.
2664 /// @brief Transforms common to all three rem instructions
2665 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2666 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2668 // 0 % X == 0, we don't need to preserve faults!
2669 if (Constant *LHS = dyn_cast<Constant>(Op0))
2670 if (LHS->isNullValue())
2671 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2673 if (isa<UndefValue>(Op0)) // undef % X -> 0
2674 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2675 if (isa<UndefValue>(Op1))
2676 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2678 // Handle cases involving: rem X, (select Cond, Y, Z)
2679 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2680 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2681 // the same basic block, then we replace the select with Y, and the
2682 // condition of the select with false (if the cond value is in the same
2683 // BB). If the select has uses other than the div, this allows them to be
2685 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2686 if (ST->isNullValue()) {
2687 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2688 if (CondI && CondI->getParent() == I.getParent())
2689 UpdateValueUsesWith(CondI, ConstantInt::getFalse());
2690 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2691 I.setOperand(1, SI->getOperand(2));
2693 UpdateValueUsesWith(SI, SI->getOperand(2));
2696 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2697 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2698 if (ST->isNullValue()) {
2699 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2700 if (CondI && CondI->getParent() == I.getParent())
2701 UpdateValueUsesWith(CondI, ConstantInt::getTrue());
2702 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2703 I.setOperand(1, SI->getOperand(1));
2705 UpdateValueUsesWith(SI, SI->getOperand(1));
2713 /// This function implements the transforms common to both integer remainder
2714 /// instructions (urem and srem). It is called by the visitors to those integer
2715 /// remainder instructions.
2716 /// @brief Common integer remainder transforms
2717 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2718 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2720 if (Instruction *common = commonRemTransforms(I))
2723 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2724 // X % 0 == undef, we don't need to preserve faults!
2725 if (RHS->equalsInt(0))
2726 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2728 if (RHS->equalsInt(1)) // X % 1 == 0
2729 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2731 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2732 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2733 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2735 } else if (isa<PHINode>(Op0I)) {
2736 if (Instruction *NV = FoldOpIntoPhi(I))
2739 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2740 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2741 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2748 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2749 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2751 if (Instruction *common = commonIRemTransforms(I))
2754 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2755 // X urem C^2 -> X and C
2756 // Check to see if this is an unsigned remainder with an exact power of 2,
2757 // if so, convert to a bitwise and.
2758 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2759 if (C->getValue().isPowerOf2())
2760 return BinaryOperator::createAnd(Op0, SubOne(C));
2763 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2764 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2765 if (RHSI->getOpcode() == Instruction::Shl &&
2766 isa<ConstantInt>(RHSI->getOperand(0))) {
2767 if (cast<ConstantInt>(RHSI->getOperand(0))->getValue().isPowerOf2()) {
2768 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2769 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2771 return BinaryOperator::createAnd(Op0, Add);
2776 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2777 // where C1&C2 are powers of two.
2778 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2779 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2780 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2781 // STO == 0 and SFO == 0 handled above.
2782 if ((STO->getValue().isPowerOf2()) &&
2783 (SFO->getValue().isPowerOf2())) {
2784 Value *TrueAnd = InsertNewInstBefore(
2785 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2786 Value *FalseAnd = InsertNewInstBefore(
2787 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2788 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2796 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2797 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2799 if (Instruction *common = commonIRemTransforms(I))
2802 if (Value *RHSNeg = dyn_castNegVal(Op1))
2803 if (!isa<ConstantInt>(RHSNeg) ||
2804 cast<ConstantInt>(RHSNeg)->getValue().isStrictlyPositive()) {
2806 AddUsesToWorkList(I);
2807 I.setOperand(1, RHSNeg);
2811 // If the top bits of both operands are zero (i.e. we can prove they are
2812 // unsigned inputs), turn this into a urem.
2813 APInt Mask(APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()));
2814 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2815 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2816 return BinaryOperator::createURem(Op0, Op1, I.getName());
2822 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2823 return commonRemTransforms(I);
2826 // isMaxValueMinusOne - return true if this is Max-1
2827 static bool isMaxValueMinusOne(const ConstantInt *C, bool isSigned) {
2828 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2830 return C->getValue() == APInt::getAllOnesValue(TypeBits) - 1;
2831 return C->getValue() == APInt::getSignedMaxValue(TypeBits)-1;
2834 // isMinValuePlusOne - return true if this is Min+1
2835 static bool isMinValuePlusOne(const ConstantInt *C, bool isSigned) {
2837 return C->getValue() == 1; // unsigned
2839 // Calculate 1111111111000000000000
2840 uint32_t TypeBits = C->getType()->getPrimitiveSizeInBits();
2841 return C->getValue() == APInt::getSignedMinValue(TypeBits)+1;
2844 // isOneBitSet - Return true if there is exactly one bit set in the specified
2846 static bool isOneBitSet(const ConstantInt *CI) {
2847 return CI->getValue().isPowerOf2();
2850 // isHighOnes - Return true if the constant is of the form 1+0+.
2851 // This is the same as lowones(~X).
2852 static bool isHighOnes(const ConstantInt *CI) {
2853 return (~CI->getValue() + 1).isPowerOf2();
2856 /// getICmpCode - Encode a icmp predicate into a three bit mask. These bits
2857 /// are carefully arranged to allow folding of expressions such as:
2859 /// (A < B) | (A > B) --> (A != B)
2861 /// Note that this is only valid if the first and second predicates have the
2862 /// same sign. Is illegal to do: (A u< B) | (A s> B)
2864 /// Three bits are used to represent the condition, as follows:
2869 /// <=> Value Definition
2870 /// 000 0 Always false
2877 /// 111 7 Always true
2879 static unsigned getICmpCode(const ICmpInst *ICI) {
2880 switch (ICI->getPredicate()) {
2882 case ICmpInst::ICMP_UGT: return 1; // 001
2883 case ICmpInst::ICMP_SGT: return 1; // 001
2884 case ICmpInst::ICMP_EQ: return 2; // 010
2885 case ICmpInst::ICMP_UGE: return 3; // 011
2886 case ICmpInst::ICMP_SGE: return 3; // 011
2887 case ICmpInst::ICMP_ULT: return 4; // 100
2888 case ICmpInst::ICMP_SLT: return 4; // 100
2889 case ICmpInst::ICMP_NE: return 5; // 101
2890 case ICmpInst::ICMP_ULE: return 6; // 110
2891 case ICmpInst::ICMP_SLE: return 6; // 110
2894 assert(0 && "Invalid ICmp predicate!");
2899 /// getICmpValue - This is the complement of getICmpCode, which turns an
2900 /// opcode and two operands into either a constant true or false, or a brand
2901 /// new /// ICmp instruction. The sign is passed in to determine which kind
2902 /// of predicate to use in new icmp instructions.
2903 static Value *getICmpValue(bool sign, unsigned code, Value *LHS, Value *RHS) {
2905 default: assert(0 && "Illegal ICmp code!");
2906 case 0: return ConstantInt::getFalse();
2909 return new ICmpInst(ICmpInst::ICMP_SGT, LHS, RHS);
2911 return new ICmpInst(ICmpInst::ICMP_UGT, LHS, RHS);
2912 case 2: return new ICmpInst(ICmpInst::ICMP_EQ, LHS, RHS);
2915 return new ICmpInst(ICmpInst::ICMP_SGE, LHS, RHS);
2917 return new ICmpInst(ICmpInst::ICMP_UGE, LHS, RHS);
2920 return new ICmpInst(ICmpInst::ICMP_SLT, LHS, RHS);
2922 return new ICmpInst(ICmpInst::ICMP_ULT, LHS, RHS);
2923 case 5: return new ICmpInst(ICmpInst::ICMP_NE, LHS, RHS);
2926 return new ICmpInst(ICmpInst::ICMP_SLE, LHS, RHS);
2928 return new ICmpInst(ICmpInst::ICMP_ULE, LHS, RHS);
2929 case 7: return ConstantInt::getTrue();
2933 static bool PredicatesFoldable(ICmpInst::Predicate p1, ICmpInst::Predicate p2) {
2934 return (ICmpInst::isSignedPredicate(p1) == ICmpInst::isSignedPredicate(p2)) ||
2935 (ICmpInst::isSignedPredicate(p1) &&
2936 (p2 == ICmpInst::ICMP_EQ || p2 == ICmpInst::ICMP_NE)) ||
2937 (ICmpInst::isSignedPredicate(p2) &&
2938 (p1 == ICmpInst::ICMP_EQ || p1 == ICmpInst::ICMP_NE));
2942 // FoldICmpLogical - Implements (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
2943 struct FoldICmpLogical {
2946 ICmpInst::Predicate pred;
2947 FoldICmpLogical(InstCombiner &ic, ICmpInst *ICI)
2948 : IC(ic), LHS(ICI->getOperand(0)), RHS(ICI->getOperand(1)),
2949 pred(ICI->getPredicate()) {}
2950 bool shouldApply(Value *V) const {
2951 if (ICmpInst *ICI = dyn_cast<ICmpInst>(V))
2952 if (PredicatesFoldable(pred, ICI->getPredicate()))
2953 return (ICI->getOperand(0) == LHS && ICI->getOperand(1) == RHS ||
2954 ICI->getOperand(0) == RHS && ICI->getOperand(1) == LHS);
2957 Instruction *apply(Instruction &Log) const {
2958 ICmpInst *ICI = cast<ICmpInst>(Log.getOperand(0));
2959 if (ICI->getOperand(0) != LHS) {
2960 assert(ICI->getOperand(1) == LHS);
2961 ICI->swapOperands(); // Swap the LHS and RHS of the ICmp
2964 ICmpInst *RHSICI = cast<ICmpInst>(Log.getOperand(1));
2965 unsigned LHSCode = getICmpCode(ICI);
2966 unsigned RHSCode = getICmpCode(RHSICI);
2968 switch (Log.getOpcode()) {
2969 case Instruction::And: Code = LHSCode & RHSCode; break;
2970 case Instruction::Or: Code = LHSCode | RHSCode; break;
2971 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2972 default: assert(0 && "Illegal logical opcode!"); return 0;
2975 bool isSigned = ICmpInst::isSignedPredicate(RHSICI->getPredicate()) ||
2976 ICmpInst::isSignedPredicate(ICI->getPredicate());
2978 Value *RV = getICmpValue(isSigned, Code, LHS, RHS);
2979 if (Instruction *I = dyn_cast<Instruction>(RV))
2981 // Otherwise, it's a constant boolean value...
2982 return IC.ReplaceInstUsesWith(Log, RV);
2985 } // end anonymous namespace
2987 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2988 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2989 // guaranteed to be a binary operator.
2990 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2992 ConstantInt *AndRHS,
2993 BinaryOperator &TheAnd) {
2994 Value *X = Op->getOperand(0);
2995 Constant *Together = 0;
2997 Together = And(AndRHS, OpRHS);
2999 switch (Op->getOpcode()) {
3000 case Instruction::Xor:
3001 if (Op->hasOneUse()) {
3002 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
3003 Instruction *And = BinaryOperator::createAnd(X, AndRHS);
3004 InsertNewInstBefore(And, TheAnd);
3006 return BinaryOperator::createXor(And, Together);
3009 case Instruction::Or:
3010 if (Together == AndRHS) // (X | C) & C --> C
3011 return ReplaceInstUsesWith(TheAnd, AndRHS);
3013 if (Op->hasOneUse() && Together != OpRHS) {
3014 // (X | C1) & C2 --> (X | (C1&C2)) & C2
3015 Instruction *Or = BinaryOperator::createOr(X, Together);
3016 InsertNewInstBefore(Or, TheAnd);
3018 return BinaryOperator::createAnd(Or, AndRHS);
3021 case Instruction::Add:
3022 if (Op->hasOneUse()) {
3023 // Adding a one to a single bit bit-field should be turned into an XOR
3024 // of the bit. First thing to check is to see if this AND is with a
3025 // single bit constant.
3026 const APInt& AndRHSV = cast<ConstantInt>(AndRHS)->getValue();
3028 // If there is only one bit set...
3029 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
3030 // Ok, at this point, we know that we are masking the result of the
3031 // ADD down to exactly one bit. If the constant we are adding has
3032 // no bits set below this bit, then we can eliminate the ADD.
3033 const APInt& AddRHS = cast<ConstantInt>(OpRHS)->getValue();
3035 // Check to see if any bits below the one bit set in AndRHSV are set.
3036 if ((AddRHS & (AndRHSV-1)) == 0) {
3037 // If not, the only thing that can effect the output of the AND is
3038 // the bit specified by AndRHSV. If that bit is set, the effect of
3039 // the XOR is to toggle the bit. If it is clear, then the ADD has
3041 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
3042 TheAnd.setOperand(0, X);
3045 // Pull the XOR out of the AND.
3046 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS);
3047 InsertNewInstBefore(NewAnd, TheAnd);
3048 NewAnd->takeName(Op);
3049 return BinaryOperator::createXor(NewAnd, AndRHS);
3056 case Instruction::Shl: {
3057 // We know that the AND will not produce any of the bits shifted in, so if
3058 // the anded constant includes them, clear them now!
3060 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3061 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3062 APInt ShlMask(APInt::getHighBitsSet(BitWidth, BitWidth-OpRHSVal));
3063 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShlMask);
3065 if (CI->getValue() == ShlMask) {
3066 // Masking out bits that the shift already masks
3067 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
3068 } else if (CI != AndRHS) { // Reducing bits set in and.
3069 TheAnd.setOperand(1, CI);
3074 case Instruction::LShr:
3076 // We know that the AND will not produce any of the bits shifted in, so if
3077 // the anded constant includes them, clear them now! This only applies to
3078 // unsigned shifts, because a signed shr may bring in set bits!
3080 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3081 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3082 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3083 ConstantInt *CI = ConstantInt::get(AndRHS->getValue() & ShrMask);
3085 if (CI->getValue() == ShrMask) {
3086 // Masking out bits that the shift already masks.
3087 return ReplaceInstUsesWith(TheAnd, Op);
3088 } else if (CI != AndRHS) {
3089 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
3094 case Instruction::AShr:
3096 // See if this is shifting in some sign extension, then masking it out
3098 if (Op->hasOneUse()) {
3099 uint32_t BitWidth = AndRHS->getType()->getBitWidth();
3100 uint32_t OpRHSVal = OpRHS->getLimitedValue(BitWidth);
3101 APInt ShrMask(APInt::getLowBitsSet(BitWidth, BitWidth - OpRHSVal));
3102 Constant *C = ConstantInt::get(AndRHS->getValue() & ShrMask);
3103 if (C == AndRHS) { // Masking out bits shifted in.
3104 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
3105 // Make the argument unsigned.
3106 Value *ShVal = Op->getOperand(0);
3107 ShVal = InsertNewInstBefore(
3108 BinaryOperator::createLShr(ShVal, OpRHS,
3109 Op->getName()), TheAnd);
3110 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
3119 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
3120 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
3121 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. isSigned indicates
3122 /// whether to treat the V, Lo and HI as signed or not. IB is the location to
3123 /// insert new instructions.
3124 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
3125 bool isSigned, bool Inside,
3127 assert(cast<ConstantInt>(ConstantExpr::getICmp((isSigned ?
3128 ICmpInst::ICMP_SLE:ICmpInst::ICMP_ULE), Lo, Hi))->getZExtValue() &&
3129 "Lo is not <= Hi in range emission code!");
3132 if (Lo == Hi) // Trivially false.
3133 return new ICmpInst(ICmpInst::ICMP_NE, V, V);
3135 // V >= Min && V < Hi --> V < Hi
3136 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3137 ICmpInst::Predicate pred = (isSigned ?
3138 ICmpInst::ICMP_SLT : ICmpInst::ICMP_ULT);
3139 return new ICmpInst(pred, V, Hi);
3142 // Emit V-Lo <u Hi-Lo
3143 Constant *NegLo = ConstantExpr::getNeg(Lo);
3144 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3145 InsertNewInstBefore(Add, IB);
3146 Constant *UpperBound = ConstantExpr::getAdd(NegLo, Hi);
3147 return new ICmpInst(ICmpInst::ICMP_ULT, Add, UpperBound);
3150 if (Lo == Hi) // Trivially true.
3151 return new ICmpInst(ICmpInst::ICMP_EQ, V, V);
3153 // V < Min || V >= Hi -> V > Hi-1
3154 Hi = SubOne(cast<ConstantInt>(Hi));
3155 if (cast<ConstantInt>(Lo)->isMinValue(isSigned)) {
3156 ICmpInst::Predicate pred = (isSigned ?
3157 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT);
3158 return new ICmpInst(pred, V, Hi);
3161 // Emit V-Lo >u Hi-1-Lo
3162 // Note that Hi has already had one subtracted from it, above.
3163 ConstantInt *NegLo = cast<ConstantInt>(ConstantExpr::getNeg(Lo));
3164 Instruction *Add = BinaryOperator::createAdd(V, NegLo, V->getName()+".off");
3165 InsertNewInstBefore(Add, IB);
3166 Constant *LowerBound = ConstantExpr::getAdd(NegLo, Hi);
3167 return new ICmpInst(ICmpInst::ICMP_UGT, Add, LowerBound);
3170 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
3171 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
3172 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
3173 // not, since all 1s are not contiguous.
3174 static bool isRunOfOnes(ConstantInt *Val, uint32_t &MB, uint32_t &ME) {
3175 const APInt& V = Val->getValue();
3176 uint32_t BitWidth = Val->getType()->getBitWidth();
3177 if (!APIntOps::isShiftedMask(BitWidth, V)) return false;
3179 // look for the first zero bit after the run of ones
3180 MB = BitWidth - ((V - 1) ^ V).countLeadingZeros();
3181 // look for the first non-zero bit
3182 ME = V.getActiveBits();
3186 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
3187 /// where isSub determines whether the operator is a sub. If we can fold one of
3188 /// the following xforms:
3190 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
3191 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3192 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
3194 /// return (A +/- B).
3196 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
3197 ConstantInt *Mask, bool isSub,
3199 Instruction *LHSI = dyn_cast<Instruction>(LHS);
3200 if (!LHSI || LHSI->getNumOperands() != 2 ||
3201 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
3203 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
3205 switch (LHSI->getOpcode()) {
3207 case Instruction::And:
3208 if (And(N, Mask) == Mask) {
3209 // If the AndRHS is a power of two minus one (0+1+), this is simple.
3210 if ((Mask->getValue().countLeadingZeros() +
3211 Mask->getValue().countPopulation()) ==
3212 Mask->getValue().getBitWidth())
3215 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
3216 // part, we don't need any explicit masks to take them out of A. If that
3217 // is all N is, ignore it.
3218 uint32_t MB = 0, ME = 0;
3219 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
3220 uint32_t BitWidth = cast<IntegerType>(RHS->getType())->getBitWidth();
3221 APInt Mask(APInt::getLowBitsSet(BitWidth, MB-1));
3222 if (MaskedValueIsZero(RHS, Mask))
3227 case Instruction::Or:
3228 case Instruction::Xor:
3229 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
3230 if ((Mask->getValue().countLeadingZeros() +
3231 Mask->getValue().countPopulation()) == Mask->getValue().getBitWidth()
3232 && And(N, Mask)->isZero())
3239 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
3241 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
3242 return InsertNewInstBefore(New, I);
3245 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3246 bool Changed = SimplifyCommutative(I);
3247 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3249 if (isa<UndefValue>(Op1)) // X & undef -> 0
3250 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3254 return ReplaceInstUsesWith(I, Op1);
3256 // See if we can simplify any instructions used by the instruction whose sole
3257 // purpose is to compute bits we don't care about.
3258 if (!isa<VectorType>(I.getType())) {
3259 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3260 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3261 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3262 KnownZero, KnownOne))
3265 if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3266 if (CP->isAllOnesValue()) // X & <-1,-1> -> X
3267 return ReplaceInstUsesWith(I, I.getOperand(0));
3268 } else if (isa<ConstantAggregateZero>(Op1)) {
3269 return ReplaceInstUsesWith(I, Op1); // X & <0,0> -> <0,0>
3273 if (ConstantInt *AndRHS = dyn_cast<ConstantInt>(Op1)) {
3274 const APInt& AndRHSMask = AndRHS->getValue();
3275 APInt NotAndRHS(~AndRHSMask);
3277 // Optimize a variety of ((val OP C1) & C2) combinations...
3278 if (isa<BinaryOperator>(Op0)) {
3279 Instruction *Op0I = cast<Instruction>(Op0);
3280 Value *Op0LHS = Op0I->getOperand(0);
3281 Value *Op0RHS = Op0I->getOperand(1);
3282 switch (Op0I->getOpcode()) {
3283 case Instruction::Xor:
3284 case Instruction::Or:
3285 // If the mask is only needed on one incoming arm, push it up.
3286 if (Op0I->hasOneUse()) {
3287 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3288 // Not masking anything out for the LHS, move to RHS.
3289 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3290 Op0RHS->getName()+".masked");
3291 InsertNewInstBefore(NewRHS, I);
3292 return BinaryOperator::create(
3293 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3295 if (!isa<Constant>(Op0RHS) &&
3296 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3297 // Not masking anything out for the RHS, move to LHS.
3298 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3299 Op0LHS->getName()+".masked");
3300 InsertNewInstBefore(NewLHS, I);
3301 return BinaryOperator::create(
3302 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3307 case Instruction::Add:
3308 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3309 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3310 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3311 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3312 return BinaryOperator::createAnd(V, AndRHS);
3313 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3314 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3317 case Instruction::Sub:
3318 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3319 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3320 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3321 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3322 return BinaryOperator::createAnd(V, AndRHS);
3326 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3327 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3329 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3330 // If this is an integer truncation or change from signed-to-unsigned, and
3331 // if the source is an and/or with immediate, transform it. This
3332 // frequently occurs for bitfield accesses.
3333 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3334 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3335 CastOp->getNumOperands() == 2)
3336 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3337 if (CastOp->getOpcode() == Instruction::And) {
3338 // Change: and (cast (and X, C1) to T), C2
3339 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3340 // This will fold the two constants together, which may allow
3341 // other simplifications.
3342 Instruction *NewCast = CastInst::createTruncOrBitCast(
3343 CastOp->getOperand(0), I.getType(),
3344 CastOp->getName()+".shrunk");
3345 NewCast = InsertNewInstBefore(NewCast, I);
3346 // trunc_or_bitcast(C1)&C2
3347 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3348 C3 = ConstantExpr::getAnd(C3, AndRHS);
3349 return BinaryOperator::createAnd(NewCast, C3);
3350 } else if (CastOp->getOpcode() == Instruction::Or) {
3351 // Change: and (cast (or X, C1) to T), C2
3352 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3353 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3354 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3355 return ReplaceInstUsesWith(I, AndRHS);
3360 // Try to fold constant and into select arguments.
3361 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3362 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3364 if (isa<PHINode>(Op0))
3365 if (Instruction *NV = FoldOpIntoPhi(I))
3369 Value *Op0NotVal = dyn_castNotVal(Op0);
3370 Value *Op1NotVal = dyn_castNotVal(Op1);
3372 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3373 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3375 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3376 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3377 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3378 I.getName()+".demorgan");
3379 InsertNewInstBefore(Or, I);
3380 return BinaryOperator::createNot(Or);
3384 Value *A = 0, *B = 0, *C = 0, *D = 0;
3385 if (match(Op0, m_Or(m_Value(A), m_Value(B)))) {
3386 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3387 return ReplaceInstUsesWith(I, Op1);
3389 // (A|B) & ~(A&B) -> A^B
3390 if (match(Op1, m_Not(m_And(m_Value(C), m_Value(D))))) {
3391 if ((A == C && B == D) || (A == D && B == C))
3392 return BinaryOperator::createXor(A, B);
3396 if (match(Op1, m_Or(m_Value(A), m_Value(B)))) {
3397 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3398 return ReplaceInstUsesWith(I, Op0);
3400 // ~(A&B) & (A|B) -> A^B
3401 if (match(Op0, m_Not(m_And(m_Value(C), m_Value(D))))) {
3402 if ((A == C && B == D) || (A == D && B == C))
3403 return BinaryOperator::createXor(A, B);
3407 if (Op0->hasOneUse() &&
3408 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3409 if (A == Op1) { // (A^B)&A -> A&(A^B)
3410 I.swapOperands(); // Simplify below
3411 std::swap(Op0, Op1);
3412 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3413 cast<BinaryOperator>(Op0)->swapOperands();
3414 I.swapOperands(); // Simplify below
3415 std::swap(Op0, Op1);
3418 if (Op1->hasOneUse() &&
3419 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3420 if (B == Op0) { // B&(A^B) -> B&(B^A)
3421 cast<BinaryOperator>(Op1)->swapOperands();
3424 if (A == Op0) { // A&(A^B) -> A & ~B
3425 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3426 InsertNewInstBefore(NotB, I);
3427 return BinaryOperator::createAnd(A, NotB);
3432 if (ICmpInst *RHS = dyn_cast<ICmpInst>(Op1)) {
3433 // (icmp1 A, B) & (icmp2 A, B) --> (icmp3 A, B)
3434 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3437 Value *LHSVal, *RHSVal;
3438 ConstantInt *LHSCst, *RHSCst;
3439 ICmpInst::Predicate LHSCC, RHSCC;
3440 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3441 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3442 if (LHSVal == RHSVal && // Found (X icmp C1) & (X icmp C2)
3443 // ICMP_[GL]E X, CST is folded to ICMP_[GL]T elsewhere.
3444 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3445 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3446 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3447 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE) {
3448 // Ensure that the larger constant is on the RHS.
3449 ICmpInst::Predicate GT = ICmpInst::isSignedPredicate(LHSCC) ?
3450 ICmpInst::ICMP_SGT : ICmpInst::ICMP_UGT;
3451 Constant *Cmp = ConstantExpr::getICmp(GT, LHSCst, RHSCst);
3452 ICmpInst *LHS = cast<ICmpInst>(Op0);
3453 if (cast<ConstantInt>(Cmp)->getZExtValue()) {
3454 std::swap(LHS, RHS);
3455 std::swap(LHSCst, RHSCst);
3456 std::swap(LHSCC, RHSCC);
3459 // At this point, we know we have have two icmp instructions
3460 // comparing a value against two constants and and'ing the result
3461 // together. Because of the above check, we know that we only have
3462 // icmp eq, icmp ne, icmp [su]lt, and icmp [SU]gt here. We also know
3463 // (from the FoldICmpLogical check above), that the two constants
3464 // are not equal and that the larger constant is on the RHS
3465 assert(LHSCst != RHSCst && "Compares not folded above?");
3468 default: assert(0 && "Unknown integer condition code!");
3469 case ICmpInst::ICMP_EQ:
3471 default: assert(0 && "Unknown integer condition code!");
3472 case ICmpInst::ICMP_EQ: // (X == 13 & X == 15) -> false
3473 case ICmpInst::ICMP_UGT: // (X == 13 & X > 15) -> false
3474 case ICmpInst::ICMP_SGT: // (X == 13 & X > 15) -> false
3475 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3476 case ICmpInst::ICMP_NE: // (X == 13 & X != 15) -> X == 13
3477 case ICmpInst::ICMP_ULT: // (X == 13 & X < 15) -> X == 13
3478 case ICmpInst::ICMP_SLT: // (X == 13 & X < 15) -> X == 13
3479 return ReplaceInstUsesWith(I, LHS);
3481 case ICmpInst::ICMP_NE:
3483 default: assert(0 && "Unknown integer condition code!");
3484 case ICmpInst::ICMP_ULT:
3485 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X u< 14) -> X < 13
3486 return new ICmpInst(ICmpInst::ICMP_ULT, LHSVal, LHSCst);
3487 break; // (X != 13 & X u< 15) -> no change
3488 case ICmpInst::ICMP_SLT:
3489 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X s< 14) -> X < 13
3490 return new ICmpInst(ICmpInst::ICMP_SLT, LHSVal, LHSCst);
3491 break; // (X != 13 & X s< 15) -> no change
3492 case ICmpInst::ICMP_EQ: // (X != 13 & X == 15) -> X == 15
3493 case ICmpInst::ICMP_UGT: // (X != 13 & X u> 15) -> X u> 15
3494 case ICmpInst::ICMP_SGT: // (X != 13 & X s> 15) -> X s> 15
3495 return ReplaceInstUsesWith(I, RHS);
3496 case ICmpInst::ICMP_NE:
3497 if (LHSCst == SubOne(RHSCst)){// (X != 13 & X != 14) -> X-13 >u 1
3498 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3499 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3500 LHSVal->getName()+".off");
3501 InsertNewInstBefore(Add, I);
3502 return new ICmpInst(ICmpInst::ICMP_UGT, Add,
3503 ConstantInt::get(Add->getType(), 1));
3505 break; // (X != 13 & X != 15) -> no change
3508 case ICmpInst::ICMP_ULT:
3510 default: assert(0 && "Unknown integer condition code!");
3511 case ICmpInst::ICMP_EQ: // (X u< 13 & X == 15) -> false
3512 case ICmpInst::ICMP_UGT: // (X u< 13 & X u> 15) -> false
3513 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3514 case ICmpInst::ICMP_SGT: // (X u< 13 & X s> 15) -> no change
3516 case ICmpInst::ICMP_NE: // (X u< 13 & X != 15) -> X u< 13
3517 case ICmpInst::ICMP_ULT: // (X u< 13 & X u< 15) -> X u< 13
3518 return ReplaceInstUsesWith(I, LHS);
3519 case ICmpInst::ICMP_SLT: // (X u< 13 & X s< 15) -> no change
3523 case ICmpInst::ICMP_SLT:
3525 default: assert(0 && "Unknown integer condition code!");
3526 case ICmpInst::ICMP_EQ: // (X s< 13 & X == 15) -> false
3527 case ICmpInst::ICMP_SGT: // (X s< 13 & X s> 15) -> false
3528 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
3529 case ICmpInst::ICMP_UGT: // (X s< 13 & X u> 15) -> no change
3531 case ICmpInst::ICMP_NE: // (X s< 13 & X != 15) -> X < 13
3532 case ICmpInst::ICMP_SLT: // (X s< 13 & X s< 15) -> X < 13
3533 return ReplaceInstUsesWith(I, LHS);
3534 case ICmpInst::ICMP_ULT: // (X s< 13 & X u< 15) -> no change
3538 case ICmpInst::ICMP_UGT:
3540 default: assert(0 && "Unknown integer condition code!");
3541 case ICmpInst::ICMP_EQ: // (X u> 13 & X == 15) -> X > 13
3542 return ReplaceInstUsesWith(I, LHS);
3543 case ICmpInst::ICMP_UGT: // (X u> 13 & X u> 15) -> X u> 15
3544 return ReplaceInstUsesWith(I, RHS);
3545 case ICmpInst::ICMP_SGT: // (X u> 13 & X s> 15) -> no change
3547 case ICmpInst::ICMP_NE:
3548 if (RHSCst == AddOne(LHSCst)) // (X u> 13 & X != 14) -> X u> 14
3549 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3550 break; // (X u> 13 & X != 15) -> no change
3551 case ICmpInst::ICMP_ULT: // (X u> 13 & X u< 15) ->(X-14) <u 1
3552 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, false,
3554 case ICmpInst::ICMP_SLT: // (X u> 13 & X s< 15) -> no change
3558 case ICmpInst::ICMP_SGT:
3560 default: assert(0 && "Unknown integer condition code!");
3561 case ICmpInst::ICMP_EQ: // (X s> 13 & X == 15) -> X s> 13
3562 return ReplaceInstUsesWith(I, LHS);
3563 case ICmpInst::ICMP_SGT: // (X s> 13 & X s> 15) -> X s> 15
3564 return ReplaceInstUsesWith(I, RHS);
3565 case ICmpInst::ICMP_UGT: // (X s> 13 & X u> 15) -> no change
3567 case ICmpInst::ICMP_NE:
3568 if (RHSCst == AddOne(LHSCst)) // (X s> 13 & X != 14) -> X s> 14
3569 return new ICmpInst(LHSCC, LHSVal, RHSCst);
3570 break; // (X s> 13 & X != 15) -> no change
3571 case ICmpInst::ICMP_SLT: // (X s> 13 & X s< 15) ->(X-14) s< 1
3572 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true,
3574 case ICmpInst::ICMP_ULT: // (X s> 13 & X u< 15) -> no change
3582 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3583 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3584 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3585 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3586 const Type *SrcTy = Op0C->getOperand(0)->getType();
3587 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
3588 // Only do this if the casts both really cause code to be generated.
3589 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
3591 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
3593 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3594 Op1C->getOperand(0),
3596 InsertNewInstBefore(NewOp, I);
3597 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3601 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3602 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3603 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3604 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3605 SI0->getOperand(1) == SI1->getOperand(1) &&
3606 (SI0->hasOneUse() || SI1->hasOneUse())) {
3607 Instruction *NewOp =
3608 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3610 SI0->getName()), I);
3611 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3612 SI1->getOperand(1));
3616 return Changed ? &I : 0;
3619 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3620 /// in the result. If it does, and if the specified byte hasn't been filled in
3621 /// yet, fill it in and return false.
3622 static bool CollectBSwapParts(Value *V, SmallVector<Value*, 8> &ByteValues) {
3623 Instruction *I = dyn_cast<Instruction>(V);
3624 if (I == 0) return true;
3626 // If this is an or instruction, it is an inner node of the bswap.
3627 if (I->getOpcode() == Instruction::Or)
3628 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3629 CollectBSwapParts(I->getOperand(1), ByteValues);
3631 uint32_t BitWidth = I->getType()->getPrimitiveSizeInBits();
3632 // If this is a shift by a constant int, and it is "24", then its operand
3633 // defines a byte. We only handle unsigned types here.
3634 if (I->isShift() && isa<ConstantInt>(I->getOperand(1))) {
3635 // Not shifting the entire input by N-1 bytes?
3636 if (cast<ConstantInt>(I->getOperand(1))->getLimitedValue(BitWidth) !=
3637 8*(ByteValues.size()-1))
3641 if (I->getOpcode() == Instruction::Shl) {
3642 // X << 24 defines the top byte with the lowest of the input bytes.
3643 DestNo = ByteValues.size()-1;
3645 // X >>u 24 defines the low byte with the highest of the input bytes.
3649 // If the destination byte value is already defined, the values are or'd
3650 // together, which isn't a bswap (unless it's an or of the same bits).
3651 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3653 ByteValues[DestNo] = I->getOperand(0);
3657 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3659 Value *Shift = 0, *ShiftLHS = 0;
3660 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3661 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3662 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3664 Instruction *SI = cast<Instruction>(Shift);
3666 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3667 if (ShiftAmt->getLimitedValue(BitWidth) & 7 ||
3668 ShiftAmt->getLimitedValue(BitWidth) > 8*ByteValues.size())
3671 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3673 if (AndAmt->getValue().getActiveBits() > 64)
3675 uint64_t AndAmtVal = AndAmt->getZExtValue();
3676 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3677 if (AndAmtVal == uint64_t(0xFF) << 8*DestByte)
3679 // Unknown mask for bswap.
3680 if (DestByte == ByteValues.size()) return true;
3682 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3684 if (SI->getOpcode() == Instruction::Shl)
3685 SrcByte = DestByte - ShiftBytes;
3687 SrcByte = DestByte + ShiftBytes;
3689 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3690 if (SrcByte != ByteValues.size()-DestByte-1)
3693 // If the destination byte value is already defined, the values are or'd
3694 // together, which isn't a bswap (unless it's an or of the same bits).
3695 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3697 ByteValues[DestByte] = SI->getOperand(0);
3701 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3702 /// If so, insert the new bswap intrinsic and return it.
3703 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3704 const IntegerType *ITy = dyn_cast<IntegerType>(I.getType());
3705 if (!ITy || ITy->getBitWidth() % 16)
3706 return 0; // Can only bswap pairs of bytes. Can't do vectors.
3708 /// ByteValues - For each byte of the result, we keep track of which value
3709 /// defines each byte.
3710 SmallVector<Value*, 8> ByteValues;
3711 ByteValues.resize(ITy->getBitWidth()/8);
3713 // Try to find all the pieces corresponding to the bswap.
3714 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3715 CollectBSwapParts(I.getOperand(1), ByteValues))
3718 // Check to see if all of the bytes come from the same value.
3719 Value *V = ByteValues[0];
3720 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3722 // Check to make sure that all of the bytes come from the same value.
3723 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3724 if (ByteValues[i] != V)
3726 const Type *Tys[] = { ITy };
3727 Module *M = I.getParent()->getParent()->getParent();
3728 Function *F = Intrinsic::getDeclaration(M, Intrinsic::bswap, Tys, 1);
3729 return new CallInst(F, V);
3733 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3734 bool Changed = SimplifyCommutative(I);
3735 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3737 if (isa<UndefValue>(Op1)) // X | undef -> -1
3738 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3742 return ReplaceInstUsesWith(I, Op0);
3744 // See if we can simplify any instructions used by the instruction whose sole
3745 // purpose is to compute bits we don't care about.
3746 if (!isa<VectorType>(I.getType())) {
3747 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
3748 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
3749 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
3750 KnownZero, KnownOne))
3752 } else if (isa<ConstantAggregateZero>(Op1)) {
3753 return ReplaceInstUsesWith(I, Op0); // X | <0,0> -> X
3754 } else if (ConstantVector *CP = dyn_cast<ConstantVector>(Op1)) {
3755 if (CP->isAllOnesValue()) // X | <-1,-1> -> <-1,-1>
3756 return ReplaceInstUsesWith(I, I.getOperand(1));
3762 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
3763 ConstantInt *C1 = 0; Value *X = 0;
3764 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3765 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3766 Instruction *Or = BinaryOperator::createOr(X, RHS);
3767 InsertNewInstBefore(Or, I);
3769 return BinaryOperator::createAnd(Or,
3770 ConstantInt::get(RHS->getValue() | C1->getValue()));
3773 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3774 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3775 Instruction *Or = BinaryOperator::createOr(X, RHS);
3776 InsertNewInstBefore(Or, I);
3778 return BinaryOperator::createXor(Or,
3779 ConstantInt::get(C1->getValue() & ~RHS->getValue()));
3782 // Try to fold constant and into select arguments.
3783 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3784 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3786 if (isa<PHINode>(Op0))
3787 if (Instruction *NV = FoldOpIntoPhi(I))
3791 Value *A = 0, *B = 0;
3792 ConstantInt *C1 = 0, *C2 = 0;
3794 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3795 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3796 return ReplaceInstUsesWith(I, Op1);
3797 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3798 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3799 return ReplaceInstUsesWith(I, Op0);
3801 // (A | B) | C and A | (B | C) -> bswap if possible.
3802 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3803 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3804 match(Op1, m_Or(m_Value(), m_Value())) ||
3805 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3806 match(Op1, m_Shift(m_Value(), m_Value())))) {
3807 if (Instruction *BSwap = MatchBSwap(I))
3811 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3812 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3813 MaskedValueIsZero(Op1, C1->getValue())) {
3814 Instruction *NOr = BinaryOperator::createOr(A, Op1);
3815 InsertNewInstBefore(NOr, I);
3817 return BinaryOperator::createXor(NOr, C1);
3820 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3821 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3822 MaskedValueIsZero(Op0, C1->getValue())) {
3823 Instruction *NOr = BinaryOperator::createOr(A, Op0);
3824 InsertNewInstBefore(NOr, I);
3826 return BinaryOperator::createXor(NOr, C1);
3830 Value *C = 0, *D = 0;
3831 if (match(Op0, m_And(m_Value(A), m_Value(C))) &&
3832 match(Op1, m_And(m_Value(B), m_Value(D)))) {
3833 Value *V1 = 0, *V2 = 0, *V3 = 0;
3834 C1 = dyn_cast<ConstantInt>(C);
3835 C2 = dyn_cast<ConstantInt>(D);
3836 if (C1 && C2) { // (A & C1)|(B & C2)
3837 // If we have: ((V + N) & C1) | (V & C2)
3838 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3839 // replace with V+N.
3840 if (C1->getValue() == ~C2->getValue()) {
3841 if ((C2->getValue() & (C2->getValue()+1)) == 0 && // C2 == 0+1+
3842 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3843 // Add commutes, try both ways.
3844 if (V1 == B && MaskedValueIsZero(V2, C2->getValue()))
3845 return ReplaceInstUsesWith(I, A);
3846 if (V2 == B && MaskedValueIsZero(V1, C2->getValue()))
3847 return ReplaceInstUsesWith(I, A);
3849 // Or commutes, try both ways.
3850 if ((C1->getValue() & (C1->getValue()+1)) == 0 &&
3851 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3852 // Add commutes, try both ways.
3853 if (V1 == A && MaskedValueIsZero(V2, C1->getValue()))
3854 return ReplaceInstUsesWith(I, B);
3855 if (V2 == A && MaskedValueIsZero(V1, C1->getValue()))
3856 return ReplaceInstUsesWith(I, B);
3859 V1 = 0; V2 = 0; V3 = 0;
3862 // Check to see if we have any common things being and'ed. If so, find the
3863 // terms for V1 & (V2|V3).
3864 if (isOnlyUse(Op0) || isOnlyUse(Op1)) {
3865 if (A == B) // (A & C)|(A & D) == A & (C|D)
3866 V1 = A, V2 = C, V3 = D;
3867 else if (A == D) // (A & C)|(B & A) == A & (B|C)
3868 V1 = A, V2 = B, V3 = C;
3869 else if (C == B) // (A & C)|(C & D) == C & (A|D)
3870 V1 = C, V2 = A, V3 = D;
3871 else if (C == D) // (A & C)|(B & C) == C & (A|B)
3872 V1 = C, V2 = A, V3 = B;
3876 InsertNewInstBefore(BinaryOperator::createOr(V2, V3, "tmp"), I);
3877 return BinaryOperator::createAnd(V1, Or);
3882 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3883 if (BinaryOperator *SI1 = dyn_cast<BinaryOperator>(Op1)) {
3884 if (BinaryOperator *SI0 = dyn_cast<BinaryOperator>(Op0))
3885 if (SI0->isShift() && SI0->getOpcode() == SI1->getOpcode() &&
3886 SI0->getOperand(1) == SI1->getOperand(1) &&
3887 (SI0->hasOneUse() || SI1->hasOneUse())) {
3888 Instruction *NewOp =
3889 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3891 SI0->getName()), I);
3892 return BinaryOperator::create(SI1->getOpcode(), NewOp,
3893 SI1->getOperand(1));
3897 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3898 if (A == Op1) // ~A | A == -1
3899 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3903 // Note, A is still live here!
3904 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3906 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
3908 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3909 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3910 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3911 I.getName()+".demorgan"), I);
3912 return BinaryOperator::createNot(And);
3916 // (icmp1 A, B) | (icmp2 A, B) --> (icmp3 A, B)
3917 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1))) {
3918 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
3921 Value *LHSVal, *RHSVal;
3922 ConstantInt *LHSCst, *RHSCst;
3923 ICmpInst::Predicate LHSCC, RHSCC;
3924 if (match(Op0, m_ICmp(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3925 if (match(RHS, m_ICmp(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3926 if (LHSVal == RHSVal && // Found (X icmp C1) | (X icmp C2)
3927 // icmp [us][gl]e x, cst is folded to icmp [us][gl]t elsewhere.
3928 LHSCC != ICmpInst::ICMP_UGE && LHSCC != ICmpInst::ICMP_ULE &&
3929 RHSCC != ICmpInst::ICMP_UGE && RHSCC != ICmpInst::ICMP_ULE &&
3930 LHSCC != ICmpInst::ICMP_SGE && LHSCC != ICmpInst::ICMP_SLE &&
3931 RHSCC != ICmpInst::ICMP_SGE && RHSCC != ICmpInst::ICMP_SLE &&
3932 // We can't fold (ugt x, C) | (sgt x, C2).
3933 PredicatesFoldable(LHSCC, RHSCC)) {
3934 // Ensure that the larger constant is on the RHS.
3935 ICmpInst *LHS = cast<ICmpInst>(Op0);
3937 if (ICmpInst::isSignedPredicate(LHSCC))
3938 NeedsSwap = LHSCst->getValue().sgt(RHSCst->getValue());
3940 NeedsSwap = LHSCst->getValue().ugt(RHSCst->getValue());
3943 std::swap(LHS, RHS);
3944 std::swap(LHSCst, RHSCst);
3945 std::swap(LHSCC, RHSCC);
3948 // At this point, we know we have have two icmp instructions
3949 // comparing a value against two constants and or'ing the result
3950 // together. Because of the above check, we know that we only have
3951 // ICMP_EQ, ICMP_NE, ICMP_LT, and ICMP_GT here. We also know (from the
3952 // FoldICmpLogical check above), that the two constants are not
3954 assert(LHSCst != RHSCst && "Compares not folded above?");
3957 default: assert(0 && "Unknown integer condition code!");
3958 case ICmpInst::ICMP_EQ:
3960 default: assert(0 && "Unknown integer condition code!");
3961 case ICmpInst::ICMP_EQ:
3962 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3963 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3964 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3965 LHSVal->getName()+".off");
3966 InsertNewInstBefore(Add, I);
3967 AddCST = Subtract(AddOne(RHSCst), LHSCst);
3968 return new ICmpInst(ICmpInst::ICMP_ULT, Add, AddCST);
3970 break; // (X == 13 | X == 15) -> no change
3971 case ICmpInst::ICMP_UGT: // (X == 13 | X u> 14) -> no change
3972 case ICmpInst::ICMP_SGT: // (X == 13 | X s> 14) -> no change
3974 case ICmpInst::ICMP_NE: // (X == 13 | X != 15) -> X != 15
3975 case ICmpInst::ICMP_ULT: // (X == 13 | X u< 15) -> X u< 15
3976 case ICmpInst::ICMP_SLT: // (X == 13 | X s< 15) -> X s< 15
3977 return ReplaceInstUsesWith(I, RHS);
3980 case ICmpInst::ICMP_NE:
3982 default: assert(0 && "Unknown integer condition code!");
3983 case ICmpInst::ICMP_EQ: // (X != 13 | X == 15) -> X != 13
3984 case ICmpInst::ICMP_UGT: // (X != 13 | X u> 15) -> X != 13
3985 case ICmpInst::ICMP_SGT: // (X != 13 | X s> 15) -> X != 13
3986 return ReplaceInstUsesWith(I, LHS);
3987 case ICmpInst::ICMP_NE: // (X != 13 | X != 15) -> true
3988 case ICmpInst::ICMP_ULT: // (X != 13 | X u< 15) -> true
3989 case ICmpInst::ICMP_SLT: // (X != 13 | X s< 15) -> true
3990 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
3993 case ICmpInst::ICMP_ULT:
3995 default: assert(0 && "Unknown integer condition code!");
3996 case ICmpInst::ICMP_EQ: // (X u< 13 | X == 14) -> no change
3998 case ICmpInst::ICMP_UGT: // (X u< 13 | X u> 15) ->(X-13) u> 2
3999 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false,
4001 case ICmpInst::ICMP_SGT: // (X u< 13 | X s> 15) -> no change
4003 case ICmpInst::ICMP_NE: // (X u< 13 | X != 15) -> X != 15
4004 case ICmpInst::ICMP_ULT: // (X u< 13 | X u< 15) -> X u< 15
4005 return ReplaceInstUsesWith(I, RHS);
4006 case ICmpInst::ICMP_SLT: // (X u< 13 | X s< 15) -> no change
4010 case ICmpInst::ICMP_SLT:
4012 default: assert(0 && "Unknown integer condition code!");
4013 case ICmpInst::ICMP_EQ: // (X s< 13 | X == 14) -> no change
4015 case ICmpInst::ICMP_SGT: // (X s< 13 | X s> 15) ->(X-13) s> 2
4016 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), true,
4018 case ICmpInst::ICMP_UGT: // (X s< 13 | X u> 15) -> no change
4020 case ICmpInst::ICMP_NE: // (X s< 13 | X != 15) -> X != 15
4021 case ICmpInst::ICMP_SLT: // (X s< 13 | X s< 15) -> X s< 15
4022 return ReplaceInstUsesWith(I, RHS);
4023 case ICmpInst::ICMP_ULT: // (X s< 13 | X u< 15) -> no change
4027 case ICmpInst::ICMP_UGT:
4029 default: assert(0 && "Unknown integer condition code!");
4030 case ICmpInst::ICMP_EQ: // (X u> 13 | X == 15) -> X u> 13
4031 case ICmpInst::ICMP_UGT: // (X u> 13 | X u> 15) -> X u> 13
4032 return ReplaceInstUsesWith(I, LHS);
4033 case ICmpInst::ICMP_SGT: // (X u> 13 | X s> 15) -> no change
4035 case ICmpInst::ICMP_NE: // (X u> 13 | X != 15) -> true
4036 case ICmpInst::ICMP_ULT: // (X u> 13 | X u< 15) -> true
4037 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4038 case ICmpInst::ICMP_SLT: // (X u> 13 | X s< 15) -> no change
4042 case ICmpInst::ICMP_SGT:
4044 default: assert(0 && "Unknown integer condition code!");
4045 case ICmpInst::ICMP_EQ: // (X s> 13 | X == 15) -> X > 13
4046 case ICmpInst::ICMP_SGT: // (X s> 13 | X s> 15) -> X > 13
4047 return ReplaceInstUsesWith(I, LHS);
4048 case ICmpInst::ICMP_UGT: // (X s> 13 | X u> 15) -> no change
4050 case ICmpInst::ICMP_NE: // (X s> 13 | X != 15) -> true
4051 case ICmpInst::ICMP_SLT: // (X s> 13 | X s< 15) -> true
4052 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4053 case ICmpInst::ICMP_ULT: // (X s> 13 | X u< 15) -> no change
4061 // fold (or (cast A), (cast B)) -> (cast (or A, B))
4062 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4063 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4064 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
4065 const Type *SrcTy = Op0C->getOperand(0)->getType();
4066 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4067 // Only do this if the casts both really cause code to be generated.
4068 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4070 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4072 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
4073 Op1C->getOperand(0),
4075 InsertNewInstBefore(NewOp, I);
4076 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4081 return Changed ? &I : 0;
4084 // XorSelf - Implements: X ^ X --> 0
4087 XorSelf(Value *rhs) : RHS(rhs) {}
4088 bool shouldApply(Value *LHS) const { return LHS == RHS; }
4089 Instruction *apply(BinaryOperator &Xor) const {
4095 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
4096 bool Changed = SimplifyCommutative(I);
4097 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4099 if (isa<UndefValue>(Op1))
4100 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
4102 // xor X, X = 0, even if X is nested in a sequence of Xor's.
4103 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
4104 assert(Result == &I && "AssociativeOpt didn't work?"); Result=Result;
4105 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
4108 // See if we can simplify any instructions used by the instruction whose sole
4109 // purpose is to compute bits we don't care about.
4110 if (!isa<VectorType>(I.getType())) {
4111 uint32_t BitWidth = cast<IntegerType>(I.getType())->getBitWidth();
4112 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4113 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(BitWidth),
4114 KnownZero, KnownOne))
4116 } else if (isa<ConstantAggregateZero>(Op1)) {
4117 return ReplaceInstUsesWith(I, Op0); // X ^ <0,0> -> X
4120 // Is this a ~ operation?
4121 if (Value *NotOp = dyn_castNotVal(&I)) {
4122 // ~(~X & Y) --> (X | ~Y) - De Morgan's Law
4123 // ~(~X | Y) === (X & ~Y) - De Morgan's Law
4124 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(NotOp)) {
4125 if (Op0I->getOpcode() == Instruction::And ||
4126 Op0I->getOpcode() == Instruction::Or) {
4127 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
4128 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
4130 BinaryOperator::createNot(Op0I->getOperand(1),
4131 Op0I->getOperand(1)->getName()+".not");
4132 InsertNewInstBefore(NotY, I);
4133 if (Op0I->getOpcode() == Instruction::And)
4134 return BinaryOperator::createOr(Op0NotVal, NotY);
4136 return BinaryOperator::createAnd(Op0NotVal, NotY);
4143 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
4144 // xor (cmp A, B), true = not (cmp A, B) = !cmp A, B
4145 if (RHS == ConstantInt::getTrue() && Op0->hasOneUse()) {
4146 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Op0))
4147 return new ICmpInst(ICI->getInversePredicate(),
4148 ICI->getOperand(0), ICI->getOperand(1));
4150 if (FCmpInst *FCI = dyn_cast<FCmpInst>(Op0))
4151 return new FCmpInst(FCI->getInversePredicate(),
4152 FCI->getOperand(0), FCI->getOperand(1));
4155 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
4156 // ~(c-X) == X-c-1 == X+(-c-1)
4157 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
4158 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
4159 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
4160 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
4161 ConstantInt::get(I.getType(), 1));
4162 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
4165 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
4166 if (Op0I->getOpcode() == Instruction::Add) {
4167 // ~(X-c) --> (-c-1)-X
4168 if (RHS->isAllOnesValue()) {
4169 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
4170 return BinaryOperator::createSub(
4171 ConstantExpr::getSub(NegOp0CI,
4172 ConstantInt::get(I.getType(), 1)),
4173 Op0I->getOperand(0));
4174 } else if (RHS->getValue().isSignBit()) {
4175 // (X + C) ^ signbit -> (X + C + signbit)
4176 Constant *C = ConstantInt::get(RHS->getValue() + Op0CI->getValue());
4177 return BinaryOperator::createAdd(Op0I->getOperand(0), C);
4180 } else if (Op0I->getOpcode() == Instruction::Or) {
4181 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
4182 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getValue())) {
4183 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
4184 // Anything in both C1 and C2 is known to be zero, remove it from
4186 Constant *CommonBits = And(Op0CI, RHS);
4187 NewRHS = ConstantExpr::getAnd(NewRHS,
4188 ConstantExpr::getNot(CommonBits));
4189 AddToWorkList(Op0I);
4190 I.setOperand(0, Op0I->getOperand(0));
4191 I.setOperand(1, NewRHS);
4197 // Try to fold constant and into select arguments.
4198 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
4199 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
4201 if (isa<PHINode>(Op0))
4202 if (Instruction *NV = FoldOpIntoPhi(I))
4206 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
4208 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4210 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
4212 return ReplaceInstUsesWith(I, Constant::getAllOnesValue(I.getType()));
4215 BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1);
4218 if (match(Op1I, m_Or(m_Value(A), m_Value(B)))) {
4219 if (A == Op0) { // B^(B|A) == (A|B)^B
4220 Op1I->swapOperands();
4222 std::swap(Op0, Op1);
4223 } else if (B == Op0) { // B^(A|B) == (A|B)^B
4224 I.swapOperands(); // Simplified below.
4225 std::swap(Op0, Op1);
4227 } else if (match(Op1I, m_Xor(m_Value(A), m_Value(B)))) {
4228 if (Op0 == A) // A^(A^B) == B
4229 return ReplaceInstUsesWith(I, B);
4230 else if (Op0 == B) // A^(B^A) == B
4231 return ReplaceInstUsesWith(I, A);
4232 } else if (match(Op1I, m_And(m_Value(A), m_Value(B))) && Op1I->hasOneUse()){
4233 if (A == Op0) { // A^(A&B) -> A^(B&A)
4234 Op1I->swapOperands();
4237 if (B == Op0) { // A^(B&A) -> (B&A)^A
4238 I.swapOperands(); // Simplified below.
4239 std::swap(Op0, Op1);
4244 BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0);
4247 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) && Op0I->hasOneUse()) {
4248 if (A == Op1) // (B|A)^B == (A|B)^B
4250 if (B == Op1) { // (A|B)^B == A & ~B
4252 InsertNewInstBefore(BinaryOperator::createNot(Op1, "tmp"), I);
4253 return BinaryOperator::createAnd(A, NotB);
4255 } else if (match(Op0I, m_Xor(m_Value(A), m_Value(B)))) {
4256 if (Op1 == A) // (A^B)^A == B
4257 return ReplaceInstUsesWith(I, B);
4258 else if (Op1 == B) // (B^A)^A == B
4259 return ReplaceInstUsesWith(I, A);
4260 } else if (match(Op0I, m_And(m_Value(A), m_Value(B))) && Op0I->hasOneUse()){
4261 if (A == Op1) // (A&B)^A -> (B&A)^A
4263 if (B == Op1 && // (B&A)^A == ~B & A
4264 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
4266 InsertNewInstBefore(BinaryOperator::createNot(A, "tmp"), I);
4267 return BinaryOperator::createAnd(N, Op1);
4272 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
4273 if (Op0I && Op1I && Op0I->isShift() &&
4274 Op0I->getOpcode() == Op1I->getOpcode() &&
4275 Op0I->getOperand(1) == Op1I->getOperand(1) &&
4276 (Op1I->hasOneUse() || Op1I->hasOneUse())) {
4277 Instruction *NewOp =
4278 InsertNewInstBefore(BinaryOperator::createXor(Op0I->getOperand(0),
4279 Op1I->getOperand(0),
4280 Op0I->getName()), I);
4281 return BinaryOperator::create(Op1I->getOpcode(), NewOp,
4282 Op1I->getOperand(1));
4286 Value *A, *B, *C, *D;
4287 // (A & B)^(A | B) -> A ^ B
4288 if (match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4289 match(Op1I, m_Or(m_Value(C), m_Value(D)))) {
4290 if ((A == C && B == D) || (A == D && B == C))
4291 return BinaryOperator::createXor(A, B);
4293 // (A | B)^(A & B) -> A ^ B
4294 if (match(Op0I, m_Or(m_Value(A), m_Value(B))) &&
4295 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4296 if ((A == C && B == D) || (A == D && B == C))
4297 return BinaryOperator::createXor(A, B);
4301 if ((Op0I->hasOneUse() || Op1I->hasOneUse()) &&
4302 match(Op0I, m_And(m_Value(A), m_Value(B))) &&
4303 match(Op1I, m_And(m_Value(C), m_Value(D)))) {
4304 // (X & Y)^(X & Y) -> (Y^Z) & X
4305 Value *X = 0, *Y = 0, *Z = 0;
4307 X = A, Y = B, Z = D;
4309 X = A, Y = B, Z = C;
4311 X = B, Y = A, Z = D;
4313 X = B, Y = A, Z = C;
4316 Instruction *NewOp =
4317 InsertNewInstBefore(BinaryOperator::createXor(Y, Z, Op0->getName()), I);
4318 return BinaryOperator::createAnd(NewOp, X);
4323 // (icmp1 A, B) ^ (icmp2 A, B) --> (icmp3 A, B)
4324 if (ICmpInst *RHS = dyn_cast<ICmpInst>(I.getOperand(1)))
4325 if (Instruction *R = AssociativeOpt(I, FoldICmpLogical(*this, RHS)))
4328 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
4329 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
4330 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
4331 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
4332 const Type *SrcTy = Op0C->getOperand(0)->getType();
4333 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isInteger() &&
4334 // Only do this if the casts both really cause code to be generated.
4335 ValueRequiresCast(Op0C->getOpcode(), Op0C->getOperand(0),
4337 ValueRequiresCast(Op1C->getOpcode(), Op1C->getOperand(0),
4339 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
4340 Op1C->getOperand(0),
4342 InsertNewInstBefore(NewOp, I);
4343 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
4347 return Changed ? &I : 0;
4350 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
4351 /// overflowed for this type.
4352 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
4353 ConstantInt *In2, bool IsSigned = false) {
4354 Result = cast<ConstantInt>(Add(In1, In2));
4357 if (In2->getValue().isNegative())
4358 return Result->getValue().sgt(In1->getValue());
4360 return Result->getValue().slt(In1->getValue());
4362 return Result->getValue().ult(In1->getValue());
4365 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
4366 /// code necessary to compute the offset from the base pointer (without adding
4367 /// in the base pointer). Return the result as a signed integer of intptr size.
4368 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
4369 TargetData &TD = IC.getTargetData();
4370 gep_type_iterator GTI = gep_type_begin(GEP);
4371 const Type *IntPtrTy = TD.getIntPtrType();
4372 Value *Result = Constant::getNullValue(IntPtrTy);
4374 // Build a mask for high order bits.
4375 unsigned IntPtrWidth = TD.getPointerSize()*8;
4376 uint64_t PtrSizeMask = ~0ULL >> (64-IntPtrWidth);
4378 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
4379 Value *Op = GEP->getOperand(i);
4380 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
4381 if (ConstantInt *OpC = dyn_cast<ConstantInt>(Op)) {
4382 if (OpC->isZero()) continue;
4384 // Handle a struct index, which adds its field offset to the pointer.
4385 if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
4386 Size = TD.getStructLayout(STy)->getElementOffset(OpC->getZExtValue());
4388 if (ConstantInt *RC = dyn_cast<ConstantInt>(Result))
4389 Result = ConstantInt::get(RC->getValue() + APInt(IntPtrWidth, Size));
4391 Result = IC.InsertNewInstBefore(
4392 BinaryOperator::createAdd(Result,
4393 ConstantInt::get(IntPtrTy, Size),
4394 GEP->getName()+".offs"), I);
4398 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4399 Constant *OC = ConstantExpr::getIntegerCast(OpC, IntPtrTy, true /*SExt*/);
4400 Scale = ConstantExpr::getMul(OC, Scale);
4401 if (Constant *RC = dyn_cast<Constant>(Result))
4402 Result = ConstantExpr::getAdd(RC, Scale);
4404 // Emit an add instruction.
4405 Result = IC.InsertNewInstBefore(
4406 BinaryOperator::createAdd(Result, Scale,
4407 GEP->getName()+".offs"), I);
4411 // Convert to correct type.
4412 if (Op->getType() != IntPtrTy) {
4413 if (Constant *OpC = dyn_cast<Constant>(Op))
4414 Op = ConstantExpr::getSExt(OpC, IntPtrTy);
4416 Op = IC.InsertNewInstBefore(new SExtInst(Op, IntPtrTy,
4417 Op->getName()+".c"), I);
4420 Constant *Scale = ConstantInt::get(IntPtrTy, Size);
4421 if (Constant *OpC = dyn_cast<Constant>(Op))
4422 Op = ConstantExpr::getMul(OpC, Scale);
4423 else // We'll let instcombine(mul) convert this to a shl if possible.
4424 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
4425 GEP->getName()+".idx"), I);
4428 // Emit an add instruction.
4429 if (isa<Constant>(Op) && isa<Constant>(Result))
4430 Result = ConstantExpr::getAdd(cast<Constant>(Op),
4431 cast<Constant>(Result));
4433 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
4434 GEP->getName()+".offs"), I);
4439 /// FoldGEPICmp - Fold comparisons between a GEP instruction and something
4440 /// else. At this point we know that the GEP is on the LHS of the comparison.
4441 Instruction *InstCombiner::FoldGEPICmp(User *GEPLHS, Value *RHS,
4442 ICmpInst::Predicate Cond,
4444 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
4446 if (CastInst *CI = dyn_cast<CastInst>(RHS))
4447 if (isa<PointerType>(CI->getOperand(0)->getType()))
4448 RHS = CI->getOperand(0);
4450 Value *PtrBase = GEPLHS->getOperand(0);
4451 if (PtrBase == RHS) {
4452 // As an optimization, we don't actually have to compute the actual value of
4453 // OFFSET if this is a icmp_eq or icmp_ne comparison, just return whether
4454 // each index is zero or not.
4455 if (Cond == ICmpInst::ICMP_EQ || Cond == ICmpInst::ICMP_NE) {
4456 Instruction *InVal = 0;
4457 gep_type_iterator GTI = gep_type_begin(GEPLHS);
4458 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
4460 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
4461 if (isa<UndefValue>(C)) // undef index -> undef.
4462 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
4463 if (C->isNullValue())
4465 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
4466 EmitIt = false; // This is indexing into a zero sized array?
4467 } else if (isa<ConstantInt>(C))
4468 return ReplaceInstUsesWith(I, // No comparison is needed here.
4469 ConstantInt::get(Type::Int1Ty,
4470 Cond == ICmpInst::ICMP_NE));
4475 new ICmpInst(Cond, GEPLHS->getOperand(i),
4476 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
4480 InVal = InsertNewInstBefore(InVal, I);
4481 InsertNewInstBefore(Comp, I);
4482 if (Cond == ICmpInst::ICMP_NE) // True if any are unequal
4483 InVal = BinaryOperator::createOr(InVal, Comp);
4484 else // True if all are equal
4485 InVal = BinaryOperator::createAnd(InVal, Comp);
4493 // No comparison is needed here, all indexes = 0
4494 ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4495 Cond == ICmpInst::ICMP_EQ));
4498 // Only lower this if the icmp is the only user of the GEP or if we expect
4499 // the result to fold to a constant!
4500 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4501 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4502 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4503 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), Offset,
4504 Constant::getNullValue(Offset->getType()));
4506 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4507 // If the base pointers are different, but the indices are the same, just
4508 // compare the base pointer.
4509 if (PtrBase != GEPRHS->getOperand(0)) {
4510 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4511 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4512 GEPRHS->getOperand(0)->getType();
4514 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4515 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4516 IndicesTheSame = false;
4520 // If all indices are the same, just compare the base pointers.
4522 return new ICmpInst(ICmpInst::getSignedPredicate(Cond),
4523 GEPLHS->getOperand(0), GEPRHS->getOperand(0));
4525 // Otherwise, the base pointers are different and the indices are
4526 // different, bail out.
4530 // If one of the GEPs has all zero indices, recurse.
4531 bool AllZeros = true;
4532 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4533 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4534 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4539 return FoldGEPICmp(GEPRHS, GEPLHS->getOperand(0),
4540 ICmpInst::getSwappedPredicate(Cond), I);
4542 // If the other GEP has all zero indices, recurse.
4544 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4545 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4546 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4551 return FoldGEPICmp(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4553 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4554 // If the GEPs only differ by one index, compare it.
4555 unsigned NumDifferences = 0; // Keep track of # differences.
4556 unsigned DiffOperand = 0; // The operand that differs.
4557 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4558 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4559 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4560 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4561 // Irreconcilable differences.
4565 if (NumDifferences++) break;
4570 if (NumDifferences == 0) // SAME GEP?
4571 return ReplaceInstUsesWith(I, // No comparison is needed here.
4572 ConstantInt::get(Type::Int1Ty,
4573 isTrueWhenEqual(Cond)));
4575 else if (NumDifferences == 1) {
4576 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4577 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4578 // Make sure we do a signed comparison here.
4579 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), LHSV, RHSV);
4583 // Only lower this if the icmp is the only user of the GEP or if we expect
4584 // the result to fold to a constant!
4585 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4586 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4587 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4588 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4589 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4590 return new ICmpInst(ICmpInst::getSignedPredicate(Cond), L, R);
4596 Instruction *InstCombiner::visitFCmpInst(FCmpInst &I) {
4597 bool Changed = SimplifyCompare(I);
4598 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4600 // Fold trivial predicates.
4601 if (I.getPredicate() == FCmpInst::FCMP_FALSE)
4602 return ReplaceInstUsesWith(I, Constant::getNullValue(Type::Int1Ty));
4603 if (I.getPredicate() == FCmpInst::FCMP_TRUE)
4604 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4606 // Simplify 'fcmp pred X, X'
4608 switch (I.getPredicate()) {
4609 default: assert(0 && "Unknown predicate!");
4610 case FCmpInst::FCMP_UEQ: // True if unordered or equal
4611 case FCmpInst::FCMP_UGE: // True if unordered, greater than, or equal
4612 case FCmpInst::FCMP_ULE: // True if unordered, less than, or equal
4613 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 1));
4614 case FCmpInst::FCMP_OGT: // True if ordered and greater than
4615 case FCmpInst::FCMP_OLT: // True if ordered and less than
4616 case FCmpInst::FCMP_ONE: // True if ordered and operands are unequal
4617 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty, 0));
4619 case FCmpInst::FCMP_UNO: // True if unordered: isnan(X) | isnan(Y)
4620 case FCmpInst::FCMP_ULT: // True if unordered or less than
4621 case FCmpInst::FCMP_UGT: // True if unordered or greater than
4622 case FCmpInst::FCMP_UNE: // True if unordered or not equal
4623 // Canonicalize these to be 'fcmp uno %X, 0.0'.
4624 I.setPredicate(FCmpInst::FCMP_UNO);
4625 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4628 case FCmpInst::FCMP_ORD: // True if ordered (no nans)
4629 case FCmpInst::FCMP_OEQ: // True if ordered and equal
4630 case FCmpInst::FCMP_OGE: // True if ordered and greater than or equal
4631 case FCmpInst::FCMP_OLE: // True if ordered and less than or equal
4632 // Canonicalize these to be 'fcmp ord %X, 0.0'.
4633 I.setPredicate(FCmpInst::FCMP_ORD);
4634 I.setOperand(1, Constant::getNullValue(Op0->getType()));
4639 if (isa<UndefValue>(Op1)) // fcmp pred X, undef -> undef
4640 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4642 // Handle fcmp with constant RHS
4643 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4644 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4645 switch (LHSI->getOpcode()) {
4646 case Instruction::PHI:
4647 if (Instruction *NV = FoldOpIntoPhi(I))
4650 case Instruction::Select:
4651 // If either operand of the select is a constant, we can fold the
4652 // comparison into the select arms, which will cause one to be
4653 // constant folded and the select turned into a bitwise or.
4654 Value *Op1 = 0, *Op2 = 0;
4655 if (LHSI->hasOneUse()) {
4656 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4657 // Fold the known value into the constant operand.
4658 Op1 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4659 // Insert a new FCmp of the other select operand.
4660 Op2 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4661 LHSI->getOperand(2), RHSC,
4663 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4664 // Fold the known value into the constant operand.
4665 Op2 = ConstantExpr::getCompare(I.getPredicate(), C, RHSC);
4666 // Insert a new FCmp of the other select operand.
4667 Op1 = InsertNewInstBefore(new FCmpInst(I.getPredicate(),
4668 LHSI->getOperand(1), RHSC,
4674 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4679 return Changed ? &I : 0;
4682 Instruction *InstCombiner::visitICmpInst(ICmpInst &I) {
4683 bool Changed = SimplifyCompare(I);
4684 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4685 const Type *Ty = Op0->getType();
4689 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4690 isTrueWhenEqual(I)));
4692 if (isa<UndefValue>(Op1)) // X icmp undef -> undef
4693 return ReplaceInstUsesWith(I, UndefValue::get(Type::Int1Ty));
4695 // icmp <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4696 // addresses never equal each other! We already know that Op0 != Op1.
4697 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4698 isa<ConstantPointerNull>(Op0)) &&
4699 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4700 isa<ConstantPointerNull>(Op1)))
4701 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4702 !isTrueWhenEqual(I)));
4704 // icmp's with boolean values can always be turned into bitwise operations
4705 if (Ty == Type::Int1Ty) {
4706 switch (I.getPredicate()) {
4707 default: assert(0 && "Invalid icmp instruction!");
4708 case ICmpInst::ICMP_EQ: { // icmp eq bool %A, %B -> ~(A^B)
4709 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4710 InsertNewInstBefore(Xor, I);
4711 return BinaryOperator::createNot(Xor);
4713 case ICmpInst::ICMP_NE: // icmp eq bool %A, %B -> A^B
4714 return BinaryOperator::createXor(Op0, Op1);
4716 case ICmpInst::ICMP_UGT:
4717 case ICmpInst::ICMP_SGT:
4718 std::swap(Op0, Op1); // Change icmp gt -> icmp lt
4720 case ICmpInst::ICMP_ULT:
4721 case ICmpInst::ICMP_SLT: { // icmp lt bool A, B -> ~X & Y
4722 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4723 InsertNewInstBefore(Not, I);
4724 return BinaryOperator::createAnd(Not, Op1);
4726 case ICmpInst::ICMP_UGE:
4727 case ICmpInst::ICMP_SGE:
4728 std::swap(Op0, Op1); // Change icmp ge -> icmp le
4730 case ICmpInst::ICMP_ULE:
4731 case ICmpInst::ICMP_SLE: { // icmp le bool %A, %B -> ~A | B
4732 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4733 InsertNewInstBefore(Not, I);
4734 return BinaryOperator::createOr(Not, Op1);
4739 // See if we are doing a comparison between a constant and an instruction that
4740 // can be folded into the comparison.
4741 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4742 switch (I.getPredicate()) {
4744 case ICmpInst::ICMP_ULT: // A <u MIN -> FALSE
4745 if (CI->isMinValue(false))
4746 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4747 if (CI->isMaxValue(false)) // A <u MAX -> A != MAX
4748 return new ICmpInst(ICmpInst::ICMP_NE, Op0,Op1);
4749 if (isMinValuePlusOne(CI,false)) // A <u MIN+1 -> A == MIN
4750 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4751 // (x <u 2147483648) -> (x >s -1) -> true if sign bit clear
4752 if (CI->isMinValue(true))
4753 return new ICmpInst(ICmpInst::ICMP_SGT, Op0,
4754 ConstantInt::getAllOnesValue(Op0->getType()));
4758 case ICmpInst::ICMP_SLT:
4759 if (CI->isMinValue(true)) // A <s MIN -> FALSE
4760 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4761 if (CI->isMaxValue(true)) // A <s MAX -> A != MAX
4762 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4763 if (isMinValuePlusOne(CI,true)) // A <s MIN+1 -> A == MIN
4764 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, SubOne(CI));
4767 case ICmpInst::ICMP_UGT:
4768 if (CI->isMaxValue(false)) // A >u MAX -> FALSE
4769 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4770 if (CI->isMinValue(false)) // A >u MIN -> A != MIN
4771 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4772 if (isMaxValueMinusOne(CI, false)) // A >u MAX-1 -> A == MAX
4773 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4775 // (x >u 2147483647) -> (x <s 0) -> true if sign bit set
4776 if (CI->isMaxValue(true))
4777 return new ICmpInst(ICmpInst::ICMP_SLT, Op0,
4778 ConstantInt::getNullValue(Op0->getType()));
4781 case ICmpInst::ICMP_SGT:
4782 if (CI->isMaxValue(true)) // A >s MAX -> FALSE
4783 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4784 if (CI->isMinValue(true)) // A >s MIN -> A != MIN
4785 return new ICmpInst(ICmpInst::ICMP_NE, Op0, Op1);
4786 if (isMaxValueMinusOne(CI, true)) // A >s MAX-1 -> A == MAX
4787 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, AddOne(CI));
4790 case ICmpInst::ICMP_ULE:
4791 if (CI->isMaxValue(false)) // A <=u MAX -> TRUE
4792 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4793 if (CI->isMinValue(false)) // A <=u MIN -> A == MIN
4794 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4795 if (isMaxValueMinusOne(CI,false)) // A <=u MAX-1 -> A != MAX
4796 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4799 case ICmpInst::ICMP_SLE:
4800 if (CI->isMaxValue(true)) // A <=s MAX -> TRUE
4801 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4802 if (CI->isMinValue(true)) // A <=s MIN -> A == MIN
4803 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4804 if (isMaxValueMinusOne(CI,true)) // A <=s MAX-1 -> A != MAX
4805 return new ICmpInst(ICmpInst::ICMP_NE, Op0, AddOne(CI));
4808 case ICmpInst::ICMP_UGE:
4809 if (CI->isMinValue(false)) // A >=u MIN -> TRUE
4810 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4811 if (CI->isMaxValue(false)) // A >=u MAX -> A == MAX
4812 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4813 if (isMinValuePlusOne(CI,false)) // A >=u MIN-1 -> A != MIN
4814 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4817 case ICmpInst::ICMP_SGE:
4818 if (CI->isMinValue(true)) // A >=s MIN -> TRUE
4819 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4820 if (CI->isMaxValue(true)) // A >=s MAX -> A == MAX
4821 return new ICmpInst(ICmpInst::ICMP_EQ, Op0, Op1);
4822 if (isMinValuePlusOne(CI,true)) // A >=s MIN-1 -> A != MIN
4823 return new ICmpInst(ICmpInst::ICMP_NE, Op0, SubOne(CI));
4827 // If we still have a icmp le or icmp ge instruction, turn it into the
4828 // appropriate icmp lt or icmp gt instruction. Since the border cases have
4829 // already been handled above, this requires little checking.
4831 switch (I.getPredicate()) {
4833 case ICmpInst::ICMP_ULE:
4834 return new ICmpInst(ICmpInst::ICMP_ULT, Op0, AddOne(CI));
4835 case ICmpInst::ICMP_SLE:
4836 return new ICmpInst(ICmpInst::ICMP_SLT, Op0, AddOne(CI));
4837 case ICmpInst::ICMP_UGE:
4838 return new ICmpInst( ICmpInst::ICMP_UGT, Op0, SubOne(CI));
4839 case ICmpInst::ICMP_SGE:
4840 return new ICmpInst(ICmpInst::ICMP_SGT, Op0, SubOne(CI));
4843 // See if we can fold the comparison based on bits known to be zero or one
4844 // in the input. If this comparison is a normal comparison, it demands all
4845 // bits, if it is a sign bit comparison, it only demands the sign bit.
4848 bool isSignBit = isSignBitCheck(I.getPredicate(), CI, UnusedBit);
4850 uint32_t BitWidth = cast<IntegerType>(Ty)->getBitWidth();
4851 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
4852 if (SimplifyDemandedBits(Op0,
4853 isSignBit ? APInt::getSignBit(BitWidth)
4854 : APInt::getAllOnesValue(BitWidth),
4855 KnownZero, KnownOne, 0))
4858 // Given the known and unknown bits, compute a range that the LHS could be
4860 if ((KnownOne | KnownZero) != 0) {
4861 // Compute the Min, Max and RHS values based on the known bits. For the
4862 // EQ and NE we use unsigned values.
4863 APInt Min(BitWidth, 0), Max(BitWidth, 0);
4864 const APInt& RHSVal = CI->getValue();
4865 if (ICmpInst::isSignedPredicate(I.getPredicate())) {
4866 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4869 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne, Min,
4872 switch (I.getPredicate()) { // LE/GE have been folded already.
4873 default: assert(0 && "Unknown icmp opcode!");
4874 case ICmpInst::ICMP_EQ:
4875 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4876 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4878 case ICmpInst::ICMP_NE:
4879 if (Max.ult(RHSVal) || Min.ugt(RHSVal))
4880 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4882 case ICmpInst::ICMP_ULT:
4883 if (Max.ult(RHSVal))
4884 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4885 if (Min.uge(RHSVal))
4886 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4888 case ICmpInst::ICMP_UGT:
4889 if (Min.ugt(RHSVal))
4890 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4891 if (Max.ule(RHSVal))
4892 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4894 case ICmpInst::ICMP_SLT:
4895 if (Max.slt(RHSVal))
4896 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4897 if (Min.sgt(RHSVal))
4898 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4900 case ICmpInst::ICMP_SGT:
4901 if (Min.sgt(RHSVal))
4902 return ReplaceInstUsesWith(I, ConstantInt::getTrue());
4903 if (Max.sle(RHSVal))
4904 return ReplaceInstUsesWith(I, ConstantInt::getFalse());
4909 // Since the RHS is a ConstantInt (CI), if the left hand side is an
4910 // instruction, see if that instruction also has constants so that the
4911 // instruction can be folded into the icmp
4912 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4913 if (Instruction *Res = visitICmpInstWithInstAndIntCst(I, LHSI, CI))
4917 // Handle icmp with constant (but not simple integer constant) RHS
4918 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4919 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4920 switch (LHSI->getOpcode()) {
4921 case Instruction::GetElementPtr:
4922 if (RHSC->isNullValue()) {
4923 // icmp pred GEP (P, int 0, int 0, int 0), null -> icmp pred P, null
4924 bool isAllZeros = true;
4925 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4926 if (!isa<Constant>(LHSI->getOperand(i)) ||
4927 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4932 return new ICmpInst(I.getPredicate(), LHSI->getOperand(0),
4933 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4937 case Instruction::PHI:
4938 if (Instruction *NV = FoldOpIntoPhi(I))
4941 case Instruction::Select: {
4942 // If either operand of the select is a constant, we can fold the
4943 // comparison into the select arms, which will cause one to be
4944 // constant folded and the select turned into a bitwise or.
4945 Value *Op1 = 0, *Op2 = 0;
4946 if (LHSI->hasOneUse()) {
4947 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4948 // Fold the known value into the constant operand.
4949 Op1 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4950 // Insert a new ICmp of the other select operand.
4951 Op2 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4952 LHSI->getOperand(2), RHSC,
4954 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4955 // Fold the known value into the constant operand.
4956 Op2 = ConstantExpr::getICmp(I.getPredicate(), C, RHSC);
4957 // Insert a new ICmp of the other select operand.
4958 Op1 = InsertNewInstBefore(new ICmpInst(I.getPredicate(),
4959 LHSI->getOperand(1), RHSC,
4965 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4968 case Instruction::Malloc:
4969 // If we have (malloc != null), and if the malloc has a single use, we
4970 // can assume it is successful and remove the malloc.
4971 if (LHSI->hasOneUse() && isa<ConstantPointerNull>(RHSC)) {
4972 AddToWorkList(LHSI);
4973 return ReplaceInstUsesWith(I, ConstantInt::get(Type::Int1Ty,
4974 !isTrueWhenEqual(I)));
4980 // If we can optimize a 'icmp GEP, P' or 'icmp P, GEP', do so now.
4981 if (User *GEP = dyn_castGetElementPtr(Op0))
4982 if (Instruction *NI = FoldGEPICmp(GEP, Op1, I.getPredicate(), I))
4984 if (User *GEP = dyn_castGetElementPtr(Op1))
4985 if (Instruction *NI = FoldGEPICmp(GEP, Op0,
4986 ICmpInst::getSwappedPredicate(I.getPredicate()), I))
4989 // Test to see if the operands of the icmp are casted versions of other
4990 // values. If the ptr->ptr cast can be stripped off both arguments, we do so
4992 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
4993 if (isa<PointerType>(Op0->getType()) &&
4994 (isa<Constant>(Op1) || isa<BitCastInst>(Op1))) {
4995 // We keep moving the cast from the left operand over to the right
4996 // operand, where it can often be eliminated completely.
4997 Op0 = CI->getOperand(0);
4999 // If operand #1 is a bitcast instruction, it must also be a ptr->ptr cast
5000 // so eliminate it as well.
5001 if (BitCastInst *CI2 = dyn_cast<BitCastInst>(Op1))
5002 Op1 = CI2->getOperand(0);
5004 // If Op1 is a constant, we can fold the cast into the constant.
5005 if (Op0->getType() != Op1->getType())
5006 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
5007 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
5009 // Otherwise, cast the RHS right before the icmp
5010 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
5012 return new ICmpInst(I.getPredicate(), Op0, Op1);
5016 if (isa<CastInst>(Op0)) {
5017 // Handle the special case of: icmp (cast bool to X), <cst>
5018 // This comes up when you have code like
5021 // For generality, we handle any zero-extension of any operand comparison
5022 // with a constant or another cast from the same type.
5023 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
5024 if (Instruction *R = visitICmpInstWithCastAndCast(I))
5028 if (I.isEquality()) {
5029 Value *A, *B, *C, *D;
5030 if (match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
5031 if (A == Op1 || B == Op1) { // (A^B) == A -> B == 0
5032 Value *OtherVal = A == Op1 ? B : A;
5033 return new ICmpInst(I.getPredicate(), OtherVal,
5034 Constant::getNullValue(A->getType()));
5037 if (match(Op1, m_Xor(m_Value(C), m_Value(D)))) {
5038 // A^c1 == C^c2 --> A == C^(c1^c2)
5039 if (ConstantInt *C1 = dyn_cast<ConstantInt>(B))
5040 if (ConstantInt *C2 = dyn_cast<ConstantInt>(D))
5041 if (Op1->hasOneUse()) {
5042 Constant *NC = ConstantInt::get(C1->getValue() ^ C2->getValue());
5043 Instruction *Xor = BinaryOperator::createXor(C, NC, "tmp");
5044 return new ICmpInst(I.getPredicate(), A,
5045 InsertNewInstBefore(Xor, I));
5048 // A^B == A^D -> B == D
5049 if (A == C) return new ICmpInst(I.getPredicate(), B, D);
5050 if (A == D) return new ICmpInst(I.getPredicate(), B, C);
5051 if (B == C) return new ICmpInst(I.getPredicate(), A, D);
5052 if (B == D) return new ICmpInst(I.getPredicate(), A, C);
5056 if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
5057 (A == Op0 || B == Op0)) {
5058 // A == (A^B) -> B == 0
5059 Value *OtherVal = A == Op0 ? B : A;
5060 return new ICmpInst(I.getPredicate(), OtherVal,
5061 Constant::getNullValue(A->getType()));
5063 if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
5064 // (A-B) == A -> B == 0
5065 return new ICmpInst(I.getPredicate(), B,
5066 Constant::getNullValue(B->getType()));
5068 if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
5069 // A == (A-B) -> B == 0
5070 return new ICmpInst(I.getPredicate(), B,
5071 Constant::getNullValue(B->getType()));
5074 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
5075 if (Op0->hasOneUse() && Op1->hasOneUse() &&
5076 match(Op0, m_And(m_Value(A), m_Value(B))) &&
5077 match(Op1, m_And(m_Value(C), m_Value(D)))) {
5078 Value *X = 0, *Y = 0, *Z = 0;
5081 X = B; Y = D; Z = A;
5082 } else if (A == D) {
5083 X = B; Y = C; Z = A;
5084 } else if (B == C) {
5085 X = A; Y = D; Z = B;
5086 } else if (B == D) {
5087 X = A; Y = C; Z = B;
5090 if (X) { // Build (X^Y) & Z
5091 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
5092 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
5093 I.setOperand(0, Op1);
5094 I.setOperand(1, Constant::getNullValue(Op1->getType()));
5099 return Changed ? &I : 0;
5103 /// FoldICmpDivCst - Fold "icmp pred, ([su]div X, DivRHS), CmpRHS" where DivRHS
5104 /// and CmpRHS are both known to be integer constants.
5105 Instruction *InstCombiner::FoldICmpDivCst(ICmpInst &ICI, BinaryOperator *DivI,
5106 ConstantInt *DivRHS) {
5107 ConstantInt *CmpRHS = cast<ConstantInt>(ICI.getOperand(1));
5108 const APInt &CmpRHSV = CmpRHS->getValue();
5110 // FIXME: If the operand types don't match the type of the divide
5111 // then don't attempt this transform. The code below doesn't have the
5112 // logic to deal with a signed divide and an unsigned compare (and
5113 // vice versa). This is because (x /s C1) <s C2 produces different
5114 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
5115 // (x /u C1) <u C2. Simply casting the operands and result won't
5116 // work. :( The if statement below tests that condition and bails
5118 bool DivIsSigned = DivI->getOpcode() == Instruction::SDiv;
5119 if (!ICI.isEquality() && DivIsSigned != ICI.isSignedPredicate())
5121 if (DivRHS->isZero())
5122 return 0; // The ProdOV computation fails on divide by zero.
5124 // Compute Prod = CI * DivRHS. We are essentially solving an equation
5125 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
5126 // C2 (CI). By solving for X we can turn this into a range check
5127 // instead of computing a divide.
5128 ConstantInt *Prod = Multiply(CmpRHS, DivRHS);
5130 // Determine if the product overflows by seeing if the product is
5131 // not equal to the divide. Make sure we do the same kind of divide
5132 // as in the LHS instruction that we're folding.
5133 bool ProdOV = (DivIsSigned ? ConstantExpr::getSDiv(Prod, DivRHS) :
5134 ConstantExpr::getUDiv(Prod, DivRHS)) != CmpRHS;
5136 // Get the ICmp opcode
5137 ICmpInst::Predicate Pred = ICI.getPredicate();
5139 // Figure out the interval that is being checked. For example, a comparison
5140 // like "X /u 5 == 0" is really checking that X is in the interval [0, 5).
5141 // Compute this interval based on the constants involved and the signedness of
5142 // the compare/divide. This computes a half-open interval, keeping track of
5143 // whether either value in the interval overflows. After analysis each
5144 // overflow variable is set to 0 if it's corresponding bound variable is valid
5145 // -1 if overflowed off the bottom end, or +1 if overflowed off the top end.
5146 int LoOverflow = 0, HiOverflow = 0;
5147 ConstantInt *LoBound = 0, *HiBound = 0;
5150 if (!DivIsSigned) { // udiv
5151 // e.g. X/5 op 3 --> [15, 20)
5153 HiOverflow = LoOverflow = ProdOV;
5155 HiOverflow = AddWithOverflow(HiBound, LoBound, DivRHS, false);
5156 } else if (DivRHS->getValue().isPositive()) { // Divisor is > 0.
5157 if (CmpRHSV == 0) { // (X / pos) op 0
5158 // Can't overflow. e.g. X/2 op 0 --> [-1, 2)
5159 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
5161 } else if (CmpRHSV.isPositive()) { // (X / pos) op pos
5162 LoBound = Prod; // e.g. X/5 op 3 --> [15, 20)
5163 HiOverflow = LoOverflow = ProdOV;
5165 HiOverflow = AddWithOverflow(HiBound, Prod, DivRHS, true);
5166 } else { // (X / pos) op neg
5167 // e.g. X/5 op -3 --> [-15-4, -15+1) --> [-19, -14)
5168 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
5169 LoOverflow = AddWithOverflow(LoBound, Prod,
5170 cast<ConstantInt>(DivRHSH), true) ? -1 : 0;
5171 HiBound = AddOne(Prod);
5172 HiOverflow = ProdOV ? -1 : 0;
5174 } else { // Divisor is < 0.
5175 if (CmpRHSV == 0) { // (X / neg) op 0
5176 // e.g. X/-5 op 0 --> [-4, 5)
5177 LoBound = AddOne(DivRHS);
5178 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
5179 if (HiBound == DivRHS) { // -INTMIN = INTMIN
5180 HiOverflow = 1; // [INTMIN+1, overflow)
5181 HiBound = 0; // e.g. X/INTMIN = 0 --> X > INTMIN
5183 } else if (CmpRHSV.isPositive()) { // (X / neg) op pos
5184 // e.g. X/-5 op 3 --> [-19, -14)
5185 HiOverflow = LoOverflow = ProdOV ? -1 : 0;
5187 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS), true) ?-1:0;
5188 HiBound = AddOne(Prod);
5189 } else { // (X / neg) op neg
5190 // e.g. X/-5 op -3 --> [15, 20)
5192 LoOverflow = HiOverflow = ProdOV ? 1 : 0;
5193 HiBound = Subtract(Prod, DivRHS);
5196 // Dividing by a negative swaps the condition. LT <-> GT
5197 Pred = ICmpInst::getSwappedPredicate(Pred);
5200 Value *X = DivI->getOperand(0);
5202 default: assert(0 && "Unhandled icmp opcode!");
5203 case ICmpInst::ICMP_EQ:
5204 if (LoOverflow && HiOverflow)
5205 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5206 else if (HiOverflow)
5207 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5208 ICmpInst::ICMP_UGE, X, LoBound);
5209 else if (LoOverflow)
5210 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5211 ICmpInst::ICMP_ULT, X, HiBound);
5213 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, true, ICI);
5214 case ICmpInst::ICMP_NE:
5215 if (LoOverflow && HiOverflow)
5216 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5217 else if (HiOverflow)
5218 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SLT :
5219 ICmpInst::ICMP_ULT, X, LoBound);
5220 else if (LoOverflow)
5221 return new ICmpInst(DivIsSigned ? ICmpInst::ICMP_SGE :
5222 ICmpInst::ICMP_UGE, X, HiBound);
5224 return InsertRangeTest(X, LoBound, HiBound, DivIsSigned, false, ICI);
5225 case ICmpInst::ICMP_ULT:
5226 case ICmpInst::ICMP_SLT:
5227 if (LoOverflow == +1) // Low bound is greater than input range.
5228 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5229 if (LoOverflow == -1) // Low bound is less than input range.
5230 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5231 return new ICmpInst(Pred, X, LoBound);
5232 case ICmpInst::ICMP_UGT:
5233 case ICmpInst::ICMP_SGT:
5234 if (HiOverflow == +1) // High bound greater than input range.
5235 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5236 else if (HiOverflow == -1) // High bound less than input range.
5237 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5238 if (Pred == ICmpInst::ICMP_UGT)
5239 return new ICmpInst(ICmpInst::ICMP_UGE, X, HiBound);
5241 return new ICmpInst(ICmpInst::ICMP_SGE, X, HiBound);
5246 /// visitICmpInstWithInstAndIntCst - Handle "icmp (instr, intcst)".
5248 Instruction *InstCombiner::visitICmpInstWithInstAndIntCst(ICmpInst &ICI,
5251 const APInt &RHSV = RHS->getValue();
5253 switch (LHSI->getOpcode()) {
5254 case Instruction::Xor: // (icmp pred (xor X, XorCST), CI)
5255 if (ConstantInt *XorCST = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
5256 // If this is a comparison that tests the signbit (X < 0) or (x > -1),
5258 if (ICI.getPredicate() == ICmpInst::ICMP_SLT && RHSV == 0 ||
5259 ICI.getPredicate() == ICmpInst::ICMP_SGT && RHSV.isAllOnesValue()) {
5260 Value *CompareVal = LHSI->getOperand(0);
5262 // If the sign bit of the XorCST is not set, there is no change to
5263 // the operation, just stop using the Xor.
5264 if (!XorCST->getValue().isNegative()) {
5265 ICI.setOperand(0, CompareVal);
5266 AddToWorkList(LHSI);
5270 // Was the old condition true if the operand is positive?
5271 bool isTrueIfPositive = ICI.getPredicate() == ICmpInst::ICMP_SGT;
5273 // If so, the new one isn't.
5274 isTrueIfPositive ^= true;
5276 if (isTrueIfPositive)
5277 return new ICmpInst(ICmpInst::ICMP_SGT, CompareVal, SubOne(RHS));
5279 return new ICmpInst(ICmpInst::ICMP_SLT, CompareVal, AddOne(RHS));
5283 case Instruction::And: // (icmp pred (and X, AndCST), RHS)
5284 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
5285 LHSI->getOperand(0)->hasOneUse()) {
5286 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
5288 // If the LHS is an AND of a truncating cast, we can widen the
5289 // and/compare to be the input width without changing the value
5290 // produced, eliminating a cast.
5291 if (TruncInst *Cast = dyn_cast<TruncInst>(LHSI->getOperand(0))) {
5292 // We can do this transformation if either the AND constant does not
5293 // have its sign bit set or if it is an equality comparison.
5294 // Extending a relational comparison when we're checking the sign
5295 // bit would not work.
5296 if (Cast->hasOneUse() &&
5297 (ICI.isEquality() || AndCST->getValue().isPositive() &&
5298 RHSV.isPositive())) {
5300 cast<IntegerType>(Cast->getOperand(0)->getType())->getBitWidth();
5301 APInt NewCST = AndCST->getValue();
5302 NewCST.zext(BitWidth);
5304 NewCI.zext(BitWidth);
5305 Instruction *NewAnd =
5306 BinaryOperator::createAnd(Cast->getOperand(0),
5307 ConstantInt::get(NewCST),LHSI->getName());
5308 InsertNewInstBefore(NewAnd, ICI);
5309 return new ICmpInst(ICI.getPredicate(), NewAnd,
5310 ConstantInt::get(NewCI));
5314 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
5315 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
5316 // happens a LOT in code produced by the C front-end, for bitfield
5318 BinaryOperator *Shift = dyn_cast<BinaryOperator>(LHSI->getOperand(0));
5319 if (Shift && !Shift->isShift())
5323 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
5324 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
5325 const Type *AndTy = AndCST->getType(); // Type of the and.
5327 // We can fold this as long as we can't shift unknown bits
5328 // into the mask. This can only happen with signed shift
5329 // rights, as they sign-extend.
5331 bool CanFold = Shift->isLogicalShift();
5333 // To test for the bad case of the signed shr, see if any
5334 // of the bits shifted in could be tested after the mask.
5335 uint32_t TyBits = Ty->getPrimitiveSizeInBits();
5336 int ShAmtVal = TyBits - ShAmt->getLimitedValue(TyBits);
5338 uint32_t BitWidth = AndTy->getPrimitiveSizeInBits();
5339 if ((APInt::getHighBitsSet(BitWidth, BitWidth-ShAmtVal) &
5340 AndCST->getValue()) == 0)
5346 if (Shift->getOpcode() == Instruction::Shl)
5347 NewCst = ConstantExpr::getLShr(RHS, ShAmt);
5349 NewCst = ConstantExpr::getShl(RHS, ShAmt);
5351 // Check to see if we are shifting out any of the bits being
5353 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != RHS) {
5354 // If we shifted bits out, the fold is not going to work out.
5355 // As a special case, check to see if this means that the
5356 // result is always true or false now.
5357 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5358 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5359 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5360 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5362 ICI.setOperand(1, NewCst);
5363 Constant *NewAndCST;
5364 if (Shift->getOpcode() == Instruction::Shl)
5365 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
5367 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
5368 LHSI->setOperand(1, NewAndCST);
5369 LHSI->setOperand(0, Shift->getOperand(0));
5370 AddToWorkList(Shift); // Shift is dead.
5371 AddUsesToWorkList(ICI);
5377 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
5378 // preferable because it allows the C<<Y expression to be hoisted out
5379 // of a loop if Y is invariant and X is not.
5380 if (Shift && Shift->hasOneUse() && RHSV == 0 &&
5381 ICI.isEquality() && !Shift->isArithmeticShift() &&
5382 isa<Instruction>(Shift->getOperand(0))) {
5385 if (Shift->getOpcode() == Instruction::LShr) {
5386 NS = BinaryOperator::createShl(AndCST,
5387 Shift->getOperand(1), "tmp");
5389 // Insert a logical shift.
5390 NS = BinaryOperator::createLShr(AndCST,
5391 Shift->getOperand(1), "tmp");
5393 InsertNewInstBefore(cast<Instruction>(NS), ICI);
5395 // Compute X & (C << Y).
5396 Instruction *NewAnd =
5397 BinaryOperator::createAnd(Shift->getOperand(0), NS, LHSI->getName());
5398 InsertNewInstBefore(NewAnd, ICI);
5400 ICI.setOperand(0, NewAnd);
5406 case Instruction::Shl: { // (icmp pred (shl X, ShAmt), CI)
5407 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5410 uint32_t TypeBits = RHSV.getBitWidth();
5412 // Check that the shift amount is in range. If not, don't perform
5413 // undefined shifts. When the shift is visited it will be
5415 if (ShAmt->uge(TypeBits))
5418 if (ICI.isEquality()) {
5419 // If we are comparing against bits always shifted out, the
5420 // comparison cannot succeed.
5422 ConstantExpr::getShl(ConstantExpr::getLShr(RHS, ShAmt), ShAmt);
5423 if (Comp != RHS) {// Comparing against a bit that we know is zero.
5424 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5425 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5426 return ReplaceInstUsesWith(ICI, Cst);
5429 if (LHSI->hasOneUse()) {
5430 // Otherwise strength reduce the shift into an and.
5431 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5433 ConstantInt::get(APInt::getLowBitsSet(TypeBits, TypeBits-ShAmtVal));
5436 BinaryOperator::createAnd(LHSI->getOperand(0),
5437 Mask, LHSI->getName()+".mask");
5438 Value *And = InsertNewInstBefore(AndI, ICI);
5439 return new ICmpInst(ICI.getPredicate(), And,
5440 ConstantInt::get(RHSV.lshr(ShAmtVal)));
5444 // Otherwise, if this is a comparison of the sign bit, simplify to and/test.
5445 bool TrueIfSigned = false;
5446 if (LHSI->hasOneUse() &&
5447 isSignBitCheck(ICI.getPredicate(), RHS, TrueIfSigned)) {
5448 // (X << 31) <s 0 --> (X&1) != 0
5449 Constant *Mask = ConstantInt::get(APInt(TypeBits, 1) <<
5450 (TypeBits-ShAmt->getZExtValue()-1));
5452 BinaryOperator::createAnd(LHSI->getOperand(0),
5453 Mask, LHSI->getName()+".mask");
5454 Value *And = InsertNewInstBefore(AndI, ICI);
5456 return new ICmpInst(TrueIfSigned ? ICmpInst::ICMP_NE : ICmpInst::ICMP_EQ,
5457 And, Constant::getNullValue(And->getType()));
5462 case Instruction::LShr: // (icmp pred (shr X, ShAmt), CI)
5463 case Instruction::AShr: {
5464 ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1));
5467 if (ICI.isEquality()) {
5468 // Check that the shift amount is in range. If not, don't perform
5469 // undefined shifts. When the shift is visited it will be
5471 uint32_t TypeBits = RHSV.getBitWidth();
5472 if (ShAmt->uge(TypeBits))
5474 uint32_t ShAmtVal = (uint32_t)ShAmt->getLimitedValue(TypeBits);
5476 // If we are comparing against bits always shifted out, the
5477 // comparison cannot succeed.
5478 APInt Comp = RHSV << ShAmtVal;
5479 if (LHSI->getOpcode() == Instruction::LShr)
5480 Comp = Comp.lshr(ShAmtVal);
5482 Comp = Comp.ashr(ShAmtVal);
5484 if (Comp != RHSV) { // Comparing against a bit that we know is zero.
5485 bool IsICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5486 Constant *Cst = ConstantInt::get(Type::Int1Ty, IsICMP_NE);
5487 return ReplaceInstUsesWith(ICI, Cst);
5490 if (LHSI->hasOneUse() || RHSV == 0) {
5491 // Otherwise strength reduce the shift into an and.
5492 APInt Val(APInt::getHighBitsSet(TypeBits, TypeBits - ShAmtVal));
5493 Constant *Mask = ConstantInt::get(Val);
5496 BinaryOperator::createAnd(LHSI->getOperand(0),
5497 Mask, LHSI->getName()+".mask");
5498 Value *And = InsertNewInstBefore(AndI, ICI);
5499 return new ICmpInst(ICI.getPredicate(), And,
5500 ConstantExpr::getShl(RHS, ShAmt));
5506 case Instruction::SDiv:
5507 case Instruction::UDiv:
5508 // Fold: icmp pred ([us]div X, C1), C2 -> range test
5509 // Fold this div into the comparison, producing a range check.
5510 // Determine, based on the divide type, what the range is being
5511 // checked. If there is an overflow on the low or high side, remember
5512 // it, otherwise compute the range [low, hi) bounding the new value.
5513 // See: InsertRangeTest above for the kinds of replacements possible.
5514 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1)))
5515 if (Instruction *R = FoldICmpDivCst(ICI, cast<BinaryOperator>(LHSI),
5521 // Simplify icmp_eq and icmp_ne instructions with integer constant RHS.
5522 if (ICI.isEquality()) {
5523 bool isICMP_NE = ICI.getPredicate() == ICmpInst::ICMP_NE;
5525 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
5526 // the second operand is a constant, simplify a bit.
5527 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(LHSI)) {
5528 switch (BO->getOpcode()) {
5529 case Instruction::SRem:
5530 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
5531 if (RHSV == 0 && isa<ConstantInt>(BO->getOperand(1)) &&BO->hasOneUse()){
5532 const APInt &V = cast<ConstantInt>(BO->getOperand(1))->getValue();
5533 if (V.sgt(APInt(V.getBitWidth(), 1)) && V.isPowerOf2()) {
5534 Instruction *NewRem =
5535 BinaryOperator::createURem(BO->getOperand(0), BO->getOperand(1),
5537 InsertNewInstBefore(NewRem, ICI);
5538 return new ICmpInst(ICI.getPredicate(), NewRem,
5539 Constant::getNullValue(BO->getType()));
5543 case Instruction::Add:
5544 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
5545 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5546 if (BO->hasOneUse())
5547 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5548 Subtract(RHS, BOp1C));
5549 } else if (RHSV == 0) {
5550 // Replace ((add A, B) != 0) with (A != -B) if A or B is
5551 // efficiently invertible, or if the add has just this one use.
5552 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
5554 if (Value *NegVal = dyn_castNegVal(BOp1))
5555 return new ICmpInst(ICI.getPredicate(), BOp0, NegVal);
5556 else if (Value *NegVal = dyn_castNegVal(BOp0))
5557 return new ICmpInst(ICI.getPredicate(), NegVal, BOp1);
5558 else if (BO->hasOneUse()) {
5559 Instruction *Neg = BinaryOperator::createNeg(BOp1);
5560 InsertNewInstBefore(Neg, ICI);
5562 return new ICmpInst(ICI.getPredicate(), BOp0, Neg);
5566 case Instruction::Xor:
5567 // For the xor case, we can xor two constants together, eliminating
5568 // the explicit xor.
5569 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
5570 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5571 ConstantExpr::getXor(RHS, BOC));
5574 case Instruction::Sub:
5575 // Replace (([sub|xor] A, B) != 0) with (A != B)
5577 return new ICmpInst(ICI.getPredicate(), BO->getOperand(0),
5581 case Instruction::Or:
5582 // If bits are being or'd in that are not present in the constant we
5583 // are comparing against, then the comparison could never succeed!
5584 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
5585 Constant *NotCI = ConstantExpr::getNot(RHS);
5586 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
5587 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5592 case Instruction::And:
5593 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
5594 // If bits are being compared against that are and'd out, then the
5595 // comparison can never succeed!
5596 if ((RHSV & ~BOC->getValue()) != 0)
5597 return ReplaceInstUsesWith(ICI, ConstantInt::get(Type::Int1Ty,
5600 // If we have ((X & C) == C), turn it into ((X & C) != 0).
5601 if (RHS == BOC && RHSV.isPowerOf2())
5602 return new ICmpInst(isICMP_NE ? ICmpInst::ICMP_EQ :
5603 ICmpInst::ICMP_NE, LHSI,
5604 Constant::getNullValue(RHS->getType()));
5606 // Replace (and X, (1 << size(X)-1) != 0) with x s< 0
5607 if (isSignBit(BOC)) {
5608 Value *X = BO->getOperand(0);
5609 Constant *Zero = Constant::getNullValue(X->getType());
5610 ICmpInst::Predicate pred = isICMP_NE ?
5611 ICmpInst::ICMP_SLT : ICmpInst::ICMP_SGE;
5612 return new ICmpInst(pred, X, Zero);
5615 // ((X & ~7) == 0) --> X < 8
5616 if (RHSV == 0 && isHighOnes(BOC)) {
5617 Value *X = BO->getOperand(0);
5618 Constant *NegX = ConstantExpr::getNeg(BOC);
5619 ICmpInst::Predicate pred = isICMP_NE ?
5620 ICmpInst::ICMP_UGE : ICmpInst::ICMP_ULT;
5621 return new ICmpInst(pred, X, NegX);
5626 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(LHSI)) {
5627 // Handle icmp {eq|ne} <intrinsic>, intcst.
5628 if (II->getIntrinsicID() == Intrinsic::bswap) {
5630 ICI.setOperand(0, II->getOperand(1));
5631 ICI.setOperand(1, ConstantInt::get(RHSV.byteSwap()));
5635 } else { // Not a ICMP_EQ/ICMP_NE
5636 // If the LHS is a cast from an integral value of the same size,
5637 // then since we know the RHS is a constant, try to simlify.
5638 if (CastInst *Cast = dyn_cast<CastInst>(LHSI)) {
5639 Value *CastOp = Cast->getOperand(0);
5640 const Type *SrcTy = CastOp->getType();
5641 uint32_t SrcTySize = SrcTy->getPrimitiveSizeInBits();
5642 if (SrcTy->isInteger() &&
5643 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
5644 // If this is an unsigned comparison, try to make the comparison use
5645 // smaller constant values.
5646 if (ICI.getPredicate() == ICmpInst::ICMP_ULT && RHSV.isSignBit()) {
5647 // X u< 128 => X s> -1
5648 return new ICmpInst(ICmpInst::ICMP_SGT, CastOp,
5649 ConstantInt::get(APInt::getAllOnesValue(SrcTySize)));
5650 } else if (ICI.getPredicate() == ICmpInst::ICMP_UGT &&
5651 RHSV == APInt::getSignedMaxValue(SrcTySize)) {
5652 // X u> 127 => X s< 0
5653 return new ICmpInst(ICmpInst::ICMP_SLT, CastOp,
5654 Constant::getNullValue(SrcTy));
5662 /// visitICmpInstWithCastAndCast - Handle icmp (cast x to y), (cast/cst).
5663 /// We only handle extending casts so far.
5665 Instruction *InstCombiner::visitICmpInstWithCastAndCast(ICmpInst &ICI) {
5666 const CastInst *LHSCI = cast<CastInst>(ICI.getOperand(0));
5667 Value *LHSCIOp = LHSCI->getOperand(0);
5668 const Type *SrcTy = LHSCIOp->getType();
5669 const Type *DestTy = LHSCI->getType();
5672 // Turn icmp (ptrtoint x), (ptrtoint/c) into a compare of the input if the
5673 // integer type is the same size as the pointer type.
5674 if (LHSCI->getOpcode() == Instruction::PtrToInt &&
5675 getTargetData().getPointerSizeInBits() ==
5676 cast<IntegerType>(DestTy)->getBitWidth()) {
5678 if (Constant *RHSC = dyn_cast<Constant>(ICI.getOperand(1))) {
5679 RHSOp = ConstantExpr::getIntToPtr(RHSC, SrcTy);
5680 } else if (PtrToIntInst *RHSC = dyn_cast<PtrToIntInst>(ICI.getOperand(1))) {
5681 RHSOp = RHSC->getOperand(0);
5682 // If the pointer types don't match, insert a bitcast.
5683 if (LHSCIOp->getType() != RHSOp->getType())
5684 RHSOp = InsertCastBefore(Instruction::BitCast, RHSOp,
5685 LHSCIOp->getType(), ICI);
5689 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSOp);
5692 // The code below only handles extension cast instructions, so far.
5694 if (LHSCI->getOpcode() != Instruction::ZExt &&
5695 LHSCI->getOpcode() != Instruction::SExt)
5698 bool isSignedExt = LHSCI->getOpcode() == Instruction::SExt;
5699 bool isSignedCmp = ICI.isSignedPredicate();
5701 if (CastInst *CI = dyn_cast<CastInst>(ICI.getOperand(1))) {
5702 // Not an extension from the same type?
5703 RHSCIOp = CI->getOperand(0);
5704 if (RHSCIOp->getType() != LHSCIOp->getType())
5707 // If the signedness of the two compares doesn't agree (i.e. one is a sext
5708 // and the other is a zext), then we can't handle this.
5709 if (CI->getOpcode() != LHSCI->getOpcode())
5712 // Likewise, if the signedness of the [sz]exts and the compare don't match,
5713 // then we can't handle this.
5714 if (isSignedExt != isSignedCmp && !ICI.isEquality())
5717 // Okay, just insert a compare of the reduced operands now!
5718 return new ICmpInst(ICI.getPredicate(), LHSCIOp, RHSCIOp);
5721 // If we aren't dealing with a constant on the RHS, exit early
5722 ConstantInt *CI = dyn_cast<ConstantInt>(ICI.getOperand(1));
5726 // Compute the constant that would happen if we truncated to SrcTy then
5727 // reextended to DestTy.
5728 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5729 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5731 // If the re-extended constant didn't change...
5733 // Make sure that sign of the Cmp and the sign of the Cast are the same.
5734 // For example, we might have:
5735 // %A = sext short %X to uint
5736 // %B = icmp ugt uint %A, 1330
5737 // It is incorrect to transform this into
5738 // %B = icmp ugt short %X, 1330
5739 // because %A may have negative value.
5741 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5742 // OR operation is EQ/NE.
5743 if (isSignedExt == isSignedCmp || SrcTy == Type::Int1Ty || ICI.isEquality())
5744 return new ICmpInst(ICI.getPredicate(), LHSCIOp, Res1);
5749 // The re-extended constant changed so the constant cannot be represented
5750 // in the shorter type. Consequently, we cannot emit a simple comparison.
5752 // First, handle some easy cases. We know the result cannot be equal at this
5753 // point so handle the ICI.isEquality() cases
5754 if (ICI.getPredicate() == ICmpInst::ICMP_EQ)
5755 return ReplaceInstUsesWith(ICI, ConstantInt::getFalse());
5756 if (ICI.getPredicate() == ICmpInst::ICMP_NE)
5757 return ReplaceInstUsesWith(ICI, ConstantInt::getTrue());
5759 // Evaluate the comparison for LT (we invert for GT below). LE and GE cases
5760 // should have been folded away previously and not enter in here.
5763 // We're performing a signed comparison.
5764 if (cast<ConstantInt>(CI)->getValue().isNegative())
5765 Result = ConstantInt::getFalse(); // X < (small) --> false
5767 Result = ConstantInt::getTrue(); // X < (large) --> true
5769 // We're performing an unsigned comparison.
5771 // We're performing an unsigned comp with a sign extended value.
5772 // This is true if the input is >= 0. [aka >s -1]
5773 Constant *NegOne = ConstantInt::getAllOnesValue(SrcTy);
5774 Result = InsertNewInstBefore(new ICmpInst(ICmpInst::ICMP_SGT, LHSCIOp,
5775 NegOne, ICI.getName()), ICI);
5777 // Unsigned extend & unsigned compare -> always true.
5778 Result = ConstantInt::getTrue();
5782 // Finally, return the value computed.
5783 if (ICI.getPredicate() == ICmpInst::ICMP_ULT ||
5784 ICI.getPredicate() == ICmpInst::ICMP_SLT) {
5785 return ReplaceInstUsesWith(ICI, Result);
5787 assert((ICI.getPredicate()==ICmpInst::ICMP_UGT ||
5788 ICI.getPredicate()==ICmpInst::ICMP_SGT) &&
5789 "ICmp should be folded!");
5790 if (Constant *CI = dyn_cast<Constant>(Result))
5791 return ReplaceInstUsesWith(ICI, ConstantExpr::getNot(CI));
5793 return BinaryOperator::createNot(Result);
5797 Instruction *InstCombiner::visitShl(BinaryOperator &I) {
5798 return commonShiftTransforms(I);
5801 Instruction *InstCombiner::visitLShr(BinaryOperator &I) {
5802 return commonShiftTransforms(I);
5805 Instruction *InstCombiner::visitAShr(BinaryOperator &I) {
5806 return commonShiftTransforms(I);
5809 Instruction *InstCombiner::commonShiftTransforms(BinaryOperator &I) {
5810 assert(I.getOperand(1)->getType() == I.getOperand(0)->getType());
5811 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5813 // shl X, 0 == X and shr X, 0 == X
5814 // shl 0, X == 0 and shr 0, X == 0
5815 if (Op1 == Constant::getNullValue(Op1->getType()) ||
5816 Op0 == Constant::getNullValue(Op0->getType()))
5817 return ReplaceInstUsesWith(I, Op0);
5819 if (isa<UndefValue>(Op0)) {
5820 if (I.getOpcode() == Instruction::AShr) // undef >>s X -> undef
5821 return ReplaceInstUsesWith(I, Op0);
5822 else // undef << X -> 0, undef >>u X -> 0
5823 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5825 if (isa<UndefValue>(Op1)) {
5826 if (I.getOpcode() == Instruction::AShr) // X >>s undef -> X
5827 return ReplaceInstUsesWith(I, Op0);
5828 else // X << undef, X >>u undef -> 0
5829 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5832 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5833 if (I.getOpcode() == Instruction::AShr)
5834 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5835 if (CSI->isAllOnesValue())
5836 return ReplaceInstUsesWith(I, CSI);
5838 // Try to fold constant and into select arguments.
5839 if (isa<Constant>(Op0))
5840 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5841 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5844 // See if we can turn a signed shr into an unsigned shr.
5845 if (I.isArithmeticShift()) {
5846 if (MaskedValueIsZero(Op0,
5847 APInt::getSignBit(I.getType()->getPrimitiveSizeInBits()))) {
5848 return BinaryOperator::createLShr(Op0, Op1, I.getName());
5852 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5853 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5858 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5859 BinaryOperator &I) {
5860 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5862 // See if we can simplify any instructions used by the instruction whose sole
5863 // purpose is to compute bits we don't care about.
5864 uint32_t TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5865 APInt KnownZero(TypeBits, 0), KnownOne(TypeBits, 0);
5866 if (SimplifyDemandedBits(&I, APInt::getAllOnesValue(TypeBits),
5867 KnownZero, KnownOne))
5870 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5871 // of a signed value.
5873 if (Op1->uge(TypeBits)) {
5874 if (I.getOpcode() != Instruction::AShr)
5875 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5877 I.setOperand(1, ConstantInt::get(I.getType(), TypeBits-1));
5882 // ((X*C1) << C2) == (X * (C1 << C2))
5883 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5884 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5885 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5886 return BinaryOperator::createMul(BO->getOperand(0),
5887 ConstantExpr::getShl(BOOp, Op1));
5889 // Try to fold constant and into select arguments.
5890 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5891 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5893 if (isa<PHINode>(Op0))
5894 if (Instruction *NV = FoldOpIntoPhi(I))
5897 if (Op0->hasOneUse()) {
5898 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5899 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5902 switch (Op0BO->getOpcode()) {
5904 case Instruction::Add:
5905 case Instruction::And:
5906 case Instruction::Or:
5907 case Instruction::Xor: {
5908 // These operators commute.
5909 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5910 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5911 match(Op0BO->getOperand(1),
5912 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5913 Instruction *YS = BinaryOperator::createShl(
5914 Op0BO->getOperand(0), Op1,
5916 InsertNewInstBefore(YS, I); // (Y << C)
5918 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5919 Op0BO->getOperand(1)->getName());
5920 InsertNewInstBefore(X, I); // (X + (Y << C))
5921 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5922 return BinaryOperator::createAnd(X, ConstantInt::get(
5923 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5926 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5927 Value *Op0BOOp1 = Op0BO->getOperand(1);
5928 if (isLeftShift && Op0BOOp1->hasOneUse() &&
5930 m_And(m_Shr(m_Value(V1), m_Value(V2)),m_ConstantInt(CC))) &&
5931 cast<BinaryOperator>(Op0BOOp1)->getOperand(0)->hasOneUse() &&
5933 Instruction *YS = BinaryOperator::createShl(
5934 Op0BO->getOperand(0), Op1,
5936 InsertNewInstBefore(YS, I); // (Y << C)
5938 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5939 V1->getName()+".mask");
5940 InsertNewInstBefore(XM, I); // X & (CC << C)
5942 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5947 case Instruction::Sub: {
5948 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5949 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5950 match(Op0BO->getOperand(0),
5951 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5952 Instruction *YS = BinaryOperator::createShl(
5953 Op0BO->getOperand(1), Op1,
5955 InsertNewInstBefore(YS, I); // (Y << C)
5957 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5958 Op0BO->getOperand(0)->getName());
5959 InsertNewInstBefore(X, I); // (X + (Y << C))
5960 uint32_t Op1Val = Op1->getLimitedValue(TypeBits);
5961 return BinaryOperator::createAnd(X, ConstantInt::get(
5962 APInt::getHighBitsSet(TypeBits, TypeBits-Op1Val)));
5965 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5966 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5967 match(Op0BO->getOperand(0),
5968 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5969 m_ConstantInt(CC))) && V2 == Op1 &&
5970 cast<BinaryOperator>(Op0BO->getOperand(0))
5971 ->getOperand(0)->hasOneUse()) {
5972 Instruction *YS = BinaryOperator::createShl(
5973 Op0BO->getOperand(1), Op1,
5975 InsertNewInstBefore(YS, I); // (Y << C)
5977 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5978 V1->getName()+".mask");
5979 InsertNewInstBefore(XM, I); // X & (CC << C)
5981 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5989 // If the operand is an bitwise operator with a constant RHS, and the
5990 // shift is the only use, we can pull it out of the shift.
5991 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5992 bool isValid = true; // Valid only for And, Or, Xor
5993 bool highBitSet = false; // Transform if high bit of constant set?
5995 switch (Op0BO->getOpcode()) {
5996 default: isValid = false; break; // Do not perform transform!
5997 case Instruction::Add:
5998 isValid = isLeftShift;
6000 case Instruction::Or:
6001 case Instruction::Xor:
6004 case Instruction::And:
6009 // If this is a signed shift right, and the high bit is modified
6010 // by the logical operation, do not perform the transformation.
6011 // The highBitSet boolean indicates the value of the high bit of
6012 // the constant which would cause it to be modified for this
6015 if (isValid && !isLeftShift && I.getOpcode() == Instruction::AShr) {
6016 isValid = Op0C->getValue()[TypeBits-1] == highBitSet;
6020 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
6022 Instruction *NewShift =
6023 BinaryOperator::create(I.getOpcode(), Op0BO->getOperand(0), Op1);
6024 InsertNewInstBefore(NewShift, I);
6025 NewShift->takeName(Op0BO);
6027 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
6034 // Find out if this is a shift of a shift by a constant.
6035 BinaryOperator *ShiftOp = dyn_cast<BinaryOperator>(Op0);
6036 if (ShiftOp && !ShiftOp->isShift())
6039 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
6040 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
6041 uint32_t ShiftAmt1 = ShiftAmt1C->getLimitedValue(TypeBits);
6042 uint32_t ShiftAmt2 = Op1->getLimitedValue(TypeBits);
6043 assert(ShiftAmt2 != 0 && "Should have been simplified earlier");
6044 if (ShiftAmt1 == 0) return 0; // Will be simplified in the future.
6045 Value *X = ShiftOp->getOperand(0);
6047 uint32_t AmtSum = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
6048 if (AmtSum > TypeBits)
6051 const IntegerType *Ty = cast<IntegerType>(I.getType());
6053 // Check for (X << c1) << c2 and (X >> c1) >> c2
6054 if (I.getOpcode() == ShiftOp->getOpcode()) {
6055 return BinaryOperator::create(I.getOpcode(), X,
6056 ConstantInt::get(Ty, AmtSum));
6057 } else if (ShiftOp->getOpcode() == Instruction::LShr &&
6058 I.getOpcode() == Instruction::AShr) {
6059 // ((X >>u C1) >>s C2) -> (X >>u (C1+C2)) since C1 != 0.
6060 return BinaryOperator::createLShr(X, ConstantInt::get(Ty, AmtSum));
6061 } else if (ShiftOp->getOpcode() == Instruction::AShr &&
6062 I.getOpcode() == Instruction::LShr) {
6063 // ((X >>s C1) >>u C2) -> ((X >>s (C1+C2)) & mask) since C1 != 0.
6064 Instruction *Shift =
6065 BinaryOperator::createAShr(X, ConstantInt::get(Ty, AmtSum));
6066 InsertNewInstBefore(Shift, I);
6068 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6069 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6072 // Okay, if we get here, one shift must be left, and the other shift must be
6073 // right. See if the amounts are equal.
6074 if (ShiftAmt1 == ShiftAmt2) {
6075 // If we have ((X >>? C) << C), turn this into X & (-1 << C).
6076 if (I.getOpcode() == Instruction::Shl) {
6077 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt1));
6078 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6080 // If we have ((X << C) >>u C), turn this into X & (-1 >>u C).
6081 if (I.getOpcode() == Instruction::LShr) {
6082 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt1));
6083 return BinaryOperator::createAnd(X, ConstantInt::get(Mask));
6085 // We can simplify ((X << C) >>s C) into a trunc + sext.
6086 // NOTE: we could do this for any C, but that would make 'unusual' integer
6087 // types. For now, just stick to ones well-supported by the code
6089 const Type *SExtType = 0;
6090 switch (Ty->getBitWidth() - ShiftAmt1) {
6097 SExtType = IntegerType::get(Ty->getBitWidth() - ShiftAmt1);
6102 Instruction *NewTrunc = new TruncInst(X, SExtType, "sext");
6103 InsertNewInstBefore(NewTrunc, I);
6104 return new SExtInst(NewTrunc, Ty);
6106 // Otherwise, we can't handle it yet.
6107 } else if (ShiftAmt1 < ShiftAmt2) {
6108 uint32_t ShiftDiff = ShiftAmt2-ShiftAmt1;
6110 // (X >>? C1) << C2 --> X << (C2-C1) & (-1 << C2)
6111 if (I.getOpcode() == Instruction::Shl) {
6112 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6113 ShiftOp->getOpcode() == Instruction::AShr);
6114 Instruction *Shift =
6115 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6116 InsertNewInstBefore(Shift, I);
6118 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6119 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6122 // (X << C1) >>u C2 --> X >>u (C2-C1) & (-1 >> C2)
6123 if (I.getOpcode() == Instruction::LShr) {
6124 assert(ShiftOp->getOpcode() == Instruction::Shl);
6125 Instruction *Shift =
6126 BinaryOperator::createLShr(X, ConstantInt::get(Ty, ShiftDiff));
6127 InsertNewInstBefore(Shift, I);
6129 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6130 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6133 // We can't handle (X << C1) >>s C2, it shifts arbitrary bits in.
6135 assert(ShiftAmt2 < ShiftAmt1);
6136 uint32_t ShiftDiff = ShiftAmt1-ShiftAmt2;
6138 // (X >>? C1) << C2 --> X >>? (C1-C2) & (-1 << C2)
6139 if (I.getOpcode() == Instruction::Shl) {
6140 assert(ShiftOp->getOpcode() == Instruction::LShr ||
6141 ShiftOp->getOpcode() == Instruction::AShr);
6142 Instruction *Shift =
6143 BinaryOperator::create(ShiftOp->getOpcode(), X,
6144 ConstantInt::get(Ty, ShiftDiff));
6145 InsertNewInstBefore(Shift, I);
6147 APInt Mask(APInt::getHighBitsSet(TypeBits, TypeBits - ShiftAmt2));
6148 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6151 // (X << C1) >>u C2 --> X << (C1-C2) & (-1 >> C2)
6152 if (I.getOpcode() == Instruction::LShr) {
6153 assert(ShiftOp->getOpcode() == Instruction::Shl);
6154 Instruction *Shift =
6155 BinaryOperator::createShl(X, ConstantInt::get(Ty, ShiftDiff));
6156 InsertNewInstBefore(Shift, I);
6158 APInt Mask(APInt::getLowBitsSet(TypeBits, TypeBits - ShiftAmt2));
6159 return BinaryOperator::createAnd(Shift, ConstantInt::get(Mask));
6162 // We can't handle (X << C1) >>a C2, it shifts arbitrary bits in.
6169 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
6170 /// expression. If so, decompose it, returning some value X, such that Val is
6173 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
6175 assert(Val->getType() == Type::Int32Ty && "Unexpected allocation size type!");
6176 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
6177 Offset = CI->getZExtValue();
6179 return ConstantInt::get(Type::Int32Ty, 0);
6180 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
6181 if (I->getNumOperands() == 2) {
6182 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6183 if (I->getOpcode() == Instruction::Shl) {
6184 // This is a value scaled by '1 << the shift amt'.
6185 Scale = 1U << CUI->getZExtValue();
6187 return I->getOperand(0);
6188 } else if (I->getOpcode() == Instruction::Mul) {
6189 // This value is scaled by 'CUI'.
6190 Scale = CUI->getZExtValue();
6192 return I->getOperand(0);
6193 } else if (I->getOpcode() == Instruction::Add) {
6194 // We have X+C. Check to see if we really have (X*C2)+C1,
6195 // where C1 is divisible by C2.
6198 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
6199 Offset += CUI->getZExtValue();
6200 if (SubScale > 1 && (Offset % SubScale == 0)) {
6209 // Otherwise, we can't look past this.
6216 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
6217 /// try to eliminate the cast by moving the type information into the alloc.
6218 Instruction *InstCombiner::PromoteCastOfAllocation(BitCastInst &CI,
6219 AllocationInst &AI) {
6220 const PointerType *PTy = cast<PointerType>(CI.getType());
6222 // Remove any uses of AI that are dead.
6223 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
6225 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
6226 Instruction *User = cast<Instruction>(*UI++);
6227 if (isInstructionTriviallyDead(User)) {
6228 while (UI != E && *UI == User)
6229 ++UI; // If this instruction uses AI more than once, don't break UI.
6232 DOUT << "IC: DCE: " << *User;
6233 EraseInstFromFunction(*User);
6237 // Get the type really allocated and the type casted to.
6238 const Type *AllocElTy = AI.getAllocatedType();
6239 const Type *CastElTy = PTy->getElementType();
6240 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
6242 unsigned AllocElTyAlign = TD->getABITypeAlignment(AllocElTy);
6243 unsigned CastElTyAlign = TD->getABITypeAlignment(CastElTy);
6244 if (CastElTyAlign < AllocElTyAlign) return 0;
6246 // If the allocation has multiple uses, only promote it if we are strictly
6247 // increasing the alignment of the resultant allocation. If we keep it the
6248 // same, we open the door to infinite loops of various kinds.
6249 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
6251 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
6252 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
6253 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
6255 // See if we can satisfy the modulus by pulling a scale out of the array
6257 unsigned ArraySizeScale;
6259 Value *NumElements = // See if the array size is a decomposable linear expr.
6260 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
6262 // If we can now satisfy the modulus, by using a non-1 scale, we really can
6264 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
6265 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
6267 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
6272 // If the allocation size is constant, form a constant mul expression
6273 Amt = ConstantInt::get(Type::Int32Ty, Scale);
6274 if (isa<ConstantInt>(NumElements))
6275 Amt = Multiply(cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
6276 // otherwise multiply the amount and the number of elements
6277 else if (Scale != 1) {
6278 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
6279 Amt = InsertNewInstBefore(Tmp, AI);
6283 if (int Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
6284 Value *Off = ConstantInt::get(Type::Int32Ty, Offset, true);
6285 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
6286 Amt = InsertNewInstBefore(Tmp, AI);
6289 AllocationInst *New;
6290 if (isa<MallocInst>(AI))
6291 New = new MallocInst(CastElTy, Amt, AI.getAlignment());
6293 New = new AllocaInst(CastElTy, Amt, AI.getAlignment());
6294 InsertNewInstBefore(New, AI);
6297 // If the allocation has multiple uses, insert a cast and change all things
6298 // that used it to use the new cast. This will also hack on CI, but it will
6300 if (!AI.hasOneUse()) {
6301 AddUsesToWorkList(AI);
6302 // New is the allocation instruction, pointer typed. AI is the original
6303 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
6304 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
6305 InsertNewInstBefore(NewCast, AI);
6306 AI.replaceAllUsesWith(NewCast);
6308 return ReplaceInstUsesWith(CI, New);
6311 /// CanEvaluateInDifferentType - Return true if we can take the specified value
6312 /// and return it as type Ty without inserting any new casts and without
6313 /// changing the computed value. This is used by code that tries to decide
6314 /// whether promoting or shrinking integer operations to wider or smaller types
6315 /// will allow us to eliminate a truncate or extend.
6317 /// This is a truncation operation if Ty is smaller than V->getType(), or an
6318 /// extension operation if Ty is larger.
6319 static bool CanEvaluateInDifferentType(Value *V, const IntegerType *Ty,
6320 unsigned CastOpc, int &NumCastsRemoved) {
6321 // We can always evaluate constants in another type.
6322 if (isa<ConstantInt>(V))
6325 Instruction *I = dyn_cast<Instruction>(V);
6326 if (!I) return false;
6328 const IntegerType *OrigTy = cast<IntegerType>(V->getType());
6330 // If this is an extension or truncate, we can often eliminate it.
6331 if (isa<TruncInst>(I) || isa<ZExtInst>(I) || isa<SExtInst>(I)) {
6332 // If this is a cast from the destination type, we can trivially eliminate
6333 // it, and this will remove a cast overall.
6334 if (I->getOperand(0)->getType() == Ty) {
6335 // If the first operand is itself a cast, and is eliminable, do not count
6336 // this as an eliminable cast. We would prefer to eliminate those two
6338 if (!isa<CastInst>(I->getOperand(0)))
6344 // We can't extend or shrink something that has multiple uses: doing so would
6345 // require duplicating the instruction in general, which isn't profitable.
6346 if (!I->hasOneUse()) return false;
6348 switch (I->getOpcode()) {
6349 case Instruction::Add:
6350 case Instruction::Sub:
6351 case Instruction::And:
6352 case Instruction::Or:
6353 case Instruction::Xor:
6354 // These operators can all arbitrarily be extended or truncated.
6355 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6357 CanEvaluateInDifferentType(I->getOperand(1), Ty, CastOpc,
6360 case Instruction::Shl:
6361 // If we are truncating the result of this SHL, and if it's a shift of a
6362 // constant amount, we can always perform a SHL in a smaller type.
6363 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6364 uint32_t BitWidth = Ty->getBitWidth();
6365 if (BitWidth < OrigTy->getBitWidth() &&
6366 CI->getLimitedValue(BitWidth) < BitWidth)
6367 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6371 case Instruction::LShr:
6372 // If this is a truncate of a logical shr, we can truncate it to a smaller
6373 // lshr iff we know that the bits we would otherwise be shifting in are
6375 if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
6376 uint32_t OrigBitWidth = OrigTy->getBitWidth();
6377 uint32_t BitWidth = Ty->getBitWidth();
6378 if (BitWidth < OrigBitWidth &&
6379 MaskedValueIsZero(I->getOperand(0),
6380 APInt::getHighBitsSet(OrigBitWidth, OrigBitWidth-BitWidth)) &&
6381 CI->getLimitedValue(BitWidth) < BitWidth) {
6382 return CanEvaluateInDifferentType(I->getOperand(0), Ty, CastOpc,
6387 case Instruction::ZExt:
6388 case Instruction::SExt:
6389 case Instruction::Trunc:
6390 // If this is the same kind of case as our original (e.g. zext+zext), we
6391 // can safely replace it. Note that replacing it does not reduce the number
6392 // of casts in the input.
6393 if (I->getOpcode() == CastOpc)
6398 // TODO: Can handle more cases here.
6405 /// EvaluateInDifferentType - Given an expression that
6406 /// CanEvaluateInDifferentType returns true for, actually insert the code to
6407 /// evaluate the expression.
6408 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
6410 if (Constant *C = dyn_cast<Constant>(V))
6411 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
6413 // Otherwise, it must be an instruction.
6414 Instruction *I = cast<Instruction>(V);
6415 Instruction *Res = 0;
6416 switch (I->getOpcode()) {
6417 case Instruction::Add:
6418 case Instruction::Sub:
6419 case Instruction::And:
6420 case Instruction::Or:
6421 case Instruction::Xor:
6422 case Instruction::AShr:
6423 case Instruction::LShr:
6424 case Instruction::Shl: {
6425 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
6426 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
6427 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
6428 LHS, RHS, I->getName());
6431 case Instruction::Trunc:
6432 case Instruction::ZExt:
6433 case Instruction::SExt:
6434 // If the source type of the cast is the type we're trying for then we can
6435 // just return the source. There's no need to insert it because it is not
6437 if (I->getOperand(0)->getType() == Ty)
6438 return I->getOperand(0);
6440 // Otherwise, must be the same type of case, so just reinsert a new one.
6441 Res = CastInst::create(cast<CastInst>(I)->getOpcode(), I->getOperand(0),
6445 // TODO: Can handle more cases here.
6446 assert(0 && "Unreachable!");
6450 return InsertNewInstBefore(Res, *I);
6453 /// @brief Implement the transforms common to all CastInst visitors.
6454 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
6455 Value *Src = CI.getOperand(0);
6457 // Many cases of "cast of a cast" are eliminable. If it's eliminable we just
6458 // eliminate it now.
6459 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6460 if (Instruction::CastOps opc =
6461 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
6462 // The first cast (CSrc) is eliminable so we need to fix up or replace
6463 // the second cast (CI). CSrc will then have a good chance of being dead.
6464 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
6468 // If we are casting a select then fold the cast into the select
6469 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
6470 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
6473 // If we are casting a PHI then fold the cast into the PHI
6474 if (isa<PHINode>(Src))
6475 if (Instruction *NV = FoldOpIntoPhi(CI))
6481 /// @brief Implement the transforms for cast of pointer (bitcast/ptrtoint)
6482 Instruction *InstCombiner::commonPointerCastTransforms(CastInst &CI) {
6483 Value *Src = CI.getOperand(0);
6485 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
6486 // If casting the result of a getelementptr instruction with no offset, turn
6487 // this into a cast of the original pointer!
6488 if (GEP->hasAllZeroIndices()) {
6489 // Changing the cast operand is usually not a good idea but it is safe
6490 // here because the pointer operand is being replaced with another
6491 // pointer operand so the opcode doesn't need to change.
6493 CI.setOperand(0, GEP->getOperand(0));
6497 // If the GEP has a single use, and the base pointer is a bitcast, and the
6498 // GEP computes a constant offset, see if we can convert these three
6499 // instructions into fewer. This typically happens with unions and other
6500 // non-type-safe code.
6501 if (GEP->hasOneUse() && isa<BitCastInst>(GEP->getOperand(0))) {
6502 if (GEP->hasAllConstantIndices()) {
6503 // We are guaranteed to get a constant from EmitGEPOffset.
6504 ConstantInt *OffsetV = cast<ConstantInt>(EmitGEPOffset(GEP, CI, *this));
6505 int64_t Offset = OffsetV->getSExtValue();
6507 // Get the base pointer input of the bitcast, and the type it points to.
6508 Value *OrigBase = cast<BitCastInst>(GEP->getOperand(0))->getOperand(0);
6509 const Type *GEPIdxTy =
6510 cast<PointerType>(OrigBase->getType())->getElementType();
6511 if (GEPIdxTy->isSized()) {
6512 SmallVector<Value*, 8> NewIndices;
6514 // Start with the index over the outer type. Note that the type size
6515 // might be zero (even if the offset isn't zero) if the indexed type
6516 // is something like [0 x {int, int}]
6517 const Type *IntPtrTy = TD->getIntPtrType();
6518 int64_t FirstIdx = 0;
6519 if (int64_t TySize = TD->getTypeSize(GEPIdxTy)) {
6520 FirstIdx = Offset/TySize;
6523 // Handle silly modulus not returning values values [0..TySize).
6527 assert(Offset >= 0);
6529 assert((uint64_t)Offset < (uint64_t)TySize &&"Out of range offset");
6532 NewIndices.push_back(ConstantInt::get(IntPtrTy, FirstIdx));
6534 // Index into the types. If we fail, set OrigBase to null.
6536 if (const StructType *STy = dyn_cast<StructType>(GEPIdxTy)) {
6537 const StructLayout *SL = TD->getStructLayout(STy);
6538 if (Offset < (int64_t)SL->getSizeInBytes()) {
6539 unsigned Elt = SL->getElementContainingOffset(Offset);
6540 NewIndices.push_back(ConstantInt::get(Type::Int32Ty, Elt));
6542 Offset -= SL->getElementOffset(Elt);
6543 GEPIdxTy = STy->getElementType(Elt);
6545 // Otherwise, we can't index into this, bail out.
6549 } else if (isa<ArrayType>(GEPIdxTy) || isa<VectorType>(GEPIdxTy)) {
6550 const SequentialType *STy = cast<SequentialType>(GEPIdxTy);
6551 if (uint64_t EltSize = TD->getTypeSize(STy->getElementType())) {
6552 NewIndices.push_back(ConstantInt::get(IntPtrTy,Offset/EltSize));
6555 NewIndices.push_back(ConstantInt::get(IntPtrTy, 0));
6557 GEPIdxTy = STy->getElementType();
6559 // Otherwise, we can't index into this, bail out.
6565 // If we were able to index down into an element, create the GEP
6566 // and bitcast the result. This eliminates one bitcast, potentially
6568 Instruction *NGEP = new GetElementPtrInst(OrigBase,
6570 NewIndices.end(), "");
6571 InsertNewInstBefore(NGEP, CI);
6572 NGEP->takeName(GEP);
6574 if (isa<BitCastInst>(CI))
6575 return new BitCastInst(NGEP, CI.getType());
6576 assert(isa<PtrToIntInst>(CI));
6577 return new PtrToIntInst(NGEP, CI.getType());
6584 return commonCastTransforms(CI);
6589 /// Only the TRUNC, ZEXT, SEXT, and BITCAST can both operand and result as
6590 /// integer types. This function implements the common transforms for all those
6592 /// @brief Implement the transforms common to CastInst with integer operands
6593 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
6594 if (Instruction *Result = commonCastTransforms(CI))
6597 Value *Src = CI.getOperand(0);
6598 const Type *SrcTy = Src->getType();
6599 const Type *DestTy = CI.getType();
6600 uint32_t SrcBitSize = SrcTy->getPrimitiveSizeInBits();
6601 uint32_t DestBitSize = DestTy->getPrimitiveSizeInBits();
6603 // See if we can simplify any instructions used by the LHS whose sole
6604 // purpose is to compute bits we don't care about.
6605 APInt KnownZero(DestBitSize, 0), KnownOne(DestBitSize, 0);
6606 if (SimplifyDemandedBits(&CI, APInt::getAllOnesValue(DestBitSize),
6607 KnownZero, KnownOne))
6610 // If the source isn't an instruction or has more than one use then we
6611 // can't do anything more.
6612 Instruction *SrcI = dyn_cast<Instruction>(Src);
6613 if (!SrcI || !Src->hasOneUse())
6616 // Attempt to propagate the cast into the instruction for int->int casts.
6617 int NumCastsRemoved = 0;
6618 if (!isa<BitCastInst>(CI) &&
6619 CanEvaluateInDifferentType(SrcI, cast<IntegerType>(DestTy),
6620 CI.getOpcode(), NumCastsRemoved)) {
6621 // If this cast is a truncate, evaluting in a different type always
6622 // eliminates the cast, so it is always a win. If this is a zero-extension,
6623 // we need to do an AND to maintain the clear top-part of the computation,
6624 // so we require that the input have eliminated at least one cast. If this
6625 // is a sign extension, we insert two new casts (to do the extension) so we
6626 // require that two casts have been eliminated.
6628 switch (CI.getOpcode()) {
6630 // All the others use floating point so we shouldn't actually
6631 // get here because of the check above.
6632 assert(0 && "Unknown cast type");
6633 case Instruction::Trunc:
6636 case Instruction::ZExt:
6637 DoXForm = NumCastsRemoved >= 1;
6639 case Instruction::SExt:
6640 DoXForm = NumCastsRemoved >= 2;
6645 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
6646 CI.getOpcode() == Instruction::SExt);
6647 assert(Res->getType() == DestTy);
6648 switch (CI.getOpcode()) {
6649 default: assert(0 && "Unknown cast type!");
6650 case Instruction::Trunc:
6651 case Instruction::BitCast:
6652 // Just replace this cast with the result.
6653 return ReplaceInstUsesWith(CI, Res);
6654 case Instruction::ZExt: {
6655 // We need to emit an AND to clear the high bits.
6656 assert(SrcBitSize < DestBitSize && "Not a zext?");
6657 Constant *C = ConstantInt::get(APInt::getLowBitsSet(DestBitSize,
6659 return BinaryOperator::createAnd(Res, C);
6661 case Instruction::SExt:
6662 // We need to emit a cast to truncate, then a cast to sext.
6663 return CastInst::create(Instruction::SExt,
6664 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
6670 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
6671 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
6673 switch (SrcI->getOpcode()) {
6674 case Instruction::Add:
6675 case Instruction::Mul:
6676 case Instruction::And:
6677 case Instruction::Or:
6678 case Instruction::Xor:
6679 // If we are discarding information, rewrite.
6680 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
6681 // Don't insert two casts if they cannot be eliminated. We allow
6682 // two casts to be inserted if the sizes are the same. This could
6683 // only be converting signedness, which is a noop.
6684 if (DestBitSize == SrcBitSize ||
6685 !ValueRequiresCast(CI.getOpcode(), Op1, DestTy,TD) ||
6686 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6687 Instruction::CastOps opcode = CI.getOpcode();
6688 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6689 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6690 return BinaryOperator::create(
6691 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6695 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
6696 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
6697 SrcI->getOpcode() == Instruction::Xor &&
6698 Op1 == ConstantInt::getTrue() &&
6699 (!Op0->hasOneUse() || !isa<CmpInst>(Op0))) {
6700 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
6701 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
6704 case Instruction::SDiv:
6705 case Instruction::UDiv:
6706 case Instruction::SRem:
6707 case Instruction::URem:
6708 // If we are just changing the sign, rewrite.
6709 if (DestBitSize == SrcBitSize) {
6710 // Don't insert two casts if they cannot be eliminated. We allow
6711 // two casts to be inserted if the sizes are the same. This could
6712 // only be converting signedness, which is a noop.
6713 if (!ValueRequiresCast(CI.getOpcode(), Op1, DestTy, TD) ||
6714 !ValueRequiresCast(CI.getOpcode(), Op0, DestTy, TD)) {
6715 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
6717 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
6719 return BinaryOperator::create(
6720 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
6725 case Instruction::Shl:
6726 // Allow changing the sign of the source operand. Do not allow
6727 // changing the size of the shift, UNLESS the shift amount is a
6728 // constant. We must not change variable sized shifts to a smaller
6729 // size, because it is undefined to shift more bits out than exist
6731 if (DestBitSize == SrcBitSize ||
6732 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
6733 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
6734 Instruction::BitCast : Instruction::Trunc);
6735 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
6736 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
6737 return BinaryOperator::createShl(Op0c, Op1c);
6740 case Instruction::AShr:
6741 // If this is a signed shr, and if all bits shifted in are about to be
6742 // truncated off, turn it into an unsigned shr to allow greater
6744 if (DestBitSize < SrcBitSize &&
6745 isa<ConstantInt>(Op1)) {
6746 uint32_t ShiftAmt = cast<ConstantInt>(Op1)->getLimitedValue(SrcBitSize);
6747 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
6748 // Insert the new logical shift right.
6749 return BinaryOperator::createLShr(Op0, Op1);
6757 Instruction *InstCombiner::visitTrunc(TruncInst &CI) {
6758 if (Instruction *Result = commonIntCastTransforms(CI))
6761 Value *Src = CI.getOperand(0);
6762 const Type *Ty = CI.getType();
6763 uint32_t DestBitWidth = Ty->getPrimitiveSizeInBits();
6764 uint32_t SrcBitWidth = cast<IntegerType>(Src->getType())->getBitWidth();
6766 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6767 switch (SrcI->getOpcode()) {
6769 case Instruction::LShr:
6770 // We can shrink lshr to something smaller if we know the bits shifted in
6771 // are already zeros.
6772 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6773 uint32_t ShAmt = ShAmtV->getLimitedValue(SrcBitWidth);
6775 // Get a mask for the bits shifting in.
6776 APInt Mask(APInt::getLowBitsSet(SrcBitWidth, ShAmt).shl(DestBitWidth));
6777 Value* SrcIOp0 = SrcI->getOperand(0);
6778 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6779 if (ShAmt >= DestBitWidth) // All zeros.
6780 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6782 // Okay, we can shrink this. Truncate the input, then return a new
6784 Value *V1 = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6785 Value *V2 = InsertCastBefore(Instruction::Trunc, SrcI->getOperand(1),
6787 return BinaryOperator::createLShr(V1, V2);
6789 } else { // This is a variable shr.
6791 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6792 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6793 // loop-invariant and CSE'd.
6794 if (CI.getType() == Type::Int1Ty && SrcI->hasOneUse()) {
6795 Value *One = ConstantInt::get(SrcI->getType(), 1);
6797 Value *V = InsertNewInstBefore(
6798 BinaryOperator::createShl(One, SrcI->getOperand(1),
6800 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6801 SrcI->getOperand(0),
6803 Value *Zero = Constant::getNullValue(V->getType());
6804 return new ICmpInst(ICmpInst::ICMP_NE, V, Zero);
6814 Instruction *InstCombiner::visitZExt(ZExtInst &CI) {
6815 // If one of the common conversion will work ..
6816 if (Instruction *Result = commonIntCastTransforms(CI))
6819 Value *Src = CI.getOperand(0);
6821 // If this is a cast of a cast
6822 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6823 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6824 // types and if the sizes are just right we can convert this into a logical
6825 // 'and' which will be much cheaper than the pair of casts.
6826 if (isa<TruncInst>(CSrc)) {
6827 // Get the sizes of the types involved
6828 Value *A = CSrc->getOperand(0);
6829 uint32_t SrcSize = A->getType()->getPrimitiveSizeInBits();
6830 uint32_t MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6831 uint32_t DstSize = CI.getType()->getPrimitiveSizeInBits();
6832 // If we're actually extending zero bits and the trunc is a no-op
6833 if (MidSize < DstSize && SrcSize == DstSize) {
6834 // Replace both of the casts with an And of the type mask.
6835 APInt AndValue(APInt::getLowBitsSet(SrcSize, MidSize));
6836 Constant *AndConst = ConstantInt::get(AndValue);
6838 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6839 // Unfortunately, if the type changed, we need to cast it back.
6840 if (And->getType() != CI.getType()) {
6841 And->setName(CSrc->getName()+".mask");
6842 InsertNewInstBefore(And, CI);
6843 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6850 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6851 // If we are just checking for a icmp eq of a single bit and zext'ing it
6852 // to an integer, then shift the bit to the appropriate place and then
6853 // cast to integer to avoid the comparison.
6854 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6855 const APInt &Op1CV = Op1C->getValue();
6857 // zext (x <s 0) to i32 --> x>>u31 true if signbit set.
6858 // zext (x >s -1) to i32 --> (x>>u31)^1 true if signbit clear.
6859 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6860 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6861 Value *In = ICI->getOperand(0);
6862 Value *Sh = ConstantInt::get(In->getType(),
6863 In->getType()->getPrimitiveSizeInBits()-1);
6864 In = InsertNewInstBefore(BinaryOperator::createLShr(In, Sh,
6865 In->getName()+".lobit"),
6867 if (In->getType() != CI.getType())
6868 In = CastInst::createIntegerCast(In, CI.getType(),
6869 false/*ZExt*/, "tmp", &CI);
6871 if (ICI->getPredicate() == ICmpInst::ICMP_SGT) {
6872 Constant *One = ConstantInt::get(In->getType(), 1);
6873 In = InsertNewInstBefore(BinaryOperator::createXor(In, One,
6874 In->getName()+".not"),
6878 return ReplaceInstUsesWith(CI, In);
6883 // zext (X == 0) to i32 --> X^1 iff X has only the low bit set.
6884 // zext (X == 0) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6885 // zext (X == 1) to i32 --> X iff X has only the low bit set.
6886 // zext (X == 2) to i32 --> X>>1 iff X has only the 2nd bit set.
6887 // zext (X != 0) to i32 --> X iff X has only the low bit set.
6888 // zext (X != 0) to i32 --> X>>1 iff X has only the 2nd bit set.
6889 // zext (X != 1) to i32 --> X^1 iff X has only the low bit set.
6890 // zext (X != 2) to i32 --> (X>>1)^1 iff X has only the 2nd bit set.
6891 if ((Op1CV == 0 || Op1CV.isPowerOf2()) &&
6892 // This only works for EQ and NE
6893 ICI->isEquality()) {
6894 // If Op1C some other power of two, convert:
6895 uint32_t BitWidth = Op1C->getType()->getBitWidth();
6896 APInt KnownZero(BitWidth, 0), KnownOne(BitWidth, 0);
6897 APInt TypeMask(APInt::getAllOnesValue(BitWidth));
6898 ComputeMaskedBits(ICI->getOperand(0), TypeMask, KnownZero, KnownOne);
6900 APInt KnownZeroMask(~KnownZero);
6901 if (KnownZeroMask.isPowerOf2()) { // Exactly 1 possible 1?
6902 bool isNE = ICI->getPredicate() == ICmpInst::ICMP_NE;
6903 if (Op1CV != 0 && (Op1CV != KnownZeroMask)) {
6904 // (X&4) == 2 --> false
6905 // (X&4) != 2 --> true
6906 Constant *Res = ConstantInt::get(Type::Int1Ty, isNE);
6907 Res = ConstantExpr::getZExt(Res, CI.getType());
6908 return ReplaceInstUsesWith(CI, Res);
6911 uint32_t ShiftAmt = KnownZeroMask.logBase2();
6912 Value *In = ICI->getOperand(0);
6914 // Perform a logical shr by shiftamt.
6915 // Insert the shift to put the result in the low bit.
6916 In = InsertNewInstBefore(
6917 BinaryOperator::createLShr(In,
6918 ConstantInt::get(In->getType(), ShiftAmt),
6919 In->getName()+".lobit"), CI);
6922 if ((Op1CV != 0) == isNE) { // Toggle the low bit.
6923 Constant *One = ConstantInt::get(In->getType(), 1);
6924 In = BinaryOperator::createXor(In, One, "tmp");
6925 InsertNewInstBefore(cast<Instruction>(In), CI);
6928 if (CI.getType() == In->getType())
6929 return ReplaceInstUsesWith(CI, In);
6931 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
6939 Instruction *InstCombiner::visitSExt(SExtInst &CI) {
6940 if (Instruction *I = commonIntCastTransforms(CI))
6943 Value *Src = CI.getOperand(0);
6945 // sext (x <s 0) -> ashr x, 31 -> all ones if signed
6946 // sext (x >s -1) -> ashr x, 31 -> all ones if not signed
6947 if (ICmpInst *ICI = dyn_cast<ICmpInst>(Src)) {
6948 // If we are just checking for a icmp eq of a single bit and zext'ing it
6949 // to an integer, then shift the bit to the appropriate place and then
6950 // cast to integer to avoid the comparison.
6951 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(ICI->getOperand(1))) {
6952 const APInt &Op1CV = Op1C->getValue();
6954 // sext (x <s 0) to i32 --> x>>s31 true if signbit set.
6955 // sext (x >s -1) to i32 --> (x>>s31)^-1 true if signbit clear.
6956 if ((ICI->getPredicate() == ICmpInst::ICMP_SLT && Op1CV == 0) ||
6957 (ICI->getPredicate() == ICmpInst::ICMP_SGT &&Op1CV.isAllOnesValue())){
6958 Value *In = ICI->getOperand(0);
6959 Value *Sh = ConstantInt::get(In->getType(),
6960 In->getType()->getPrimitiveSizeInBits()-1);
6961 In = InsertNewInstBefore(BinaryOperator::createAShr(In, Sh,
6962 In->getName()+".lobit"),
6964 if (In->getType() != CI.getType())
6965 In = CastInst::createIntegerCast(In, CI.getType(),
6966 true/*SExt*/, "tmp", &CI);
6968 if (ICI->getPredicate() == ICmpInst::ICMP_SGT)
6969 In = InsertNewInstBefore(BinaryOperator::createNot(In,
6970 In->getName()+".not"), CI);
6972 return ReplaceInstUsesWith(CI, In);
6980 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6981 return commonCastTransforms(CI);
6984 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6985 return commonCastTransforms(CI);
6988 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6989 return commonCastTransforms(CI);
6992 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6993 return commonCastTransforms(CI);
6996 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6997 return commonCastTransforms(CI);
7000 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
7001 return commonCastTransforms(CI);
7004 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
7005 return commonPointerCastTransforms(CI);
7008 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
7009 return commonCastTransforms(CI);
7012 Instruction *InstCombiner::visitBitCast(BitCastInst &CI) {
7013 // If the operands are integer typed then apply the integer transforms,
7014 // otherwise just apply the common ones.
7015 Value *Src = CI.getOperand(0);
7016 const Type *SrcTy = Src->getType();
7017 const Type *DestTy = CI.getType();
7019 if (SrcTy->isInteger() && DestTy->isInteger()) {
7020 if (Instruction *Result = commonIntCastTransforms(CI))
7022 } else if (isa<PointerType>(SrcTy)) {
7023 if (Instruction *I = commonPointerCastTransforms(CI))
7026 if (Instruction *Result = commonCastTransforms(CI))
7031 // Get rid of casts from one type to the same type. These are useless and can
7032 // be replaced by the operand.
7033 if (DestTy == Src->getType())
7034 return ReplaceInstUsesWith(CI, Src);
7036 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
7037 const PointerType *SrcPTy = cast<PointerType>(SrcTy);
7038 const Type *DstElTy = DstPTy->getElementType();
7039 const Type *SrcElTy = SrcPTy->getElementType();
7041 // If we are casting a malloc or alloca to a pointer to a type of the same
7042 // size, rewrite the allocation instruction to allocate the "right" type.
7043 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
7044 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
7047 // If the source and destination are pointers, and this cast is equivalent
7048 // to a getelementptr X, 0, 0, 0... turn it into the appropriate gep.
7049 // This can enhance SROA and other transforms that want type-safe pointers.
7050 Constant *ZeroUInt = Constant::getNullValue(Type::Int32Ty);
7051 unsigned NumZeros = 0;
7052 while (SrcElTy != DstElTy &&
7053 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
7054 SrcElTy->getNumContainedTypes() /* not "{}" */) {
7055 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
7059 // If we found a path from the src to dest, create the getelementptr now.
7060 if (SrcElTy == DstElTy) {
7061 SmallVector<Value*, 8> Idxs(NumZeros+1, ZeroUInt);
7062 return new GetElementPtrInst(Src, Idxs.begin(), Idxs.end(), "",
7063 ((Instruction*) NULL));
7067 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
7068 if (SVI->hasOneUse()) {
7069 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
7070 // a bitconvert to a vector with the same # elts.
7071 if (isa<VectorType>(DestTy) &&
7072 cast<VectorType>(DestTy)->getNumElements() ==
7073 SVI->getType()->getNumElements()) {
7075 // If either of the operands is a cast from CI.getType(), then
7076 // evaluating the shuffle in the casted destination's type will allow
7077 // us to eliminate at least one cast.
7078 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
7079 Tmp->getOperand(0)->getType() == DestTy) ||
7080 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
7081 Tmp->getOperand(0)->getType() == DestTy)) {
7082 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
7083 SVI->getOperand(0), DestTy, &CI);
7084 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
7085 SVI->getOperand(1), DestTy, &CI);
7086 // Return a new shuffle vector. Use the same element ID's, as we
7087 // know the vector types match #elts.
7088 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
7096 /// GetSelectFoldableOperands - We want to turn code that looks like this:
7098 /// %D = select %cond, %C, %A
7100 /// %C = select %cond, %B, 0
7103 /// Assuming that the specified instruction is an operand to the select, return
7104 /// a bitmask indicating which operands of this instruction are foldable if they
7105 /// equal the other incoming value of the select.
7107 static unsigned GetSelectFoldableOperands(Instruction *I) {
7108 switch (I->getOpcode()) {
7109 case Instruction::Add:
7110 case Instruction::Mul:
7111 case Instruction::And:
7112 case Instruction::Or:
7113 case Instruction::Xor:
7114 return 3; // Can fold through either operand.
7115 case Instruction::Sub: // Can only fold on the amount subtracted.
7116 case Instruction::Shl: // Can only fold on the shift amount.
7117 case Instruction::LShr:
7118 case Instruction::AShr:
7121 return 0; // Cannot fold
7125 /// GetSelectFoldableConstant - For the same transformation as the previous
7126 /// function, return the identity constant that goes into the select.
7127 static Constant *GetSelectFoldableConstant(Instruction *I) {
7128 switch (I->getOpcode()) {
7129 default: assert(0 && "This cannot happen!"); abort();
7130 case Instruction::Add:
7131 case Instruction::Sub:
7132 case Instruction::Or:
7133 case Instruction::Xor:
7134 case Instruction::Shl:
7135 case Instruction::LShr:
7136 case Instruction::AShr:
7137 return Constant::getNullValue(I->getType());
7138 case Instruction::And:
7139 return Constant::getAllOnesValue(I->getType());
7140 case Instruction::Mul:
7141 return ConstantInt::get(I->getType(), 1);
7145 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
7146 /// have the same opcode and only one use each. Try to simplify this.
7147 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
7149 if (TI->getNumOperands() == 1) {
7150 // If this is a non-volatile load or a cast from the same type,
7153 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
7156 return 0; // unknown unary op.
7159 // Fold this by inserting a select from the input values.
7160 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
7161 FI->getOperand(0), SI.getName()+".v");
7162 InsertNewInstBefore(NewSI, SI);
7163 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
7167 // Only handle binary operators here.
7168 if (!isa<BinaryOperator>(TI))
7171 // Figure out if the operations have any operands in common.
7172 Value *MatchOp, *OtherOpT, *OtherOpF;
7174 if (TI->getOperand(0) == FI->getOperand(0)) {
7175 MatchOp = TI->getOperand(0);
7176 OtherOpT = TI->getOperand(1);
7177 OtherOpF = FI->getOperand(1);
7178 MatchIsOpZero = true;
7179 } else if (TI->getOperand(1) == FI->getOperand(1)) {
7180 MatchOp = TI->getOperand(1);
7181 OtherOpT = TI->getOperand(0);
7182 OtherOpF = FI->getOperand(0);
7183 MatchIsOpZero = false;
7184 } else if (!TI->isCommutative()) {
7186 } else if (TI->getOperand(0) == FI->getOperand(1)) {
7187 MatchOp = TI->getOperand(0);
7188 OtherOpT = TI->getOperand(1);
7189 OtherOpF = FI->getOperand(0);
7190 MatchIsOpZero = true;
7191 } else if (TI->getOperand(1) == FI->getOperand(0)) {
7192 MatchOp = TI->getOperand(1);
7193 OtherOpT = TI->getOperand(0);
7194 OtherOpF = FI->getOperand(1);
7195 MatchIsOpZero = true;
7200 // If we reach here, they do have operations in common.
7201 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
7202 OtherOpF, SI.getName()+".v");
7203 InsertNewInstBefore(NewSI, SI);
7205 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
7207 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
7209 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
7211 assert(0 && "Shouldn't get here");
7215 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
7216 Value *CondVal = SI.getCondition();
7217 Value *TrueVal = SI.getTrueValue();
7218 Value *FalseVal = SI.getFalseValue();
7220 // select true, X, Y -> X
7221 // select false, X, Y -> Y
7222 if (ConstantInt *C = dyn_cast<ConstantInt>(CondVal))
7223 return ReplaceInstUsesWith(SI, C->getZExtValue() ? TrueVal : FalseVal);
7225 // select C, X, X -> X
7226 if (TrueVal == FalseVal)
7227 return ReplaceInstUsesWith(SI, TrueVal);
7229 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
7230 return ReplaceInstUsesWith(SI, FalseVal);
7231 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
7232 return ReplaceInstUsesWith(SI, TrueVal);
7233 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
7234 if (isa<Constant>(TrueVal))
7235 return ReplaceInstUsesWith(SI, TrueVal);
7237 return ReplaceInstUsesWith(SI, FalseVal);
7240 if (SI.getType() == Type::Int1Ty) {
7241 if (ConstantInt *C = dyn_cast<ConstantInt>(TrueVal)) {
7242 if (C->getZExtValue()) {
7243 // Change: A = select B, true, C --> A = or B, C
7244 return BinaryOperator::createOr(CondVal, FalseVal);
7246 // Change: A = select B, false, C --> A = and !B, C
7248 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7249 "not."+CondVal->getName()), SI);
7250 return BinaryOperator::createAnd(NotCond, FalseVal);
7252 } else if (ConstantInt *C = dyn_cast<ConstantInt>(FalseVal)) {
7253 if (C->getZExtValue() == false) {
7254 // Change: A = select B, C, false --> A = and B, C
7255 return BinaryOperator::createAnd(CondVal, TrueVal);
7257 // Change: A = select B, C, true --> A = or !B, C
7259 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7260 "not."+CondVal->getName()), SI);
7261 return BinaryOperator::createOr(NotCond, TrueVal);
7266 // Selecting between two integer constants?
7267 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
7268 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
7269 // select C, 1, 0 -> zext C to int
7270 if (FalseValC->isZero() && TrueValC->getValue() == 1) {
7271 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
7272 } else if (TrueValC->isZero() && FalseValC->getValue() == 1) {
7273 // select C, 0, 1 -> zext !C to int
7275 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
7276 "not."+CondVal->getName()), SI);
7277 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
7280 // FIXME: Turn select 0/-1 and -1/0 into sext from condition!
7282 if (ICmpInst *IC = dyn_cast<ICmpInst>(SI.getCondition())) {
7284 // (x <s 0) ? -1 : 0 -> ashr x, 31
7285 if (TrueValC->isAllOnesValue() && FalseValC->isZero())
7286 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
7287 if (IC->getPredicate() == ICmpInst::ICMP_SLT && CmpCst->isZero()) {
7288 // The comparison constant and the result are not neccessarily the
7289 // same width. Make an all-ones value by inserting a AShr.
7290 Value *X = IC->getOperand(0);
7291 uint32_t Bits = X->getType()->getPrimitiveSizeInBits();
7292 Constant *ShAmt = ConstantInt::get(X->getType(), Bits-1);
7293 Instruction *SRA = BinaryOperator::create(Instruction::AShr, X,
7295 InsertNewInstBefore(SRA, SI);
7297 // Finally, convert to the type of the select RHS. We figure out
7298 // if this requires a SExt, Trunc or BitCast based on the sizes.
7299 Instruction::CastOps opc = Instruction::BitCast;
7300 uint32_t SRASize = SRA->getType()->getPrimitiveSizeInBits();
7301 uint32_t SISize = SI.getType()->getPrimitiveSizeInBits();
7302 if (SRASize < SISize)
7303 opc = Instruction::SExt;
7304 else if (SRASize > SISize)
7305 opc = Instruction::Trunc;
7306 return CastInst::create(opc, SRA, SI.getType());
7311 // If one of the constants is zero (we know they can't both be) and we
7312 // have an icmp instruction with zero, and we have an 'and' with the
7313 // non-constant value, eliminate this whole mess. This corresponds to
7314 // cases like this: ((X & 27) ? 27 : 0)
7315 if (TrueValC->isZero() || FalseValC->isZero())
7316 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
7317 cast<Constant>(IC->getOperand(1))->isNullValue())
7318 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
7319 if (ICA->getOpcode() == Instruction::And &&
7320 isa<ConstantInt>(ICA->getOperand(1)) &&
7321 (ICA->getOperand(1) == TrueValC ||
7322 ICA->getOperand(1) == FalseValC) &&
7323 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
7324 // Okay, now we know that everything is set up, we just don't
7325 // know whether we have a icmp_ne or icmp_eq and whether the
7326 // true or false val is the zero.
7327 bool ShouldNotVal = !TrueValC->isZero();
7328 ShouldNotVal ^= IC->getPredicate() == ICmpInst::ICMP_NE;
7331 V = InsertNewInstBefore(BinaryOperator::create(
7332 Instruction::Xor, V, ICA->getOperand(1)), SI);
7333 return ReplaceInstUsesWith(SI, V);
7338 // See if we are selecting two values based on a comparison of the two values.
7339 if (FCmpInst *FCI = dyn_cast<FCmpInst>(CondVal)) {
7340 if (FCI->getOperand(0) == TrueVal && FCI->getOperand(1) == FalseVal) {
7341 // Transform (X == Y) ? X : Y -> Y
7342 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7343 return ReplaceInstUsesWith(SI, FalseVal);
7344 // Transform (X != Y) ? X : Y -> X
7345 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7346 return ReplaceInstUsesWith(SI, TrueVal);
7347 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7349 } else if (FCI->getOperand(0) == FalseVal && FCI->getOperand(1) == TrueVal){
7350 // Transform (X == Y) ? Y : X -> X
7351 if (FCI->getPredicate() == FCmpInst::FCMP_OEQ)
7352 return ReplaceInstUsesWith(SI, FalseVal);
7353 // Transform (X != Y) ? Y : X -> Y
7354 if (FCI->getPredicate() == FCmpInst::FCMP_ONE)
7355 return ReplaceInstUsesWith(SI, TrueVal);
7356 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7360 // See if we are selecting two values based on a comparison of the two values.
7361 if (ICmpInst *ICI = dyn_cast<ICmpInst>(CondVal)) {
7362 if (ICI->getOperand(0) == TrueVal && ICI->getOperand(1) == FalseVal) {
7363 // Transform (X == Y) ? X : Y -> Y
7364 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7365 return ReplaceInstUsesWith(SI, FalseVal);
7366 // Transform (X != Y) ? X : Y -> X
7367 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7368 return ReplaceInstUsesWith(SI, TrueVal);
7369 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7371 } else if (ICI->getOperand(0) == FalseVal && ICI->getOperand(1) == TrueVal){
7372 // Transform (X == Y) ? Y : X -> X
7373 if (ICI->getPredicate() == ICmpInst::ICMP_EQ)
7374 return ReplaceInstUsesWith(SI, FalseVal);
7375 // Transform (X != Y) ? Y : X -> Y
7376 if (ICI->getPredicate() == ICmpInst::ICMP_NE)
7377 return ReplaceInstUsesWith(SI, TrueVal);
7378 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
7382 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
7383 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
7384 if (TI->hasOneUse() && FI->hasOneUse()) {
7385 Instruction *AddOp = 0, *SubOp = 0;
7387 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
7388 if (TI->getOpcode() == FI->getOpcode())
7389 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
7392 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
7393 // even legal for FP.
7394 if (TI->getOpcode() == Instruction::Sub &&
7395 FI->getOpcode() == Instruction::Add) {
7396 AddOp = FI; SubOp = TI;
7397 } else if (FI->getOpcode() == Instruction::Sub &&
7398 TI->getOpcode() == Instruction::Add) {
7399 AddOp = TI; SubOp = FI;
7403 Value *OtherAddOp = 0;
7404 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
7405 OtherAddOp = AddOp->getOperand(1);
7406 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
7407 OtherAddOp = AddOp->getOperand(0);
7411 // So at this point we know we have (Y -> OtherAddOp):
7412 // select C, (add X, Y), (sub X, Z)
7413 Value *NegVal; // Compute -Z
7414 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
7415 NegVal = ConstantExpr::getNeg(C);
7417 NegVal = InsertNewInstBefore(
7418 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
7421 Value *NewTrueOp = OtherAddOp;
7422 Value *NewFalseOp = NegVal;
7424 std::swap(NewTrueOp, NewFalseOp);
7425 Instruction *NewSel =
7426 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
7428 NewSel = InsertNewInstBefore(NewSel, SI);
7429 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
7434 // See if we can fold the select into one of our operands.
7435 if (SI.getType()->isInteger()) {
7436 // See the comment above GetSelectFoldableOperands for a description of the
7437 // transformation we are doing here.
7438 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
7439 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
7440 !isa<Constant>(FalseVal))
7441 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
7442 unsigned OpToFold = 0;
7443 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
7445 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
7450 Constant *C = GetSelectFoldableConstant(TVI);
7451 Instruction *NewSel =
7452 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C);
7453 InsertNewInstBefore(NewSel, SI);
7454 NewSel->takeName(TVI);
7455 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
7456 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
7458 assert(0 && "Unknown instruction!!");
7463 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
7464 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
7465 !isa<Constant>(TrueVal))
7466 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
7467 unsigned OpToFold = 0;
7468 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
7470 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
7475 Constant *C = GetSelectFoldableConstant(FVI);
7476 Instruction *NewSel =
7477 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold));
7478 InsertNewInstBefore(NewSel, SI);
7479 NewSel->takeName(FVI);
7480 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
7481 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
7483 assert(0 && "Unknown instruction!!");
7488 if (BinaryOperator::isNot(CondVal)) {
7489 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
7490 SI.setOperand(1, FalseVal);
7491 SI.setOperand(2, TrueVal);
7498 /// GetOrEnforceKnownAlignment - If the specified pointer has an alignment that
7499 /// we can determine, return it, otherwise return 0. If PrefAlign is specified,
7500 /// and it is more than the alignment of the ultimate object, see if we can
7501 /// increase the alignment of the ultimate object, making this check succeed.
7502 static unsigned GetOrEnforceKnownAlignment(Value *V, TargetData *TD,
7503 unsigned PrefAlign = 0) {
7504 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
7505 unsigned Align = GV->getAlignment();
7506 if (Align == 0 && TD)
7507 Align = TD->getPrefTypeAlignment(GV->getType()->getElementType());
7509 // If there is a large requested alignment and we can, bump up the alignment
7511 if (PrefAlign > Align && GV->hasInitializer()) {
7512 GV->setAlignment(PrefAlign);
7516 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
7517 unsigned Align = AI->getAlignment();
7518 if (Align == 0 && TD) {
7519 if (isa<AllocaInst>(AI))
7520 Align = TD->getPrefTypeAlignment(AI->getType()->getElementType());
7521 else if (isa<MallocInst>(AI)) {
7522 // Malloc returns maximally aligned memory.
7523 Align = TD->getABITypeAlignment(AI->getType()->getElementType());
7526 (unsigned)TD->getABITypeAlignment(Type::DoubleTy));
7529 (unsigned)TD->getABITypeAlignment(Type::Int64Ty));
7533 // If there is a requested alignment and if this is an alloca, round up. We
7534 // don't do this for malloc, because some systems can't respect the request.
7535 if (PrefAlign > Align && isa<AllocaInst>(AI)) {
7536 AI->setAlignment(PrefAlign);
7540 } else if (isa<BitCastInst>(V) ||
7541 (isa<ConstantExpr>(V) &&
7542 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
7543 return GetOrEnforceKnownAlignment(cast<User>(V)->getOperand(0),
7545 } else if (User *GEPI = dyn_castGetElementPtr(V)) {
7546 // If all indexes are zero, it is just the alignment of the base pointer.
7547 bool AllZeroOperands = true;
7548 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
7549 if (!isa<Constant>(GEPI->getOperand(i)) ||
7550 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
7551 AllZeroOperands = false;
7555 if (AllZeroOperands) {
7556 // Treat this like a bitcast.
7557 return GetOrEnforceKnownAlignment(GEPI->getOperand(0), TD, PrefAlign);
7560 unsigned BaseAlignment = GetOrEnforceKnownAlignment(GEPI->getOperand(0),TD);
7561 if (BaseAlignment == 0) return 0;
7563 // Otherwise, if the base alignment is >= the alignment we expect for the
7564 // base pointer type, then we know that the resultant pointer is aligned at
7565 // least as much as its type requires.
7568 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
7569 const PointerType *PtrTy = cast<PointerType>(BasePtrTy);
7570 unsigned Align = TD->getABITypeAlignment(PtrTy->getElementType());
7571 if (Align <= BaseAlignment) {
7572 const Type *GEPTy = GEPI->getType();
7573 const PointerType *GEPPtrTy = cast<PointerType>(GEPTy);
7574 Align = std::min(Align, (unsigned)
7575 TD->getABITypeAlignment(GEPPtrTy->getElementType()));
7584 /// visitCallInst - CallInst simplification. This mostly only handles folding
7585 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
7586 /// the heavy lifting.
7588 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
7589 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
7590 if (!II) return visitCallSite(&CI);
7592 // Intrinsics cannot occur in an invoke, so handle them here instead of in
7594 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
7595 bool Changed = false;
7597 // memmove/cpy/set of zero bytes is a noop.
7598 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
7599 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
7601 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
7602 if (CI->getZExtValue() == 1) {
7603 // Replace the instruction with just byte operations. We would
7604 // transform other cases to loads/stores, but we don't know if
7605 // alignment is sufficient.
7609 // If we have a memmove and the source operation is a constant global,
7610 // then the source and dest pointers can't alias, so we can change this
7611 // into a call to memcpy.
7612 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
7613 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
7614 if (GVSrc->isConstant()) {
7615 Module *M = CI.getParent()->getParent()->getParent();
7617 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
7619 Name = "llvm.memcpy.i32";
7621 Name = "llvm.memcpy.i64";
7622 Constant *MemCpy = M->getOrInsertFunction(Name,
7623 CI.getCalledFunction()->getFunctionType());
7624 CI.setOperand(0, MemCpy);
7629 // If we can determine a pointer alignment that is bigger than currently
7630 // set, update the alignment.
7631 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
7632 unsigned Alignment1 = GetOrEnforceKnownAlignment(MI->getOperand(1), TD);
7633 unsigned Alignment2 = GetOrEnforceKnownAlignment(MI->getOperand(2), TD);
7634 unsigned Align = std::min(Alignment1, Alignment2);
7635 if (MI->getAlignment()->getZExtValue() < Align) {
7636 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Align));
7639 } else if (isa<MemSetInst>(MI)) {
7640 unsigned Alignment = GetOrEnforceKnownAlignment(MI->getDest(), TD);
7641 if (MI->getAlignment()->getZExtValue() < Alignment) {
7642 MI->setAlignment(ConstantInt::get(Type::Int32Ty, Alignment));
7647 if (Changed) return II;
7649 switch (II->getIntrinsicID()) {
7651 case Intrinsic::ppc_altivec_lvx:
7652 case Intrinsic::ppc_altivec_lvxl:
7653 case Intrinsic::x86_sse_loadu_ps:
7654 case Intrinsic::x86_sse2_loadu_pd:
7655 case Intrinsic::x86_sse2_loadu_dq:
7656 // Turn PPC lvx -> load if the pointer is known aligned.
7657 // Turn X86 loadups -> load if the pointer is known aligned.
7658 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7659 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7660 PointerType::get(II->getType()), CI);
7661 return new LoadInst(Ptr);
7664 case Intrinsic::ppc_altivec_stvx:
7665 case Intrinsic::ppc_altivec_stvxl:
7666 // Turn stvx -> store if the pointer is known aligned.
7667 if (GetOrEnforceKnownAlignment(II->getOperand(2), TD, 16) >= 16) {
7668 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
7669 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
7671 return new StoreInst(II->getOperand(1), Ptr);
7674 case Intrinsic::x86_sse_storeu_ps:
7675 case Intrinsic::x86_sse2_storeu_pd:
7676 case Intrinsic::x86_sse2_storeu_dq:
7677 case Intrinsic::x86_sse2_storel_dq:
7678 // Turn X86 storeu -> store if the pointer is known aligned.
7679 if (GetOrEnforceKnownAlignment(II->getOperand(1), TD, 16) >= 16) {
7680 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
7681 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
7683 return new StoreInst(II->getOperand(2), Ptr);
7687 case Intrinsic::x86_sse_cvttss2si: {
7688 // These intrinsics only demands the 0th element of its input vector. If
7689 // we can simplify the input based on that, do so now.
7691 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
7693 II->setOperand(1, V);
7699 case Intrinsic::ppc_altivec_vperm:
7700 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
7701 if (ConstantVector *Mask = dyn_cast<ConstantVector>(II->getOperand(3))) {
7702 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
7704 // Check that all of the elements are integer constants or undefs.
7705 bool AllEltsOk = true;
7706 for (unsigned i = 0; i != 16; ++i) {
7707 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
7708 !isa<UndefValue>(Mask->getOperand(i))) {
7715 // Cast the input vectors to byte vectors.
7716 Value *Op0 = InsertCastBefore(Instruction::BitCast,
7717 II->getOperand(1), Mask->getType(), CI);
7718 Value *Op1 = InsertCastBefore(Instruction::BitCast,
7719 II->getOperand(2), Mask->getType(), CI);
7720 Value *Result = UndefValue::get(Op0->getType());
7722 // Only extract each element once.
7723 Value *ExtractedElts[32];
7724 memset(ExtractedElts, 0, sizeof(ExtractedElts));
7726 for (unsigned i = 0; i != 16; ++i) {
7727 if (isa<UndefValue>(Mask->getOperand(i)))
7729 unsigned Idx=cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
7730 Idx &= 31; // Match the hardware behavior.
7732 if (ExtractedElts[Idx] == 0) {
7734 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
7735 InsertNewInstBefore(Elt, CI);
7736 ExtractedElts[Idx] = Elt;
7739 // Insert this value into the result vector.
7740 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
7741 InsertNewInstBefore(cast<Instruction>(Result), CI);
7743 return CastInst::create(Instruction::BitCast, Result, CI.getType());
7748 case Intrinsic::stackrestore: {
7749 // If the save is right next to the restore, remove the restore. This can
7750 // happen when variable allocas are DCE'd.
7751 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
7752 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
7753 BasicBlock::iterator BI = SS;
7755 return EraseInstFromFunction(CI);
7759 // If the stack restore is in a return/unwind block and if there are no
7760 // allocas or calls between the restore and the return, nuke the restore.
7761 TerminatorInst *TI = II->getParent()->getTerminator();
7762 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
7763 BasicBlock::iterator BI = II;
7764 bool CannotRemove = false;
7765 for (++BI; &*BI != TI; ++BI) {
7766 if (isa<AllocaInst>(BI) ||
7767 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
7768 CannotRemove = true;
7773 return EraseInstFromFunction(CI);
7780 return visitCallSite(II);
7783 // InvokeInst simplification
7785 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
7786 return visitCallSite(&II);
7789 // visitCallSite - Improvements for call and invoke instructions.
7791 Instruction *InstCombiner::visitCallSite(CallSite CS) {
7792 bool Changed = false;
7794 // If the callee is a constexpr cast of a function, attempt to move the cast
7795 // to the arguments of the call/invoke.
7796 if (transformConstExprCastCall(CS)) return 0;
7798 Value *Callee = CS.getCalledValue();
7800 if (Function *CalleeF = dyn_cast<Function>(Callee))
7801 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
7802 Instruction *OldCall = CS.getInstruction();
7803 // If the call and callee calling conventions don't match, this call must
7804 // be unreachable, as the call is undefined.
7805 new StoreInst(ConstantInt::getTrue(),
7806 UndefValue::get(PointerType::get(Type::Int1Ty)), OldCall);
7807 if (!OldCall->use_empty())
7808 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
7809 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
7810 return EraseInstFromFunction(*OldCall);
7814 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
7815 // This instruction is not reachable, just remove it. We insert a store to
7816 // undef so that we know that this code is not reachable, despite the fact
7817 // that we can't modify the CFG here.
7818 new StoreInst(ConstantInt::getTrue(),
7819 UndefValue::get(PointerType::get(Type::Int1Ty)),
7820 CS.getInstruction());
7822 if (!CS.getInstruction()->use_empty())
7823 CS.getInstruction()->
7824 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
7826 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
7827 // Don't break the CFG, insert a dummy cond branch.
7828 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
7829 ConstantInt::getTrue(), II);
7831 return EraseInstFromFunction(*CS.getInstruction());
7834 const PointerType *PTy = cast<PointerType>(Callee->getType());
7835 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
7836 if (FTy->isVarArg()) {
7837 // See if we can optimize any arguments passed through the varargs area of
7839 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
7840 E = CS.arg_end(); I != E; ++I)
7841 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
7842 // If this cast does not effect the value passed through the varargs
7843 // area, we can eliminate the use of the cast.
7844 Value *Op = CI->getOperand(0);
7845 if (CI->isLosslessCast()) {
7852 if (BitCastInst *BC = dyn_cast<BitCastInst>(Callee)) {
7853 IntrinsicInst *In = dyn_cast<IntrinsicInst>(BC->getOperand(0));
7854 if (In && In->getIntrinsicID() == Intrinsic::init_trampoline) {
7856 cast<Function>(IntrinsicInst::StripPointerCasts(In->getOperand(2)));
7857 const PointerType *NestFPTy = cast<PointerType>(NestF->getType());
7858 const FunctionType *NestFTy =
7859 cast<FunctionType>(NestFPTy->getElementType());
7861 if (const ParamAttrsList *NestAttrs = NestFTy->getParamAttrs()) {
7862 unsigned NestIdx = 1;
7863 const Type *NestTy = 0;
7864 uint16_t NestAttr = 0;
7866 Instruction *Caller = CS.getInstruction();
7868 // Look for a parameter marked with the 'nest' attribute.
7869 for (FunctionType::param_iterator I = NestFTy->param_begin(),
7870 E = NestFTy->param_end(); I != E; ++NestIdx, ++I)
7871 if (NestAttrs->paramHasAttr(NestIdx, ParamAttr::Nest)) {
7872 // Record the parameter type and any other attributes.
7874 NestAttr = NestAttrs->getParamAttrs(NestIdx);
7879 std::vector<Value*> NewArgs;
7880 NewArgs.reserve(unsigned(CS.arg_end()-CS.arg_begin())+1);
7882 // Insert the nest argument into the call argument list, which may
7883 // mean appending it.
7886 CallSite::arg_iterator I = CS.arg_begin(), E = CS.arg_end();
7888 if (Idx == NestIdx) {
7889 // Add the chain argument.
7890 Value *NestVal = In->getOperand(3);
7891 if (NestVal->getType() != NestTy)
7892 NestVal = new BitCastInst(NestVal, NestTy, "nest", Caller);
7893 NewArgs.push_back(NestVal);
7899 // Add the original argument.
7900 NewArgs.push_back(*I);
7906 // The trampoline may have been bitcast to a bogus type (FTy).
7907 // Handle this by synthesizing a new function type, equal to FTy
7908 // with the chain parameter inserted. Likewise for attributes.
7910 const ParamAttrsList *Attrs = FTy->getParamAttrs();
7911 std::vector<const Type*> NewTypes;
7912 ParamAttrsVector NewAttrs;
7913 NewTypes.reserve(FTy->getNumParams()+1);
7915 // Add any function result attributes.
7916 uint16_t Attr = Attrs ? Attrs->getParamAttrs(0) : 0;
7918 NewAttrs.push_back (ParamAttrsWithIndex::get(0, Attr));
7920 // Insert the chain's type into the list of parameter types, which may
7921 // mean appending it. Likewise for the chain's attributes.
7924 FunctionType::param_iterator I = FTy->param_begin(),
7925 E = FTy->param_end();
7928 if (Idx == NestIdx) {
7929 // Add the chain's type and attributes.
7930 NewTypes.push_back(NestTy);
7931 NewAttrs.push_back(ParamAttrsWithIndex::get(NestIdx, NestAttr));
7937 // Add the original type and attributes.
7938 NewTypes.push_back(*I);
7939 Attr = Attrs ? Attrs->getParamAttrs(Idx) : 0;
7942 (ParamAttrsWithIndex::get(Idx + (Idx >= NestIdx), Attr));
7948 // Replace the trampoline call with a direct call. Let the generic
7949 // code sort out any function type mismatches.
7950 FunctionType *NewFTy =
7951 FunctionType::get(FTy->getReturnType(), NewTypes, FTy->isVarArg(),
7952 ParamAttrsList::get(NewAttrs));
7953 Constant *NewCallee = NestF->getType() == PointerType::get(NewFTy) ?
7954 NestF : ConstantExpr::getBitCast(NestF, PointerType::get(NewFTy));
7956 Instruction *NewCaller;
7957 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7958 NewCaller = new InvokeInst(NewCallee, II->getNormalDest(),
7959 II->getUnwindDest(), NewArgs.begin(),
7960 NewArgs.end(), Caller->getName(),
7962 cast<InvokeInst>(NewCaller)->setCallingConv(II->getCallingConv());
7964 NewCaller = new CallInst(NewCallee, NewArgs.begin(), NewArgs.end(),
7965 Caller->getName(), Caller);
7966 if (cast<CallInst>(Caller)->isTailCall())
7967 cast<CallInst>(NewCaller)->setTailCall();
7968 cast<CallInst>(NewCaller)->
7969 setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7971 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7972 Caller->replaceAllUsesWith(NewCaller);
7973 Caller->eraseFromParent();
7974 RemoveFromWorkList(Caller);
7979 // Replace the trampoline call with a direct call. Since there is no
7980 // 'nest' parameter, there is no need to adjust the argument list. Let
7981 // the generic code sort out any function type mismatches.
7982 Constant *NewCallee = NestF->getType() == PTy ?
7983 NestF : ConstantExpr::getBitCast(NestF, PTy);
7984 CS.setCalledFunction(NewCallee);
7989 return Changed ? CS.getInstruction() : 0;
7992 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
7993 // attempt to move the cast to the arguments of the call/invoke.
7995 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
7996 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
7997 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
7998 if (CE->getOpcode() != Instruction::BitCast ||
7999 !isa<Function>(CE->getOperand(0)))
8001 Function *Callee = cast<Function>(CE->getOperand(0));
8002 Instruction *Caller = CS.getInstruction();
8004 // Okay, this is a cast from a function to a different type. Unless doing so
8005 // would cause a type conversion of one of our arguments, change this call to
8006 // be a direct call with arguments casted to the appropriate types.
8008 const FunctionType *FT = Callee->getFunctionType();
8009 const Type *OldRetTy = Caller->getType();
8011 const FunctionType *ActualFT =
8012 cast<FunctionType>(cast<PointerType>(CE->getType())->getElementType());
8014 // If the parameter attributes don't match up, don't do the xform. We don't
8015 // want to lose an sret attribute or something.
8016 if (FT->getParamAttrs() != ActualFT->getParamAttrs())
8019 // Check to see if we are changing the return type...
8020 if (OldRetTy != FT->getReturnType()) {
8021 if (Callee->isDeclaration() && !Caller->use_empty() &&
8022 // Conversion is ok if changing from pointer to int of same size.
8023 !(isa<PointerType>(FT->getReturnType()) &&
8024 TD->getIntPtrType() == OldRetTy))
8025 return false; // Cannot transform this return value.
8027 // If the callsite is an invoke instruction, and the return value is used by
8028 // a PHI node in a successor, we cannot change the return type of the call
8029 // because there is no place to put the cast instruction (without breaking
8030 // the critical edge). Bail out in this case.
8031 if (!Caller->use_empty())
8032 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
8033 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
8035 if (PHINode *PN = dyn_cast<PHINode>(*UI))
8036 if (PN->getParent() == II->getNormalDest() ||
8037 PN->getParent() == II->getUnwindDest())
8041 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
8042 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
8044 CallSite::arg_iterator AI = CS.arg_begin();
8045 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
8046 const Type *ParamTy = FT->getParamType(i);
8047 const Type *ActTy = (*AI)->getType();
8048 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
8049 //Some conversions are safe even if we do not have a body.
8050 //Either we can cast directly, or we can upconvert the argument
8051 bool isConvertible = ActTy == ParamTy ||
8052 (isa<PointerType>(ParamTy) && isa<PointerType>(ActTy)) ||
8053 (ParamTy->isInteger() && ActTy->isInteger() &&
8054 ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()) ||
8055 (c && ParamTy->getPrimitiveSizeInBits() >= ActTy->getPrimitiveSizeInBits()
8056 && c->getValue().isStrictlyPositive());
8057 if (Callee->isDeclaration() && !isConvertible) return false;
8059 // Most other conversions can be done if we have a body, even if these
8060 // lose information, e.g. int->short.
8061 // Some conversions cannot be done at all, e.g. float to pointer.
8062 // Logic here parallels CastInst::getCastOpcode (the design there
8063 // requires legality checks like this be done before calling it).
8064 if (ParamTy->isInteger()) {
8065 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8066 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
8069 if (!ActTy->isInteger() && !ActTy->isFloatingPoint() &&
8070 !isa<PointerType>(ActTy))
8072 } else if (ParamTy->isFloatingPoint()) {
8073 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8074 if (VActTy->getBitWidth() != ParamTy->getPrimitiveSizeInBits())
8077 if (!ActTy->isInteger() && !ActTy->isFloatingPoint())
8079 } else if (const VectorType *VParamTy = dyn_cast<VectorType>(ParamTy)) {
8080 if (const VectorType *VActTy = dyn_cast<VectorType>(ActTy)) {
8081 if (VActTy->getBitWidth() != VParamTy->getBitWidth())
8084 if (VParamTy->getBitWidth() != ActTy->getPrimitiveSizeInBits())
8086 } else if (isa<PointerType>(ParamTy)) {
8087 if (!ActTy->isInteger() && !isa<PointerType>(ActTy))
8094 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
8095 Callee->isDeclaration())
8096 return false; // Do not delete arguments unless we have a function body...
8098 // Okay, we decided that this is a safe thing to do: go ahead and start
8099 // inserting cast instructions as necessary...
8100 std::vector<Value*> Args;
8101 Args.reserve(NumActualArgs);
8103 AI = CS.arg_begin();
8104 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
8105 const Type *ParamTy = FT->getParamType(i);
8106 if ((*AI)->getType() == ParamTy) {
8107 Args.push_back(*AI);
8109 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
8110 false, ParamTy, false);
8111 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
8112 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
8116 // If the function takes more arguments than the call was taking, add them
8118 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
8119 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
8121 // If we are removing arguments to the function, emit an obnoxious warning...
8122 if (FT->getNumParams() < NumActualArgs)
8123 if (!FT->isVarArg()) {
8124 cerr << "WARNING: While resolving call to function '"
8125 << Callee->getName() << "' arguments were dropped!\n";
8127 // Add all of the arguments in their promoted form to the arg list...
8128 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
8129 const Type *PTy = getPromotedType((*AI)->getType());
8130 if (PTy != (*AI)->getType()) {
8131 // Must promote to pass through va_arg area!
8132 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI, false,
8134 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
8135 InsertNewInstBefore(Cast, *Caller);
8136 Args.push_back(Cast);
8138 Args.push_back(*AI);
8143 if (FT->getReturnType() == Type::VoidTy)
8144 Caller->setName(""); // Void type should not have a name.
8147 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8148 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
8149 Args.begin(), Args.end(), Caller->getName(), Caller);
8150 cast<InvokeInst>(NC)->setCallingConv(II->getCallingConv());
8152 NC = new CallInst(Callee, Args.begin(), Args.end(),
8153 Caller->getName(), Caller);
8154 if (cast<CallInst>(Caller)->isTailCall())
8155 cast<CallInst>(NC)->setTailCall();
8156 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
8159 // Insert a cast of the return type as necessary.
8161 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
8162 if (NV->getType() != Type::VoidTy) {
8163 const Type *CallerTy = Caller->getType();
8164 Instruction::CastOps opcode = CastInst::getCastOpcode(NC, false,
8166 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
8168 // If this is an invoke instruction, we should insert it after the first
8169 // non-phi, instruction in the normal successor block.
8170 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
8171 BasicBlock::iterator I = II->getNormalDest()->begin();
8172 while (isa<PHINode>(I)) ++I;
8173 InsertNewInstBefore(NC, *I);
8175 // Otherwise, it's a call, just insert cast right after the call instr
8176 InsertNewInstBefore(NC, *Caller);
8178 AddUsersToWorkList(*Caller);
8180 NV = UndefValue::get(Caller->getType());
8184 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
8185 Caller->replaceAllUsesWith(NV);
8186 Caller->eraseFromParent();
8187 RemoveFromWorkList(Caller);
8191 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
8192 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
8193 /// and a single binop.
8194 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
8195 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8196 assert(isa<BinaryOperator>(FirstInst) || isa<GetElementPtrInst>(FirstInst) ||
8197 isa<CmpInst>(FirstInst));
8198 unsigned Opc = FirstInst->getOpcode();
8199 Value *LHSVal = FirstInst->getOperand(0);
8200 Value *RHSVal = FirstInst->getOperand(1);
8202 const Type *LHSType = LHSVal->getType();
8203 const Type *RHSType = RHSVal->getType();
8205 // Scan to see if all operands are the same opcode, all have one use, and all
8206 // kill their operands (i.e. the operands have one use).
8207 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
8208 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
8209 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
8210 // Verify type of the LHS matches so we don't fold cmp's of different
8211 // types or GEP's with different index types.
8212 I->getOperand(0)->getType() != LHSType ||
8213 I->getOperand(1)->getType() != RHSType)
8216 // If they are CmpInst instructions, check their predicates
8217 if (Opc == Instruction::ICmp || Opc == Instruction::FCmp)
8218 if (cast<CmpInst>(I)->getPredicate() !=
8219 cast<CmpInst>(FirstInst)->getPredicate())
8222 // Keep track of which operand needs a phi node.
8223 if (I->getOperand(0) != LHSVal) LHSVal = 0;
8224 if (I->getOperand(1) != RHSVal) RHSVal = 0;
8227 // Otherwise, this is safe to transform, determine if it is profitable.
8229 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
8230 // Indexes are often folded into load/store instructions, so we don't want to
8231 // hide them behind a phi.
8232 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
8235 Value *InLHS = FirstInst->getOperand(0);
8236 Value *InRHS = FirstInst->getOperand(1);
8237 PHINode *NewLHS = 0, *NewRHS = 0;
8239 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
8240 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
8241 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
8242 InsertNewInstBefore(NewLHS, PN);
8247 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
8248 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
8249 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
8250 InsertNewInstBefore(NewRHS, PN);
8254 // Add all operands to the new PHIs.
8255 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8257 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8258 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
8261 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
8262 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
8266 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8267 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
8268 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8269 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(), LHSVal,
8272 assert(isa<GetElementPtrInst>(FirstInst));
8273 return new GetElementPtrInst(LHSVal, RHSVal);
8277 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
8278 /// of the block that defines it. This means that it must be obvious the value
8279 /// of the load is not changed from the point of the load to the end of the
8282 /// Finally, it is safe, but not profitable, to sink a load targetting a
8283 /// non-address-taken alloca. Doing so will cause us to not promote the alloca
8285 static bool isSafeToSinkLoad(LoadInst *L) {
8286 BasicBlock::iterator BBI = L, E = L->getParent()->end();
8288 for (++BBI; BBI != E; ++BBI)
8289 if (BBI->mayWriteToMemory())
8292 // Check for non-address taken alloca. If not address-taken already, it isn't
8293 // profitable to do this xform.
8294 if (AllocaInst *AI = dyn_cast<AllocaInst>(L->getOperand(0))) {
8295 bool isAddressTaken = false;
8296 for (Value::use_iterator UI = AI->use_begin(), E = AI->use_end();
8298 if (isa<LoadInst>(UI)) continue;
8299 if (StoreInst *SI = dyn_cast<StoreInst>(*UI)) {
8300 // If storing TO the alloca, then the address isn't taken.
8301 if (SI->getOperand(1) == AI) continue;
8303 isAddressTaken = true;
8307 if (!isAddressTaken)
8315 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
8316 // operator and they all are only used by the PHI, PHI together their
8317 // inputs, and do the operation once, to the result of the PHI.
8318 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
8319 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
8321 // Scan the instruction, looking for input operations that can be folded away.
8322 // If all input operands to the phi are the same instruction (e.g. a cast from
8323 // the same type or "+42") we can pull the operation through the PHI, reducing
8324 // code size and simplifying code.
8325 Constant *ConstantOp = 0;
8326 const Type *CastSrcTy = 0;
8327 bool isVolatile = false;
8328 if (isa<CastInst>(FirstInst)) {
8329 CastSrcTy = FirstInst->getOperand(0)->getType();
8330 } else if (isa<BinaryOperator>(FirstInst) || isa<CmpInst>(FirstInst)) {
8331 // Can fold binop, compare or shift here if the RHS is a constant,
8332 // otherwise call FoldPHIArgBinOpIntoPHI.
8333 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
8334 if (ConstantOp == 0)
8335 return FoldPHIArgBinOpIntoPHI(PN);
8336 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
8337 isVolatile = LI->isVolatile();
8338 // We can't sink the load if the loaded value could be modified between the
8339 // load and the PHI.
8340 if (LI->getParent() != PN.getIncomingBlock(0) ||
8341 !isSafeToSinkLoad(LI))
8343 } else if (isa<GetElementPtrInst>(FirstInst)) {
8344 if (FirstInst->getNumOperands() == 2)
8345 return FoldPHIArgBinOpIntoPHI(PN);
8346 // Can't handle general GEPs yet.
8349 return 0; // Cannot fold this operation.
8352 // Check to see if all arguments are the same operation.
8353 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8354 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
8355 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
8356 if (!I->hasOneUse() || !I->isSameOperationAs(FirstInst))
8359 if (I->getOperand(0)->getType() != CastSrcTy)
8360 return 0; // Cast operation must match.
8361 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8362 // We can't sink the load if the loaded value could be modified between
8363 // the load and the PHI.
8364 if (LI->isVolatile() != isVolatile ||
8365 LI->getParent() != PN.getIncomingBlock(i) ||
8366 !isSafeToSinkLoad(LI))
8368 } else if (I->getOperand(1) != ConstantOp) {
8373 // Okay, they are all the same operation. Create a new PHI node of the
8374 // correct type, and PHI together all of the LHS's of the instructions.
8375 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
8376 PN.getName()+".in");
8377 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
8379 Value *InVal = FirstInst->getOperand(0);
8380 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
8382 // Add all operands to the new PHI.
8383 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
8384 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
8385 if (NewInVal != InVal)
8387 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
8392 // The new PHI unions all of the same values together. This is really
8393 // common, so we handle it intelligently here for compile-time speed.
8397 InsertNewInstBefore(NewPN, PN);
8401 // Insert and return the new operation.
8402 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
8403 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
8404 else if (isa<LoadInst>(FirstInst))
8405 return new LoadInst(PhiVal, "", isVolatile);
8406 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
8407 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
8408 else if (CmpInst *CIOp = dyn_cast<CmpInst>(FirstInst))
8409 return CmpInst::create(CIOp->getOpcode(), CIOp->getPredicate(),
8410 PhiVal, ConstantOp);
8412 assert(0 && "Unknown operation");
8416 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
8418 static bool DeadPHICycle(PHINode *PN,
8419 SmallPtrSet<PHINode*, 16> &PotentiallyDeadPHIs) {
8420 if (PN->use_empty()) return true;
8421 if (!PN->hasOneUse()) return false;
8423 // Remember this node, and if we find the cycle, return.
8424 if (!PotentiallyDeadPHIs.insert(PN))
8427 // Don't scan crazily complex things.
8428 if (PotentiallyDeadPHIs.size() == 16)
8431 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
8432 return DeadPHICycle(PU, PotentiallyDeadPHIs);
8437 // PHINode simplification
8439 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
8440 // If LCSSA is around, don't mess with Phi nodes
8441 if (MustPreserveLCSSA) return 0;
8443 if (Value *V = PN.hasConstantValue())
8444 return ReplaceInstUsesWith(PN, V);
8446 // If all PHI operands are the same operation, pull them through the PHI,
8447 // reducing code size.
8448 if (isa<Instruction>(PN.getIncomingValue(0)) &&
8449 PN.getIncomingValue(0)->hasOneUse())
8450 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
8453 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
8454 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
8455 // PHI)... break the cycle.
8456 if (PN.hasOneUse()) {
8457 Instruction *PHIUser = cast<Instruction>(PN.use_back());
8458 if (PHINode *PU = dyn_cast<PHINode>(PHIUser)) {
8459 SmallPtrSet<PHINode*, 16> PotentiallyDeadPHIs;
8460 PotentiallyDeadPHIs.insert(&PN);
8461 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
8462 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8465 // If this phi has a single use, and if that use just computes a value for
8466 // the next iteration of a loop, delete the phi. This occurs with unused
8467 // induction variables, e.g. "for (int j = 0; ; ++j);". Detecting this
8468 // common case here is good because the only other things that catch this
8469 // are induction variable analysis (sometimes) and ADCE, which is only run
8471 if (PHIUser->hasOneUse() &&
8472 (isa<BinaryOperator>(PHIUser) || isa<GetElementPtrInst>(PHIUser)) &&
8473 PHIUser->use_back() == &PN) {
8474 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
8481 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
8482 Instruction *InsertPoint,
8484 unsigned PtrSize = DTy->getPrimitiveSizeInBits();
8485 unsigned VTySize = V->getType()->getPrimitiveSizeInBits();
8486 // We must cast correctly to the pointer type. Ensure that we
8487 // sign extend the integer value if it is smaller as this is
8488 // used for address computation.
8489 Instruction::CastOps opcode =
8490 (VTySize < PtrSize ? Instruction::SExt :
8491 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
8492 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
8496 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
8497 Value *PtrOp = GEP.getOperand(0);
8498 // Is it 'getelementptr %P, i32 0' or 'getelementptr %P'
8499 // If so, eliminate the noop.
8500 if (GEP.getNumOperands() == 1)
8501 return ReplaceInstUsesWith(GEP, PtrOp);
8503 if (isa<UndefValue>(GEP.getOperand(0)))
8504 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
8506 bool HasZeroPointerIndex = false;
8507 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
8508 HasZeroPointerIndex = C->isNullValue();
8510 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
8511 return ReplaceInstUsesWith(GEP, PtrOp);
8513 // Eliminate unneeded casts for indices.
8514 bool MadeChange = false;
8516 gep_type_iterator GTI = gep_type_begin(GEP);
8517 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI) {
8518 if (isa<SequentialType>(*GTI)) {
8519 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
8520 if (CI->getOpcode() == Instruction::ZExt ||
8521 CI->getOpcode() == Instruction::SExt) {
8522 const Type *SrcTy = CI->getOperand(0)->getType();
8523 // We can eliminate a cast from i32 to i64 iff the target
8524 // is a 32-bit pointer target.
8525 if (SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
8527 GEP.setOperand(i, CI->getOperand(0));
8531 // If we are using a wider index than needed for this platform, shrink it
8532 // to what we need. If the incoming value needs a cast instruction,
8533 // insert it. This explicit cast can make subsequent optimizations more
8535 Value *Op = GEP.getOperand(i);
8536 if (TD->getTypeSize(Op->getType()) > TD->getPointerSize())
8537 if (Constant *C = dyn_cast<Constant>(Op)) {
8538 GEP.setOperand(i, ConstantExpr::getTrunc(C, TD->getIntPtrType()));
8541 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
8543 GEP.setOperand(i, Op);
8548 if (MadeChange) return &GEP;
8550 // If this GEP instruction doesn't move the pointer, and if the input operand
8551 // is a bitcast of another pointer, just replace the GEP with a bitcast of the
8552 // real input to the dest type.
8553 if (GEP.hasAllZeroIndices() && isa<BitCastInst>(GEP.getOperand(0)))
8554 return new BitCastInst(cast<BitCastInst>(GEP.getOperand(0))->getOperand(0),
8557 // Combine Indices - If the source pointer to this getelementptr instruction
8558 // is a getelementptr instruction, combine the indices of the two
8559 // getelementptr instructions into a single instruction.
8561 SmallVector<Value*, 8> SrcGEPOperands;
8562 if (User *Src = dyn_castGetElementPtr(PtrOp))
8563 SrcGEPOperands.append(Src->op_begin(), Src->op_end());
8565 if (!SrcGEPOperands.empty()) {
8566 // Note that if our source is a gep chain itself that we wait for that
8567 // chain to be resolved before we perform this transformation. This
8568 // avoids us creating a TON of code in some cases.
8570 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
8571 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
8572 return 0; // Wait until our source is folded to completion.
8574 SmallVector<Value*, 8> Indices;
8576 // Find out whether the last index in the source GEP is a sequential idx.
8577 bool EndsWithSequential = false;
8578 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
8579 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
8580 EndsWithSequential = !isa<StructType>(*I);
8582 // Can we combine the two pointer arithmetics offsets?
8583 if (EndsWithSequential) {
8584 // Replace: gep (gep %P, long B), long A, ...
8585 // With: T = long A+B; gep %P, T, ...
8587 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
8588 if (SO1 == Constant::getNullValue(SO1->getType())) {
8590 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
8593 // If they aren't the same type, convert both to an integer of the
8594 // target's pointer size.
8595 if (SO1->getType() != GO1->getType()) {
8596 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
8597 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
8598 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
8599 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
8601 unsigned PS = TD->getPointerSize();
8602 if (TD->getTypeSize(SO1->getType()) == PS) {
8603 // Convert GO1 to SO1's type.
8604 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
8606 } else if (TD->getTypeSize(GO1->getType()) == PS) {
8607 // Convert SO1 to GO1's type.
8608 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
8610 const Type *PT = TD->getIntPtrType();
8611 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
8612 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
8616 if (isa<Constant>(SO1) && isa<Constant>(GO1))
8617 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
8619 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
8620 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
8624 // Recycle the GEP we already have if possible.
8625 if (SrcGEPOperands.size() == 2) {
8626 GEP.setOperand(0, SrcGEPOperands[0]);
8627 GEP.setOperand(1, Sum);
8630 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8631 SrcGEPOperands.end()-1);
8632 Indices.push_back(Sum);
8633 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
8635 } else if (isa<Constant>(*GEP.idx_begin()) &&
8636 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
8637 SrcGEPOperands.size() != 1) {
8638 // Otherwise we can do the fold if the first index of the GEP is a zero
8639 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
8640 SrcGEPOperands.end());
8641 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
8644 if (!Indices.empty())
8645 return new GetElementPtrInst(SrcGEPOperands[0], Indices.begin(),
8646 Indices.end(), GEP.getName());
8648 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
8649 // GEP of global variable. If all of the indices for this GEP are
8650 // constants, we can promote this to a constexpr instead of an instruction.
8652 // Scan for nonconstants...
8653 SmallVector<Constant*, 8> Indices;
8654 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
8655 for (; I != E && isa<Constant>(*I); ++I)
8656 Indices.push_back(cast<Constant>(*I));
8658 if (I == E) { // If they are all constants...
8659 Constant *CE = ConstantExpr::getGetElementPtr(GV,
8660 &Indices[0],Indices.size());
8662 // Replace all uses of the GEP with the new constexpr...
8663 return ReplaceInstUsesWith(GEP, CE);
8665 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
8666 if (!isa<PointerType>(X->getType())) {
8667 // Not interesting. Source pointer must be a cast from pointer.
8668 } else if (HasZeroPointerIndex) {
8669 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
8670 // into : GEP [10 x ubyte]* X, long 0, ...
8672 // This occurs when the program declares an array extern like "int X[];"
8674 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
8675 const PointerType *XTy = cast<PointerType>(X->getType());
8676 if (const ArrayType *XATy =
8677 dyn_cast<ArrayType>(XTy->getElementType()))
8678 if (const ArrayType *CATy =
8679 dyn_cast<ArrayType>(CPTy->getElementType()))
8680 if (CATy->getElementType() == XATy->getElementType()) {
8681 // At this point, we know that the cast source type is a pointer
8682 // to an array of the same type as the destination pointer
8683 // array. Because the array type is never stepped over (there
8684 // is a leading zero) we can fold the cast into this GEP.
8685 GEP.setOperand(0, X);
8688 } else if (GEP.getNumOperands() == 2) {
8689 // Transform things like:
8690 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
8691 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
8692 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
8693 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
8694 if (isa<ArrayType>(SrcElTy) &&
8695 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
8696 TD->getTypeSize(ResElTy)) {
8698 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8699 Idx[1] = GEP.getOperand(1);
8700 Value *V = InsertNewInstBefore(
8701 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName()), GEP);
8702 // V and GEP are both pointer types --> BitCast
8703 return new BitCastInst(V, GEP.getType());
8706 // Transform things like:
8707 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
8708 // (where tmp = 8*tmp2) into:
8709 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
8711 if (isa<ArrayType>(SrcElTy) &&
8712 (ResElTy == Type::Int8Ty || ResElTy == Type::Int8Ty)) {
8713 uint64_t ArrayEltSize =
8714 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
8716 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
8717 // allow either a mul, shift, or constant here.
8719 ConstantInt *Scale = 0;
8720 if (ArrayEltSize == 1) {
8721 NewIdx = GEP.getOperand(1);
8722 Scale = ConstantInt::get(NewIdx->getType(), 1);
8723 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
8724 NewIdx = ConstantInt::get(CI->getType(), 1);
8726 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
8727 if (Inst->getOpcode() == Instruction::Shl &&
8728 isa<ConstantInt>(Inst->getOperand(1))) {
8729 ConstantInt *ShAmt = cast<ConstantInt>(Inst->getOperand(1));
8730 uint32_t ShAmtVal = ShAmt->getLimitedValue(64);
8731 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmtVal);
8732 NewIdx = Inst->getOperand(0);
8733 } else if (Inst->getOpcode() == Instruction::Mul &&
8734 isa<ConstantInt>(Inst->getOperand(1))) {
8735 Scale = cast<ConstantInt>(Inst->getOperand(1));
8736 NewIdx = Inst->getOperand(0);
8740 // If the index will be to exactly the right offset with the scale taken
8741 // out, perform the transformation.
8742 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
8743 if (isa<ConstantInt>(Scale))
8744 Scale = ConstantInt::get(Scale->getType(),
8745 Scale->getZExtValue() / ArrayEltSize);
8746 if (Scale->getZExtValue() != 1) {
8747 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
8749 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
8750 NewIdx = InsertNewInstBefore(Sc, GEP);
8753 // Insert the new GEP instruction.
8755 Idx[0] = Constant::getNullValue(Type::Int32Ty);
8757 Instruction *NewGEP =
8758 new GetElementPtrInst(X, Idx, Idx + 2, GEP.getName());
8759 NewGEP = InsertNewInstBefore(NewGEP, GEP);
8760 // The NewGEP must be pointer typed, so must the old one -> BitCast
8761 return new BitCastInst(NewGEP, GEP.getType());
8770 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
8771 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
8772 if (AI.isArrayAllocation()) // Check C != 1
8773 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
8775 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
8776 AllocationInst *New = 0;
8778 // Create and insert the replacement instruction...
8779 if (isa<MallocInst>(AI))
8780 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
8782 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
8783 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
8786 InsertNewInstBefore(New, AI);
8788 // Scan to the end of the allocation instructions, to skip over a block of
8789 // allocas if possible...
8791 BasicBlock::iterator It = New;
8792 while (isa<AllocationInst>(*It)) ++It;
8794 // Now that I is pointing to the first non-allocation-inst in the block,
8795 // insert our getelementptr instruction...
8797 Value *NullIdx = Constant::getNullValue(Type::Int32Ty);
8801 Value *V = new GetElementPtrInst(New, Idx, Idx + 2,
8802 New->getName()+".sub", It);
8804 // Now make everything use the getelementptr instead of the original
8806 return ReplaceInstUsesWith(AI, V);
8807 } else if (isa<UndefValue>(AI.getArraySize())) {
8808 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8811 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
8812 // Note that we only do this for alloca's, because malloc should allocate and
8813 // return a unique pointer, even for a zero byte allocation.
8814 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
8815 TD->getTypeSize(AI.getAllocatedType()) == 0)
8816 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
8821 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
8822 Value *Op = FI.getOperand(0);
8824 // free undef -> unreachable.
8825 if (isa<UndefValue>(Op)) {
8826 // Insert a new store to null because we cannot modify the CFG here.
8827 new StoreInst(ConstantInt::getTrue(),
8828 UndefValue::get(PointerType::get(Type::Int1Ty)), &FI);
8829 return EraseInstFromFunction(FI);
8832 // If we have 'free null' delete the instruction. This can happen in stl code
8833 // when lots of inlining happens.
8834 if (isa<ConstantPointerNull>(Op))
8835 return EraseInstFromFunction(FI);
8837 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
8838 if (BitCastInst *CI = dyn_cast<BitCastInst>(Op)) {
8839 FI.setOperand(0, CI->getOperand(0));
8843 // Change free (gep X, 0,0,0,0) into free(X)
8844 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op)) {
8845 if (GEPI->hasAllZeroIndices()) {
8846 AddToWorkList(GEPI);
8847 FI.setOperand(0, GEPI->getOperand(0));
8852 // Change free(malloc) into nothing, if the malloc has a single use.
8853 if (MallocInst *MI = dyn_cast<MallocInst>(Op))
8854 if (MI->hasOneUse()) {
8855 EraseInstFromFunction(FI);
8856 return EraseInstFromFunction(*MI);
8863 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
8864 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
8865 User *CI = cast<User>(LI.getOperand(0));
8866 Value *CastOp = CI->getOperand(0);
8868 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
8869 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
8870 const Type *SrcPTy = SrcTy->getElementType();
8872 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
8873 isa<VectorType>(DestPTy)) {
8874 // If the source is an array, the code below will not succeed. Check to
8875 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
8877 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
8878 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
8879 if (ASrcTy->getNumElements() != 0) {
8881 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
8882 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
8883 SrcTy = cast<PointerType>(CastOp->getType());
8884 SrcPTy = SrcTy->getElementType();
8887 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
8888 isa<VectorType>(SrcPTy)) &&
8889 // Do not allow turning this into a load of an integer, which is then
8890 // casted to a pointer, this pessimizes pointer analysis a lot.
8891 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
8892 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
8893 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
8895 // Okay, we are casting from one integer or pointer type to another of
8896 // the same size. Instead of casting the pointer before the load, cast
8897 // the result of the loaded value.
8898 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
8900 LI.isVolatile()),LI);
8901 // Now cast the result of the load.
8902 return new BitCastInst(NewLoad, LI.getType());
8909 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
8910 /// from this value cannot trap. If it is not obviously safe to load from the
8911 /// specified pointer, we do a quick local scan of the basic block containing
8912 /// ScanFrom, to determine if the address is already accessed.
8913 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
8914 // If it is an alloca or global variable, it is always safe to load from.
8915 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
8917 // Otherwise, be a little bit agressive by scanning the local block where we
8918 // want to check to see if the pointer is already being loaded or stored
8919 // from/to. If so, the previous load or store would have already trapped,
8920 // so there is no harm doing an extra load (also, CSE will later eliminate
8921 // the load entirely).
8922 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
8927 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
8928 if (LI->getOperand(0) == V) return true;
8929 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8930 if (SI->getOperand(1) == V) return true;
8936 /// GetUnderlyingObject - Trace through a series of getelementptrs and bitcasts
8937 /// until we find the underlying object a pointer is referring to or something
8938 /// we don't understand. Note that the returned pointer may be offset from the
8939 /// input, because we ignore GEP indices.
8940 static Value *GetUnderlyingObject(Value *Ptr) {
8942 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr)) {
8943 if (CE->getOpcode() == Instruction::BitCast ||
8944 CE->getOpcode() == Instruction::GetElementPtr)
8945 Ptr = CE->getOperand(0);
8948 } else if (BitCastInst *BCI = dyn_cast<BitCastInst>(Ptr)) {
8949 Ptr = BCI->getOperand(0);
8950 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr)) {
8951 Ptr = GEP->getOperand(0);
8958 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
8959 Value *Op = LI.getOperand(0);
8961 // Attempt to improve the alignment.
8962 unsigned KnownAlign = GetOrEnforceKnownAlignment(Op, TD);
8963 if (KnownAlign > LI.getAlignment())
8964 LI.setAlignment(KnownAlign);
8966 // load (cast X) --> cast (load X) iff safe
8967 if (isa<CastInst>(Op))
8968 if (Instruction *Res = InstCombineLoadCast(*this, LI))
8971 // None of the following transforms are legal for volatile loads.
8972 if (LI.isVolatile()) return 0;
8974 if (&LI.getParent()->front() != &LI) {
8975 BasicBlock::iterator BBI = &LI; --BBI;
8976 // If the instruction immediately before this is a store to the same
8977 // address, do a simple form of store->load forwarding.
8978 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
8979 if (SI->getOperand(1) == LI.getOperand(0))
8980 return ReplaceInstUsesWith(LI, SI->getOperand(0));
8981 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
8982 if (LIB->getOperand(0) == LI.getOperand(0))
8983 return ReplaceInstUsesWith(LI, LIB);
8986 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
8987 if (isa<ConstantPointerNull>(GEPI->getOperand(0))) {
8988 // Insert a new store to null instruction before the load to indicate
8989 // that this code is not reachable. We do this instead of inserting
8990 // an unreachable instruction directly because we cannot modify the
8992 new StoreInst(UndefValue::get(LI.getType()),
8993 Constant::getNullValue(Op->getType()), &LI);
8994 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
8997 if (Constant *C = dyn_cast<Constant>(Op)) {
8998 // load null/undef -> undef
8999 if ((C->isNullValue() || isa<UndefValue>(C))) {
9000 // Insert a new store to null instruction before the load to indicate that
9001 // this code is not reachable. We do this instead of inserting an
9002 // unreachable instruction directly because we cannot modify the CFG.
9003 new StoreInst(UndefValue::get(LI.getType()),
9004 Constant::getNullValue(Op->getType()), &LI);
9005 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9008 // Instcombine load (constant global) into the value loaded.
9009 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
9010 if (GV->isConstant() && !GV->isDeclaration())
9011 return ReplaceInstUsesWith(LI, GV->getInitializer());
9013 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
9014 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
9015 if (CE->getOpcode() == Instruction::GetElementPtr) {
9016 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
9017 if (GV->isConstant() && !GV->isDeclaration())
9019 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
9020 return ReplaceInstUsesWith(LI, V);
9021 if (CE->getOperand(0)->isNullValue()) {
9022 // Insert a new store to null instruction before the load to indicate
9023 // that this code is not reachable. We do this instead of inserting
9024 // an unreachable instruction directly because we cannot modify the
9026 new StoreInst(UndefValue::get(LI.getType()),
9027 Constant::getNullValue(Op->getType()), &LI);
9028 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9031 } else if (CE->isCast()) {
9032 if (Instruction *Res = InstCombineLoadCast(*this, LI))
9037 // If this load comes from anywhere in a constant global, and if the global
9038 // is all undef or zero, we know what it loads.
9039 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(GetUnderlyingObject(Op))) {
9040 if (GV->isConstant() && GV->hasInitializer()) {
9041 if (GV->getInitializer()->isNullValue())
9042 return ReplaceInstUsesWith(LI, Constant::getNullValue(LI.getType()));
9043 else if (isa<UndefValue>(GV->getInitializer()))
9044 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
9048 if (Op->hasOneUse()) {
9049 // Change select and PHI nodes to select values instead of addresses: this
9050 // helps alias analysis out a lot, allows many others simplifications, and
9051 // exposes redundancy in the code.
9053 // Note that we cannot do the transformation unless we know that the
9054 // introduced loads cannot trap! Something like this is valid as long as
9055 // the condition is always false: load (select bool %C, int* null, int* %G),
9056 // but it would not be valid if we transformed it to load from null
9059 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
9060 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
9061 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
9062 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
9063 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
9064 SI->getOperand(1)->getName()+".val"), LI);
9065 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
9066 SI->getOperand(2)->getName()+".val"), LI);
9067 return new SelectInst(SI->getCondition(), V1, V2);
9070 // load (select (cond, null, P)) -> load P
9071 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
9072 if (C->isNullValue()) {
9073 LI.setOperand(0, SI->getOperand(2));
9077 // load (select (cond, P, null)) -> load P
9078 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
9079 if (C->isNullValue()) {
9080 LI.setOperand(0, SI->getOperand(1));
9088 /// InstCombineStoreToCast - Fold store V, (cast P) -> store (cast V), P
9090 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
9091 User *CI = cast<User>(SI.getOperand(1));
9092 Value *CastOp = CI->getOperand(0);
9094 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
9095 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
9096 const Type *SrcPTy = SrcTy->getElementType();
9098 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
9099 // If the source is an array, the code below will not succeed. Check to
9100 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
9102 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
9103 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
9104 if (ASrcTy->getNumElements() != 0) {
9106 Idxs[0] = Idxs[1] = Constant::getNullValue(Type::Int32Ty);
9107 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs, 2);
9108 SrcTy = cast<PointerType>(CastOp->getType());
9109 SrcPTy = SrcTy->getElementType();
9112 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
9113 IC.getTargetData().getTypeSizeInBits(SrcPTy) ==
9114 IC.getTargetData().getTypeSizeInBits(DestPTy)) {
9116 // Okay, we are casting from one integer or pointer type to another of
9117 // the same size. Instead of casting the pointer before
9118 // the store, cast the value to be stored.
9120 Value *SIOp0 = SI.getOperand(0);
9121 Instruction::CastOps opcode = Instruction::BitCast;
9122 const Type* CastSrcTy = SIOp0->getType();
9123 const Type* CastDstTy = SrcPTy;
9124 if (isa<PointerType>(CastDstTy)) {
9125 if (CastSrcTy->isInteger())
9126 opcode = Instruction::IntToPtr;
9127 } else if (isa<IntegerType>(CastDstTy)) {
9128 if (isa<PointerType>(SIOp0->getType()))
9129 opcode = Instruction::PtrToInt;
9131 if (Constant *C = dyn_cast<Constant>(SIOp0))
9132 NewCast = ConstantExpr::getCast(opcode, C, CastDstTy);
9134 NewCast = IC.InsertNewInstBefore(
9135 CastInst::create(opcode, SIOp0, CastDstTy, SIOp0->getName()+".c"),
9137 return new StoreInst(NewCast, CastOp);
9144 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
9145 Value *Val = SI.getOperand(0);
9146 Value *Ptr = SI.getOperand(1);
9148 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
9149 EraseInstFromFunction(SI);
9154 // If the RHS is an alloca with a single use, zapify the store, making the
9156 if (Ptr->hasOneUse()) {
9157 if (isa<AllocaInst>(Ptr)) {
9158 EraseInstFromFunction(SI);
9163 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Ptr))
9164 if (isa<AllocaInst>(GEP->getOperand(0)) &&
9165 GEP->getOperand(0)->hasOneUse()) {
9166 EraseInstFromFunction(SI);
9172 // Attempt to improve the alignment.
9173 unsigned KnownAlign = GetOrEnforceKnownAlignment(Ptr, TD);
9174 if (KnownAlign > SI.getAlignment())
9175 SI.setAlignment(KnownAlign);
9177 // Do really simple DSE, to catch cases where there are several consequtive
9178 // stores to the same location, separated by a few arithmetic operations. This
9179 // situation often occurs with bitfield accesses.
9180 BasicBlock::iterator BBI = &SI;
9181 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
9185 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
9186 // Prev store isn't volatile, and stores to the same location?
9187 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
9190 EraseInstFromFunction(*PrevSI);
9196 // If this is a load, we have to stop. However, if the loaded value is from
9197 // the pointer we're loading and is producing the pointer we're storing,
9198 // then *this* store is dead (X = load P; store X -> P).
9199 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
9200 if (LI == Val && LI->getOperand(0) == Ptr && !SI.isVolatile()) {
9201 EraseInstFromFunction(SI);
9205 // Otherwise, this is a load from some other location. Stores before it
9210 // Don't skip over loads or things that can modify memory.
9211 if (BBI->mayWriteToMemory())
9216 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
9218 // store X, null -> turns into 'unreachable' in SimplifyCFG
9219 if (isa<ConstantPointerNull>(Ptr)) {
9220 if (!isa<UndefValue>(Val)) {
9221 SI.setOperand(0, UndefValue::get(Val->getType()));
9222 if (Instruction *U = dyn_cast<Instruction>(Val))
9223 AddToWorkList(U); // Dropped a use.
9226 return 0; // Do not modify these!
9229 // store undef, Ptr -> noop
9230 if (isa<UndefValue>(Val)) {
9231 EraseInstFromFunction(SI);
9236 // If the pointer destination is a cast, see if we can fold the cast into the
9238 if (isa<CastInst>(Ptr))
9239 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9241 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
9243 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
9247 // If this store is the last instruction in the basic block, and if the block
9248 // ends with an unconditional branch, try to move it to the successor block.
9250 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
9251 if (BI->isUnconditional())
9252 if (SimplifyStoreAtEndOfBlock(SI))
9253 return 0; // xform done!
9258 /// SimplifyStoreAtEndOfBlock - Turn things like:
9259 /// if () { *P = v1; } else { *P = v2 }
9260 /// into a phi node with a store in the successor.
9262 /// Simplify things like:
9263 /// *P = v1; if () { *P = v2; }
9264 /// into a phi node with a store in the successor.
9266 bool InstCombiner::SimplifyStoreAtEndOfBlock(StoreInst &SI) {
9267 BasicBlock *StoreBB = SI.getParent();
9269 // Check to see if the successor block has exactly two incoming edges. If
9270 // so, see if the other predecessor contains a store to the same location.
9271 // if so, insert a PHI node (if needed) and move the stores down.
9272 BasicBlock *DestBB = StoreBB->getTerminator()->getSuccessor(0);
9274 // Determine whether Dest has exactly two predecessors and, if so, compute
9275 // the other predecessor.
9276 pred_iterator PI = pred_begin(DestBB);
9277 BasicBlock *OtherBB = 0;
9281 if (PI == pred_end(DestBB))
9284 if (*PI != StoreBB) {
9289 if (++PI != pred_end(DestBB))
9293 // Verify that the other block ends in a branch and is not otherwise empty.
9294 BasicBlock::iterator BBI = OtherBB->getTerminator();
9295 BranchInst *OtherBr = dyn_cast<BranchInst>(BBI);
9296 if (!OtherBr || BBI == OtherBB->begin())
9299 // If the other block ends in an unconditional branch, check for the 'if then
9300 // else' case. there is an instruction before the branch.
9301 StoreInst *OtherStore = 0;
9302 if (OtherBr->isUnconditional()) {
9303 // If this isn't a store, or isn't a store to the same location, bail out.
9305 OtherStore = dyn_cast<StoreInst>(BBI);
9306 if (!OtherStore || OtherStore->getOperand(1) != SI.getOperand(1))
9309 // Otherwise, the other block ended with a conditional branch. If one of the
9310 // destinations is StoreBB, then we have the if/then case.
9311 if (OtherBr->getSuccessor(0) != StoreBB &&
9312 OtherBr->getSuccessor(1) != StoreBB)
9315 // Okay, we know that OtherBr now goes to Dest and StoreBB, so this is an
9316 // if/then triangle. See if there is a store to the same ptr as SI that
9317 // lives in OtherBB.
9319 // Check to see if we find the matching store.
9320 if ((OtherStore = dyn_cast<StoreInst>(BBI))) {
9321 if (OtherStore->getOperand(1) != SI.getOperand(1))
9325 // If we find something that may be using the stored value, or if we run
9326 // out of instructions, we can't do the xform.
9327 if (isa<LoadInst>(BBI) || BBI->mayWriteToMemory() ||
9328 BBI == OtherBB->begin())
9332 // In order to eliminate the store in OtherBr, we have to
9333 // make sure nothing reads the stored value in StoreBB.
9334 for (BasicBlock::iterator I = StoreBB->begin(); &*I != &SI; ++I) {
9335 // FIXME: This should really be AA driven.
9336 if (isa<LoadInst>(I) || I->mayWriteToMemory())
9341 // Insert a PHI node now if we need it.
9342 Value *MergedVal = OtherStore->getOperand(0);
9343 if (MergedVal != SI.getOperand(0)) {
9344 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
9345 PN->reserveOperandSpace(2);
9346 PN->addIncoming(SI.getOperand(0), SI.getParent());
9347 PN->addIncoming(OtherStore->getOperand(0), OtherBB);
9348 MergedVal = InsertNewInstBefore(PN, DestBB->front());
9351 // Advance to a place where it is safe to insert the new store and
9353 BBI = DestBB->begin();
9354 while (isa<PHINode>(BBI)) ++BBI;
9355 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
9356 OtherStore->isVolatile()), *BBI);
9358 // Nuke the old stores.
9359 EraseInstFromFunction(SI);
9360 EraseInstFromFunction(*OtherStore);
9366 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
9367 // Change br (not X), label True, label False to: br X, label False, True
9369 BasicBlock *TrueDest;
9370 BasicBlock *FalseDest;
9371 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
9372 !isa<Constant>(X)) {
9373 // Swap Destinations and condition...
9375 BI.setSuccessor(0, FalseDest);
9376 BI.setSuccessor(1, TrueDest);
9380 // Cannonicalize fcmp_one -> fcmp_oeq
9381 FCmpInst::Predicate FPred; Value *Y;
9382 if (match(&BI, m_Br(m_FCmp(FPred, m_Value(X), m_Value(Y)),
9383 TrueDest, FalseDest)))
9384 if ((FPred == FCmpInst::FCMP_ONE || FPred == FCmpInst::FCMP_OLE ||
9385 FPred == FCmpInst::FCMP_OGE) && BI.getCondition()->hasOneUse()) {
9386 FCmpInst *I = cast<FCmpInst>(BI.getCondition());
9387 FCmpInst::Predicate NewPred = FCmpInst::getInversePredicate(FPred);
9388 Instruction *NewSCC = new FCmpInst(NewPred, X, Y, "", I);
9389 NewSCC->takeName(I);
9390 // Swap Destinations and condition...
9391 BI.setCondition(NewSCC);
9392 BI.setSuccessor(0, FalseDest);
9393 BI.setSuccessor(1, TrueDest);
9394 RemoveFromWorkList(I);
9395 I->eraseFromParent();
9396 AddToWorkList(NewSCC);
9400 // Cannonicalize icmp_ne -> icmp_eq
9401 ICmpInst::Predicate IPred;
9402 if (match(&BI, m_Br(m_ICmp(IPred, m_Value(X), m_Value(Y)),
9403 TrueDest, FalseDest)))
9404 if ((IPred == ICmpInst::ICMP_NE || IPred == ICmpInst::ICMP_ULE ||
9405 IPred == ICmpInst::ICMP_SLE || IPred == ICmpInst::ICMP_UGE ||
9406 IPred == ICmpInst::ICMP_SGE) && BI.getCondition()->hasOneUse()) {
9407 ICmpInst *I = cast<ICmpInst>(BI.getCondition());
9408 ICmpInst::Predicate NewPred = ICmpInst::getInversePredicate(IPred);
9409 Instruction *NewSCC = new ICmpInst(NewPred, X, Y, "", I);
9410 NewSCC->takeName(I);
9411 // Swap Destinations and condition...
9412 BI.setCondition(NewSCC);
9413 BI.setSuccessor(0, FalseDest);
9414 BI.setSuccessor(1, TrueDest);
9415 RemoveFromWorkList(I);
9416 I->eraseFromParent();;
9417 AddToWorkList(NewSCC);
9424 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
9425 Value *Cond = SI.getCondition();
9426 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
9427 if (I->getOpcode() == Instruction::Add)
9428 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
9429 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
9430 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
9431 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
9433 SI.setOperand(0, I->getOperand(0));
9441 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
9442 /// is to leave as a vector operation.
9443 static bool CheapToScalarize(Value *V, bool isConstant) {
9444 if (isa<ConstantAggregateZero>(V))
9446 if (ConstantVector *C = dyn_cast<ConstantVector>(V)) {
9447 if (isConstant) return true;
9448 // If all elts are the same, we can extract.
9449 Constant *Op0 = C->getOperand(0);
9450 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9451 if (C->getOperand(i) != Op0)
9455 Instruction *I = dyn_cast<Instruction>(V);
9456 if (!I) return false;
9458 // Insert element gets simplified to the inserted element or is deleted if
9459 // this is constant idx extract element and its a constant idx insertelt.
9460 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
9461 isa<ConstantInt>(I->getOperand(2)))
9463 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
9465 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
9466 if (BO->hasOneUse() &&
9467 (CheapToScalarize(BO->getOperand(0), isConstant) ||
9468 CheapToScalarize(BO->getOperand(1), isConstant)))
9470 if (CmpInst *CI = dyn_cast<CmpInst>(I))
9471 if (CI->hasOneUse() &&
9472 (CheapToScalarize(CI->getOperand(0), isConstant) ||
9473 CheapToScalarize(CI->getOperand(1), isConstant)))
9479 /// Read and decode a shufflevector mask.
9481 /// It turns undef elements into values that are larger than the number of
9482 /// elements in the input.
9483 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
9484 unsigned NElts = SVI->getType()->getNumElements();
9485 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
9486 return std::vector<unsigned>(NElts, 0);
9487 if (isa<UndefValue>(SVI->getOperand(2)))
9488 return std::vector<unsigned>(NElts, 2*NElts);
9490 std::vector<unsigned> Result;
9491 const ConstantVector *CP = cast<ConstantVector>(SVI->getOperand(2));
9492 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
9493 if (isa<UndefValue>(CP->getOperand(i)))
9494 Result.push_back(NElts*2); // undef -> 8
9496 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
9500 /// FindScalarElement - Given a vector and an element number, see if the scalar
9501 /// value is already around as a register, for example if it were inserted then
9502 /// extracted from the vector.
9503 static Value *FindScalarElement(Value *V, unsigned EltNo) {
9504 assert(isa<VectorType>(V->getType()) && "Not looking at a vector?");
9505 const VectorType *PTy = cast<VectorType>(V->getType());
9506 unsigned Width = PTy->getNumElements();
9507 if (EltNo >= Width) // Out of range access.
9508 return UndefValue::get(PTy->getElementType());
9510 if (isa<UndefValue>(V))
9511 return UndefValue::get(PTy->getElementType());
9512 else if (isa<ConstantAggregateZero>(V))
9513 return Constant::getNullValue(PTy->getElementType());
9514 else if (ConstantVector *CP = dyn_cast<ConstantVector>(V))
9515 return CP->getOperand(EltNo);
9516 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
9517 // If this is an insert to a variable element, we don't know what it is.
9518 if (!isa<ConstantInt>(III->getOperand(2)))
9520 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
9522 // If this is an insert to the element we are looking for, return the
9525 return III->getOperand(1);
9527 // Otherwise, the insertelement doesn't modify the value, recurse on its
9529 return FindScalarElement(III->getOperand(0), EltNo);
9530 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
9531 unsigned InEl = getShuffleMask(SVI)[EltNo];
9533 return FindScalarElement(SVI->getOperand(0), InEl);
9534 else if (InEl < Width*2)
9535 return FindScalarElement(SVI->getOperand(1), InEl - Width);
9537 return UndefValue::get(PTy->getElementType());
9540 // Otherwise, we don't know.
9544 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
9546 // If vector val is undef, replace extract with scalar undef.
9547 if (isa<UndefValue>(EI.getOperand(0)))
9548 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9550 // If vector val is constant 0, replace extract with scalar 0.
9551 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
9552 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
9554 if (ConstantVector *C = dyn_cast<ConstantVector>(EI.getOperand(0))) {
9555 // If vector val is constant with uniform operands, replace EI
9556 // with that operand
9557 Constant *op0 = C->getOperand(0);
9558 for (unsigned i = 1; i < C->getNumOperands(); ++i)
9559 if (C->getOperand(i) != op0) {
9564 return ReplaceInstUsesWith(EI, op0);
9567 // If extracting a specified index from the vector, see if we can recursively
9568 // find a previously computed scalar that was inserted into the vector.
9569 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9570 unsigned IndexVal = IdxC->getZExtValue();
9571 unsigned VectorWidth =
9572 cast<VectorType>(EI.getOperand(0)->getType())->getNumElements();
9574 // If this is extracting an invalid index, turn this into undef, to avoid
9575 // crashing the code below.
9576 if (IndexVal >= VectorWidth)
9577 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9579 // This instruction only demands the single element from the input vector.
9580 // If the input vector has a single use, simplify it based on this use
9582 if (EI.getOperand(0)->hasOneUse() && VectorWidth != 1) {
9584 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
9587 EI.setOperand(0, V);
9592 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
9593 return ReplaceInstUsesWith(EI, Elt);
9595 // If the this extractelement is directly using a bitcast from a vector of
9596 // the same number of elements, see if we can find the source element from
9597 // it. In this case, we will end up needing to bitcast the scalars.
9598 if (BitCastInst *BCI = dyn_cast<BitCastInst>(EI.getOperand(0))) {
9599 if (const VectorType *VT =
9600 dyn_cast<VectorType>(BCI->getOperand(0)->getType()))
9601 if (VT->getNumElements() == VectorWidth)
9602 if (Value *Elt = FindScalarElement(BCI->getOperand(0), IndexVal))
9603 return new BitCastInst(Elt, EI.getType());
9607 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
9608 if (I->hasOneUse()) {
9609 // Push extractelement into predecessor operation if legal and
9610 // profitable to do so
9611 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
9612 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
9613 if (CheapToScalarize(BO, isConstantElt)) {
9614 ExtractElementInst *newEI0 =
9615 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
9616 EI.getName()+".lhs");
9617 ExtractElementInst *newEI1 =
9618 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
9619 EI.getName()+".rhs");
9620 InsertNewInstBefore(newEI0, EI);
9621 InsertNewInstBefore(newEI1, EI);
9622 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
9624 } else if (isa<LoadInst>(I)) {
9625 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
9626 PointerType::get(EI.getType()), EI);
9627 GetElementPtrInst *GEP =
9628 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
9629 InsertNewInstBefore(GEP, EI);
9630 return new LoadInst(GEP);
9633 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
9634 // Extracting the inserted element?
9635 if (IE->getOperand(2) == EI.getOperand(1))
9636 return ReplaceInstUsesWith(EI, IE->getOperand(1));
9637 // If the inserted and extracted elements are constants, they must not
9638 // be the same value, extract from the pre-inserted value instead.
9639 if (isa<Constant>(IE->getOperand(2)) &&
9640 isa<Constant>(EI.getOperand(1))) {
9641 AddUsesToWorkList(EI);
9642 EI.setOperand(0, IE->getOperand(0));
9645 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
9646 // If this is extracting an element from a shufflevector, figure out where
9647 // it came from and extract from the appropriate input element instead.
9648 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
9649 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
9651 if (SrcIdx < SVI->getType()->getNumElements())
9652 Src = SVI->getOperand(0);
9653 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
9654 SrcIdx -= SVI->getType()->getNumElements();
9655 Src = SVI->getOperand(1);
9657 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
9659 return new ExtractElementInst(Src, SrcIdx);
9666 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
9667 /// elements from either LHS or RHS, return the shuffle mask and true.
9668 /// Otherwise, return false.
9669 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
9670 std::vector<Constant*> &Mask) {
9671 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
9672 "Invalid CollectSingleShuffleElements");
9673 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9675 if (isa<UndefValue>(V)) {
9676 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9678 } else if (V == LHS) {
9679 for (unsigned i = 0; i != NumElts; ++i)
9680 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9682 } else if (V == RHS) {
9683 for (unsigned i = 0; i != NumElts; ++i)
9684 Mask.push_back(ConstantInt::get(Type::Int32Ty, i+NumElts));
9686 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9687 // If this is an insert of an extract from some other vector, include it.
9688 Value *VecOp = IEI->getOperand(0);
9689 Value *ScalarOp = IEI->getOperand(1);
9690 Value *IdxOp = IEI->getOperand(2);
9692 if (!isa<ConstantInt>(IdxOp))
9694 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9696 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
9697 // Okay, we can handle this if the vector we are insertinting into is
9699 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9700 // If so, update the mask to reflect the inserted undef.
9701 Mask[InsertedIdx] = UndefValue::get(Type::Int32Ty);
9704 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
9705 if (isa<ConstantInt>(EI->getOperand(1)) &&
9706 EI->getOperand(0)->getType() == V->getType()) {
9707 unsigned ExtractedIdx =
9708 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9710 // This must be extracting from either LHS or RHS.
9711 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
9712 // Okay, we can handle this if the vector we are insertinting into is
9714 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
9715 // If so, update the mask to reflect the inserted value.
9716 if (EI->getOperand(0) == LHS) {
9717 Mask[InsertedIdx & (NumElts-1)] =
9718 ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9720 assert(EI->getOperand(0) == RHS);
9721 Mask[InsertedIdx & (NumElts-1)] =
9722 ConstantInt::get(Type::Int32Ty, ExtractedIdx+NumElts);
9731 // TODO: Handle shufflevector here!
9736 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
9737 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
9738 /// that computes V and the LHS value of the shuffle.
9739 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
9741 assert(isa<VectorType>(V->getType()) &&
9742 (RHS == 0 || V->getType() == RHS->getType()) &&
9743 "Invalid shuffle!");
9744 unsigned NumElts = cast<VectorType>(V->getType())->getNumElements();
9746 if (isa<UndefValue>(V)) {
9747 Mask.assign(NumElts, UndefValue::get(Type::Int32Ty));
9749 } else if (isa<ConstantAggregateZero>(V)) {
9750 Mask.assign(NumElts, ConstantInt::get(Type::Int32Ty, 0));
9752 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
9753 // If this is an insert of an extract from some other vector, include it.
9754 Value *VecOp = IEI->getOperand(0);
9755 Value *ScalarOp = IEI->getOperand(1);
9756 Value *IdxOp = IEI->getOperand(2);
9758 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9759 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9760 EI->getOperand(0)->getType() == V->getType()) {
9761 unsigned ExtractedIdx =
9762 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9763 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9765 // Either the extracted from or inserted into vector must be RHSVec,
9766 // otherwise we'd end up with a shuffle of three inputs.
9767 if (EI->getOperand(0) == RHS || RHS == 0) {
9768 RHS = EI->getOperand(0);
9769 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
9770 Mask[InsertedIdx & (NumElts-1)] =
9771 ConstantInt::get(Type::Int32Ty, NumElts+ExtractedIdx);
9776 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
9777 // Everything but the extracted element is replaced with the RHS.
9778 for (unsigned i = 0; i != NumElts; ++i) {
9779 if (i != InsertedIdx)
9780 Mask[i] = ConstantInt::get(Type::Int32Ty, NumElts+i);
9785 // If this insertelement is a chain that comes from exactly these two
9786 // vectors, return the vector and the effective shuffle.
9787 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
9788 return EI->getOperand(0);
9793 // TODO: Handle shufflevector here!
9795 // Otherwise, can't do anything fancy. Return an identity vector.
9796 for (unsigned i = 0; i != NumElts; ++i)
9797 Mask.push_back(ConstantInt::get(Type::Int32Ty, i));
9801 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
9802 Value *VecOp = IE.getOperand(0);
9803 Value *ScalarOp = IE.getOperand(1);
9804 Value *IdxOp = IE.getOperand(2);
9806 // Inserting an undef or into an undefined place, remove this.
9807 if (isa<UndefValue>(ScalarOp) || isa<UndefValue>(IdxOp))
9808 ReplaceInstUsesWith(IE, VecOp);
9810 // If the inserted element was extracted from some other vector, and if the
9811 // indexes are constant, try to turn this into a shufflevector operation.
9812 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
9813 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
9814 EI->getOperand(0)->getType() == IE.getType()) {
9815 unsigned NumVectorElts = IE.getType()->getNumElements();
9816 unsigned ExtractedIdx =
9817 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
9818 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
9820 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
9821 return ReplaceInstUsesWith(IE, VecOp);
9823 if (InsertedIdx >= NumVectorElts) // Out of range insert.
9824 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
9826 // If we are extracting a value from a vector, then inserting it right
9827 // back into the same place, just use the input vector.
9828 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
9829 return ReplaceInstUsesWith(IE, VecOp);
9831 // We could theoretically do this for ANY input. However, doing so could
9832 // turn chains of insertelement instructions into a chain of shufflevector
9833 // instructions, and right now we do not merge shufflevectors. As such,
9834 // only do this in a situation where it is clear that there is benefit.
9835 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
9836 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
9837 // the values of VecOp, except then one read from EIOp0.
9838 // Build a new shuffle mask.
9839 std::vector<Constant*> Mask;
9840 if (isa<UndefValue>(VecOp))
9841 Mask.assign(NumVectorElts, UndefValue::get(Type::Int32Ty));
9843 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
9844 Mask.assign(NumVectorElts, ConstantInt::get(Type::Int32Ty,
9847 Mask[InsertedIdx] = ConstantInt::get(Type::Int32Ty, ExtractedIdx);
9848 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
9849 ConstantVector::get(Mask));
9852 // If this insertelement isn't used by some other insertelement, turn it
9853 // (and any insertelements it points to), into one big shuffle.
9854 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
9855 std::vector<Constant*> Mask;
9857 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
9858 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
9859 // We now have a shuffle of LHS, RHS, Mask.
9860 return new ShuffleVectorInst(LHS, RHS, ConstantVector::get(Mask));
9869 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
9870 Value *LHS = SVI.getOperand(0);
9871 Value *RHS = SVI.getOperand(1);
9872 std::vector<unsigned> Mask = getShuffleMask(&SVI);
9874 bool MadeChange = false;
9876 // Undefined shuffle mask -> undefined value.
9877 if (isa<UndefValue>(SVI.getOperand(2)))
9878 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
9880 // If we have shuffle(x, undef, mask) and any elements of mask refer to
9881 // the undef, change them to undefs.
9882 if (isa<UndefValue>(SVI.getOperand(1))) {
9883 // Scan to see if there are any references to the RHS. If so, replace them
9884 // with undef element refs and set MadeChange to true.
9885 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9886 if (Mask[i] >= e && Mask[i] != 2*e) {
9893 // Remap any references to RHS to use LHS.
9894 std::vector<Constant*> Elts;
9895 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9897 Elts.push_back(UndefValue::get(Type::Int32Ty));
9899 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9901 SVI.setOperand(2, ConstantVector::get(Elts));
9905 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
9906 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
9907 if (LHS == RHS || isa<UndefValue>(LHS)) {
9908 if (isa<UndefValue>(LHS) && LHS == RHS) {
9909 // shuffle(undef,undef,mask) -> undef.
9910 return ReplaceInstUsesWith(SVI, LHS);
9913 // Remap any references to RHS to use LHS.
9914 std::vector<Constant*> Elts;
9915 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9917 Elts.push_back(UndefValue::get(Type::Int32Ty));
9919 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
9920 (Mask[i] < e && isa<UndefValue>(LHS)))
9921 Mask[i] = 2*e; // Turn into undef.
9923 Mask[i] &= (e-1); // Force to LHS.
9924 Elts.push_back(ConstantInt::get(Type::Int32Ty, Mask[i]));
9927 SVI.setOperand(0, SVI.getOperand(1));
9928 SVI.setOperand(1, UndefValue::get(RHS->getType()));
9929 SVI.setOperand(2, ConstantVector::get(Elts));
9930 LHS = SVI.getOperand(0);
9931 RHS = SVI.getOperand(1);
9935 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
9936 bool isLHSID = true, isRHSID = true;
9938 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
9939 if (Mask[i] >= e*2) continue; // Ignore undef values.
9940 // Is this an identity shuffle of the LHS value?
9941 isLHSID &= (Mask[i] == i);
9943 // Is this an identity shuffle of the RHS value?
9944 isRHSID &= (Mask[i]-e == i);
9947 // Eliminate identity shuffles.
9948 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
9949 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
9951 // If the LHS is a shufflevector itself, see if we can combine it with this
9952 // one without producing an unusual shuffle. Here we are really conservative:
9953 // we are absolutely afraid of producing a shuffle mask not in the input
9954 // program, because the code gen may not be smart enough to turn a merged
9955 // shuffle into two specific shuffles: it may produce worse code. As such,
9956 // we only merge two shuffles if the result is one of the two input shuffle
9957 // masks. In this case, merging the shuffles just removes one instruction,
9958 // which we know is safe. This is good for things like turning:
9959 // (splat(splat)) -> splat.
9960 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
9961 if (isa<UndefValue>(RHS)) {
9962 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
9964 std::vector<unsigned> NewMask;
9965 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
9967 NewMask.push_back(2*e);
9969 NewMask.push_back(LHSMask[Mask[i]]);
9971 // If the result mask is equal to the src shuffle or this shuffle mask, do
9973 if (NewMask == LHSMask || NewMask == Mask) {
9974 std::vector<Constant*> Elts;
9975 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
9976 if (NewMask[i] >= e*2) {
9977 Elts.push_back(UndefValue::get(Type::Int32Ty));
9979 Elts.push_back(ConstantInt::get(Type::Int32Ty, NewMask[i]));
9982 return new ShuffleVectorInst(LHSSVI->getOperand(0),
9983 LHSSVI->getOperand(1),
9984 ConstantVector::get(Elts));
9989 return MadeChange ? &SVI : 0;
9995 /// TryToSinkInstruction - Try to move the specified instruction from its
9996 /// current block into the beginning of DestBlock, which can only happen if it's
9997 /// safe to move the instruction past all of the instructions between it and the
9998 /// end of its block.
9999 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
10000 assert(I->hasOneUse() && "Invariants didn't hold!");
10002 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
10003 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
10005 // Do not sink alloca instructions out of the entry block.
10006 if (isa<AllocaInst>(I) && I->getParent() ==
10007 &DestBlock->getParent()->getEntryBlock())
10010 // We can only sink load instructions if there is nothing between the load and
10011 // the end of block that could change the value.
10012 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
10013 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
10015 if (Scan->mayWriteToMemory())
10019 BasicBlock::iterator InsertPos = DestBlock->begin();
10020 while (isa<PHINode>(InsertPos)) ++InsertPos;
10022 I->moveBefore(InsertPos);
10028 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
10029 /// all reachable code to the worklist.
10031 /// This has a couple of tricks to make the code faster and more powerful. In
10032 /// particular, we constant fold and DCE instructions as we go, to avoid adding
10033 /// them to the worklist (this significantly speeds up instcombine on code where
10034 /// many instructions are dead or constant). Additionally, if we find a branch
10035 /// whose condition is a known constant, we only visit the reachable successors.
10037 static void AddReachableCodeToWorklist(BasicBlock *BB,
10038 SmallPtrSet<BasicBlock*, 64> &Visited,
10040 const TargetData *TD) {
10041 std::vector<BasicBlock*> Worklist;
10042 Worklist.push_back(BB);
10044 while (!Worklist.empty()) {
10045 BB = Worklist.back();
10046 Worklist.pop_back();
10048 // We have now visited this block! If we've already been here, ignore it.
10049 if (!Visited.insert(BB)) continue;
10051 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
10052 Instruction *Inst = BBI++;
10054 // DCE instruction if trivially dead.
10055 if (isInstructionTriviallyDead(Inst)) {
10057 DOUT << "IC: DCE: " << *Inst;
10058 Inst->eraseFromParent();
10062 // ConstantProp instruction if trivially constant.
10063 if (Constant *C = ConstantFoldInstruction(Inst, TD)) {
10064 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
10065 Inst->replaceAllUsesWith(C);
10067 Inst->eraseFromParent();
10071 IC.AddToWorkList(Inst);
10074 // Recursively visit successors. If this is a branch or switch on a
10075 // constant, only visit the reachable successor.
10076 TerminatorInst *TI = BB->getTerminator();
10077 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
10078 if (BI->isConditional() && isa<ConstantInt>(BI->getCondition())) {
10079 bool CondVal = cast<ConstantInt>(BI->getCondition())->getZExtValue();
10080 Worklist.push_back(BI->getSuccessor(!CondVal));
10083 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
10084 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
10085 // See if this is an explicit destination.
10086 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
10087 if (SI->getCaseValue(i) == Cond) {
10088 Worklist.push_back(SI->getSuccessor(i));
10092 // Otherwise it is the default destination.
10093 Worklist.push_back(SI->getSuccessor(0));
10098 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
10099 Worklist.push_back(TI->getSuccessor(i));
10103 bool InstCombiner::DoOneIteration(Function &F, unsigned Iteration) {
10104 bool Changed = false;
10105 TD = &getAnalysis<TargetData>();
10107 DEBUG(DOUT << "\n\nINSTCOMBINE ITERATION #" << Iteration << " on "
10108 << F.getNameStr() << "\n");
10111 // Do a depth-first traversal of the function, populate the worklist with
10112 // the reachable instructions. Ignore blocks that are not reachable. Keep
10113 // track of which blocks we visit.
10114 SmallPtrSet<BasicBlock*, 64> Visited;
10115 AddReachableCodeToWorklist(F.begin(), Visited, *this, TD);
10117 // Do a quick scan over the function. If we find any blocks that are
10118 // unreachable, remove any instructions inside of them. This prevents
10119 // the instcombine code from having to deal with some bad special cases.
10120 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
10121 if (!Visited.count(BB)) {
10122 Instruction *Term = BB->getTerminator();
10123 while (Term != BB->begin()) { // Remove instrs bottom-up
10124 BasicBlock::iterator I = Term; --I;
10126 DOUT << "IC: DCE: " << *I;
10129 if (!I->use_empty())
10130 I->replaceAllUsesWith(UndefValue::get(I->getType()));
10131 I->eraseFromParent();
10136 while (!Worklist.empty()) {
10137 Instruction *I = RemoveOneFromWorkList();
10138 if (I == 0) continue; // skip null values.
10140 // Check to see if we can DCE the instruction.
10141 if (isInstructionTriviallyDead(I)) {
10142 // Add operands to the worklist.
10143 if (I->getNumOperands() < 4)
10144 AddUsesToWorkList(*I);
10147 DOUT << "IC: DCE: " << *I;
10149 I->eraseFromParent();
10150 RemoveFromWorkList(I);
10154 // Instruction isn't dead, see if we can constant propagate it.
10155 if (Constant *C = ConstantFoldInstruction(I, TD)) {
10156 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
10158 // Add operands to the worklist.
10159 AddUsesToWorkList(*I);
10160 ReplaceInstUsesWith(*I, C);
10163 I->eraseFromParent();
10164 RemoveFromWorkList(I);
10168 // See if we can trivially sink this instruction to a successor basic block.
10169 if (I->hasOneUse()) {
10170 BasicBlock *BB = I->getParent();
10171 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
10172 if (UserParent != BB) {
10173 bool UserIsSuccessor = false;
10174 // See if the user is one of our successors.
10175 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
10176 if (*SI == UserParent) {
10177 UserIsSuccessor = true;
10181 // If the user is one of our immediate successors, and if that successor
10182 // only has us as a predecessors (we'd have to split the critical edge
10183 // otherwise), we can keep going.
10184 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
10185 next(pred_begin(UserParent)) == pred_end(UserParent))
10186 // Okay, the CFG is simple enough, try to sink this instruction.
10187 Changed |= TryToSinkInstruction(I, UserParent);
10191 // Now that we have an instruction, try combining it to simplify it...
10195 DEBUG(std::ostringstream SS; I->print(SS); OrigI = SS.str(););
10196 if (Instruction *Result = visit(*I)) {
10198 // Should we replace the old instruction with a new one?
10200 DOUT << "IC: Old = " << *I
10201 << " New = " << *Result;
10203 // Everything uses the new instruction now.
10204 I->replaceAllUsesWith(Result);
10206 // Push the new instruction and any users onto the worklist.
10207 AddToWorkList(Result);
10208 AddUsersToWorkList(*Result);
10210 // Move the name to the new instruction first.
10211 Result->takeName(I);
10213 // Insert the new instruction into the basic block...
10214 BasicBlock *InstParent = I->getParent();
10215 BasicBlock::iterator InsertPos = I;
10217 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
10218 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
10221 InstParent->getInstList().insert(InsertPos, Result);
10223 // Make sure that we reprocess all operands now that we reduced their
10225 AddUsesToWorkList(*I);
10227 // Instructions can end up on the worklist more than once. Make sure
10228 // we do not process an instruction that has been deleted.
10229 RemoveFromWorkList(I);
10231 // Erase the old instruction.
10232 InstParent->getInstList().erase(I);
10235 DOUT << "IC: Mod = " << OrigI
10236 << " New = " << *I;
10239 // If the instruction was modified, it's possible that it is now dead.
10240 // if so, remove it.
10241 if (isInstructionTriviallyDead(I)) {
10242 // Make sure we process all operands now that we are reducing their
10244 AddUsesToWorkList(*I);
10246 // Instructions may end up in the worklist more than once. Erase all
10247 // occurrences of this instruction.
10248 RemoveFromWorkList(I);
10249 I->eraseFromParent();
10252 AddUsersToWorkList(*I);
10259 assert(WorklistMap.empty() && "Worklist empty, but map not?");
10261 // Do an explicit clear, this shrinks the map if needed.
10262 WorklistMap.clear();
10267 bool InstCombiner::runOnFunction(Function &F) {
10268 MustPreserveLCSSA = mustPreserveAnalysisID(LCSSAID);
10270 bool EverMadeChange = false;
10272 // Iterate while there is work to do.
10273 unsigned Iteration = 0;
10274 while (DoOneIteration(F, Iteration++))
10275 EverMadeChange = true;
10276 return EverMadeChange;
10279 FunctionPass *llvm::createInstructionCombiningPass() {
10280 return new InstCombiner();