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. SetCC instructions are converted from <,>,<=,>= to ==,!= if possible
28 // 4. All SetCC 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/Target/TargetData.h"
43 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
44 #include "llvm/Transforms/Utils/Local.h"
45 #include "llvm/Support/CallSite.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/GetElementPtrTypeIterator.h"
48 #include "llvm/Support/InstVisitor.h"
49 #include "llvm/Support/MathExtras.h"
50 #include "llvm/Support/PatternMatch.h"
51 #include "llvm/Support/Compiler.h"
52 #include "llvm/ADT/Statistic.h"
53 #include "llvm/ADT/STLExtras.h"
56 using namespace llvm::PatternMatch;
58 STATISTIC(NumCombined , "Number of insts combined");
59 STATISTIC(NumConstProp, "Number of constant folds");
60 STATISTIC(NumDeadInst , "Number of dead inst eliminated");
61 STATISTIC(NumDeadStore, "Number of dead stores eliminated");
62 STATISTIC(NumSunkInst , "Number of instructions sunk");
65 class VISIBILITY_HIDDEN InstCombiner
66 : public FunctionPass,
67 public InstVisitor<InstCombiner, Instruction*> {
68 // Worklist of all of the instructions that need to be simplified.
69 std::vector<Instruction*> WorkList;
72 /// AddUsersToWorkList - When an instruction is simplified, add all users of
73 /// the instruction to the work lists because they might get more simplified
76 void AddUsersToWorkList(Value &I) {
77 for (Value::use_iterator UI = I.use_begin(), UE = I.use_end();
79 WorkList.push_back(cast<Instruction>(*UI));
82 /// AddUsesToWorkList - When an instruction is simplified, add operands to
83 /// the work lists because they might get more simplified now.
85 void AddUsesToWorkList(Instruction &I) {
86 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
87 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i)))
88 WorkList.push_back(Op);
91 /// AddSoonDeadInstToWorklist - The specified instruction is about to become
92 /// dead. Add all of its operands to the worklist, turning them into
93 /// undef's to reduce the number of uses of those instructions.
95 /// Return the specified operand before it is turned into an undef.
97 Value *AddSoonDeadInstToWorklist(Instruction &I, unsigned op) {
98 Value *R = I.getOperand(op);
100 for (unsigned i = 0, e = I.getNumOperands(); i != e; ++i)
101 if (Instruction *Op = dyn_cast<Instruction>(I.getOperand(i))) {
102 WorkList.push_back(Op);
103 // Set the operand to undef to drop the use.
104 I.setOperand(i, UndefValue::get(Op->getType()));
110 // removeFromWorkList - remove all instances of I from the worklist.
111 void removeFromWorkList(Instruction *I);
113 virtual bool runOnFunction(Function &F);
115 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
116 AU.addRequired<TargetData>();
117 AU.addPreservedID(LCSSAID);
118 AU.setPreservesCFG();
121 TargetData &getTargetData() const { return *TD; }
123 // Visitation implementation - Implement instruction combining for different
124 // instruction types. The semantics are as follows:
126 // null - No change was made
127 // I - Change was made, I is still valid, I may be dead though
128 // otherwise - Change was made, replace I with returned instruction
130 Instruction *visitAdd(BinaryOperator &I);
131 Instruction *visitSub(BinaryOperator &I);
132 Instruction *visitMul(BinaryOperator &I);
133 Instruction *visitURem(BinaryOperator &I);
134 Instruction *visitSRem(BinaryOperator &I);
135 Instruction *visitFRem(BinaryOperator &I);
136 Instruction *commonRemTransforms(BinaryOperator &I);
137 Instruction *commonIRemTransforms(BinaryOperator &I);
138 Instruction *commonDivTransforms(BinaryOperator &I);
139 Instruction *commonIDivTransforms(BinaryOperator &I);
140 Instruction *visitUDiv(BinaryOperator &I);
141 Instruction *visitSDiv(BinaryOperator &I);
142 Instruction *visitFDiv(BinaryOperator &I);
143 Instruction *visitAnd(BinaryOperator &I);
144 Instruction *visitOr (BinaryOperator &I);
145 Instruction *visitXor(BinaryOperator &I);
146 Instruction *visitSetCondInst(SetCondInst &I);
147 Instruction *visitSetCondInstWithCastAndCast(SetCondInst &SCI);
149 Instruction *FoldGEPSetCC(User *GEPLHS, Value *RHS,
150 Instruction::BinaryOps Cond, Instruction &I);
151 Instruction *visitShiftInst(ShiftInst &I);
152 Instruction *FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
154 Instruction *commonCastTransforms(CastInst &CI);
155 Instruction *commonIntCastTransforms(CastInst &CI);
156 Instruction *visitTrunc(CastInst &CI);
157 Instruction *visitZExt(CastInst &CI);
158 Instruction *visitSExt(CastInst &CI);
159 Instruction *visitFPTrunc(CastInst &CI);
160 Instruction *visitFPExt(CastInst &CI);
161 Instruction *visitFPToUI(CastInst &CI);
162 Instruction *visitFPToSI(CastInst &CI);
163 Instruction *visitUIToFP(CastInst &CI);
164 Instruction *visitSIToFP(CastInst &CI);
165 Instruction *visitPtrToInt(CastInst &CI);
166 Instruction *visitIntToPtr(CastInst &CI);
167 Instruction *visitBitCast(CastInst &CI);
168 Instruction *FoldSelectOpOp(SelectInst &SI, Instruction *TI,
170 Instruction *visitSelectInst(SelectInst &CI);
171 Instruction *visitCallInst(CallInst &CI);
172 Instruction *visitInvokeInst(InvokeInst &II);
173 Instruction *visitPHINode(PHINode &PN);
174 Instruction *visitGetElementPtrInst(GetElementPtrInst &GEP);
175 Instruction *visitAllocationInst(AllocationInst &AI);
176 Instruction *visitFreeInst(FreeInst &FI);
177 Instruction *visitLoadInst(LoadInst &LI);
178 Instruction *visitStoreInst(StoreInst &SI);
179 Instruction *visitBranchInst(BranchInst &BI);
180 Instruction *visitSwitchInst(SwitchInst &SI);
181 Instruction *visitInsertElementInst(InsertElementInst &IE);
182 Instruction *visitExtractElementInst(ExtractElementInst &EI);
183 Instruction *visitShuffleVectorInst(ShuffleVectorInst &SVI);
185 // visitInstruction - Specify what to return for unhandled instructions...
186 Instruction *visitInstruction(Instruction &I) { return 0; }
189 Instruction *visitCallSite(CallSite CS);
190 bool transformConstExprCastCall(CallSite CS);
193 // InsertNewInstBefore - insert an instruction New before instruction Old
194 // in the program. Add the new instruction to the worklist.
196 Instruction *InsertNewInstBefore(Instruction *New, Instruction &Old) {
197 assert(New && New->getParent() == 0 &&
198 "New instruction already inserted into a basic block!");
199 BasicBlock *BB = Old.getParent();
200 BB->getInstList().insert(&Old, New); // Insert inst
201 WorkList.push_back(New); // Add to worklist
205 /// InsertCastBefore - Insert a cast of V to TY before the instruction POS.
206 /// This also adds the cast to the worklist. Finally, this returns the
208 Value *InsertCastBefore(Instruction::CastOps opc, Value *V, const Type *Ty,
210 if (V->getType() == Ty) return V;
212 if (Constant *CV = dyn_cast<Constant>(V))
213 return ConstantExpr::getCast(opc, CV, Ty);
215 Instruction *C = CastInst::create(opc, V, Ty, V->getName(), &Pos);
216 WorkList.push_back(C);
220 // ReplaceInstUsesWith - This method is to be used when an instruction is
221 // found to be dead, replacable with another preexisting expression. Here
222 // we add all uses of I to the worklist, replace all uses of I with the new
223 // value, then return I, so that the inst combiner will know that I was
226 Instruction *ReplaceInstUsesWith(Instruction &I, Value *V) {
227 AddUsersToWorkList(I); // Add all modified instrs to worklist
229 I.replaceAllUsesWith(V);
232 // If we are replacing the instruction with itself, this must be in a
233 // segment of unreachable code, so just clobber the instruction.
234 I.replaceAllUsesWith(UndefValue::get(I.getType()));
239 // UpdateValueUsesWith - This method is to be used when an value is
240 // found to be replacable with another preexisting expression or was
241 // updated. Here we add all uses of I to the worklist, replace all uses of
242 // I with the new value (unless the instruction was just updated), then
243 // return true, so that the inst combiner will know that I was modified.
245 bool UpdateValueUsesWith(Value *Old, Value *New) {
246 AddUsersToWorkList(*Old); // Add all modified instrs to worklist
248 Old->replaceAllUsesWith(New);
249 if (Instruction *I = dyn_cast<Instruction>(Old))
250 WorkList.push_back(I);
251 if (Instruction *I = dyn_cast<Instruction>(New))
252 WorkList.push_back(I);
256 // EraseInstFromFunction - When dealing with an instruction that has side
257 // effects or produces a void value, we can't rely on DCE to delete the
258 // instruction. Instead, visit methods should return the value returned by
260 Instruction *EraseInstFromFunction(Instruction &I) {
261 assert(I.use_empty() && "Cannot erase instruction that is used!");
262 AddUsesToWorkList(I);
263 removeFromWorkList(&I);
265 return 0; // Don't do anything with FI
269 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
270 /// InsertBefore instruction. This is specialized a bit to avoid inserting
271 /// casts that are known to not do anything...
273 Value *InsertOperandCastBefore(Instruction::CastOps opcode,
274 Value *V, const Type *DestTy,
275 Instruction *InsertBefore);
277 // SimplifyCommutative - This performs a few simplifications for commutative
279 bool SimplifyCommutative(BinaryOperator &I);
281 bool SimplifyDemandedBits(Value *V, uint64_t Mask,
282 uint64_t &KnownZero, uint64_t &KnownOne,
285 Value *SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
286 uint64_t &UndefElts, unsigned Depth = 0);
288 // FoldOpIntoPhi - Given a binary operator or cast instruction which has a
289 // PHI node as operand #0, see if we can fold the instruction into the PHI
290 // (which is only possible if all operands to the PHI are constants).
291 Instruction *FoldOpIntoPhi(Instruction &I);
293 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
294 // operator and they all are only used by the PHI, PHI together their
295 // inputs, and do the operation once, to the result of the PHI.
296 Instruction *FoldPHIArgOpIntoPHI(PHINode &PN);
297 Instruction *FoldPHIArgBinOpIntoPHI(PHINode &PN);
300 Instruction *OptAndOp(Instruction *Op, ConstantIntegral *OpRHS,
301 ConstantIntegral *AndRHS, BinaryOperator &TheAnd);
303 Value *FoldLogicalPlusAnd(Value *LHS, Value *RHS, ConstantIntegral *Mask,
304 bool isSub, Instruction &I);
305 Instruction *InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
306 bool Inside, Instruction &IB);
307 Instruction *PromoteCastOfAllocation(CastInst &CI, AllocationInst &AI);
308 Instruction *MatchBSwap(BinaryOperator &I);
310 Value *EvaluateInDifferentType(Value *V, const Type *Ty, bool isSigned);
313 RegisterPass<InstCombiner> X("instcombine", "Combine redundant instructions");
316 // getComplexity: Assign a complexity or rank value to LLVM Values...
317 // 0 -> undef, 1 -> Const, 2 -> Other, 3 -> Arg, 3 -> Unary, 4 -> OtherInst
318 static unsigned getComplexity(Value *V) {
319 if (isa<Instruction>(V)) {
320 if (BinaryOperator::isNeg(V) || BinaryOperator::isNot(V))
324 if (isa<Argument>(V)) return 3;
325 return isa<Constant>(V) ? (isa<UndefValue>(V) ? 0 : 1) : 2;
328 // isOnlyUse - Return true if this instruction will be deleted if we stop using
330 static bool isOnlyUse(Value *V) {
331 return V->hasOneUse() || isa<Constant>(V);
334 // getPromotedType - Return the specified type promoted as it would be to pass
335 // though a va_arg area...
336 static const Type *getPromotedType(const Type *Ty) {
337 switch (Ty->getTypeID()) {
338 case Type::SByteTyID:
339 case Type::ShortTyID: return Type::IntTy;
340 case Type::UByteTyID:
341 case Type::UShortTyID: return Type::UIntTy;
342 case Type::FloatTyID: return Type::DoubleTy;
347 /// getBitCastOperand - If the specified operand is a CastInst or a constant
348 /// expression bitcast, return the operand value, otherwise return null.
349 static Value *getBitCastOperand(Value *V) {
350 if (BitCastInst *I = dyn_cast<BitCastInst>(V))
351 return I->getOperand(0);
352 else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
353 if (CE->getOpcode() == Instruction::BitCast)
354 return CE->getOperand(0);
358 /// This function is a wrapper around CastInst::isEliminableCastPair. It
359 /// simply extracts arguments and returns what that function returns.
360 /// @Determine if it is valid to eliminate a Convert pair
361 static Instruction::CastOps
362 isEliminableCastPair(
363 const CastInst *CI, ///< The first cast instruction
364 unsigned opcode, ///< The opcode of the second cast instruction
365 const Type *DstTy, ///< The target type for the second cast instruction
366 TargetData *TD ///< The target data for pointer size
369 const Type *SrcTy = CI->getOperand(0)->getType(); // A from above
370 const Type *MidTy = CI->getType(); // B from above
372 // Get the opcodes of the two Cast instructions
373 Instruction::CastOps firstOp = Instruction::CastOps(CI->getOpcode());
374 Instruction::CastOps secondOp = Instruction::CastOps(opcode);
376 return Instruction::CastOps(
377 CastInst::isEliminableCastPair(firstOp, secondOp, SrcTy, MidTy,
378 DstTy, TD->getIntPtrType()));
381 /// ValueRequiresCast - Return true if the cast from "V to Ty" actually results
382 /// in any code being generated. It does not require codegen if V is simple
383 /// enough or if the cast can be folded into other casts.
384 static bool ValueRequiresCast(const Value *V, const Type *Ty, TargetData *TD) {
385 if (V->getType() == Ty || isa<Constant>(V)) return false;
387 // If this is a noop cast, it isn't real codegen.
388 if (V->getType()->canLosslesslyBitCastTo(Ty))
391 // If this is another cast that can be eliminated, it isn't codegen either.
392 if (const CastInst *CI = dyn_cast<CastInst>(V))
393 if (isEliminableCastPair(CI, CastInst::getCastOpcode(
394 V, V->getType()->isSigned(), Ty, Ty->isSigned()), Ty, TD))
399 /// InsertOperandCastBefore - This inserts a cast of V to DestTy before the
400 /// InsertBefore instruction. This is specialized a bit to avoid inserting
401 /// casts that are known to not do anything...
403 Value *InstCombiner::InsertOperandCastBefore(Instruction::CastOps opcode,
404 Value *V, const Type *DestTy,
405 Instruction *InsertBefore) {
406 if (V->getType() == DestTy) return V;
407 if (Constant *C = dyn_cast<Constant>(V))
408 return ConstantExpr::getCast(opcode, C, DestTy);
410 return InsertCastBefore(opcode, V, DestTy, *InsertBefore);
413 // SimplifyCommutative - This performs a few simplifications for commutative
416 // 1. Order operands such that they are listed from right (least complex) to
417 // left (most complex). This puts constants before unary operators before
420 // 2. Transform: (op (op V, C1), C2) ==> (op V, (op C1, C2))
421 // 3. Transform: (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
423 bool InstCombiner::SimplifyCommutative(BinaryOperator &I) {
424 bool Changed = false;
425 if (getComplexity(I.getOperand(0)) < getComplexity(I.getOperand(1)))
426 Changed = !I.swapOperands();
428 if (!I.isAssociative()) return Changed;
429 Instruction::BinaryOps Opcode = I.getOpcode();
430 if (BinaryOperator *Op = dyn_cast<BinaryOperator>(I.getOperand(0)))
431 if (Op->getOpcode() == Opcode && isa<Constant>(Op->getOperand(1))) {
432 if (isa<Constant>(I.getOperand(1))) {
433 Constant *Folded = ConstantExpr::get(I.getOpcode(),
434 cast<Constant>(I.getOperand(1)),
435 cast<Constant>(Op->getOperand(1)));
436 I.setOperand(0, Op->getOperand(0));
437 I.setOperand(1, Folded);
439 } else if (BinaryOperator *Op1=dyn_cast<BinaryOperator>(I.getOperand(1)))
440 if (Op1->getOpcode() == Opcode && isa<Constant>(Op1->getOperand(1)) &&
441 isOnlyUse(Op) && isOnlyUse(Op1)) {
442 Constant *C1 = cast<Constant>(Op->getOperand(1));
443 Constant *C2 = cast<Constant>(Op1->getOperand(1));
445 // Fold (op (op V1, C1), (op V2, C2)) ==> (op (op V1, V2), (op C1,C2))
446 Constant *Folded = ConstantExpr::get(I.getOpcode(), C1, C2);
447 Instruction *New = BinaryOperator::create(Opcode, Op->getOperand(0),
450 WorkList.push_back(New);
451 I.setOperand(0, New);
452 I.setOperand(1, Folded);
459 // dyn_castNegVal - Given a 'sub' instruction, return the RHS of the instruction
460 // if the LHS is a constant zero (which is the 'negate' form).
462 static inline Value *dyn_castNegVal(Value *V) {
463 if (BinaryOperator::isNeg(V))
464 return BinaryOperator::getNegArgument(V);
466 // Constants can be considered to be negated values if they can be folded.
467 if (ConstantInt *C = dyn_cast<ConstantInt>(V))
468 return ConstantExpr::getNeg(C);
472 static inline Value *dyn_castNotVal(Value *V) {
473 if (BinaryOperator::isNot(V))
474 return BinaryOperator::getNotArgument(V);
476 // Constants can be considered to be not'ed values...
477 if (ConstantIntegral *C = dyn_cast<ConstantIntegral>(V))
478 return ConstantExpr::getNot(C);
482 // dyn_castFoldableMul - If this value is a multiply that can be folded into
483 // other computations (because it has a constant operand), return the
484 // non-constant operand of the multiply, and set CST to point to the multiplier.
485 // Otherwise, return null.
487 static inline Value *dyn_castFoldableMul(Value *V, ConstantInt *&CST) {
488 if (V->hasOneUse() && V->getType()->isInteger())
489 if (Instruction *I = dyn_cast<Instruction>(V)) {
490 if (I->getOpcode() == Instruction::Mul)
491 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1))))
492 return I->getOperand(0);
493 if (I->getOpcode() == Instruction::Shl)
494 if ((CST = dyn_cast<ConstantInt>(I->getOperand(1)))) {
495 // The multiplier is really 1 << CST.
496 Constant *One = ConstantInt::get(V->getType(), 1);
497 CST = cast<ConstantInt>(ConstantExpr::getShl(One, CST));
498 return I->getOperand(0);
504 /// dyn_castGetElementPtr - If this is a getelementptr instruction or constant
505 /// expression, return it.
506 static User *dyn_castGetElementPtr(Value *V) {
507 if (isa<GetElementPtrInst>(V)) return cast<User>(V);
508 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
509 if (CE->getOpcode() == Instruction::GetElementPtr)
510 return cast<User>(V);
514 // AddOne, SubOne - Add or subtract a constant one from an integer constant...
515 static ConstantInt *AddOne(ConstantInt *C) {
516 return cast<ConstantInt>(ConstantExpr::getAdd(C,
517 ConstantInt::get(C->getType(), 1)));
519 static ConstantInt *SubOne(ConstantInt *C) {
520 return cast<ConstantInt>(ConstantExpr::getSub(C,
521 ConstantInt::get(C->getType(), 1)));
524 /// GetConstantInType - Return a ConstantInt with the specified type and value.
526 static ConstantIntegral *GetConstantInType(const Type *Ty, uint64_t Val) {
527 if (Ty->isUnsigned())
528 return ConstantInt::get(Ty, Val);
529 else if (Ty->getTypeID() == Type::BoolTyID)
530 return ConstantBool::get(Val);
532 SVal <<= 64-Ty->getPrimitiveSizeInBits();
533 SVal >>= 64-Ty->getPrimitiveSizeInBits();
534 return ConstantInt::get(Ty, SVal);
538 /// ComputeMaskedBits - Determine which of the bits specified in Mask are
539 /// known to be either zero or one and return them in the KnownZero/KnownOne
540 /// bitsets. This code only analyzes bits in Mask, in order to short-circuit
542 static void ComputeMaskedBits(Value *V, uint64_t Mask, uint64_t &KnownZero,
543 uint64_t &KnownOne, unsigned Depth = 0) {
544 // Note, we cannot consider 'undef' to be "IsZero" here. The problem is that
545 // we cannot optimize based on the assumption that it is zero without changing
546 // it to be an explicit zero. If we don't change it to zero, other code could
547 // optimized based on the contradictory assumption that it is non-zero.
548 // Because instcombine aggressively folds operations with undef args anyway,
549 // this won't lose us code quality.
550 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
551 // We know all of the bits for a constant!
552 KnownOne = CI->getZExtValue() & Mask;
553 KnownZero = ~KnownOne & Mask;
557 KnownZero = KnownOne = 0; // Don't know anything.
558 if (Depth == 6 || Mask == 0)
559 return; // Limit search depth.
561 uint64_t KnownZero2, KnownOne2;
562 Instruction *I = dyn_cast<Instruction>(V);
565 Mask &= V->getType()->getIntegralTypeMask();
567 switch (I->getOpcode()) {
568 case Instruction::And:
569 // If either the LHS or the RHS are Zero, the result is zero.
570 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
572 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
573 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
574 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
576 // Output known-1 bits are only known if set in both the LHS & RHS.
577 KnownOne &= KnownOne2;
578 // Output known-0 are known to be clear if zero in either the LHS | RHS.
579 KnownZero |= KnownZero2;
581 case Instruction::Or:
582 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
584 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
585 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
586 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
588 // Output known-0 bits are only known if clear in both the LHS & RHS.
589 KnownZero &= KnownZero2;
590 // Output known-1 are known to be set if set in either the LHS | RHS.
591 KnownOne |= KnownOne2;
593 case Instruction::Xor: {
594 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero, KnownOne, Depth+1);
595 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero2, KnownOne2, Depth+1);
596 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
597 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
599 // Output known-0 bits are known if clear or set in both the LHS & RHS.
600 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
601 // Output known-1 are known to be set if set in only one of the LHS, RHS.
602 KnownOne = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
603 KnownZero = KnownZeroOut;
606 case Instruction::Select:
607 ComputeMaskedBits(I->getOperand(2), Mask, KnownZero, KnownOne, Depth+1);
608 ComputeMaskedBits(I->getOperand(1), Mask, KnownZero2, KnownOne2, Depth+1);
609 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
610 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
612 // Only known if known in both the LHS and RHS.
613 KnownOne &= KnownOne2;
614 KnownZero &= KnownZero2;
616 case Instruction::FPTrunc:
617 case Instruction::FPExt:
618 case Instruction::FPToUI:
619 case Instruction::FPToSI:
620 case Instruction::SIToFP:
621 case Instruction::PtrToInt:
622 case Instruction::UIToFP:
623 case Instruction::IntToPtr:
624 return; // Can't work with floating point or pointers
625 case Instruction::Trunc:
626 // All these have integer operands
627 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
629 case Instruction::BitCast: {
630 const Type *SrcTy = I->getOperand(0)->getType();
631 if (SrcTy->isIntegral()) {
632 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
637 case Instruction::ZExt: {
638 // Compute the bits in the result that are not present in the input.
639 const Type *SrcTy = I->getOperand(0)->getType();
640 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
641 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
643 Mask &= SrcTy->getIntegralTypeMask();
644 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
645 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
646 // The top bits are known to be zero.
647 KnownZero |= NewBits;
650 case Instruction::SExt: {
651 // Compute the bits in the result that are not present in the input.
652 const Type *SrcTy = I->getOperand(0)->getType();
653 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
654 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
656 Mask &= SrcTy->getIntegralTypeMask();
657 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
658 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
660 // If the sign bit of the input is known set or clear, then we know the
661 // top bits of the result.
662 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
663 if (KnownZero & InSignBit) { // Input sign bit known zero
664 KnownZero |= NewBits;
665 KnownOne &= ~NewBits;
666 } else if (KnownOne & InSignBit) { // Input sign bit known set
668 KnownZero &= ~NewBits;
669 } else { // Input sign bit unknown
670 KnownZero &= ~NewBits;
671 KnownOne &= ~NewBits;
675 case Instruction::Shl:
676 // (shl X, C1) & C2 == 0 iff (X & C2 >>u C1) == 0
677 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
678 uint64_t ShiftAmt = SA->getZExtValue();
680 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero, KnownOne, Depth+1);
681 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
682 KnownZero <<= ShiftAmt;
683 KnownOne <<= ShiftAmt;
684 KnownZero |= (1ULL << ShiftAmt)-1; // low bits known zero.
688 case Instruction::LShr:
689 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
690 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
691 // Compute the new bits that are at the top now.
692 uint64_t ShiftAmt = SA->getZExtValue();
693 uint64_t HighBits = (1ULL << ShiftAmt)-1;
694 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
696 // Unsigned shift right.
698 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
699 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
700 KnownZero >>= ShiftAmt;
701 KnownOne >>= ShiftAmt;
702 KnownZero |= HighBits; // high bits known zero.
706 case Instruction::AShr:
707 // (ushr X, C1) & C2 == 0 iff (-1 >> C1) & C2 == 0
708 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
709 // Compute the new bits that are at the top now.
710 uint64_t ShiftAmt = SA->getZExtValue();
711 uint64_t HighBits = (1ULL << ShiftAmt)-1;
712 HighBits <<= I->getType()->getPrimitiveSizeInBits()-ShiftAmt;
714 // Signed shift right.
716 ComputeMaskedBits(I->getOperand(0), Mask, KnownZero,KnownOne,Depth+1);
717 assert((KnownZero & KnownOne) == 0&&"Bits known to be one AND zero?");
718 KnownZero >>= ShiftAmt;
719 KnownOne >>= ShiftAmt;
721 // Handle the sign bits.
722 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
723 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
725 if (KnownZero & SignBit) { // New bits are known zero.
726 KnownZero |= HighBits;
727 } else if (KnownOne & SignBit) { // New bits are known one.
728 KnownOne |= HighBits;
736 /// MaskedValueIsZero - Return true if 'V & Mask' is known to be zero. We use
737 /// this predicate to simplify operations downstream. Mask is known to be zero
738 /// for bits that V cannot have.
739 static bool MaskedValueIsZero(Value *V, uint64_t Mask, unsigned Depth = 0) {
740 uint64_t KnownZero, KnownOne;
741 ComputeMaskedBits(V, Mask, KnownZero, KnownOne, Depth);
742 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
743 return (KnownZero & Mask) == Mask;
746 /// ShrinkDemandedConstant - Check to see if the specified operand of the
747 /// specified instruction is a constant integer. If so, check to see if there
748 /// are any bits set in the constant that are not demanded. If so, shrink the
749 /// constant and return true.
750 static bool ShrinkDemandedConstant(Instruction *I, unsigned OpNo,
752 ConstantInt *OpC = dyn_cast<ConstantInt>(I->getOperand(OpNo));
753 if (!OpC) return false;
755 // If there are no bits set that aren't demanded, nothing to do.
756 if ((~Demanded & OpC->getZExtValue()) == 0)
759 // This is producing any bits that are not needed, shrink the RHS.
760 uint64_t Val = Demanded & OpC->getZExtValue();
761 I->setOperand(OpNo, GetConstantInType(OpC->getType(), Val));
765 // ComputeSignedMinMaxValuesFromKnownBits - Given a signed integer type and a
766 // set of known zero and one bits, compute the maximum and minimum values that
767 // could have the specified known zero and known one bits, returning them in
769 static void ComputeSignedMinMaxValuesFromKnownBits(const Type *Ty,
772 int64_t &Min, int64_t &Max) {
773 uint64_t TypeBits = Ty->getIntegralTypeMask();
774 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
776 uint64_t SignBit = 1ULL << (Ty->getPrimitiveSizeInBits()-1);
778 // The minimum value is when all unknown bits are zeros, EXCEPT for the sign
779 // bit if it is unknown.
781 Max = KnownOne|UnknownBits;
783 if (SignBit & UnknownBits) { // Sign bit is unknown
788 // Sign extend the min/max values.
789 int ShAmt = 64-Ty->getPrimitiveSizeInBits();
790 Min = (Min << ShAmt) >> ShAmt;
791 Max = (Max << ShAmt) >> ShAmt;
794 // ComputeUnsignedMinMaxValuesFromKnownBits - Given an unsigned integer type and
795 // a set of known zero and one bits, compute the maximum and minimum values that
796 // could have the specified known zero and known one bits, returning them in
798 static void ComputeUnsignedMinMaxValuesFromKnownBits(const Type *Ty,
803 uint64_t TypeBits = Ty->getIntegralTypeMask();
804 uint64_t UnknownBits = ~(KnownZero|KnownOne) & TypeBits;
806 // The minimum value is when the unknown bits are all zeros.
808 // The maximum value is when the unknown bits are all ones.
809 Max = KnownOne|UnknownBits;
813 /// SimplifyDemandedBits - Look at V. At this point, we know that only the
814 /// DemandedMask bits of the result of V are ever used downstream. If we can
815 /// use this information to simplify V, do so and return true. Otherwise,
816 /// analyze the expression and return a mask of KnownOne and KnownZero bits for
817 /// the expression (used to simplify the caller). The KnownZero/One bits may
818 /// only be accurate for those bits in the DemandedMask.
819 bool InstCombiner::SimplifyDemandedBits(Value *V, uint64_t DemandedMask,
820 uint64_t &KnownZero, uint64_t &KnownOne,
822 if (ConstantIntegral *CI = dyn_cast<ConstantIntegral>(V)) {
823 // We know all of the bits for a constant!
824 KnownOne = CI->getZExtValue() & DemandedMask;
825 KnownZero = ~KnownOne & DemandedMask;
829 KnownZero = KnownOne = 0;
830 if (!V->hasOneUse()) { // Other users may use these bits.
831 if (Depth != 0) { // Not at the root.
832 // Just compute the KnownZero/KnownOne bits to simplify things downstream.
833 ComputeMaskedBits(V, DemandedMask, KnownZero, KnownOne, Depth);
836 // If this is the root being simplified, allow it to have multiple uses,
837 // just set the DemandedMask to all bits.
838 DemandedMask = V->getType()->getIntegralTypeMask();
839 } else if (DemandedMask == 0) { // Not demanding any bits from V.
840 if (V != UndefValue::get(V->getType()))
841 return UpdateValueUsesWith(V, UndefValue::get(V->getType()));
843 } else if (Depth == 6) { // Limit search depth.
847 Instruction *I = dyn_cast<Instruction>(V);
848 if (!I) return false; // Only analyze instructions.
850 DemandedMask &= V->getType()->getIntegralTypeMask();
852 uint64_t KnownZero2 = 0, KnownOne2 = 0;
853 switch (I->getOpcode()) {
855 case Instruction::And:
856 // If either the LHS or the RHS are Zero, the result is zero.
857 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
858 KnownZero, KnownOne, Depth+1))
860 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
862 // If something is known zero on the RHS, the bits aren't demanded on the
864 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownZero,
865 KnownZero2, KnownOne2, Depth+1))
867 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
869 // If all of the demanded bits are known 1 on one side, return the other.
870 // These bits cannot contribute to the result of the 'and'.
871 if ((DemandedMask & ~KnownZero2 & KnownOne) == (DemandedMask & ~KnownZero2))
872 return UpdateValueUsesWith(I, I->getOperand(0));
873 if ((DemandedMask & ~KnownZero & KnownOne2) == (DemandedMask & ~KnownZero))
874 return UpdateValueUsesWith(I, I->getOperand(1));
876 // If all of the demanded bits in the inputs are known zeros, return zero.
877 if ((DemandedMask & (KnownZero|KnownZero2)) == DemandedMask)
878 return UpdateValueUsesWith(I, Constant::getNullValue(I->getType()));
880 // If the RHS is a constant, see if we can simplify it.
881 if (ShrinkDemandedConstant(I, 1, DemandedMask & ~KnownZero2))
882 return UpdateValueUsesWith(I, I);
884 // Output known-1 bits are only known if set in both the LHS & RHS.
885 KnownOne &= KnownOne2;
886 // Output known-0 are known to be clear if zero in either the LHS | RHS.
887 KnownZero |= KnownZero2;
889 case Instruction::Or:
890 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
891 KnownZero, KnownOne, Depth+1))
893 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
894 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask & ~KnownOne,
895 KnownZero2, KnownOne2, Depth+1))
897 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
899 // If all of the demanded bits are known zero on one side, return the other.
900 // These bits cannot contribute to the result of the 'or'.
901 if ((DemandedMask & ~KnownOne2 & KnownZero) == (DemandedMask & ~KnownOne2))
902 return UpdateValueUsesWith(I, I->getOperand(0));
903 if ((DemandedMask & ~KnownOne & KnownZero2) == (DemandedMask & ~KnownOne))
904 return UpdateValueUsesWith(I, I->getOperand(1));
906 // If all of the potentially set bits on one side are known to be set on
907 // the other side, just use the 'other' side.
908 if ((DemandedMask & (~KnownZero) & KnownOne2) ==
909 (DemandedMask & (~KnownZero)))
910 return UpdateValueUsesWith(I, I->getOperand(0));
911 if ((DemandedMask & (~KnownZero2) & KnownOne) ==
912 (DemandedMask & (~KnownZero2)))
913 return UpdateValueUsesWith(I, I->getOperand(1));
915 // If the RHS is a constant, see if we can simplify it.
916 if (ShrinkDemandedConstant(I, 1, DemandedMask))
917 return UpdateValueUsesWith(I, I);
919 // Output known-0 bits are only known if clear in both the LHS & RHS.
920 KnownZero &= KnownZero2;
921 // Output known-1 are known to be set if set in either the LHS | RHS.
922 KnownOne |= KnownOne2;
924 case Instruction::Xor: {
925 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
926 KnownZero, KnownOne, Depth+1))
928 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
929 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
930 KnownZero2, KnownOne2, Depth+1))
932 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
934 // If all of the demanded bits are known zero on one side, return the other.
935 // These bits cannot contribute to the result of the 'xor'.
936 if ((DemandedMask & KnownZero) == DemandedMask)
937 return UpdateValueUsesWith(I, I->getOperand(0));
938 if ((DemandedMask & KnownZero2) == DemandedMask)
939 return UpdateValueUsesWith(I, I->getOperand(1));
941 // Output known-0 bits are known if clear or set in both the LHS & RHS.
942 uint64_t KnownZeroOut = (KnownZero & KnownZero2) | (KnownOne & KnownOne2);
943 // Output known-1 are known to be set if set in only one of the LHS, RHS.
944 uint64_t KnownOneOut = (KnownZero & KnownOne2) | (KnownOne & KnownZero2);
946 // If all of the demanded bits are known to be zero on one side or the
947 // other, turn this into an *inclusive* or.
948 // e.g. (A & C1)^(B & C2) -> (A & C1)|(B & C2) iff C1&C2 == 0
949 if ((DemandedMask & ~KnownZero & ~KnownZero2) == 0) {
951 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
953 InsertNewInstBefore(Or, *I);
954 return UpdateValueUsesWith(I, Or);
957 // If all of the demanded bits on one side are known, and all of the set
958 // bits on that side are also known to be set on the other side, turn this
959 // into an AND, as we know the bits will be cleared.
960 // e.g. (X | C1) ^ C2 --> (X | C1) & ~C2 iff (C1&C2) == C2
961 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask) { // all known
962 if ((KnownOne & KnownOne2) == KnownOne) {
963 Constant *AndC = GetConstantInType(I->getType(),
964 ~KnownOne & DemandedMask);
966 BinaryOperator::createAnd(I->getOperand(0), AndC, "tmp");
967 InsertNewInstBefore(And, *I);
968 return UpdateValueUsesWith(I, And);
972 // If the RHS is a constant, see if we can simplify it.
973 // FIXME: for XOR, we prefer to force bits to 1 if they will make a -1.
974 if (ShrinkDemandedConstant(I, 1, DemandedMask))
975 return UpdateValueUsesWith(I, I);
977 KnownZero = KnownZeroOut;
978 KnownOne = KnownOneOut;
981 case Instruction::Select:
982 if (SimplifyDemandedBits(I->getOperand(2), DemandedMask,
983 KnownZero, KnownOne, Depth+1))
985 if (SimplifyDemandedBits(I->getOperand(1), DemandedMask,
986 KnownZero2, KnownOne2, Depth+1))
988 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
989 assert((KnownZero2 & KnownOne2) == 0 && "Bits known to be one AND zero?");
991 // If the operands are constants, see if we can simplify them.
992 if (ShrinkDemandedConstant(I, 1, DemandedMask))
993 return UpdateValueUsesWith(I, I);
994 if (ShrinkDemandedConstant(I, 2, DemandedMask))
995 return UpdateValueUsesWith(I, I);
997 // Only known if known in both the LHS and RHS.
998 KnownOne &= KnownOne2;
999 KnownZero &= KnownZero2;
1001 case Instruction::Trunc:
1002 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1003 KnownZero, KnownOne, Depth+1))
1005 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1007 case Instruction::BitCast:
1008 if (!I->getOperand(0)->getType()->isIntegral())
1011 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1012 KnownZero, KnownOne, Depth+1))
1014 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1016 case Instruction::ZExt: {
1017 // Compute the bits in the result that are not present in the input.
1018 const Type *SrcTy = I->getOperand(0)->getType();
1019 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1020 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1022 DemandedMask &= SrcTy->getIntegralTypeMask();
1023 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask,
1024 KnownZero, KnownOne, Depth+1))
1026 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1027 // The top bits are known to be zero.
1028 KnownZero |= NewBits;
1031 case Instruction::SExt: {
1032 // Compute the bits in the result that are not present in the input.
1033 const Type *SrcTy = I->getOperand(0)->getType();
1034 uint64_t NotIn = ~SrcTy->getIntegralTypeMask();
1035 uint64_t NewBits = I->getType()->getIntegralTypeMask() & NotIn;
1037 // Get the sign bit for the source type
1038 uint64_t InSignBit = 1ULL << (SrcTy->getPrimitiveSizeInBits()-1);
1039 int64_t InputDemandedBits = DemandedMask & SrcTy->getIntegralTypeMask();
1041 // If any of the sign extended bits are demanded, we know that the sign
1043 if (NewBits & DemandedMask)
1044 InputDemandedBits |= InSignBit;
1046 if (SimplifyDemandedBits(I->getOperand(0), InputDemandedBits,
1047 KnownZero, KnownOne, Depth+1))
1049 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1051 // If the sign bit of the input is known set or clear, then we know the
1052 // top bits of the result.
1054 // If the input sign bit is known zero, or if the NewBits are not demanded
1055 // convert this into a zero extension.
1056 if ((KnownZero & InSignBit) || (NewBits & ~DemandedMask) == NewBits) {
1057 // Convert to ZExt cast
1058 CastInst *NewCast = CastInst::create(
1059 Instruction::ZExt, I->getOperand(0), I->getType(), I->getName(), I);
1060 return UpdateValueUsesWith(I, NewCast);
1061 } else if (KnownOne & InSignBit) { // Input sign bit known set
1062 KnownOne |= NewBits;
1063 KnownZero &= ~NewBits;
1064 } else { // Input sign bit unknown
1065 KnownZero &= ~NewBits;
1066 KnownOne &= ~NewBits;
1070 case Instruction::Add:
1071 // If there is a constant on the RHS, there are a variety of xformations
1073 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
1074 // If null, this should be simplified elsewhere. Some of the xforms here
1075 // won't work if the RHS is zero.
1076 if (RHS->isNullValue())
1079 // Figure out what the input bits are. If the top bits of the and result
1080 // are not demanded, then the add doesn't demand them from its input
1083 // Shift the demanded mask up so that it's at the top of the uint64_t.
1084 unsigned BitWidth = I->getType()->getPrimitiveSizeInBits();
1085 unsigned NLZ = CountLeadingZeros_64(DemandedMask << (64-BitWidth));
1087 // If the top bit of the output is demanded, demand everything from the
1088 // input. Otherwise, we demand all the input bits except NLZ top bits.
1089 uint64_t InDemandedBits = ~0ULL >> 64-BitWidth+NLZ;
1091 // Find information about known zero/one bits in the input.
1092 if (SimplifyDemandedBits(I->getOperand(0), InDemandedBits,
1093 KnownZero2, KnownOne2, Depth+1))
1096 // If the RHS of the add has bits set that can't affect the input, reduce
1098 if (ShrinkDemandedConstant(I, 1, InDemandedBits))
1099 return UpdateValueUsesWith(I, I);
1101 // Avoid excess work.
1102 if (KnownZero2 == 0 && KnownOne2 == 0)
1105 // Turn it into OR if input bits are zero.
1106 if ((KnownZero2 & RHS->getZExtValue()) == RHS->getZExtValue()) {
1108 BinaryOperator::createOr(I->getOperand(0), I->getOperand(1),
1110 InsertNewInstBefore(Or, *I);
1111 return UpdateValueUsesWith(I, Or);
1114 // We can say something about the output known-zero and known-one bits,
1115 // depending on potential carries from the input constant and the
1116 // unknowns. For example if the LHS is known to have at most the 0x0F0F0
1117 // bits set and the RHS constant is 0x01001, then we know we have a known
1118 // one mask of 0x00001 and a known zero mask of 0xE0F0E.
1120 // To compute this, we first compute the potential carry bits. These are
1121 // the bits which may be modified. I'm not aware of a better way to do
1123 uint64_t RHSVal = RHS->getZExtValue();
1125 bool CarryIn = false;
1126 uint64_t CarryBits = 0;
1127 uint64_t CurBit = 1;
1128 for (unsigned i = 0; i != BitWidth; ++i, CurBit <<= 1) {
1129 // Record the current carry in.
1130 if (CarryIn) CarryBits |= CurBit;
1134 // This bit has a carry out unless it is "zero + zero" or
1135 // "zero + anything" with no carry in.
1136 if ((KnownZero2 & CurBit) && ((RHSVal & CurBit) == 0)) {
1137 CarryOut = false; // 0 + 0 has no carry out, even with carry in.
1138 } else if (!CarryIn &&
1139 ((KnownZero2 & CurBit) || ((RHSVal & CurBit) == 0))) {
1140 CarryOut = false; // 0 + anything has no carry out if no carry in.
1142 // Otherwise, we have to assume we have a carry out.
1146 // This stage's carry out becomes the next stage's carry-in.
1150 // Now that we know which bits have carries, compute the known-1/0 sets.
1152 // Bits are known one if they are known zero in one operand and one in the
1153 // other, and there is no input carry.
1154 KnownOne = ((KnownZero2 & RHSVal) | (KnownOne2 & ~RHSVal)) & ~CarryBits;
1156 // Bits are known zero if they are known zero in both operands and there
1157 // is no input carry.
1158 KnownZero = KnownZero2 & ~RHSVal & ~CarryBits;
1161 case Instruction::Shl:
1162 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1163 uint64_t ShiftAmt = SA->getZExtValue();
1164 if (SimplifyDemandedBits(I->getOperand(0), DemandedMask >> ShiftAmt,
1165 KnownZero, KnownOne, Depth+1))
1167 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1168 KnownZero <<= ShiftAmt;
1169 KnownOne <<= ShiftAmt;
1170 KnownZero |= (1ULL << ShiftAmt) - 1; // low bits known zero.
1173 case Instruction::LShr:
1174 // For a logical shift right
1175 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1176 unsigned ShiftAmt = SA->getZExtValue();
1178 // Compute the new bits that are at the top now.
1179 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1180 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1181 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1182 // Unsigned shift right.
1183 if (SimplifyDemandedBits(I->getOperand(0),
1184 (DemandedMask << ShiftAmt) & TypeMask,
1185 KnownZero, KnownOne, Depth+1))
1187 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1188 KnownZero &= TypeMask;
1189 KnownOne &= TypeMask;
1190 KnownZero >>= ShiftAmt;
1191 KnownOne >>= ShiftAmt;
1192 KnownZero |= HighBits; // high bits known zero.
1195 case Instruction::AShr:
1196 // If this is an arithmetic shift right and only the low-bit is set, we can
1197 // always convert this into a logical shr, even if the shift amount is
1198 // variable. The low bit of the shift cannot be an input sign bit unless
1199 // the shift amount is >= the size of the datatype, which is undefined.
1200 if (DemandedMask == 1) {
1201 // Perform the logical shift right.
1202 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1203 I->getOperand(1), I->getName());
1204 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1205 return UpdateValueUsesWith(I, NewVal);
1208 if (ConstantInt *SA = dyn_cast<ConstantInt>(I->getOperand(1))) {
1209 unsigned ShiftAmt = SA->getZExtValue();
1211 // Compute the new bits that are at the top now.
1212 uint64_t HighBits = (1ULL << ShiftAmt)-1;
1213 HighBits <<= I->getType()->getPrimitiveSizeInBits() - ShiftAmt;
1214 uint64_t TypeMask = I->getType()->getIntegralTypeMask();
1215 // Signed shift right.
1216 if (SimplifyDemandedBits(I->getOperand(0),
1217 (DemandedMask << ShiftAmt) & TypeMask,
1218 KnownZero, KnownOne, Depth+1))
1220 assert((KnownZero & KnownOne) == 0 && "Bits known to be one AND zero?");
1221 KnownZero &= TypeMask;
1222 KnownOne &= TypeMask;
1223 KnownZero >>= ShiftAmt;
1224 KnownOne >>= ShiftAmt;
1226 // Handle the sign bits.
1227 uint64_t SignBit = 1ULL << (I->getType()->getPrimitiveSizeInBits()-1);
1228 SignBit >>= ShiftAmt; // Adjust to where it is now in the mask.
1230 // If the input sign bit is known to be zero, or if none of the top bits
1231 // are demanded, turn this into an unsigned shift right.
1232 if ((KnownZero & SignBit) || (HighBits & ~DemandedMask) == HighBits) {
1233 // Perform the logical shift right.
1234 Value *NewVal = new ShiftInst(Instruction::LShr, I->getOperand(0),
1236 InsertNewInstBefore(cast<Instruction>(NewVal), *I);
1237 return UpdateValueUsesWith(I, NewVal);
1238 } else if (KnownOne & SignBit) { // New bits are known one.
1239 KnownOne |= HighBits;
1245 // If the client is only demanding bits that we know, return the known
1247 if ((DemandedMask & (KnownZero|KnownOne)) == DemandedMask)
1248 return UpdateValueUsesWith(I, GetConstantInType(I->getType(), KnownOne));
1253 /// SimplifyDemandedVectorElts - The specified value producecs a vector with
1254 /// 64 or fewer elements. DemandedElts contains the set of elements that are
1255 /// actually used by the caller. This method analyzes which elements of the
1256 /// operand are undef and returns that information in UndefElts.
1258 /// If the information about demanded elements can be used to simplify the
1259 /// operation, the operation is simplified, then the resultant value is
1260 /// returned. This returns null if no change was made.
1261 Value *InstCombiner::SimplifyDemandedVectorElts(Value *V, uint64_t DemandedElts,
1262 uint64_t &UndefElts,
1264 unsigned VWidth = cast<PackedType>(V->getType())->getNumElements();
1265 assert(VWidth <= 64 && "Vector too wide to analyze!");
1266 uint64_t EltMask = ~0ULL >> (64-VWidth);
1267 assert(DemandedElts != EltMask && (DemandedElts & ~EltMask) == 0 &&
1268 "Invalid DemandedElts!");
1270 if (isa<UndefValue>(V)) {
1271 // If the entire vector is undefined, just return this info.
1272 UndefElts = EltMask;
1274 } else if (DemandedElts == 0) { // If nothing is demanded, provide undef.
1275 UndefElts = EltMask;
1276 return UndefValue::get(V->getType());
1280 if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V)) {
1281 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1282 Constant *Undef = UndefValue::get(EltTy);
1284 std::vector<Constant*> Elts;
1285 for (unsigned i = 0; i != VWidth; ++i)
1286 if (!(DemandedElts & (1ULL << i))) { // If not demanded, set to undef.
1287 Elts.push_back(Undef);
1288 UndefElts |= (1ULL << i);
1289 } else if (isa<UndefValue>(CP->getOperand(i))) { // Already undef.
1290 Elts.push_back(Undef);
1291 UndefElts |= (1ULL << i);
1292 } else { // Otherwise, defined.
1293 Elts.push_back(CP->getOperand(i));
1296 // If we changed the constant, return it.
1297 Constant *NewCP = ConstantPacked::get(Elts);
1298 return NewCP != CP ? NewCP : 0;
1299 } else if (isa<ConstantAggregateZero>(V)) {
1300 // Simplify the CAZ to a ConstantPacked where the non-demanded elements are
1302 const Type *EltTy = cast<PackedType>(V->getType())->getElementType();
1303 Constant *Zero = Constant::getNullValue(EltTy);
1304 Constant *Undef = UndefValue::get(EltTy);
1305 std::vector<Constant*> Elts;
1306 for (unsigned i = 0; i != VWidth; ++i)
1307 Elts.push_back((DemandedElts & (1ULL << i)) ? Zero : Undef);
1308 UndefElts = DemandedElts ^ EltMask;
1309 return ConstantPacked::get(Elts);
1312 if (!V->hasOneUse()) { // Other users may use these bits.
1313 if (Depth != 0) { // Not at the root.
1314 // TODO: Just compute the UndefElts information recursively.
1318 } else if (Depth == 10) { // Limit search depth.
1322 Instruction *I = dyn_cast<Instruction>(V);
1323 if (!I) return false; // Only analyze instructions.
1325 bool MadeChange = false;
1326 uint64_t UndefElts2;
1328 switch (I->getOpcode()) {
1331 case Instruction::InsertElement: {
1332 // If this is a variable index, we don't know which element it overwrites.
1333 // demand exactly the same input as we produce.
1334 ConstantInt *Idx = dyn_cast<ConstantInt>(I->getOperand(2));
1336 // Note that we can't propagate undef elt info, because we don't know
1337 // which elt is getting updated.
1338 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1339 UndefElts2, Depth+1);
1340 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1344 // If this is inserting an element that isn't demanded, remove this
1346 unsigned IdxNo = Idx->getZExtValue();
1347 if (IdxNo >= VWidth || (DemandedElts & (1ULL << IdxNo)) == 0)
1348 return AddSoonDeadInstToWorklist(*I, 0);
1350 // Otherwise, the element inserted overwrites whatever was there, so the
1351 // input demanded set is simpler than the output set.
1352 TmpV = SimplifyDemandedVectorElts(I->getOperand(0),
1353 DemandedElts & ~(1ULL << IdxNo),
1354 UndefElts, Depth+1);
1355 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1357 // The inserted element is defined.
1358 UndefElts |= 1ULL << IdxNo;
1362 case Instruction::And:
1363 case Instruction::Or:
1364 case Instruction::Xor:
1365 case Instruction::Add:
1366 case Instruction::Sub:
1367 case Instruction::Mul:
1368 // div/rem demand all inputs, because they don't want divide by zero.
1369 TmpV = SimplifyDemandedVectorElts(I->getOperand(0), DemandedElts,
1370 UndefElts, Depth+1);
1371 if (TmpV) { I->setOperand(0, TmpV); MadeChange = true; }
1372 TmpV = SimplifyDemandedVectorElts(I->getOperand(1), DemandedElts,
1373 UndefElts2, Depth+1);
1374 if (TmpV) { I->setOperand(1, TmpV); MadeChange = true; }
1376 // Output elements are undefined if both are undefined. Consider things
1377 // like undef&0. The result is known zero, not undef.
1378 UndefElts &= UndefElts2;
1381 case Instruction::Call: {
1382 IntrinsicInst *II = dyn_cast<IntrinsicInst>(I);
1384 switch (II->getIntrinsicID()) {
1387 // Binary vector operations that work column-wise. A dest element is a
1388 // function of the corresponding input elements from the two inputs.
1389 case Intrinsic::x86_sse_sub_ss:
1390 case Intrinsic::x86_sse_mul_ss:
1391 case Intrinsic::x86_sse_min_ss:
1392 case Intrinsic::x86_sse_max_ss:
1393 case Intrinsic::x86_sse2_sub_sd:
1394 case Intrinsic::x86_sse2_mul_sd:
1395 case Intrinsic::x86_sse2_min_sd:
1396 case Intrinsic::x86_sse2_max_sd:
1397 TmpV = SimplifyDemandedVectorElts(II->getOperand(1), DemandedElts,
1398 UndefElts, Depth+1);
1399 if (TmpV) { II->setOperand(1, TmpV); MadeChange = true; }
1400 TmpV = SimplifyDemandedVectorElts(II->getOperand(2), DemandedElts,
1401 UndefElts2, Depth+1);
1402 if (TmpV) { II->setOperand(2, TmpV); MadeChange = true; }
1404 // If only the low elt is demanded and this is a scalarizable intrinsic,
1405 // scalarize it now.
1406 if (DemandedElts == 1) {
1407 switch (II->getIntrinsicID()) {
1409 case Intrinsic::x86_sse_sub_ss:
1410 case Intrinsic::x86_sse_mul_ss:
1411 case Intrinsic::x86_sse2_sub_sd:
1412 case Intrinsic::x86_sse2_mul_sd:
1413 // TODO: Lower MIN/MAX/ABS/etc
1414 Value *LHS = II->getOperand(1);
1415 Value *RHS = II->getOperand(2);
1416 // Extract the element as scalars.
1417 LHS = InsertNewInstBefore(new ExtractElementInst(LHS, 0U,"tmp"), *II);
1418 RHS = InsertNewInstBefore(new ExtractElementInst(RHS, 0U,"tmp"), *II);
1420 switch (II->getIntrinsicID()) {
1421 default: assert(0 && "Case stmts out of sync!");
1422 case Intrinsic::x86_sse_sub_ss:
1423 case Intrinsic::x86_sse2_sub_sd:
1424 TmpV = InsertNewInstBefore(BinaryOperator::createSub(LHS, RHS,
1425 II->getName()), *II);
1427 case Intrinsic::x86_sse_mul_ss:
1428 case Intrinsic::x86_sse2_mul_sd:
1429 TmpV = InsertNewInstBefore(BinaryOperator::createMul(LHS, RHS,
1430 II->getName()), *II);
1435 new InsertElementInst(UndefValue::get(II->getType()), TmpV, 0U,
1437 InsertNewInstBefore(New, *II);
1438 AddSoonDeadInstToWorklist(*II, 0);
1443 // Output elements are undefined if both are undefined. Consider things
1444 // like undef&0. The result is known zero, not undef.
1445 UndefElts &= UndefElts2;
1451 return MadeChange ? I : 0;
1454 // isTrueWhenEqual - Return true if the specified setcondinst instruction is
1455 // true when both operands are equal...
1457 static bool isTrueWhenEqual(Instruction &I) {
1458 return I.getOpcode() == Instruction::SetEQ ||
1459 I.getOpcode() == Instruction::SetGE ||
1460 I.getOpcode() == Instruction::SetLE;
1463 /// AssociativeOpt - Perform an optimization on an associative operator. This
1464 /// function is designed to check a chain of associative operators for a
1465 /// potential to apply a certain optimization. Since the optimization may be
1466 /// applicable if the expression was reassociated, this checks the chain, then
1467 /// reassociates the expression as necessary to expose the optimization
1468 /// opportunity. This makes use of a special Functor, which must define
1469 /// 'shouldApply' and 'apply' methods.
1471 template<typename Functor>
1472 Instruction *AssociativeOpt(BinaryOperator &Root, const Functor &F) {
1473 unsigned Opcode = Root.getOpcode();
1474 Value *LHS = Root.getOperand(0);
1476 // Quick check, see if the immediate LHS matches...
1477 if (F.shouldApply(LHS))
1478 return F.apply(Root);
1480 // Otherwise, if the LHS is not of the same opcode as the root, return.
1481 Instruction *LHSI = dyn_cast<Instruction>(LHS);
1482 while (LHSI && LHSI->getOpcode() == Opcode && LHSI->hasOneUse()) {
1483 // Should we apply this transform to the RHS?
1484 bool ShouldApply = F.shouldApply(LHSI->getOperand(1));
1486 // If not to the RHS, check to see if we should apply to the LHS...
1487 if (!ShouldApply && F.shouldApply(LHSI->getOperand(0))) {
1488 cast<BinaryOperator>(LHSI)->swapOperands(); // Make the LHS the RHS
1492 // If the functor wants to apply the optimization to the RHS of LHSI,
1493 // reassociate the expression from ((? op A) op B) to (? op (A op B))
1495 BasicBlock *BB = Root.getParent();
1497 // Now all of the instructions are in the current basic block, go ahead
1498 // and perform the reassociation.
1499 Instruction *TmpLHSI = cast<Instruction>(Root.getOperand(0));
1501 // First move the selected RHS to the LHS of the root...
1502 Root.setOperand(0, LHSI->getOperand(1));
1504 // Make what used to be the LHS of the root be the user of the root...
1505 Value *ExtraOperand = TmpLHSI->getOperand(1);
1506 if (&Root == TmpLHSI) {
1507 Root.replaceAllUsesWith(Constant::getNullValue(TmpLHSI->getType()));
1510 Root.replaceAllUsesWith(TmpLHSI); // Users now use TmpLHSI
1511 TmpLHSI->setOperand(1, &Root); // TmpLHSI now uses the root
1512 TmpLHSI->getParent()->getInstList().remove(TmpLHSI);
1513 BasicBlock::iterator ARI = &Root; ++ARI;
1514 BB->getInstList().insert(ARI, TmpLHSI); // Move TmpLHSI to after Root
1517 // Now propagate the ExtraOperand down the chain of instructions until we
1519 while (TmpLHSI != LHSI) {
1520 Instruction *NextLHSI = cast<Instruction>(TmpLHSI->getOperand(0));
1521 // Move the instruction to immediately before the chain we are
1522 // constructing to avoid breaking dominance properties.
1523 NextLHSI->getParent()->getInstList().remove(NextLHSI);
1524 BB->getInstList().insert(ARI, NextLHSI);
1527 Value *NextOp = NextLHSI->getOperand(1);
1528 NextLHSI->setOperand(1, ExtraOperand);
1530 ExtraOperand = NextOp;
1533 // Now that the instructions are reassociated, have the functor perform
1534 // the transformation...
1535 return F.apply(Root);
1538 LHSI = dyn_cast<Instruction>(LHSI->getOperand(0));
1544 // AddRHS - Implements: X + X --> X << 1
1547 AddRHS(Value *rhs) : RHS(rhs) {}
1548 bool shouldApply(Value *LHS) const { return LHS == RHS; }
1549 Instruction *apply(BinaryOperator &Add) const {
1550 return new ShiftInst(Instruction::Shl, Add.getOperand(0),
1551 ConstantInt::get(Type::UByteTy, 1));
1555 // AddMaskingAnd - Implements (A & C1)+(B & C2) --> (A & C1)|(B & C2)
1557 struct AddMaskingAnd {
1559 AddMaskingAnd(Constant *c) : C2(c) {}
1560 bool shouldApply(Value *LHS) const {
1562 return match(LHS, m_And(m_Value(), m_ConstantInt(C1))) &&
1563 ConstantExpr::getAnd(C1, C2)->isNullValue();
1565 Instruction *apply(BinaryOperator &Add) const {
1566 return BinaryOperator::createOr(Add.getOperand(0), Add.getOperand(1));
1570 static Value *FoldOperationIntoSelectOperand(Instruction &I, Value *SO,
1572 if (CastInst *CI = dyn_cast<CastInst>(&I)) {
1573 if (Constant *SOC = dyn_cast<Constant>(SO))
1574 return ConstantExpr::getCast(CI->getOpcode(), SOC, I.getType());
1576 return IC->InsertNewInstBefore(CastInst::create(
1577 CI->getOpcode(), SO, I.getType(), SO->getName() + ".cast"), I);
1580 // Figure out if the constant is the left or the right argument.
1581 bool ConstIsRHS = isa<Constant>(I.getOperand(1));
1582 Constant *ConstOperand = cast<Constant>(I.getOperand(ConstIsRHS));
1584 if (Constant *SOC = dyn_cast<Constant>(SO)) {
1586 return ConstantExpr::get(I.getOpcode(), SOC, ConstOperand);
1587 return ConstantExpr::get(I.getOpcode(), ConstOperand, SOC);
1590 Value *Op0 = SO, *Op1 = ConstOperand;
1592 std::swap(Op0, Op1);
1594 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1595 New = BinaryOperator::create(BO->getOpcode(), Op0, Op1,SO->getName()+".op");
1596 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1597 New = new ShiftInst(SI->getOpcode(), Op0, Op1, SO->getName()+".sh");
1599 assert(0 && "Unknown binary instruction type!");
1602 return IC->InsertNewInstBefore(New, I);
1605 // FoldOpIntoSelect - Given an instruction with a select as one operand and a
1606 // constant as the other operand, try to fold the binary operator into the
1607 // select arguments. This also works for Cast instructions, which obviously do
1608 // not have a second operand.
1609 static Instruction *FoldOpIntoSelect(Instruction &Op, SelectInst *SI,
1611 // Don't modify shared select instructions
1612 if (!SI->hasOneUse()) return 0;
1613 Value *TV = SI->getOperand(1);
1614 Value *FV = SI->getOperand(2);
1616 if (isa<Constant>(TV) || isa<Constant>(FV)) {
1617 // Bool selects with constant operands can be folded to logical ops.
1618 if (SI->getType() == Type::BoolTy) return 0;
1620 Value *SelectTrueVal = FoldOperationIntoSelectOperand(Op, TV, IC);
1621 Value *SelectFalseVal = FoldOperationIntoSelectOperand(Op, FV, IC);
1623 return new SelectInst(SI->getCondition(), SelectTrueVal,
1630 /// FoldOpIntoPhi - Given a binary operator or cast instruction which has a PHI
1631 /// node as operand #0, see if we can fold the instruction into the PHI (which
1632 /// is only possible if all operands to the PHI are constants).
1633 Instruction *InstCombiner::FoldOpIntoPhi(Instruction &I) {
1634 PHINode *PN = cast<PHINode>(I.getOperand(0));
1635 unsigned NumPHIValues = PN->getNumIncomingValues();
1636 if (!PN->hasOneUse() || NumPHIValues == 0) return 0;
1638 // Check to see if all of the operands of the PHI are constants. If there is
1639 // one non-constant value, remember the BB it is. If there is more than one
1641 BasicBlock *NonConstBB = 0;
1642 for (unsigned i = 0; i != NumPHIValues; ++i)
1643 if (!isa<Constant>(PN->getIncomingValue(i))) {
1644 if (NonConstBB) return 0; // More than one non-const value.
1645 NonConstBB = PN->getIncomingBlock(i);
1647 // If the incoming non-constant value is in I's block, we have an infinite
1649 if (NonConstBB == I.getParent())
1653 // If there is exactly one non-constant value, we can insert a copy of the
1654 // operation in that block. However, if this is a critical edge, we would be
1655 // inserting the computation one some other paths (e.g. inside a loop). Only
1656 // do this if the pred block is unconditionally branching into the phi block.
1658 BranchInst *BI = dyn_cast<BranchInst>(NonConstBB->getTerminator());
1659 if (!BI || !BI->isUnconditional()) return 0;
1662 // Okay, we can do the transformation: create the new PHI node.
1663 PHINode *NewPN = new PHINode(I.getType(), I.getName());
1665 NewPN->reserveOperandSpace(PN->getNumOperands()/2);
1666 InsertNewInstBefore(NewPN, *PN);
1668 // Next, add all of the operands to the PHI.
1669 if (I.getNumOperands() == 2) {
1670 Constant *C = cast<Constant>(I.getOperand(1));
1671 for (unsigned i = 0; i != NumPHIValues; ++i) {
1673 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1674 InV = ConstantExpr::get(I.getOpcode(), InC, C);
1676 assert(PN->getIncomingBlock(i) == NonConstBB);
1677 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(&I))
1678 InV = BinaryOperator::create(BO->getOpcode(),
1679 PN->getIncomingValue(i), C, "phitmp",
1680 NonConstBB->getTerminator());
1681 else if (ShiftInst *SI = dyn_cast<ShiftInst>(&I))
1682 InV = new ShiftInst(SI->getOpcode(),
1683 PN->getIncomingValue(i), C, "phitmp",
1684 NonConstBB->getTerminator());
1686 assert(0 && "Unknown binop!");
1688 WorkList.push_back(cast<Instruction>(InV));
1690 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1693 CastInst *CI = cast<CastInst>(&I);
1694 const Type *RetTy = CI->getType();
1695 for (unsigned i = 0; i != NumPHIValues; ++i) {
1697 if (Constant *InC = dyn_cast<Constant>(PN->getIncomingValue(i))) {
1698 InV = ConstantExpr::getCast(CI->getOpcode(), InC, RetTy);
1700 assert(PN->getIncomingBlock(i) == NonConstBB);
1701 InV = CastInst::create(CI->getOpcode(), PN->getIncomingValue(i),
1702 I.getType(), "phitmp",
1703 NonConstBB->getTerminator());
1704 WorkList.push_back(cast<Instruction>(InV));
1706 NewPN->addIncoming(InV, PN->getIncomingBlock(i));
1709 return ReplaceInstUsesWith(I, NewPN);
1712 Instruction *InstCombiner::visitAdd(BinaryOperator &I) {
1713 bool Changed = SimplifyCommutative(I);
1714 Value *LHS = I.getOperand(0), *RHS = I.getOperand(1);
1716 if (Constant *RHSC = dyn_cast<Constant>(RHS)) {
1717 // X + undef -> undef
1718 if (isa<UndefValue>(RHS))
1719 return ReplaceInstUsesWith(I, RHS);
1722 if (!I.getType()->isFPOrFPVector()) { // NOTE: -0 + +0 = +0.
1723 if (RHSC->isNullValue())
1724 return ReplaceInstUsesWith(I, LHS);
1725 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(RHSC)) {
1726 if (CFP->isExactlyValue(-0.0))
1727 return ReplaceInstUsesWith(I, LHS);
1730 if (ConstantInt *CI = dyn_cast<ConstantInt>(RHSC)) {
1731 // X + (signbit) --> X ^ signbit
1732 uint64_t Val = CI->getZExtValue();
1733 if (Val == (1ULL << (CI->getType()->getPrimitiveSizeInBits()-1)))
1734 return BinaryOperator::createXor(LHS, RHS);
1736 // See if SimplifyDemandedBits can simplify this. This handles stuff like
1737 // (X & 254)+1 -> (X&254)|1
1738 uint64_t KnownZero, KnownOne;
1739 if (!isa<PackedType>(I.getType()) &&
1740 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
1741 KnownZero, KnownOne))
1745 if (isa<PHINode>(LHS))
1746 if (Instruction *NV = FoldOpIntoPhi(I))
1749 ConstantInt *XorRHS = 0;
1751 if (match(LHS, m_Xor(m_Value(XorLHS), m_ConstantInt(XorRHS)))) {
1752 unsigned TySizeBits = I.getType()->getPrimitiveSizeInBits();
1753 int64_t RHSSExt = cast<ConstantInt>(RHSC)->getSExtValue();
1754 uint64_t RHSZExt = cast<ConstantInt>(RHSC)->getZExtValue();
1756 uint64_t C0080Val = 1ULL << 31;
1757 int64_t CFF80Val = -C0080Val;
1760 if (TySizeBits > Size) {
1762 // If we have ADD(XOR(AND(X, 0xFF), 0x80), 0xF..F80), it's a sext.
1763 // If we have ADD(XOR(AND(X, 0xFF), 0xF..F80), 0x80), it's a sext.
1764 if (RHSSExt == CFF80Val) {
1765 if (XorRHS->getZExtValue() == C0080Val)
1767 } else if (RHSZExt == C0080Val) {
1768 if (XorRHS->getSExtValue() == CFF80Val)
1772 // This is a sign extend if the top bits are known zero.
1773 uint64_t Mask = ~0ULL;
1774 Mask <<= 64-(TySizeBits-Size);
1775 Mask &= XorLHS->getType()->getIntegralTypeMask();
1776 if (!MaskedValueIsZero(XorLHS, Mask))
1777 Size = 0; // Not a sign ext, but can't be any others either.
1784 } while (Size >= 8);
1787 const Type *MiddleType = 0;
1790 case 32: MiddleType = Type::IntTy; break;
1791 case 16: MiddleType = Type::ShortTy; break;
1792 case 8: MiddleType = Type::SByteTy; break;
1795 Instruction *NewTrunc = new TruncInst(XorLHS, MiddleType, "sext");
1796 InsertNewInstBefore(NewTrunc, I);
1797 return new SExtInst(NewTrunc, I.getType());
1803 if (I.getType()->isInteger()) {
1804 if (Instruction *Result = AssociativeOpt(I, AddRHS(RHS))) return Result;
1806 if (Instruction *RHSI = dyn_cast<Instruction>(RHS)) {
1807 if (RHSI->getOpcode() == Instruction::Sub)
1808 if (LHS == RHSI->getOperand(1)) // A + (B - A) --> B
1809 return ReplaceInstUsesWith(I, RHSI->getOperand(0));
1811 if (Instruction *LHSI = dyn_cast<Instruction>(LHS)) {
1812 if (LHSI->getOpcode() == Instruction::Sub)
1813 if (RHS == LHSI->getOperand(1)) // (B - A) + A --> B
1814 return ReplaceInstUsesWith(I, LHSI->getOperand(0));
1819 if (Value *V = dyn_castNegVal(LHS))
1820 return BinaryOperator::createSub(RHS, V);
1823 if (!isa<Constant>(RHS))
1824 if (Value *V = dyn_castNegVal(RHS))
1825 return BinaryOperator::createSub(LHS, V);
1829 if (Value *X = dyn_castFoldableMul(LHS, C2)) {
1830 if (X == RHS) // X*C + X --> X * (C+1)
1831 return BinaryOperator::createMul(RHS, AddOne(C2));
1833 // X*C1 + X*C2 --> X * (C1+C2)
1835 if (X == dyn_castFoldableMul(RHS, C1))
1836 return BinaryOperator::createMul(X, ConstantExpr::getAdd(C1, C2));
1839 // X + X*C --> X * (C+1)
1840 if (dyn_castFoldableMul(RHS, C2) == LHS)
1841 return BinaryOperator::createMul(LHS, AddOne(C2));
1844 // (A & C1)+(B & C2) --> (A & C1)|(B & C2) iff C1&C2 == 0
1845 if (match(RHS, m_And(m_Value(), m_ConstantInt(C2))))
1846 if (Instruction *R = AssociativeOpt(I, AddMaskingAnd(C2))) return R;
1848 if (ConstantInt *CRHS = dyn_cast<ConstantInt>(RHS)) {
1850 if (match(LHS, m_Not(m_Value(X)))) { // ~X + C --> (C-1) - X
1851 Constant *C= ConstantExpr::getSub(CRHS, ConstantInt::get(I.getType(), 1));
1852 return BinaryOperator::createSub(C, X);
1855 // (X & FF00) + xx00 -> (X+xx00) & FF00
1856 if (LHS->hasOneUse() && match(LHS, m_And(m_Value(X), m_ConstantInt(C2)))) {
1857 Constant *Anded = ConstantExpr::getAnd(CRHS, C2);
1858 if (Anded == CRHS) {
1859 // See if all bits from the first bit set in the Add RHS up are included
1860 // in the mask. First, get the rightmost bit.
1861 uint64_t AddRHSV = CRHS->getZExtValue();
1863 // Form a mask of all bits from the lowest bit added through the top.
1864 uint64_t AddRHSHighBits = ~((AddRHSV & -AddRHSV)-1);
1865 AddRHSHighBits &= C2->getType()->getIntegralTypeMask();
1867 // See if the and mask includes all of these bits.
1868 uint64_t AddRHSHighBitsAnd = AddRHSHighBits & C2->getZExtValue();
1870 if (AddRHSHighBits == AddRHSHighBitsAnd) {
1871 // Okay, the xform is safe. Insert the new add pronto.
1872 Value *NewAdd = InsertNewInstBefore(BinaryOperator::createAdd(X, CRHS,
1873 LHS->getName()), I);
1874 return BinaryOperator::createAnd(NewAdd, C2);
1879 // Try to fold constant add into select arguments.
1880 if (SelectInst *SI = dyn_cast<SelectInst>(LHS))
1881 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1885 // add (cast *A to intptrtype) B ->
1886 // cast (GEP (cast *A to sbyte*) B) ->
1889 CastInst *CI = dyn_cast<CastInst>(LHS);
1892 CI = dyn_cast<CastInst>(RHS);
1895 if (CI && CI->getType()->isSized() &&
1896 (CI->getType()->getPrimitiveSize() ==
1897 TD->getIntPtrType()->getPrimitiveSize())
1898 && isa<PointerType>(CI->getOperand(0)->getType())) {
1899 Value *I2 = InsertCastBefore(Instruction::BitCast, CI->getOperand(0),
1900 PointerType::get(Type::SByteTy), I);
1901 I2 = InsertNewInstBefore(new GetElementPtrInst(I2, Other, "ctg2"), I);
1902 return new PtrToIntInst(I2, CI->getType());
1906 return Changed ? &I : 0;
1909 // isSignBit - Return true if the value represented by the constant only has the
1910 // highest order bit set.
1911 static bool isSignBit(ConstantInt *CI) {
1912 unsigned NumBits = CI->getType()->getPrimitiveSizeInBits();
1913 return (CI->getZExtValue() & (~0ULL >> (64-NumBits))) == (1ULL << (NumBits-1));
1916 /// RemoveNoopCast - Strip off nonconverting casts from the value.
1918 static Value *RemoveNoopCast(Value *V) {
1919 if (CastInst *CI = dyn_cast<CastInst>(V)) {
1920 const Type *CTy = CI->getType();
1921 const Type *OpTy = CI->getOperand(0)->getType();
1922 if (CTy->isInteger() && OpTy->isInteger()) {
1923 if (CTy->getPrimitiveSizeInBits() == OpTy->getPrimitiveSizeInBits())
1924 return RemoveNoopCast(CI->getOperand(0));
1925 } else if (isa<PointerType>(CTy) && isa<PointerType>(OpTy))
1926 return RemoveNoopCast(CI->getOperand(0));
1931 Instruction *InstCombiner::visitSub(BinaryOperator &I) {
1932 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
1934 if (Op0 == Op1) // sub X, X -> 0
1935 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
1937 // If this is a 'B = x-(-A)', change to B = x+A...
1938 if (Value *V = dyn_castNegVal(Op1))
1939 return BinaryOperator::createAdd(Op0, V);
1941 if (isa<UndefValue>(Op0))
1942 return ReplaceInstUsesWith(I, Op0); // undef - X -> undef
1943 if (isa<UndefValue>(Op1))
1944 return ReplaceInstUsesWith(I, Op1); // X - undef -> undef
1946 if (ConstantInt *C = dyn_cast<ConstantInt>(Op0)) {
1947 // Replace (-1 - A) with (~A)...
1948 if (C->isAllOnesValue())
1949 return BinaryOperator::createNot(Op1);
1951 // C - ~X == X + (1+C)
1953 if (match(Op1, m_Not(m_Value(X))))
1954 return BinaryOperator::createAdd(X,
1955 ConstantExpr::getAdd(C, ConstantInt::get(I.getType(), 1)));
1956 // -((uint)X >> 31) -> ((int)X >> 31)
1957 // -((int)X >> 31) -> ((uint)X >> 31)
1958 if (C->isNullValue()) {
1959 Value *NoopCastedRHS = RemoveNoopCast(Op1);
1960 if (ShiftInst *SI = dyn_cast<ShiftInst>(NoopCastedRHS))
1961 if (SI->getOpcode() == Instruction::LShr) {
1962 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1963 // Check to see if we are shifting out everything but the sign bit.
1964 if (CU->getZExtValue() ==
1965 SI->getType()->getPrimitiveSizeInBits()-1) {
1966 // Ok, the transformation is safe. Insert AShr.
1967 return new ShiftInst(Instruction::AShr, SI->getOperand(0),
1972 else if (SI->getOpcode() == Instruction::AShr) {
1973 if (ConstantInt *CU = dyn_cast<ConstantInt>(SI->getOperand(1))) {
1974 // Check to see if we are shifting out everything but the sign bit.
1975 if (CU->getZExtValue() ==
1976 SI->getType()->getPrimitiveSizeInBits()-1) {
1977 // Ok, the transformation is safe. Insert LShr.
1978 return new ShiftInst(Instruction::LShr, SI->getOperand(0),
1985 // Try to fold constant sub into select arguments.
1986 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
1987 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
1990 if (isa<PHINode>(Op0))
1991 if (Instruction *NV = FoldOpIntoPhi(I))
1995 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1)) {
1996 if (Op1I->getOpcode() == Instruction::Add &&
1997 !Op0->getType()->isFPOrFPVector()) {
1998 if (Op1I->getOperand(0) == Op0) // X-(X+Y) == -Y
1999 return BinaryOperator::createNeg(Op1I->getOperand(1), I.getName());
2000 else if (Op1I->getOperand(1) == Op0) // X-(Y+X) == -Y
2001 return BinaryOperator::createNeg(Op1I->getOperand(0), I.getName());
2002 else if (ConstantInt *CI1 = dyn_cast<ConstantInt>(I.getOperand(0))) {
2003 if (ConstantInt *CI2 = dyn_cast<ConstantInt>(Op1I->getOperand(1)))
2004 // C1-(X+C2) --> (C1-C2)-X
2005 return BinaryOperator::createSub(ConstantExpr::getSub(CI1, CI2),
2006 Op1I->getOperand(0));
2010 if (Op1I->hasOneUse()) {
2011 // Replace (x - (y - z)) with (x + (z - y)) if the (y - z) subexpression
2012 // is not used by anyone else...
2014 if (Op1I->getOpcode() == Instruction::Sub &&
2015 !Op1I->getType()->isFPOrFPVector()) {
2016 // Swap the two operands of the subexpr...
2017 Value *IIOp0 = Op1I->getOperand(0), *IIOp1 = Op1I->getOperand(1);
2018 Op1I->setOperand(0, IIOp1);
2019 Op1I->setOperand(1, IIOp0);
2021 // Create the new top level add instruction...
2022 return BinaryOperator::createAdd(Op0, Op1);
2025 // Replace (A - (A & B)) with (A & ~B) if this is the only use of (A&B)...
2027 if (Op1I->getOpcode() == Instruction::And &&
2028 (Op1I->getOperand(0) == Op0 || Op1I->getOperand(1) == Op0)) {
2029 Value *OtherOp = Op1I->getOperand(Op1I->getOperand(0) == Op0);
2032 InsertNewInstBefore(BinaryOperator::createNot(OtherOp, "B.not"), I);
2033 return BinaryOperator::createAnd(Op0, NewNot);
2036 // 0 - (X sdiv C) -> (X sdiv -C)
2037 if (Op1I->getOpcode() == Instruction::SDiv)
2038 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
2039 if (CSI->isNullValue())
2040 if (Constant *DivRHS = dyn_cast<Constant>(Op1I->getOperand(1)))
2041 return BinaryOperator::createSDiv(Op1I->getOperand(0),
2042 ConstantExpr::getNeg(DivRHS));
2044 // X - X*C --> X * (1-C)
2045 ConstantInt *C2 = 0;
2046 if (dyn_castFoldableMul(Op1I, C2) == Op0) {
2048 ConstantExpr::getSub(ConstantInt::get(I.getType(), 1), C2);
2049 return BinaryOperator::createMul(Op0, CP1);
2054 if (!Op0->getType()->isFPOrFPVector())
2055 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2056 if (Op0I->getOpcode() == Instruction::Add) {
2057 if (Op0I->getOperand(0) == Op1) // (Y+X)-Y == X
2058 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
2059 else if (Op0I->getOperand(1) == Op1) // (X+Y)-Y == X
2060 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
2061 } else if (Op0I->getOpcode() == Instruction::Sub) {
2062 if (Op0I->getOperand(0) == Op1) // (X-Y)-X == -Y
2063 return BinaryOperator::createNeg(Op0I->getOperand(1), I.getName());
2067 if (Value *X = dyn_castFoldableMul(Op0, C1)) {
2068 if (X == Op1) { // X*C - X --> X * (C-1)
2069 Constant *CP1 = ConstantExpr::getSub(C1, ConstantInt::get(I.getType(),1));
2070 return BinaryOperator::createMul(Op1, CP1);
2073 ConstantInt *C2; // X*C1 - X*C2 -> X * (C1-C2)
2074 if (X == dyn_castFoldableMul(Op1, C2))
2075 return BinaryOperator::createMul(Op1, ConstantExpr::getSub(C1, C2));
2080 /// isSignBitCheck - Given an exploded setcc instruction, return true if it is
2081 /// really just returns true if the most significant (sign) bit is set.
2082 static bool isSignBitCheck(unsigned Opcode, Value *LHS, ConstantInt *RHS) {
2083 if (RHS->getType()->isSigned()) {
2084 // True if source is LHS < 0 or LHS <= -1
2085 return Opcode == Instruction::SetLT && RHS->isNullValue() ||
2086 Opcode == Instruction::SetLE && RHS->isAllOnesValue();
2088 ConstantInt *RHSC = cast<ConstantInt>(RHS);
2089 // True if source is LHS > 127 or LHS >= 128, where the constants depend on
2090 // the size of the integer type.
2091 if (Opcode == Instruction::SetGE)
2092 return RHSC->getZExtValue() ==
2093 1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1);
2094 if (Opcode == Instruction::SetGT)
2095 return RHSC->getZExtValue() ==
2096 (1ULL << (RHS->getType()->getPrimitiveSizeInBits()-1))-1;
2101 Instruction *InstCombiner::visitMul(BinaryOperator &I) {
2102 bool Changed = SimplifyCommutative(I);
2103 Value *Op0 = I.getOperand(0);
2105 if (isa<UndefValue>(I.getOperand(1))) // undef * X -> 0
2106 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2108 // Simplify mul instructions with a constant RHS...
2109 if (Constant *Op1 = dyn_cast<Constant>(I.getOperand(1))) {
2110 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2112 // ((X << C1)*C2) == (X * (C2 << C1))
2113 if (ShiftInst *SI = dyn_cast<ShiftInst>(Op0))
2114 if (SI->getOpcode() == Instruction::Shl)
2115 if (Constant *ShOp = dyn_cast<Constant>(SI->getOperand(1)))
2116 return BinaryOperator::createMul(SI->getOperand(0),
2117 ConstantExpr::getShl(CI, ShOp));
2119 if (CI->isNullValue())
2120 return ReplaceInstUsesWith(I, Op1); // X * 0 == 0
2121 if (CI->equalsInt(1)) // X * 1 == X
2122 return ReplaceInstUsesWith(I, Op0);
2123 if (CI->isAllOnesValue()) // X * -1 == 0 - X
2124 return BinaryOperator::createNeg(Op0, I.getName());
2126 int64_t Val = (int64_t)cast<ConstantInt>(CI)->getZExtValue();
2127 if (isPowerOf2_64(Val)) { // Replace X*(2^C) with X << C
2128 uint64_t C = Log2_64(Val);
2129 return new ShiftInst(Instruction::Shl, Op0,
2130 ConstantInt::get(Type::UByteTy, C));
2132 } else if (ConstantFP *Op1F = dyn_cast<ConstantFP>(Op1)) {
2133 if (Op1F->isNullValue())
2134 return ReplaceInstUsesWith(I, Op1);
2136 // "In IEEE floating point, x*1 is not equivalent to x for nans. However,
2137 // ANSI says we can drop signals, so we can do this anyway." (from GCC)
2138 if (Op1F->getValue() == 1.0)
2139 return ReplaceInstUsesWith(I, Op0); // Eliminate 'mul double %X, 1.0'
2142 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
2143 if (Op0I->getOpcode() == Instruction::Add && Op0I->hasOneUse() &&
2144 isa<ConstantInt>(Op0I->getOperand(1))) {
2145 // Canonicalize (X+C1)*C2 -> X*C2+C1*C2.
2146 Instruction *Add = BinaryOperator::createMul(Op0I->getOperand(0),
2148 InsertNewInstBefore(Add, I);
2149 Value *C1C2 = ConstantExpr::getMul(Op1,
2150 cast<Constant>(Op0I->getOperand(1)));
2151 return BinaryOperator::createAdd(Add, C1C2);
2155 // Try to fold constant mul into select arguments.
2156 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2157 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2160 if (isa<PHINode>(Op0))
2161 if (Instruction *NV = FoldOpIntoPhi(I))
2165 if (Value *Op0v = dyn_castNegVal(Op0)) // -X * -Y = X*Y
2166 if (Value *Op1v = dyn_castNegVal(I.getOperand(1)))
2167 return BinaryOperator::createMul(Op0v, Op1v);
2169 // If one of the operands of the multiply is a cast from a boolean value, then
2170 // we know the bool is either zero or one, so this is a 'masking' multiply.
2171 // See if we can simplify things based on how the boolean was originally
2173 CastInst *BoolCast = 0;
2174 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(0)))
2175 if (CI->getOperand(0)->getType() == Type::BoolTy)
2178 if (ZExtInst *CI = dyn_cast<ZExtInst>(I.getOperand(1)))
2179 if (CI->getOperand(0)->getType() == Type::BoolTy)
2182 if (SetCondInst *SCI = dyn_cast<SetCondInst>(BoolCast->getOperand(0))) {
2183 Value *SCIOp0 = SCI->getOperand(0), *SCIOp1 = SCI->getOperand(1);
2184 const Type *SCOpTy = SCIOp0->getType();
2186 // If the setcc is true iff the sign bit of X is set, then convert this
2187 // multiply into a shift/and combination.
2188 if (isa<ConstantInt>(SCIOp1) &&
2189 isSignBitCheck(SCI->getOpcode(), SCIOp0, cast<ConstantInt>(SCIOp1))) {
2190 // Shift the X value right to turn it into "all signbits".
2191 Constant *Amt = ConstantInt::get(Type::UByteTy,
2192 SCOpTy->getPrimitiveSizeInBits()-1);
2194 InsertNewInstBefore(new ShiftInst(Instruction::AShr, SCIOp0, Amt,
2195 BoolCast->getOperand(0)->getName()+
2198 // If the multiply type is not the same as the source type, sign extend
2199 // or truncate to the multiply type.
2200 if (I.getType() != V->getType()) {
2201 unsigned SrcBits = V->getType()->getPrimitiveSizeInBits();
2202 unsigned DstBits = I.getType()->getPrimitiveSizeInBits();
2203 Instruction::CastOps opcode =
2204 (SrcBits == DstBits ? Instruction::BitCast :
2205 (SrcBits < DstBits ? Instruction::SExt : Instruction::Trunc));
2206 V = InsertCastBefore(opcode, V, I.getType(), I);
2209 Value *OtherOp = Op0 == BoolCast ? I.getOperand(1) : Op0;
2210 return BinaryOperator::createAnd(V, OtherOp);
2215 return Changed ? &I : 0;
2218 /// This function implements the transforms on div instructions that work
2219 /// regardless of the kind of div instruction it is (udiv, sdiv, or fdiv). It is
2220 /// used by the visitors to those instructions.
2221 /// @brief Transforms common to all three div instructions
2222 Instruction *InstCombiner::commonDivTransforms(BinaryOperator &I) {
2223 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2226 if (isa<UndefValue>(Op0))
2227 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2229 // X / undef -> undef
2230 if (isa<UndefValue>(Op1))
2231 return ReplaceInstUsesWith(I, Op1);
2233 // Handle cases involving: div X, (select Cond, Y, Z)
2234 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2235 // div X, (Cond ? 0 : Y) -> div X, Y. If the div and the select are in the
2236 // same basic block, then we replace the select with Y, and the condition
2237 // of the select with false (if the cond value is in the same BB). If the
2238 // select has uses other than the div, this allows them to be simplified
2239 // also. Note that div X, Y is just as good as div X, 0 (undef)
2240 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2241 if (ST->isNullValue()) {
2242 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2243 if (CondI && CondI->getParent() == I.getParent())
2244 UpdateValueUsesWith(CondI, ConstantBool::getFalse());
2245 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2246 I.setOperand(1, SI->getOperand(2));
2248 UpdateValueUsesWith(SI, SI->getOperand(2));
2252 // Likewise for: div X, (Cond ? Y : 0) -> div X, Y
2253 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2254 if (ST->isNullValue()) {
2255 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2256 if (CondI && CondI->getParent() == I.getParent())
2257 UpdateValueUsesWith(CondI, ConstantBool::getTrue());
2258 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2259 I.setOperand(1, SI->getOperand(1));
2261 UpdateValueUsesWith(SI, SI->getOperand(1));
2269 /// This function implements the transforms common to both integer division
2270 /// instructions (udiv and sdiv). It is called by the visitors to those integer
2271 /// division instructions.
2272 /// @brief Common integer divide transforms
2273 Instruction *InstCombiner::commonIDivTransforms(BinaryOperator &I) {
2274 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2276 if (Instruction *Common = commonDivTransforms(I))
2279 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2281 if (RHS->equalsInt(1))
2282 return ReplaceInstUsesWith(I, Op0);
2284 // (X / C1) / C2 -> X / (C1*C2)
2285 if (Instruction *LHS = dyn_cast<Instruction>(Op0))
2286 if (Instruction::BinaryOps(LHS->getOpcode()) == I.getOpcode())
2287 if (ConstantInt *LHSRHS = dyn_cast<ConstantInt>(LHS->getOperand(1))) {
2288 return BinaryOperator::create(I.getOpcode(), LHS->getOperand(0),
2289 ConstantExpr::getMul(RHS, LHSRHS));
2292 if (!RHS->isNullValue()) { // avoid X udiv 0
2293 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
2294 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2296 if (isa<PHINode>(Op0))
2297 if (Instruction *NV = FoldOpIntoPhi(I))
2302 // 0 / X == 0, we don't need to preserve faults!
2303 if (ConstantInt *LHS = dyn_cast<ConstantInt>(Op0))
2304 if (LHS->equalsInt(0))
2305 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2310 Instruction *InstCombiner::visitUDiv(BinaryOperator &I) {
2311 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2313 // Handle the integer div common cases
2314 if (Instruction *Common = commonIDivTransforms(I))
2317 // X udiv C^2 -> X >> C
2318 // Check to see if this is an unsigned division with an exact power of 2,
2319 // if so, convert to a right shift.
2320 if (ConstantInt *C = dyn_cast<ConstantInt>(Op1)) {
2321 if (uint64_t Val = C->getZExtValue()) // Don't break X / 0
2322 if (isPowerOf2_64(Val)) {
2323 uint64_t ShiftAmt = Log2_64(Val);
2324 return new ShiftInst(Instruction::LShr, Op0,
2325 ConstantInt::get(Type::UByteTy, ShiftAmt));
2329 // X udiv (C1 << N), where C1 is "1<<C2" --> X >> (N+C2)
2330 if (ShiftInst *RHSI = dyn_cast<ShiftInst>(I.getOperand(1))) {
2331 if (RHSI->getOpcode() == Instruction::Shl &&
2332 isa<ConstantInt>(RHSI->getOperand(0))) {
2333 uint64_t C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2334 if (isPowerOf2_64(C1)) {
2335 Value *N = RHSI->getOperand(1);
2336 const Type *NTy = N->getType();
2337 if (uint64_t C2 = Log2_64(C1)) {
2338 Constant *C2V = ConstantInt::get(NTy, C2);
2339 N = InsertNewInstBefore(BinaryOperator::createAdd(N, C2V, "tmp"), I);
2341 return new ShiftInst(Instruction::LShr, Op0, N);
2346 // udiv X, (Select Cond, C1, C2) --> Select Cond, (shr X, C1), (shr X, C2)
2347 // where C1&C2 are powers of two.
2348 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2349 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2350 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2)))
2351 if (!STO->isNullValue() && !STO->isNullValue()) {
2352 uint64_t TVA = STO->getZExtValue(), FVA = SFO->getZExtValue();
2353 if (isPowerOf2_64(TVA) && isPowerOf2_64(FVA)) {
2354 // Compute the shift amounts
2355 unsigned TSA = Log2_64(TVA), FSA = Log2_64(FVA);
2356 // Construct the "on true" case of the select
2357 Constant *TC = ConstantInt::get(Type::UByteTy, TSA);
2359 new ShiftInst(Instruction::LShr, Op0, TC, SI->getName()+".t");
2360 TSI = InsertNewInstBefore(TSI, I);
2362 // Construct the "on false" case of the select
2363 Constant *FC = ConstantInt::get(Type::UByteTy, FSA);
2365 new ShiftInst(Instruction::LShr, Op0, FC, SI->getName()+".f");
2366 FSI = InsertNewInstBefore(FSI, I);
2368 // construct the select instruction and return it.
2369 return new SelectInst(SI->getOperand(0), TSI, FSI, SI->getName());
2376 Instruction *InstCombiner::visitSDiv(BinaryOperator &I) {
2377 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2379 // Handle the integer div common cases
2380 if (Instruction *Common = commonIDivTransforms(I))
2383 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2385 if (RHS->isAllOnesValue())
2386 return BinaryOperator::createNeg(Op0);
2389 if (Value *LHSNeg = dyn_castNegVal(Op0))
2390 return BinaryOperator::createSDiv(LHSNeg, ConstantExpr::getNeg(RHS));
2393 // If the sign bits of both operands are zero (i.e. we can prove they are
2394 // unsigned inputs), turn this into a udiv.
2395 if (I.getType()->isInteger()) {
2396 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2397 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2398 return BinaryOperator::createUDiv(Op0, Op1, I.getName());
2405 Instruction *InstCombiner::visitFDiv(BinaryOperator &I) {
2406 return commonDivTransforms(I);
2409 /// GetFactor - If we can prove that the specified value is at least a multiple
2410 /// of some factor, return that factor.
2411 static Constant *GetFactor(Value *V) {
2412 if (ConstantInt *CI = dyn_cast<ConstantInt>(V))
2415 // Unless we can be tricky, we know this is a multiple of 1.
2416 Constant *Result = ConstantInt::get(V->getType(), 1);
2418 Instruction *I = dyn_cast<Instruction>(V);
2419 if (!I) return Result;
2421 if (I->getOpcode() == Instruction::Mul) {
2422 // Handle multiplies by a constant, etc.
2423 return ConstantExpr::getMul(GetFactor(I->getOperand(0)),
2424 GetFactor(I->getOperand(1)));
2425 } else if (I->getOpcode() == Instruction::Shl) {
2426 // (X<<C) -> X * (1 << C)
2427 if (Constant *ShRHS = dyn_cast<Constant>(I->getOperand(1))) {
2428 ShRHS = ConstantExpr::getShl(Result, ShRHS);
2429 return ConstantExpr::getMul(GetFactor(I->getOperand(0)), ShRHS);
2431 } else if (I->getOpcode() == Instruction::And) {
2432 if (ConstantInt *RHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
2433 // X & 0xFFF0 is known to be a multiple of 16.
2434 unsigned Zeros = CountTrailingZeros_64(RHS->getZExtValue());
2435 if (Zeros != V->getType()->getPrimitiveSizeInBits())
2436 return ConstantExpr::getShl(Result,
2437 ConstantInt::get(Type::UByteTy, Zeros));
2439 } else if (CastInst *CI = dyn_cast<CastInst>(I)) {
2440 // Only handle int->int casts.
2441 if (!CI->isIntegerCast())
2443 Value *Op = CI->getOperand(0);
2444 return ConstantExpr::getCast(CI->getOpcode(), GetFactor(Op), V->getType());
2449 /// This function implements the transforms on rem instructions that work
2450 /// regardless of the kind of rem instruction it is (urem, srem, or frem). It
2451 /// is used by the visitors to those instructions.
2452 /// @brief Transforms common to all three rem instructions
2453 Instruction *InstCombiner::commonRemTransforms(BinaryOperator &I) {
2454 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2456 // 0 % X == 0, we don't need to preserve faults!
2457 if (Constant *LHS = dyn_cast<Constant>(Op0))
2458 if (LHS->isNullValue())
2459 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2461 if (isa<UndefValue>(Op0)) // undef % X -> 0
2462 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2463 if (isa<UndefValue>(Op1))
2464 return ReplaceInstUsesWith(I, Op1); // X % undef -> undef
2466 // Handle cases involving: rem X, (select Cond, Y, Z)
2467 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2468 // rem X, (Cond ? 0 : Y) -> rem X, Y. If the rem and the select are in
2469 // the same basic block, then we replace the select with Y, and the
2470 // condition of the select with false (if the cond value is in the same
2471 // BB). If the select has uses other than the div, this allows them to be
2473 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(1)))
2474 if (ST->isNullValue()) {
2475 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2476 if (CondI && CondI->getParent() == I.getParent())
2477 UpdateValueUsesWith(CondI, ConstantBool::getFalse());
2478 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2479 I.setOperand(1, SI->getOperand(2));
2481 UpdateValueUsesWith(SI, SI->getOperand(2));
2484 // Likewise for: rem X, (Cond ? Y : 0) -> rem X, Y
2485 if (Constant *ST = dyn_cast<Constant>(SI->getOperand(2)))
2486 if (ST->isNullValue()) {
2487 Instruction *CondI = dyn_cast<Instruction>(SI->getOperand(0));
2488 if (CondI && CondI->getParent() == I.getParent())
2489 UpdateValueUsesWith(CondI, ConstantBool::getTrue());
2490 else if (I.getParent() != SI->getParent() || SI->hasOneUse())
2491 I.setOperand(1, SI->getOperand(1));
2493 UpdateValueUsesWith(SI, SI->getOperand(1));
2501 /// This function implements the transforms common to both integer remainder
2502 /// instructions (urem and srem). It is called by the visitors to those integer
2503 /// remainder instructions.
2504 /// @brief Common integer remainder transforms
2505 Instruction *InstCombiner::commonIRemTransforms(BinaryOperator &I) {
2506 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2508 if (Instruction *common = commonRemTransforms(I))
2511 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2512 // X % 0 == undef, we don't need to preserve faults!
2513 if (RHS->equalsInt(0))
2514 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
2516 if (RHS->equalsInt(1)) // X % 1 == 0
2517 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2519 if (Instruction *Op0I = dyn_cast<Instruction>(Op0)) {
2520 if (SelectInst *SI = dyn_cast<SelectInst>(Op0I)) {
2521 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
2523 } else if (isa<PHINode>(Op0I)) {
2524 if (Instruction *NV = FoldOpIntoPhi(I))
2527 // (X * C1) % C2 --> 0 iff C1 % C2 == 0
2528 if (ConstantExpr::getSRem(GetFactor(Op0I), RHS)->isNullValue())
2529 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
2536 Instruction *InstCombiner::visitURem(BinaryOperator &I) {
2537 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2539 if (Instruction *common = commonIRemTransforms(I))
2542 if (ConstantInt *RHS = dyn_cast<ConstantInt>(Op1)) {
2543 // X urem C^2 -> X and C
2544 // Check to see if this is an unsigned remainder with an exact power of 2,
2545 // if so, convert to a bitwise and.
2546 if (ConstantInt *C = dyn_cast<ConstantInt>(RHS))
2547 if (isPowerOf2_64(C->getZExtValue()))
2548 return BinaryOperator::createAnd(Op0, SubOne(C));
2551 if (Instruction *RHSI = dyn_cast<Instruction>(I.getOperand(1))) {
2552 // Turn A % (C << N), where C is 2^k, into A & ((C << N)-1)
2553 if (RHSI->getOpcode() == Instruction::Shl &&
2554 isa<ConstantInt>(RHSI->getOperand(0))) {
2555 unsigned C1 = cast<ConstantInt>(RHSI->getOperand(0))->getZExtValue();
2556 if (isPowerOf2_64(C1)) {
2557 Constant *N1 = ConstantInt::getAllOnesValue(I.getType());
2558 Value *Add = InsertNewInstBefore(BinaryOperator::createAdd(RHSI, N1,
2560 return BinaryOperator::createAnd(Op0, Add);
2565 // urem X, (select Cond, 2^C1, 2^C2) --> select Cond, (and X, C1), (and X, C2)
2566 // where C1&C2 are powers of two.
2567 if (SelectInst *SI = dyn_cast<SelectInst>(Op1)) {
2568 if (ConstantInt *STO = dyn_cast<ConstantInt>(SI->getOperand(1)))
2569 if (ConstantInt *SFO = dyn_cast<ConstantInt>(SI->getOperand(2))) {
2570 // STO == 0 and SFO == 0 handled above.
2571 if (isPowerOf2_64(STO->getZExtValue()) &&
2572 isPowerOf2_64(SFO->getZExtValue())) {
2573 Value *TrueAnd = InsertNewInstBefore(
2574 BinaryOperator::createAnd(Op0, SubOne(STO), SI->getName()+".t"), I);
2575 Value *FalseAnd = InsertNewInstBefore(
2576 BinaryOperator::createAnd(Op0, SubOne(SFO), SI->getName()+".f"), I);
2577 return new SelectInst(SI->getOperand(0), TrueAnd, FalseAnd);
2585 Instruction *InstCombiner::visitSRem(BinaryOperator &I) {
2586 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2588 if (Instruction *common = commonIRemTransforms(I))
2591 if (Value *RHSNeg = dyn_castNegVal(Op1))
2592 if (!isa<ConstantInt>(RHSNeg) ||
2593 cast<ConstantInt>(RHSNeg)->getSExtValue() > 0) {
2595 AddUsesToWorkList(I);
2596 I.setOperand(1, RHSNeg);
2600 // If the top bits of both operands are zero (i.e. we can prove they are
2601 // unsigned inputs), turn this into a urem.
2602 uint64_t Mask = 1ULL << (I.getType()->getPrimitiveSizeInBits()-1);
2603 if (MaskedValueIsZero(Op1, Mask) && MaskedValueIsZero(Op0, Mask)) {
2604 // X srem Y -> X urem Y, iff X and Y don't have sign bit set
2605 return BinaryOperator::createURem(Op0, Op1, I.getName());
2611 Instruction *InstCombiner::visitFRem(BinaryOperator &I) {
2612 return commonRemTransforms(I);
2615 // isMaxValueMinusOne - return true if this is Max-1
2616 static bool isMaxValueMinusOne(const ConstantInt *C) {
2617 if (C->getType()->isUnsigned())
2618 return C->getZExtValue() == C->getType()->getIntegralTypeMask()-1;
2620 // Calculate 0111111111..11111
2621 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2622 int64_t Val = INT64_MAX; // All ones
2623 Val >>= 64-TypeBits; // Shift out unwanted 1 bits...
2624 return C->getSExtValue() == Val-1;
2627 // isMinValuePlusOne - return true if this is Min+1
2628 static bool isMinValuePlusOne(const ConstantInt *C) {
2629 if (C->getType()->isUnsigned())
2630 return C->getZExtValue() == 1;
2632 // Calculate 1111111111000000000000
2633 unsigned TypeBits = C->getType()->getPrimitiveSizeInBits();
2634 int64_t Val = -1; // All ones
2635 Val <<= TypeBits-1; // Shift over to the right spot
2636 return C->getSExtValue() == Val+1;
2639 // isOneBitSet - Return true if there is exactly one bit set in the specified
2641 static bool isOneBitSet(const ConstantInt *CI) {
2642 uint64_t V = CI->getZExtValue();
2643 return V && (V & (V-1)) == 0;
2646 #if 0 // Currently unused
2647 // isLowOnes - Return true if the constant is of the form 0+1+.
2648 static bool isLowOnes(const ConstantInt *CI) {
2649 uint64_t V = CI->getZExtValue();
2651 // There won't be bits set in parts that the type doesn't contain.
2652 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2654 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2655 return U && V && (U & V) == 0;
2659 // isHighOnes - Return true if the constant is of the form 1+0+.
2660 // This is the same as lowones(~X).
2661 static bool isHighOnes(const ConstantInt *CI) {
2662 uint64_t V = ~CI->getZExtValue();
2663 if (~V == 0) return false; // 0's does not match "1+"
2665 // There won't be bits set in parts that the type doesn't contain.
2666 V &= ConstantInt::getAllOnesValue(CI->getType())->getZExtValue();
2668 uint64_t U = V+1; // If it is low ones, this should be a power of two.
2669 return U && V && (U & V) == 0;
2673 /// getSetCondCode - Encode a setcc opcode into a three bit mask. These bits
2674 /// are carefully arranged to allow folding of expressions such as:
2676 /// (A < B) | (A > B) --> (A != B)
2678 /// Bit value '4' represents that the comparison is true if A > B, bit value '2'
2679 /// represents that the comparison is true if A == B, and bit value '1' is true
2682 static unsigned getSetCondCode(const SetCondInst *SCI) {
2683 switch (SCI->getOpcode()) {
2685 case Instruction::SetGT: return 1;
2686 case Instruction::SetEQ: return 2;
2687 case Instruction::SetGE: return 3;
2688 case Instruction::SetLT: return 4;
2689 case Instruction::SetNE: return 5;
2690 case Instruction::SetLE: return 6;
2693 assert(0 && "Invalid SetCC opcode!");
2698 /// getSetCCValue - This is the complement of getSetCondCode, which turns an
2699 /// opcode and two operands into either a constant true or false, or a brand new
2700 /// SetCC instruction.
2701 static Value *getSetCCValue(unsigned Opcode, Value *LHS, Value *RHS) {
2703 case 0: return ConstantBool::getFalse();
2704 case 1: return new SetCondInst(Instruction::SetGT, LHS, RHS);
2705 case 2: return new SetCondInst(Instruction::SetEQ, LHS, RHS);
2706 case 3: return new SetCondInst(Instruction::SetGE, LHS, RHS);
2707 case 4: return new SetCondInst(Instruction::SetLT, LHS, RHS);
2708 case 5: return new SetCondInst(Instruction::SetNE, LHS, RHS);
2709 case 6: return new SetCondInst(Instruction::SetLE, LHS, RHS);
2710 case 7: return ConstantBool::getTrue();
2711 default: assert(0 && "Illegal SetCCCode!"); return 0;
2715 // FoldSetCCLogical - Implements (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
2717 struct FoldSetCCLogical {
2720 FoldSetCCLogical(InstCombiner &ic, SetCondInst *SCI)
2721 : IC(ic), LHS(SCI->getOperand(0)), RHS(SCI->getOperand(1)) {}
2722 bool shouldApply(Value *V) const {
2723 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
2724 return (SCI->getOperand(0) == LHS && SCI->getOperand(1) == RHS ||
2725 SCI->getOperand(0) == RHS && SCI->getOperand(1) == LHS);
2728 Instruction *apply(BinaryOperator &Log) const {
2729 SetCondInst *SCI = cast<SetCondInst>(Log.getOperand(0));
2730 if (SCI->getOperand(0) != LHS) {
2731 assert(SCI->getOperand(1) == LHS);
2732 SCI->swapOperands(); // Swap the LHS and RHS of the SetCC
2735 unsigned LHSCode = getSetCondCode(SCI);
2736 unsigned RHSCode = getSetCondCode(cast<SetCondInst>(Log.getOperand(1)));
2738 switch (Log.getOpcode()) {
2739 case Instruction::And: Code = LHSCode & RHSCode; break;
2740 case Instruction::Or: Code = LHSCode | RHSCode; break;
2741 case Instruction::Xor: Code = LHSCode ^ RHSCode; break;
2742 default: assert(0 && "Illegal logical opcode!"); return 0;
2745 Value *RV = getSetCCValue(Code, LHS, RHS);
2746 if (Instruction *I = dyn_cast<Instruction>(RV))
2748 // Otherwise, it's a constant boolean value...
2749 return IC.ReplaceInstUsesWith(Log, RV);
2752 } // end anonymous namespace
2754 // OptAndOp - This handles expressions of the form ((val OP C1) & C2). Where
2755 // the Op parameter is 'OP', OpRHS is 'C1', and AndRHS is 'C2'. Op is
2756 // guaranteed to be either a shift instruction or a binary operator.
2757 Instruction *InstCombiner::OptAndOp(Instruction *Op,
2758 ConstantIntegral *OpRHS,
2759 ConstantIntegral *AndRHS,
2760 BinaryOperator &TheAnd) {
2761 Value *X = Op->getOperand(0);
2762 Constant *Together = 0;
2763 if (!isa<ShiftInst>(Op))
2764 Together = ConstantExpr::getAnd(AndRHS, OpRHS);
2766 switch (Op->getOpcode()) {
2767 case Instruction::Xor:
2768 if (Op->hasOneUse()) {
2769 // (X ^ C1) & C2 --> (X & C2) ^ (C1&C2)
2770 std::string OpName = Op->getName(); Op->setName("");
2771 Instruction *And = BinaryOperator::createAnd(X, AndRHS, OpName);
2772 InsertNewInstBefore(And, TheAnd);
2773 return BinaryOperator::createXor(And, Together);
2776 case Instruction::Or:
2777 if (Together == AndRHS) // (X | C) & C --> C
2778 return ReplaceInstUsesWith(TheAnd, AndRHS);
2780 if (Op->hasOneUse() && Together != OpRHS) {
2781 // (X | C1) & C2 --> (X | (C1&C2)) & C2
2782 std::string Op0Name = Op->getName(); Op->setName("");
2783 Instruction *Or = BinaryOperator::createOr(X, Together, Op0Name);
2784 InsertNewInstBefore(Or, TheAnd);
2785 return BinaryOperator::createAnd(Or, AndRHS);
2788 case Instruction::Add:
2789 if (Op->hasOneUse()) {
2790 // Adding a one to a single bit bit-field should be turned into an XOR
2791 // of the bit. First thing to check is to see if this AND is with a
2792 // single bit constant.
2793 uint64_t AndRHSV = cast<ConstantInt>(AndRHS)->getZExtValue();
2795 // Clear bits that are not part of the constant.
2796 AndRHSV &= AndRHS->getType()->getIntegralTypeMask();
2798 // If there is only one bit set...
2799 if (isOneBitSet(cast<ConstantInt>(AndRHS))) {
2800 // Ok, at this point, we know that we are masking the result of the
2801 // ADD down to exactly one bit. If the constant we are adding has
2802 // no bits set below this bit, then we can eliminate the ADD.
2803 uint64_t AddRHS = cast<ConstantInt>(OpRHS)->getZExtValue();
2805 // Check to see if any bits below the one bit set in AndRHSV are set.
2806 if ((AddRHS & (AndRHSV-1)) == 0) {
2807 // If not, the only thing that can effect the output of the AND is
2808 // the bit specified by AndRHSV. If that bit is set, the effect of
2809 // the XOR is to toggle the bit. If it is clear, then the ADD has
2811 if ((AddRHS & AndRHSV) == 0) { // Bit is not set, noop
2812 TheAnd.setOperand(0, X);
2815 std::string Name = Op->getName(); Op->setName("");
2816 // Pull the XOR out of the AND.
2817 Instruction *NewAnd = BinaryOperator::createAnd(X, AndRHS, Name);
2818 InsertNewInstBefore(NewAnd, TheAnd);
2819 return BinaryOperator::createXor(NewAnd, AndRHS);
2826 case Instruction::Shl: {
2827 // We know that the AND will not produce any of the bits shifted in, so if
2828 // the anded constant includes them, clear them now!
2830 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2831 Constant *ShlMask = ConstantExpr::getShl(AllOne, OpRHS);
2832 Constant *CI = ConstantExpr::getAnd(AndRHS, ShlMask);
2834 if (CI == ShlMask) { // Masking out bits that the shift already masks
2835 return ReplaceInstUsesWith(TheAnd, Op); // No need for the and.
2836 } else if (CI != AndRHS) { // Reducing bits set in and.
2837 TheAnd.setOperand(1, CI);
2842 case Instruction::LShr:
2844 // We know that the AND will not produce any of the bits shifted in, so if
2845 // the anded constant includes them, clear them now! This only applies to
2846 // unsigned shifts, because a signed shr may bring in set bits!
2848 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2849 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2850 Constant *CI = ConstantExpr::getAnd(AndRHS, ShrMask);
2852 if (CI == ShrMask) { // Masking out bits that the shift already masks.
2853 return ReplaceInstUsesWith(TheAnd, Op);
2854 } else if (CI != AndRHS) {
2855 TheAnd.setOperand(1, CI); // Reduce bits set in and cst.
2860 case Instruction::AShr:
2862 // See if this is shifting in some sign extension, then masking it out
2864 if (Op->hasOneUse()) {
2865 Constant *AllOne = ConstantIntegral::getAllOnesValue(AndRHS->getType());
2866 Constant *ShrMask = ConstantExpr::getLShr(AllOne, OpRHS);
2867 Constant *C = ConstantExpr::getAnd(AndRHS, ShrMask);
2868 if (C == AndRHS) { // Masking out bits shifted in.
2869 // (Val ashr C1) & C2 -> (Val lshr C1) & C2
2870 // Make the argument unsigned.
2871 Value *ShVal = Op->getOperand(0);
2872 ShVal = InsertNewInstBefore(new ShiftInst(Instruction::LShr, ShVal,
2873 OpRHS, Op->getName()), TheAnd);
2874 return BinaryOperator::createAnd(ShVal, AndRHS, TheAnd.getName());
2883 /// InsertRangeTest - Emit a computation of: (V >= Lo && V < Hi) if Inside is
2884 /// true, otherwise (V < Lo || V >= Hi). In pratice, we emit the more efficient
2885 /// (V-Lo) <u Hi-Lo. This method expects that Lo <= Hi. IB is the location to
2886 /// insert new instructions.
2887 Instruction *InstCombiner::InsertRangeTest(Value *V, Constant *Lo, Constant *Hi,
2888 bool Inside, Instruction &IB) {
2889 assert(cast<ConstantBool>(ConstantExpr::getSetLE(Lo, Hi))->getValue() &&
2890 "Lo is not <= Hi in range emission code!");
2892 if (Lo == Hi) // Trivially false.
2893 return new SetCondInst(Instruction::SetNE, V, V);
2894 if (cast<ConstantIntegral>(Lo)->isMinValue(Lo->getType()->isSigned()))
2895 return new SetCondInst(Instruction::SetLT, V, Hi);
2897 Constant *AddCST = ConstantExpr::getNeg(Lo);
2898 Instruction *Add = BinaryOperator::createAdd(V, AddCST,V->getName()+".off");
2899 InsertNewInstBefore(Add, IB);
2900 // Convert to unsigned for the comparison.
2901 const Type *UnsType = Add->getType()->getUnsignedVersion();
2902 Value *OffsetVal = InsertCastBefore(Instruction::BitCast, Add, UnsType, IB);
2903 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2904 AddCST = ConstantExpr::getBitCast(AddCST, UnsType);
2905 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
2908 if (Lo == Hi) // Trivially true.
2909 return new SetCondInst(Instruction::SetEQ, V, V);
2911 Hi = SubOne(cast<ConstantInt>(Hi));
2913 // V < 0 || V >= Hi ->'V > Hi-1'
2914 if (cast<ConstantIntegral>(Lo)->isMinValue(Lo->getType()->isSigned()))
2915 return new SetCondInst(Instruction::SetGT, V, Hi);
2917 // Emit X-Lo > Hi-Lo-1
2918 Constant *AddCST = ConstantExpr::getNeg(Lo);
2919 Instruction *Add = BinaryOperator::createAdd(V, AddCST, V->getName()+".off");
2920 InsertNewInstBefore(Add, IB);
2921 // Convert to unsigned for the comparison.
2922 const Type *UnsType = Add->getType()->getUnsignedVersion();
2923 Value *OffsetVal = InsertCastBefore(Instruction::BitCast, Add, UnsType, IB);
2924 AddCST = ConstantExpr::getAdd(AddCST, Hi);
2925 AddCST = ConstantExpr::getBitCast(AddCST, UnsType);
2926 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
2929 // isRunOfOnes - Returns true iff Val consists of one contiguous run of 1s with
2930 // any number of 0s on either side. The 1s are allowed to wrap from LSB to
2931 // MSB, so 0x000FFF0, 0x0000FFFF, and 0xFF0000FF are all runs. 0x0F0F0000 is
2932 // not, since all 1s are not contiguous.
2933 static bool isRunOfOnes(ConstantIntegral *Val, unsigned &MB, unsigned &ME) {
2934 uint64_t V = Val->getZExtValue();
2935 if (!isShiftedMask_64(V)) return false;
2937 // look for the first zero bit after the run of ones
2938 MB = 64-CountLeadingZeros_64((V - 1) ^ V);
2939 // look for the first non-zero bit
2940 ME = 64-CountLeadingZeros_64(V);
2946 /// FoldLogicalPlusAnd - This is part of an expression (LHS +/- RHS) & Mask,
2947 /// where isSub determines whether the operator is a sub. If we can fold one of
2948 /// the following xforms:
2950 /// ((A & N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == Mask
2951 /// ((A | N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2952 /// ((A ^ N) +/- B) & Mask -> (A +/- B) & Mask iff N&Mask == 0
2954 /// return (A +/- B).
2956 Value *InstCombiner::FoldLogicalPlusAnd(Value *LHS, Value *RHS,
2957 ConstantIntegral *Mask, bool isSub,
2959 Instruction *LHSI = dyn_cast<Instruction>(LHS);
2960 if (!LHSI || LHSI->getNumOperands() != 2 ||
2961 !isa<ConstantInt>(LHSI->getOperand(1))) return 0;
2963 ConstantInt *N = cast<ConstantInt>(LHSI->getOperand(1));
2965 switch (LHSI->getOpcode()) {
2967 case Instruction::And:
2968 if (ConstantExpr::getAnd(N, Mask) == Mask) {
2969 // If the AndRHS is a power of two minus one (0+1+), this is simple.
2970 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0)
2973 // Otherwise, if Mask is 0+1+0+, and if B is known to have the low 0+
2974 // part, we don't need any explicit masks to take them out of A. If that
2975 // is all N is, ignore it.
2977 if (isRunOfOnes(Mask, MB, ME)) { // begin/end bit of run, inclusive
2978 uint64_t Mask = RHS->getType()->getIntegralTypeMask();
2980 if (MaskedValueIsZero(RHS, Mask))
2985 case Instruction::Or:
2986 case Instruction::Xor:
2987 // If the AndRHS is a power of two minus one (0+1+), and N&Mask == 0
2988 if ((Mask->getZExtValue() & Mask->getZExtValue()+1) == 0 &&
2989 ConstantExpr::getAnd(N, Mask)->isNullValue())
2996 New = BinaryOperator::createSub(LHSI->getOperand(0), RHS, "fold");
2998 New = BinaryOperator::createAdd(LHSI->getOperand(0), RHS, "fold");
2999 return InsertNewInstBefore(New, I);
3002 Instruction *InstCombiner::visitAnd(BinaryOperator &I) {
3003 bool Changed = SimplifyCommutative(I);
3004 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3006 if (isa<UndefValue>(Op1)) // X & undef -> 0
3007 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3011 return ReplaceInstUsesWith(I, Op1);
3013 // See if we can simplify any instructions used by the instruction whose sole
3014 // purpose is to compute bits we don't care about.
3015 uint64_t KnownZero, KnownOne;
3016 if (!isa<PackedType>(I.getType()) &&
3017 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3018 KnownZero, KnownOne))
3021 if (ConstantIntegral *AndRHS = dyn_cast<ConstantIntegral>(Op1)) {
3022 uint64_t AndRHSMask = AndRHS->getZExtValue();
3023 uint64_t TypeMask = Op0->getType()->getIntegralTypeMask();
3024 uint64_t NotAndRHS = AndRHSMask^TypeMask;
3026 // Optimize a variety of ((val OP C1) & C2) combinations...
3027 if (isa<BinaryOperator>(Op0) || isa<ShiftInst>(Op0)) {
3028 Instruction *Op0I = cast<Instruction>(Op0);
3029 Value *Op0LHS = Op0I->getOperand(0);
3030 Value *Op0RHS = Op0I->getOperand(1);
3031 switch (Op0I->getOpcode()) {
3032 case Instruction::Xor:
3033 case Instruction::Or:
3034 // If the mask is only needed on one incoming arm, push it up.
3035 if (Op0I->hasOneUse()) {
3036 if (MaskedValueIsZero(Op0LHS, NotAndRHS)) {
3037 // Not masking anything out for the LHS, move to RHS.
3038 Instruction *NewRHS = BinaryOperator::createAnd(Op0RHS, AndRHS,
3039 Op0RHS->getName()+".masked");
3040 InsertNewInstBefore(NewRHS, I);
3041 return BinaryOperator::create(
3042 cast<BinaryOperator>(Op0I)->getOpcode(), Op0LHS, NewRHS);
3044 if (!isa<Constant>(Op0RHS) &&
3045 MaskedValueIsZero(Op0RHS, NotAndRHS)) {
3046 // Not masking anything out for the RHS, move to LHS.
3047 Instruction *NewLHS = BinaryOperator::createAnd(Op0LHS, AndRHS,
3048 Op0LHS->getName()+".masked");
3049 InsertNewInstBefore(NewLHS, I);
3050 return BinaryOperator::create(
3051 cast<BinaryOperator>(Op0I)->getOpcode(), NewLHS, Op0RHS);
3056 case Instruction::Add:
3057 // ((A & N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == AndRHS.
3058 // ((A | N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3059 // ((A ^ N) + B) & AndRHS -> (A + B) & AndRHS iff N&AndRHS == 0
3060 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, false, I))
3061 return BinaryOperator::createAnd(V, AndRHS);
3062 if (Value *V = FoldLogicalPlusAnd(Op0RHS, Op0LHS, AndRHS, false, I))
3063 return BinaryOperator::createAnd(V, AndRHS); // Add commutes
3066 case Instruction::Sub:
3067 // ((A & N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == AndRHS.
3068 // ((A | N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3069 // ((A ^ N) - B) & AndRHS -> (A - B) & AndRHS iff N&AndRHS == 0
3070 if (Value *V = FoldLogicalPlusAnd(Op0LHS, Op0RHS, AndRHS, true, I))
3071 return BinaryOperator::createAnd(V, AndRHS);
3075 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3076 if (Instruction *Res = OptAndOp(Op0I, Op0CI, AndRHS, I))
3078 } else if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
3079 // If this is an integer truncation or change from signed-to-unsigned, and
3080 // if the source is an and/or with immediate, transform it. This
3081 // frequently occurs for bitfield accesses.
3082 if (Instruction *CastOp = dyn_cast<Instruction>(CI->getOperand(0))) {
3083 if ((isa<TruncInst>(CI) || isa<BitCastInst>(CI)) &&
3084 CastOp->getNumOperands() == 2)
3085 if (ConstantInt *AndCI = dyn_cast<ConstantInt>(CastOp->getOperand(1)))
3086 if (CastOp->getOpcode() == Instruction::And) {
3087 // Change: and (cast (and X, C1) to T), C2
3088 // into : and (cast X to T), trunc_or_bitcast(C1)&C2
3089 // This will fold the two constants together, which may allow
3090 // other simplifications.
3091 Instruction *NewCast = CastInst::createTruncOrBitCast(
3092 CastOp->getOperand(0), I.getType(),
3093 CastOp->getName()+".shrunk");
3094 NewCast = InsertNewInstBefore(NewCast, I);
3095 // trunc_or_bitcast(C1)&C2
3096 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3097 C3 = ConstantExpr::getAnd(C3, AndRHS);
3098 return BinaryOperator::createAnd(NewCast, C3);
3099 } else if (CastOp->getOpcode() == Instruction::Or) {
3100 // Change: and (cast (or X, C1) to T), C2
3101 // into : trunc(C1)&C2 iff trunc(C1)&C2 == C2
3102 Constant *C3 = ConstantExpr::getTruncOrBitCast(AndCI,I.getType());
3103 if (ConstantExpr::getAnd(C3, AndRHS) == AndRHS) // trunc(C1)&C2
3104 return ReplaceInstUsesWith(I, AndRHS);
3109 // Try to fold constant and into select arguments.
3110 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3111 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3113 if (isa<PHINode>(Op0))
3114 if (Instruction *NV = FoldOpIntoPhi(I))
3118 Value *Op0NotVal = dyn_castNotVal(Op0);
3119 Value *Op1NotVal = dyn_castNotVal(Op1);
3121 if (Op0NotVal == Op1 || Op1NotVal == Op0) // A & ~A == ~A & A == 0
3122 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3124 // (~A & ~B) == (~(A | B)) - De Morgan's Law
3125 if (Op0NotVal && Op1NotVal && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3126 Instruction *Or = BinaryOperator::createOr(Op0NotVal, Op1NotVal,
3127 I.getName()+".demorgan");
3128 InsertNewInstBefore(Or, I);
3129 return BinaryOperator::createNot(Or);
3133 Value *A = 0, *B = 0;
3134 if (match(Op0, m_Or(m_Value(A), m_Value(B))))
3135 if (A == Op1 || B == Op1) // (A | ?) & A --> A
3136 return ReplaceInstUsesWith(I, Op1);
3137 if (match(Op1, m_Or(m_Value(A), m_Value(B))))
3138 if (A == Op0 || B == Op0) // A & (A | ?) --> A
3139 return ReplaceInstUsesWith(I, Op0);
3141 if (Op0->hasOneUse() &&
3142 match(Op0, m_Xor(m_Value(A), m_Value(B)))) {
3143 if (A == Op1) { // (A^B)&A -> A&(A^B)
3144 I.swapOperands(); // Simplify below
3145 std::swap(Op0, Op1);
3146 } else if (B == Op1) { // (A^B)&B -> B&(B^A)
3147 cast<BinaryOperator>(Op0)->swapOperands();
3148 I.swapOperands(); // Simplify below
3149 std::swap(Op0, Op1);
3152 if (Op1->hasOneUse() &&
3153 match(Op1, m_Xor(m_Value(A), m_Value(B)))) {
3154 if (B == Op0) { // B&(A^B) -> B&(B^A)
3155 cast<BinaryOperator>(Op1)->swapOperands();
3158 if (A == Op0) { // A&(A^B) -> A & ~B
3159 Instruction *NotB = BinaryOperator::createNot(B, "tmp");
3160 InsertNewInstBefore(NotB, I);
3161 return BinaryOperator::createAnd(A, NotB);
3167 if (SetCondInst *RHS = dyn_cast<SetCondInst>(Op1)) {
3168 // (setcc1 A, B) & (setcc2 A, B) --> (setcc3 A, B)
3169 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3172 Value *LHSVal, *RHSVal;
3173 ConstantInt *LHSCst, *RHSCst;
3174 Instruction::BinaryOps LHSCC, RHSCC;
3175 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3176 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3177 if (LHSVal == RHSVal && // Found (X setcc C1) & (X setcc C2)
3178 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
3179 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
3180 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
3181 // Ensure that the larger constant is on the RHS.
3182 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
3183 SetCondInst *LHS = cast<SetCondInst>(Op0);
3184 if (cast<ConstantBool>(Cmp)->getValue()) {
3185 std::swap(LHS, RHS);
3186 std::swap(LHSCst, RHSCst);
3187 std::swap(LHSCC, RHSCC);
3190 // At this point, we know we have have two setcc instructions
3191 // comparing a value against two constants and and'ing the result
3192 // together. Because of the above check, we know that we only have
3193 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
3194 // FoldSetCCLogical check above), that the two constants are not
3196 assert(LHSCst != RHSCst && "Compares not folded above?");
3199 default: assert(0 && "Unknown integer condition code!");
3200 case Instruction::SetEQ:
3202 default: assert(0 && "Unknown integer condition code!");
3203 case Instruction::SetEQ: // (X == 13 & X == 15) -> false
3204 case Instruction::SetGT: // (X == 13 & X > 15) -> false
3205 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3206 case Instruction::SetNE: // (X == 13 & X != 15) -> X == 13
3207 case Instruction::SetLT: // (X == 13 & X < 15) -> X == 13
3208 return ReplaceInstUsesWith(I, LHS);
3210 case Instruction::SetNE:
3212 default: assert(0 && "Unknown integer condition code!");
3213 case Instruction::SetLT:
3214 if (LHSCst == SubOne(RHSCst)) // (X != 13 & X < 14) -> X < 13
3215 return new SetCondInst(Instruction::SetLT, LHSVal, LHSCst);
3216 break; // (X != 13 & X < 15) -> no change
3217 case Instruction::SetEQ: // (X != 13 & X == 15) -> X == 15
3218 case Instruction::SetGT: // (X != 13 & X > 15) -> X > 15
3219 return ReplaceInstUsesWith(I, RHS);
3220 case Instruction::SetNE:
3221 if (LHSCst == SubOne(RHSCst)) {// (X != 13 & X != 14) -> X-13 >u 1
3222 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3223 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3224 LHSVal->getName()+".off");
3225 InsertNewInstBefore(Add, I);
3226 const Type *UnsType = Add->getType()->getUnsignedVersion();
3227 Value *OffsetVal = InsertCastBefore(Instruction::BitCast, Add,
3229 AddCST = ConstantExpr::getSub(RHSCst, LHSCst);
3230 AddCST = ConstantExpr::getBitCast(AddCST, UnsType);
3231 return new SetCondInst(Instruction::SetGT, OffsetVal, AddCST);
3233 break; // (X != 13 & X != 15) -> no change
3236 case Instruction::SetLT:
3238 default: assert(0 && "Unknown integer condition code!");
3239 case Instruction::SetEQ: // (X < 13 & X == 15) -> false
3240 case Instruction::SetGT: // (X < 13 & X > 15) -> false
3241 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
3242 case Instruction::SetNE: // (X < 13 & X != 15) -> X < 13
3243 case Instruction::SetLT: // (X < 13 & X < 15) -> X < 13
3244 return ReplaceInstUsesWith(I, LHS);
3246 case Instruction::SetGT:
3248 default: assert(0 && "Unknown integer condition code!");
3249 case Instruction::SetEQ: // (X > 13 & X == 15) -> X > 13
3250 return ReplaceInstUsesWith(I, LHS);
3251 case Instruction::SetGT: // (X > 13 & X > 15) -> X > 15
3252 return ReplaceInstUsesWith(I, RHS);
3253 case Instruction::SetNE:
3254 if (RHSCst == AddOne(LHSCst)) // (X > 13 & X != 14) -> X > 14
3255 return new SetCondInst(Instruction::SetGT, LHSVal, RHSCst);
3256 break; // (X > 13 & X != 15) -> no change
3257 case Instruction::SetLT: // (X > 13 & X < 15) -> (X-14) <u 1
3258 return InsertRangeTest(LHSVal, AddOne(LHSCst), RHSCst, true, I);
3264 // fold (and (cast A), (cast B)) -> (cast (and A, B))
3265 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3266 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3267 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind ?
3268 const Type *SrcTy = Op0C->getOperand(0)->getType();
3269 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3270 // Only do this if the casts both really cause code to be generated.
3271 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3272 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3273 Instruction *NewOp = BinaryOperator::createAnd(Op0C->getOperand(0),
3274 Op1C->getOperand(0),
3276 InsertNewInstBefore(NewOp, I);
3277 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3281 // (X >> Z) & (Y >> Z) -> (X&Y) >> Z for all shifts.
3282 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3283 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3284 if (SI0->getOpcode() == SI1->getOpcode() &&
3285 SI0->getOperand(1) == SI1->getOperand(1) &&
3286 (SI0->hasOneUse() || SI1->hasOneUse())) {
3287 Instruction *NewOp =
3288 InsertNewInstBefore(BinaryOperator::createAnd(SI0->getOperand(0),
3290 SI0->getName()), I);
3291 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3295 return Changed ? &I : 0;
3298 /// CollectBSwapParts - Look to see if the specified value defines a single byte
3299 /// in the result. If it does, and if the specified byte hasn't been filled in
3300 /// yet, fill it in and return false.
3301 static bool CollectBSwapParts(Value *V, std::vector<Value*> &ByteValues) {
3302 Instruction *I = dyn_cast<Instruction>(V);
3303 if (I == 0) return true;
3305 // If this is an or instruction, it is an inner node of the bswap.
3306 if (I->getOpcode() == Instruction::Or)
3307 return CollectBSwapParts(I->getOperand(0), ByteValues) ||
3308 CollectBSwapParts(I->getOperand(1), ByteValues);
3310 // If this is a shift by a constant int, and it is "24", then its operand
3311 // defines a byte. We only handle unsigned types here.
3312 if (isa<ShiftInst>(I) && isa<ConstantInt>(I->getOperand(1))) {
3313 // Not shifting the entire input by N-1 bytes?
3314 if (cast<ConstantInt>(I->getOperand(1))->getZExtValue() !=
3315 8*(ByteValues.size()-1))
3319 if (I->getOpcode() == Instruction::Shl) {
3320 // X << 24 defines the top byte with the lowest of the input bytes.
3321 DestNo = ByteValues.size()-1;
3323 // X >>u 24 defines the low byte with the highest of the input bytes.
3327 // If the destination byte value is already defined, the values are or'd
3328 // together, which isn't a bswap (unless it's an or of the same bits).
3329 if (ByteValues[DestNo] && ByteValues[DestNo] != I->getOperand(0))
3331 ByteValues[DestNo] = I->getOperand(0);
3335 // Otherwise, we can only handle and(shift X, imm), imm). Bail out of if we
3337 Value *Shift = 0, *ShiftLHS = 0;
3338 ConstantInt *AndAmt = 0, *ShiftAmt = 0;
3339 if (!match(I, m_And(m_Value(Shift), m_ConstantInt(AndAmt))) ||
3340 !match(Shift, m_Shift(m_Value(ShiftLHS), m_ConstantInt(ShiftAmt))))
3342 Instruction *SI = cast<Instruction>(Shift);
3344 // Make sure that the shift amount is by a multiple of 8 and isn't too big.
3345 if (ShiftAmt->getZExtValue() & 7 ||
3346 ShiftAmt->getZExtValue() > 8*ByteValues.size())
3349 // Turn 0xFF -> 0, 0xFF00 -> 1, 0xFF0000 -> 2, etc.
3351 for (DestByte = 0; DestByte != ByteValues.size(); ++DestByte)
3352 if (AndAmt->getZExtValue() == uint64_t(0xFF) << 8*DestByte)
3354 // Unknown mask for bswap.
3355 if (DestByte == ByteValues.size()) return true;
3357 unsigned ShiftBytes = ShiftAmt->getZExtValue()/8;
3359 if (SI->getOpcode() == Instruction::Shl)
3360 SrcByte = DestByte - ShiftBytes;
3362 SrcByte = DestByte + ShiftBytes;
3364 // If the SrcByte isn't a bswapped value from the DestByte, reject it.
3365 if (SrcByte != ByteValues.size()-DestByte-1)
3368 // If the destination byte value is already defined, the values are or'd
3369 // together, which isn't a bswap (unless it's an or of the same bits).
3370 if (ByteValues[DestByte] && ByteValues[DestByte] != SI->getOperand(0))
3372 ByteValues[DestByte] = SI->getOperand(0);
3376 /// MatchBSwap - Given an OR instruction, check to see if this is a bswap idiom.
3377 /// If so, insert the new bswap intrinsic and return it.
3378 Instruction *InstCombiner::MatchBSwap(BinaryOperator &I) {
3379 // We can only handle bswap of unsigned integers, and cannot bswap one byte.
3380 if (!I.getType()->isUnsigned() || I.getType() == Type::UByteTy)
3383 /// ByteValues - For each byte of the result, we keep track of which value
3384 /// defines each byte.
3385 std::vector<Value*> ByteValues;
3386 ByteValues.resize(I.getType()->getPrimitiveSize());
3388 // Try to find all the pieces corresponding to the bswap.
3389 if (CollectBSwapParts(I.getOperand(0), ByteValues) ||
3390 CollectBSwapParts(I.getOperand(1), ByteValues))
3393 // Check to see if all of the bytes come from the same value.
3394 Value *V = ByteValues[0];
3395 if (V == 0) return 0; // Didn't find a byte? Must be zero.
3397 // Check to make sure that all of the bytes come from the same value.
3398 for (unsigned i = 1, e = ByteValues.size(); i != e; ++i)
3399 if (ByteValues[i] != V)
3402 // If they do then *success* we can turn this into a bswap. Figure out what
3403 // bswap to make it into.
3404 Module *M = I.getParent()->getParent()->getParent();
3405 const char *FnName = 0;
3406 if (I.getType() == Type::UShortTy)
3407 FnName = "llvm.bswap.i16";
3408 else if (I.getType() == Type::UIntTy)
3409 FnName = "llvm.bswap.i32";
3410 else if (I.getType() == Type::ULongTy)
3411 FnName = "llvm.bswap.i64";
3413 assert(0 && "Unknown integer type!");
3414 Function *F = M->getOrInsertFunction(FnName, I.getType(), I.getType(), NULL);
3416 return new CallInst(F, V);
3420 Instruction *InstCombiner::visitOr(BinaryOperator &I) {
3421 bool Changed = SimplifyCommutative(I);
3422 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3424 if (isa<UndefValue>(Op1))
3425 return ReplaceInstUsesWith(I, // X | undef -> -1
3426 ConstantIntegral::getAllOnesValue(I.getType()));
3430 return ReplaceInstUsesWith(I, Op0);
3432 // See if we can simplify any instructions used by the instruction whose sole
3433 // purpose is to compute bits we don't care about.
3434 uint64_t KnownZero, KnownOne;
3435 if (!isa<PackedType>(I.getType()) &&
3436 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3437 KnownZero, KnownOne))
3441 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3442 ConstantInt *C1 = 0; Value *X = 0;
3443 // (X & C1) | C2 --> (X | C2) & (C1|C2)
3444 if (match(Op0, m_And(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3445 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0->getName());
3447 InsertNewInstBefore(Or, I);
3448 return BinaryOperator::createAnd(Or, ConstantExpr::getOr(RHS, C1));
3451 // (X ^ C1) | C2 --> (X | C2) ^ (C1&~C2)
3452 if (match(Op0, m_Xor(m_Value(X), m_ConstantInt(C1))) && isOnlyUse(Op0)) {
3453 std::string Op0Name = Op0->getName(); Op0->setName("");
3454 Instruction *Or = BinaryOperator::createOr(X, RHS, Op0Name);
3455 InsertNewInstBefore(Or, I);
3456 return BinaryOperator::createXor(Or,
3457 ConstantExpr::getAnd(C1, ConstantExpr::getNot(RHS)));
3460 // Try to fold constant and into select arguments.
3461 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3462 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3464 if (isa<PHINode>(Op0))
3465 if (Instruction *NV = FoldOpIntoPhi(I))
3469 Value *A = 0, *B = 0;
3470 ConstantInt *C1 = 0, *C2 = 0;
3472 if (match(Op0, m_And(m_Value(A), m_Value(B))))
3473 if (A == Op1 || B == Op1) // (A & ?) | A --> A
3474 return ReplaceInstUsesWith(I, Op1);
3475 if (match(Op1, m_And(m_Value(A), m_Value(B))))
3476 if (A == Op0 || B == Op0) // A | (A & ?) --> A
3477 return ReplaceInstUsesWith(I, Op0);
3479 // (A | B) | C and A | (B | C) -> bswap if possible.
3480 // (A >> B) | (C << D) and (A << B) | (B >> C) -> bswap if possible.
3481 if (match(Op0, m_Or(m_Value(), m_Value())) ||
3482 match(Op1, m_Or(m_Value(), m_Value())) ||
3483 (match(Op0, m_Shift(m_Value(), m_Value())) &&
3484 match(Op1, m_Shift(m_Value(), m_Value())))) {
3485 if (Instruction *BSwap = MatchBSwap(I))
3489 // (X^C)|Y -> (X|Y)^C iff Y&C == 0
3490 if (Op0->hasOneUse() && match(Op0, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3491 MaskedValueIsZero(Op1, C1->getZExtValue())) {
3492 Instruction *NOr = BinaryOperator::createOr(A, Op1, Op0->getName());
3494 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3497 // Y|(X^C) -> (X|Y)^C iff Y&C == 0
3498 if (Op1->hasOneUse() && match(Op1, m_Xor(m_Value(A), m_ConstantInt(C1))) &&
3499 MaskedValueIsZero(Op0, C1->getZExtValue())) {
3500 Instruction *NOr = BinaryOperator::createOr(A, Op0, Op1->getName());
3502 return BinaryOperator::createXor(InsertNewInstBefore(NOr, I), C1);
3505 // (A & C1)|(B & C2)
3506 if (match(Op0, m_And(m_Value(A), m_ConstantInt(C1))) &&
3507 match(Op1, m_And(m_Value(B), m_ConstantInt(C2)))) {
3509 if (A == B) // (A & C1)|(A & C2) == A & (C1|C2)
3510 return BinaryOperator::createAnd(A, ConstantExpr::getOr(C1, C2));
3513 // If we have: ((V + N) & C1) | (V & C2)
3514 // .. and C2 = ~C1 and C2 is 0+1+ and (N & C2) == 0
3515 // replace with V+N.
3516 if (C1 == ConstantExpr::getNot(C2)) {
3517 Value *V1 = 0, *V2 = 0;
3518 if ((C2->getZExtValue() & (C2->getZExtValue()+1)) == 0 && // C2 == 0+1+
3519 match(A, m_Add(m_Value(V1), m_Value(V2)))) {
3520 // Add commutes, try both ways.
3521 if (V1 == B && MaskedValueIsZero(V2, C2->getZExtValue()))
3522 return ReplaceInstUsesWith(I, A);
3523 if (V2 == B && MaskedValueIsZero(V1, C2->getZExtValue()))
3524 return ReplaceInstUsesWith(I, A);
3526 // Or commutes, try both ways.
3527 if ((C1->getZExtValue() & (C1->getZExtValue()+1)) == 0 &&
3528 match(B, m_Add(m_Value(V1), m_Value(V2)))) {
3529 // Add commutes, try both ways.
3530 if (V1 == A && MaskedValueIsZero(V2, C1->getZExtValue()))
3531 return ReplaceInstUsesWith(I, B);
3532 if (V2 == A && MaskedValueIsZero(V1, C1->getZExtValue()))
3533 return ReplaceInstUsesWith(I, B);
3538 // (X >> Z) | (Y >> Z) -> (X|Y) >> Z for all shifts.
3539 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3540 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3541 if (SI0->getOpcode() == SI1->getOpcode() &&
3542 SI0->getOperand(1) == SI1->getOperand(1) &&
3543 (SI0->hasOneUse() || SI1->hasOneUse())) {
3544 Instruction *NewOp =
3545 InsertNewInstBefore(BinaryOperator::createOr(SI0->getOperand(0),
3547 SI0->getName()), I);
3548 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3552 if (match(Op0, m_Not(m_Value(A)))) { // ~A | Op1
3553 if (A == Op1) // ~A | A == -1
3554 return ReplaceInstUsesWith(I,
3555 ConstantIntegral::getAllOnesValue(I.getType()));
3559 // Note, A is still live here!
3560 if (match(Op1, m_Not(m_Value(B)))) { // Op0 | ~B
3562 return ReplaceInstUsesWith(I,
3563 ConstantIntegral::getAllOnesValue(I.getType()));
3565 // (~A | ~B) == (~(A & B)) - De Morgan's Law
3566 if (A && isOnlyUse(Op0) && isOnlyUse(Op1)) {
3567 Value *And = InsertNewInstBefore(BinaryOperator::createAnd(A, B,
3568 I.getName()+".demorgan"), I);
3569 return BinaryOperator::createNot(And);
3573 // (setcc1 A, B) | (setcc2 A, B) --> (setcc3 A, B)
3574 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1))) {
3575 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3578 Value *LHSVal, *RHSVal;
3579 ConstantInt *LHSCst, *RHSCst;
3580 Instruction::BinaryOps LHSCC, RHSCC;
3581 if (match(Op0, m_SetCond(LHSCC, m_Value(LHSVal), m_ConstantInt(LHSCst))))
3582 if (match(RHS, m_SetCond(RHSCC, m_Value(RHSVal), m_ConstantInt(RHSCst))))
3583 if (LHSVal == RHSVal && // Found (X setcc C1) | (X setcc C2)
3584 // Set[GL]E X, CST is folded to Set[GL]T elsewhere.
3585 LHSCC != Instruction::SetGE && LHSCC != Instruction::SetLE &&
3586 RHSCC != Instruction::SetGE && RHSCC != Instruction::SetLE) {
3587 // Ensure that the larger constant is on the RHS.
3588 Constant *Cmp = ConstantExpr::getSetGT(LHSCst, RHSCst);
3589 SetCondInst *LHS = cast<SetCondInst>(Op0);
3590 if (cast<ConstantBool>(Cmp)->getValue()) {
3591 std::swap(LHS, RHS);
3592 std::swap(LHSCst, RHSCst);
3593 std::swap(LHSCC, RHSCC);
3596 // At this point, we know we have have two setcc instructions
3597 // comparing a value against two constants and or'ing the result
3598 // together. Because of the above check, we know that we only have
3599 // SetEQ, SetNE, SetLT, and SetGT here. We also know (from the
3600 // FoldSetCCLogical check above), that the two constants are not
3602 assert(LHSCst != RHSCst && "Compares not folded above?");
3605 default: assert(0 && "Unknown integer condition code!");
3606 case Instruction::SetEQ:
3608 default: assert(0 && "Unknown integer condition code!");
3609 case Instruction::SetEQ:
3610 if (LHSCst == SubOne(RHSCst)) {// (X == 13 | X == 14) -> X-13 <u 2
3611 Constant *AddCST = ConstantExpr::getNeg(LHSCst);
3612 Instruction *Add = BinaryOperator::createAdd(LHSVal, AddCST,
3613 LHSVal->getName()+".off");
3614 InsertNewInstBefore(Add, I);
3615 const Type *UnsType = Add->getType()->getUnsignedVersion();
3616 Value *OffsetVal = InsertCastBefore(Instruction::BitCast, Add,
3618 AddCST = ConstantExpr::getSub(AddOne(RHSCst), LHSCst);
3619 AddCST = ConstantExpr::getBitCast(AddCST, UnsType);
3620 return new SetCondInst(Instruction::SetLT, OffsetVal, AddCST);
3622 break; // (X == 13 | X == 15) -> no change
3624 case Instruction::SetGT: // (X == 13 | X > 14) -> no change
3626 case Instruction::SetNE: // (X == 13 | X != 15) -> X != 15
3627 case Instruction::SetLT: // (X == 13 | X < 15) -> X < 15
3628 return ReplaceInstUsesWith(I, RHS);
3631 case Instruction::SetNE:
3633 default: assert(0 && "Unknown integer condition code!");
3634 case Instruction::SetEQ: // (X != 13 | X == 15) -> X != 13
3635 case Instruction::SetGT: // (X != 13 | X > 15) -> X != 13
3636 return ReplaceInstUsesWith(I, LHS);
3637 case Instruction::SetNE: // (X != 13 | X != 15) -> true
3638 case Instruction::SetLT: // (X != 13 | X < 15) -> true
3639 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
3642 case Instruction::SetLT:
3644 default: assert(0 && "Unknown integer condition code!");
3645 case Instruction::SetEQ: // (X < 13 | X == 14) -> no change
3647 case Instruction::SetGT: // (X < 13 | X > 15) -> (X-13) > 2
3648 return InsertRangeTest(LHSVal, LHSCst, AddOne(RHSCst), false, I);
3649 case Instruction::SetNE: // (X < 13 | X != 15) -> X != 15
3650 case Instruction::SetLT: // (X < 13 | X < 15) -> X < 15
3651 return ReplaceInstUsesWith(I, RHS);
3654 case Instruction::SetGT:
3656 default: assert(0 && "Unknown integer condition code!");
3657 case Instruction::SetEQ: // (X > 13 | X == 15) -> X > 13
3658 case Instruction::SetGT: // (X > 13 | X > 15) -> X > 13
3659 return ReplaceInstUsesWith(I, LHS);
3660 case Instruction::SetNE: // (X > 13 | X != 15) -> true
3661 case Instruction::SetLT: // (X > 13 | X < 15) -> true
3662 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
3668 // fold (or (cast A), (cast B)) -> (cast (or A, B))
3669 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3670 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3671 if (Op0C->getOpcode() == Op1C->getOpcode()) {// same cast kind ?
3672 const Type *SrcTy = Op0C->getOperand(0)->getType();
3673 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3674 // Only do this if the casts both really cause code to be generated.
3675 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3676 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3677 Instruction *NewOp = BinaryOperator::createOr(Op0C->getOperand(0),
3678 Op1C->getOperand(0),
3680 InsertNewInstBefore(NewOp, I);
3681 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3686 return Changed ? &I : 0;
3689 // XorSelf - Implements: X ^ X --> 0
3692 XorSelf(Value *rhs) : RHS(rhs) {}
3693 bool shouldApply(Value *LHS) const { return LHS == RHS; }
3694 Instruction *apply(BinaryOperator &Xor) const {
3700 Instruction *InstCombiner::visitXor(BinaryOperator &I) {
3701 bool Changed = SimplifyCommutative(I);
3702 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
3704 if (isa<UndefValue>(Op1))
3705 return ReplaceInstUsesWith(I, Op1); // X ^ undef -> undef
3707 // xor X, X = 0, even if X is nested in a sequence of Xor's.
3708 if (Instruction *Result = AssociativeOpt(I, XorSelf(Op1))) {
3709 assert(Result == &I && "AssociativeOpt didn't work?");
3710 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
3713 // See if we can simplify any instructions used by the instruction whose sole
3714 // purpose is to compute bits we don't care about.
3715 uint64_t KnownZero, KnownOne;
3716 if (!isa<PackedType>(I.getType()) &&
3717 SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
3718 KnownZero, KnownOne))
3721 if (ConstantIntegral *RHS = dyn_cast<ConstantIntegral>(Op1)) {
3722 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0)) {
3723 // xor (setcc A, B), true = not (setcc A, B) = setncc A, B
3724 if (SetCondInst *SCI = dyn_cast<SetCondInst>(Op0I))
3725 if (RHS == ConstantBool::getTrue() && SCI->hasOneUse())
3726 return new SetCondInst(SCI->getInverseCondition(),
3727 SCI->getOperand(0), SCI->getOperand(1));
3729 // ~(c-X) == X-c-1 == X+(-c-1)
3730 if (Op0I->getOpcode() == Instruction::Sub && RHS->isAllOnesValue())
3731 if (Constant *Op0I0C = dyn_cast<Constant>(Op0I->getOperand(0))) {
3732 Constant *NegOp0I0C = ConstantExpr::getNeg(Op0I0C);
3733 Constant *ConstantRHS = ConstantExpr::getSub(NegOp0I0C,
3734 ConstantInt::get(I.getType(), 1));
3735 return BinaryOperator::createAdd(Op0I->getOperand(1), ConstantRHS);
3738 // ~(~X & Y) --> (X | ~Y)
3739 if (Op0I->getOpcode() == Instruction::And && RHS->isAllOnesValue()) {
3740 if (dyn_castNotVal(Op0I->getOperand(1))) Op0I->swapOperands();
3741 if (Value *Op0NotVal = dyn_castNotVal(Op0I->getOperand(0))) {
3743 BinaryOperator::createNot(Op0I->getOperand(1),
3744 Op0I->getOperand(1)->getName()+".not");
3745 InsertNewInstBefore(NotY, I);
3746 return BinaryOperator::createOr(Op0NotVal, NotY);
3750 if (ConstantInt *Op0CI = dyn_cast<ConstantInt>(Op0I->getOperand(1)))
3751 if (Op0I->getOpcode() == Instruction::Add) {
3752 // ~(X-c) --> (-c-1)-X
3753 if (RHS->isAllOnesValue()) {
3754 Constant *NegOp0CI = ConstantExpr::getNeg(Op0CI);
3755 return BinaryOperator::createSub(
3756 ConstantExpr::getSub(NegOp0CI,
3757 ConstantInt::get(I.getType(), 1)),
3758 Op0I->getOperand(0));
3760 } else if (Op0I->getOpcode() == Instruction::Or) {
3761 // (X|C1)^C2 -> X^(C1|C2) iff X&~C1 == 0
3762 if (MaskedValueIsZero(Op0I->getOperand(0), Op0CI->getZExtValue())) {
3763 Constant *NewRHS = ConstantExpr::getOr(Op0CI, RHS);
3764 // Anything in both C1 and C2 is known to be zero, remove it from
3766 Constant *CommonBits = ConstantExpr::getAnd(Op0CI, RHS);
3767 NewRHS = ConstantExpr::getAnd(NewRHS,
3768 ConstantExpr::getNot(CommonBits));
3769 WorkList.push_back(Op0I);
3770 I.setOperand(0, Op0I->getOperand(0));
3771 I.setOperand(1, NewRHS);
3777 // Try to fold constant and into select arguments.
3778 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
3779 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
3781 if (isa<PHINode>(Op0))
3782 if (Instruction *NV = FoldOpIntoPhi(I))
3786 if (Value *X = dyn_castNotVal(Op0)) // ~A ^ A == -1
3788 return ReplaceInstUsesWith(I,
3789 ConstantIntegral::getAllOnesValue(I.getType()));
3791 if (Value *X = dyn_castNotVal(Op1)) // A ^ ~A == -1
3793 return ReplaceInstUsesWith(I,
3794 ConstantIntegral::getAllOnesValue(I.getType()));
3796 if (BinaryOperator *Op1I = dyn_cast<BinaryOperator>(Op1))
3797 if (Op1I->getOpcode() == Instruction::Or) {
3798 if (Op1I->getOperand(0) == Op0) { // B^(B|A) == (A|B)^B
3799 Op1I->swapOperands();
3801 std::swap(Op0, Op1);
3802 } else if (Op1I->getOperand(1) == Op0) { // B^(A|B) == (A|B)^B
3803 I.swapOperands(); // Simplified below.
3804 std::swap(Op0, Op1);
3806 } else if (Op1I->getOpcode() == Instruction::Xor) {
3807 if (Op0 == Op1I->getOperand(0)) // A^(A^B) == B
3808 return ReplaceInstUsesWith(I, Op1I->getOperand(1));
3809 else if (Op0 == Op1I->getOperand(1)) // A^(B^A) == B
3810 return ReplaceInstUsesWith(I, Op1I->getOperand(0));
3811 } else if (Op1I->getOpcode() == Instruction::And && Op1I->hasOneUse()) {
3812 if (Op1I->getOperand(0) == Op0) // A^(A&B) -> A^(B&A)
3813 Op1I->swapOperands();
3814 if (Op0 == Op1I->getOperand(1)) { // A^(B&A) -> (B&A)^A
3815 I.swapOperands(); // Simplified below.
3816 std::swap(Op0, Op1);
3820 if (BinaryOperator *Op0I = dyn_cast<BinaryOperator>(Op0))
3821 if (Op0I->getOpcode() == Instruction::Or && Op0I->hasOneUse()) {
3822 if (Op0I->getOperand(0) == Op1) // (B|A)^B == (A|B)^B
3823 Op0I->swapOperands();
3824 if (Op0I->getOperand(1) == Op1) { // (A|B)^B == A & ~B
3825 Instruction *NotB = BinaryOperator::createNot(Op1, "tmp");
3826 InsertNewInstBefore(NotB, I);
3827 return BinaryOperator::createAnd(Op0I->getOperand(0), NotB);
3829 } else if (Op0I->getOpcode() == Instruction::Xor) {
3830 if (Op1 == Op0I->getOperand(0)) // (A^B)^A == B
3831 return ReplaceInstUsesWith(I, Op0I->getOperand(1));
3832 else if (Op1 == Op0I->getOperand(1)) // (B^A)^A == B
3833 return ReplaceInstUsesWith(I, Op0I->getOperand(0));
3834 } else if (Op0I->getOpcode() == Instruction::And && Op0I->hasOneUse()) {
3835 if (Op0I->getOperand(0) == Op1) // (A&B)^A -> (B&A)^A
3836 Op0I->swapOperands();
3837 if (Op0I->getOperand(1) == Op1 && // (B&A)^A == ~B & A
3838 !isa<ConstantInt>(Op1)) { // Canonical form is (B&C)^C
3839 Instruction *N = BinaryOperator::createNot(Op0I->getOperand(0), "tmp");
3840 InsertNewInstBefore(N, I);
3841 return BinaryOperator::createAnd(N, Op1);
3845 // (setcc1 A, B) ^ (setcc2 A, B) --> (setcc3 A, B)
3846 if (SetCondInst *RHS = dyn_cast<SetCondInst>(I.getOperand(1)))
3847 if (Instruction *R = AssociativeOpt(I, FoldSetCCLogical(*this, RHS)))
3850 // fold (xor (cast A), (cast B)) -> (cast (xor A, B))
3851 if (CastInst *Op0C = dyn_cast<CastInst>(Op0))
3852 if (CastInst *Op1C = dyn_cast<CastInst>(Op1))
3853 if (Op0C->getOpcode() == Op1C->getOpcode()) { // same cast kind?
3854 const Type *SrcTy = Op0C->getOperand(0)->getType();
3855 if (SrcTy == Op1C->getOperand(0)->getType() && SrcTy->isIntegral() &&
3856 // Only do this if the casts both really cause code to be generated.
3857 ValueRequiresCast(Op0C->getOperand(0), I.getType(), TD) &&
3858 ValueRequiresCast(Op1C->getOperand(0), I.getType(), TD)) {
3859 Instruction *NewOp = BinaryOperator::createXor(Op0C->getOperand(0),
3860 Op1C->getOperand(0),
3862 InsertNewInstBefore(NewOp, I);
3863 return CastInst::create(Op0C->getOpcode(), NewOp, I.getType());
3867 // (X >> Z) ^ (Y >> Z) -> (X^Y) >> Z for all shifts.
3868 if (ShiftInst *SI1 = dyn_cast<ShiftInst>(Op1)) {
3869 if (ShiftInst *SI0 = dyn_cast<ShiftInst>(Op0))
3870 if (SI0->getOpcode() == SI1->getOpcode() &&
3871 SI0->getOperand(1) == SI1->getOperand(1) &&
3872 (SI0->hasOneUse() || SI1->hasOneUse())) {
3873 Instruction *NewOp =
3874 InsertNewInstBefore(BinaryOperator::createXor(SI0->getOperand(0),
3876 SI0->getName()), I);
3877 return new ShiftInst(SI1->getOpcode(), NewOp, SI1->getOperand(1));
3881 return Changed ? &I : 0;
3884 static bool isPositive(ConstantInt *C) {
3885 return C->getSExtValue() >= 0;
3888 /// AddWithOverflow - Compute Result = In1+In2, returning true if the result
3889 /// overflowed for this type.
3890 static bool AddWithOverflow(ConstantInt *&Result, ConstantInt *In1,
3892 Result = cast<ConstantInt>(ConstantExpr::getAdd(In1, In2));
3894 if (In1->getType()->isUnsigned())
3895 return cast<ConstantInt>(Result)->getZExtValue() <
3896 cast<ConstantInt>(In1)->getZExtValue();
3897 if (isPositive(In1) != isPositive(In2))
3899 if (isPositive(In1))
3900 return cast<ConstantInt>(Result)->getSExtValue() <
3901 cast<ConstantInt>(In1)->getSExtValue();
3902 return cast<ConstantInt>(Result)->getSExtValue() >
3903 cast<ConstantInt>(In1)->getSExtValue();
3906 /// EmitGEPOffset - Given a getelementptr instruction/constantexpr, emit the
3907 /// code necessary to compute the offset from the base pointer (without adding
3908 /// in the base pointer). Return the result as a signed integer of intptr size.
3909 static Value *EmitGEPOffset(User *GEP, Instruction &I, InstCombiner &IC) {
3910 TargetData &TD = IC.getTargetData();
3911 gep_type_iterator GTI = gep_type_begin(GEP);
3912 const Type *UIntPtrTy = TD.getIntPtrType();
3913 const Type *SIntPtrTy = UIntPtrTy->getSignedVersion();
3914 Value *Result = Constant::getNullValue(SIntPtrTy);
3916 // Build a mask for high order bits.
3917 uint64_t PtrSizeMask = ~0ULL >> (64-TD.getPointerSize()*8);
3919 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i, ++GTI) {
3920 Value *Op = GEP->getOperand(i);
3921 uint64_t Size = TD.getTypeSize(GTI.getIndexedType()) & PtrSizeMask;
3922 Constant *Scale = ConstantInt::get(SIntPtrTy, Size);
3923 if (Constant *OpC = dyn_cast<Constant>(Op)) {
3924 if (!OpC->isNullValue()) {
3925 OpC = ConstantExpr::getIntegerCast(OpC, SIntPtrTy, true /*SExt*/);
3926 Scale = ConstantExpr::getMul(OpC, Scale);
3927 if (Constant *RC = dyn_cast<Constant>(Result))
3928 Result = ConstantExpr::getAdd(RC, Scale);
3930 // Emit an add instruction.
3931 Result = IC.InsertNewInstBefore(
3932 BinaryOperator::createAdd(Result, Scale,
3933 GEP->getName()+".offs"), I);
3937 // Convert to correct type.
3938 Op = IC.InsertNewInstBefore(CastInst::createSExtOrBitCast(Op, SIntPtrTy,
3939 Op->getName()+".c"), I);
3941 // We'll let instcombine(mul) convert this to a shl if possible.
3942 Op = IC.InsertNewInstBefore(BinaryOperator::createMul(Op, Scale,
3943 GEP->getName()+".idx"), I);
3945 // Emit an add instruction.
3946 Result = IC.InsertNewInstBefore(BinaryOperator::createAdd(Op, Result,
3947 GEP->getName()+".offs"), I);
3953 /// FoldGEPSetCC - Fold comparisons between a GEP instruction and something
3954 /// else. At this point we know that the GEP is on the LHS of the comparison.
3955 Instruction *InstCombiner::FoldGEPSetCC(User *GEPLHS, Value *RHS,
3956 Instruction::BinaryOps Cond,
3958 assert(dyn_castGetElementPtr(GEPLHS) && "LHS is not a getelementptr!");
3960 if (CastInst *CI = dyn_cast<CastInst>(RHS))
3961 if (isa<PointerType>(CI->getOperand(0)->getType()))
3962 RHS = CI->getOperand(0);
3964 Value *PtrBase = GEPLHS->getOperand(0);
3965 if (PtrBase == RHS) {
3966 // As an optimization, we don't actually have to compute the actual value of
3967 // OFFSET if this is a seteq or setne comparison, just return whether each
3968 // index is zero or not.
3969 if (Cond == Instruction::SetEQ || Cond == Instruction::SetNE) {
3970 Instruction *InVal = 0;
3971 gep_type_iterator GTI = gep_type_begin(GEPLHS);
3972 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i, ++GTI) {
3974 if (Constant *C = dyn_cast<Constant>(GEPLHS->getOperand(i))) {
3975 if (isa<UndefValue>(C)) // undef index -> undef.
3976 return ReplaceInstUsesWith(I, UndefValue::get(I.getType()));
3977 if (C->isNullValue())
3979 else if (TD->getTypeSize(GTI.getIndexedType()) == 0) {
3980 EmitIt = false; // This is indexing into a zero sized array?
3981 } else if (isa<ConstantInt>(C))
3982 return ReplaceInstUsesWith(I, // No comparison is needed here.
3983 ConstantBool::get(Cond == Instruction::SetNE));
3988 new SetCondInst(Cond, GEPLHS->getOperand(i),
3989 Constant::getNullValue(GEPLHS->getOperand(i)->getType()));
3993 InVal = InsertNewInstBefore(InVal, I);
3994 InsertNewInstBefore(Comp, I);
3995 if (Cond == Instruction::SetNE) // True if any are unequal
3996 InVal = BinaryOperator::createOr(InVal, Comp);
3997 else // True if all are equal
3998 InVal = BinaryOperator::createAnd(InVal, Comp);
4006 ReplaceInstUsesWith(I, // No comparison is needed here, all indexes = 0
4007 ConstantBool::get(Cond == Instruction::SetEQ));
4010 // Only lower this if the setcc is the only user of the GEP or if we expect
4011 // the result to fold to a constant!
4012 if (isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) {
4013 // ((gep Ptr, OFFSET) cmp Ptr) ---> (OFFSET cmp 0).
4014 Value *Offset = EmitGEPOffset(GEPLHS, I, *this);
4015 return new SetCondInst(Cond, Offset,
4016 Constant::getNullValue(Offset->getType()));
4018 } else if (User *GEPRHS = dyn_castGetElementPtr(RHS)) {
4019 // If the base pointers are different, but the indices are the same, just
4020 // compare the base pointer.
4021 if (PtrBase != GEPRHS->getOperand(0)) {
4022 bool IndicesTheSame = GEPLHS->getNumOperands()==GEPRHS->getNumOperands();
4023 IndicesTheSame &= GEPLHS->getOperand(0)->getType() ==
4024 GEPRHS->getOperand(0)->getType();
4026 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4027 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4028 IndicesTheSame = false;
4032 // If all indices are the same, just compare the base pointers.
4034 return new SetCondInst(Cond, GEPLHS->getOperand(0),
4035 GEPRHS->getOperand(0));
4037 // Otherwise, the base pointers are different and the indices are
4038 // different, bail out.
4042 // If one of the GEPs has all zero indices, recurse.
4043 bool AllZeros = true;
4044 for (unsigned i = 1, e = GEPLHS->getNumOperands(); i != e; ++i)
4045 if (!isa<Constant>(GEPLHS->getOperand(i)) ||
4046 !cast<Constant>(GEPLHS->getOperand(i))->isNullValue()) {
4051 return FoldGEPSetCC(GEPRHS, GEPLHS->getOperand(0),
4052 SetCondInst::getSwappedCondition(Cond), I);
4054 // If the other GEP has all zero indices, recurse.
4056 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4057 if (!isa<Constant>(GEPRHS->getOperand(i)) ||
4058 !cast<Constant>(GEPRHS->getOperand(i))->isNullValue()) {
4063 return FoldGEPSetCC(GEPLHS, GEPRHS->getOperand(0), Cond, I);
4065 if (GEPLHS->getNumOperands() == GEPRHS->getNumOperands()) {
4066 // If the GEPs only differ by one index, compare it.
4067 unsigned NumDifferences = 0; // Keep track of # differences.
4068 unsigned DiffOperand = 0; // The operand that differs.
4069 for (unsigned i = 1, e = GEPRHS->getNumOperands(); i != e; ++i)
4070 if (GEPLHS->getOperand(i) != GEPRHS->getOperand(i)) {
4071 if (GEPLHS->getOperand(i)->getType()->getPrimitiveSizeInBits() !=
4072 GEPRHS->getOperand(i)->getType()->getPrimitiveSizeInBits()) {
4073 // Irreconcilable differences.
4077 if (NumDifferences++) break;
4082 if (NumDifferences == 0) // SAME GEP?
4083 return ReplaceInstUsesWith(I, // No comparison is needed here.
4084 ConstantBool::get(Cond == Instruction::SetEQ));
4085 else if (NumDifferences == 1) {
4086 Value *LHSV = GEPLHS->getOperand(DiffOperand);
4087 Value *RHSV = GEPRHS->getOperand(DiffOperand);
4089 // Convert the operands to signed values to make sure to perform a
4090 // signed comparison.
4091 const Type *NewTy = LHSV->getType()->getSignedVersion();
4092 if (LHSV->getType() != NewTy)
4093 LHSV = InsertCastBefore(Instruction::BitCast, LHSV, NewTy, I);
4094 if (RHSV->getType() != NewTy)
4095 RHSV = InsertCastBefore(Instruction::BitCast, RHSV, NewTy, I);
4096 return new SetCondInst(Cond, LHSV, RHSV);
4100 // Only lower this if the setcc is the only user of the GEP or if we expect
4101 // the result to fold to a constant!
4102 if ((isa<ConstantExpr>(GEPLHS) || GEPLHS->hasOneUse()) &&
4103 (isa<ConstantExpr>(GEPRHS) || GEPRHS->hasOneUse())) {
4104 // ((gep Ptr, OFFSET1) cmp (gep Ptr, OFFSET2) ---> (OFFSET1 cmp OFFSET2)
4105 Value *L = EmitGEPOffset(GEPLHS, I, *this);
4106 Value *R = EmitGEPOffset(GEPRHS, I, *this);
4107 return new SetCondInst(Cond, L, R);
4114 Instruction *InstCombiner::visitSetCondInst(SetCondInst &I) {
4115 bool Changed = SimplifyCommutative(I);
4116 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
4117 const Type *Ty = Op0->getType();
4121 return ReplaceInstUsesWith(I, ConstantBool::get(isTrueWhenEqual(I)));
4123 if (isa<UndefValue>(Op1)) // X setcc undef -> undef
4124 return ReplaceInstUsesWith(I, UndefValue::get(Type::BoolTy));
4126 // setcc <global/alloca*/null>, <global/alloca*/null> - Global/Stack value
4127 // addresses never equal each other! We already know that Op0 != Op1.
4128 if ((isa<GlobalValue>(Op0) || isa<AllocaInst>(Op0) ||
4129 isa<ConstantPointerNull>(Op0)) &&
4130 (isa<GlobalValue>(Op1) || isa<AllocaInst>(Op1) ||
4131 isa<ConstantPointerNull>(Op1)))
4132 return ReplaceInstUsesWith(I, ConstantBool::get(!isTrueWhenEqual(I)));
4134 // setcc's with boolean values can always be turned into bitwise operations
4135 if (Ty == Type::BoolTy) {
4136 switch (I.getOpcode()) {
4137 default: assert(0 && "Invalid setcc instruction!");
4138 case Instruction::SetEQ: { // seteq bool %A, %B -> ~(A^B)
4139 Instruction *Xor = BinaryOperator::createXor(Op0, Op1, I.getName()+"tmp");
4140 InsertNewInstBefore(Xor, I);
4141 return BinaryOperator::createNot(Xor);
4143 case Instruction::SetNE:
4144 return BinaryOperator::createXor(Op0, Op1);
4146 case Instruction::SetGT:
4147 std::swap(Op0, Op1); // Change setgt -> setlt
4149 case Instruction::SetLT: { // setlt bool A, B -> ~X & Y
4150 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4151 InsertNewInstBefore(Not, I);
4152 return BinaryOperator::createAnd(Not, Op1);
4154 case Instruction::SetGE:
4155 std::swap(Op0, Op1); // Change setge -> setle
4157 case Instruction::SetLE: { // setle bool %A, %B -> ~A | B
4158 Instruction *Not = BinaryOperator::createNot(Op0, I.getName()+"tmp");
4159 InsertNewInstBefore(Not, I);
4160 return BinaryOperator::createOr(Not, Op1);
4165 // See if we are doing a comparison between a constant and an instruction that
4166 // can be folded into the comparison.
4167 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
4168 // Check to see if we are comparing against the minimum or maximum value...
4169 if (CI->isMinValue(CI->getType()->isSigned())) {
4170 if (I.getOpcode() == Instruction::SetLT) // A < MIN -> FALSE
4171 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4172 if (I.getOpcode() == Instruction::SetGE) // A >= MIN -> TRUE
4173 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4174 if (I.getOpcode() == Instruction::SetLE) // A <= MIN -> A == MIN
4175 return BinaryOperator::createSetEQ(Op0, Op1);
4176 if (I.getOpcode() == Instruction::SetGT) // A > MIN -> A != MIN
4177 return BinaryOperator::createSetNE(Op0, Op1);
4179 } else if (CI->isMaxValue(CI->getType()->isSigned())) {
4180 if (I.getOpcode() == Instruction::SetGT) // A > MAX -> FALSE
4181 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4182 if (I.getOpcode() == Instruction::SetLE) // A <= MAX -> TRUE
4183 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4184 if (I.getOpcode() == Instruction::SetGE) // A >= MAX -> A == MAX
4185 return BinaryOperator::createSetEQ(Op0, Op1);
4186 if (I.getOpcode() == Instruction::SetLT) // A < MAX -> A != MAX
4187 return BinaryOperator::createSetNE(Op0, Op1);
4189 // Comparing against a value really close to min or max?
4190 } else if (isMinValuePlusOne(CI)) {
4191 if (I.getOpcode() == Instruction::SetLT) // A < MIN+1 -> A == MIN
4192 return BinaryOperator::createSetEQ(Op0, SubOne(CI));
4193 if (I.getOpcode() == Instruction::SetGE) // A >= MIN-1 -> A != MIN
4194 return BinaryOperator::createSetNE(Op0, SubOne(CI));
4196 } else if (isMaxValueMinusOne(CI)) {
4197 if (I.getOpcode() == Instruction::SetGT) // A > MAX-1 -> A == MAX
4198 return BinaryOperator::createSetEQ(Op0, AddOne(CI));
4199 if (I.getOpcode() == Instruction::SetLE) // A <= MAX-1 -> A != MAX
4200 return BinaryOperator::createSetNE(Op0, AddOne(CI));
4203 // If we still have a setle or setge instruction, turn it into the
4204 // appropriate setlt or setgt instruction. Since the border cases have
4205 // already been handled above, this requires little checking.
4207 if (I.getOpcode() == Instruction::SetLE)
4208 return BinaryOperator::createSetLT(Op0, AddOne(CI));
4209 if (I.getOpcode() == Instruction::SetGE)
4210 return BinaryOperator::createSetGT(Op0, SubOne(CI));
4213 // See if we can fold the comparison based on bits known to be zero or one
4215 uint64_t KnownZero, KnownOne;
4216 if (SimplifyDemandedBits(Op0, Ty->getIntegralTypeMask(),
4217 KnownZero, KnownOne, 0))
4220 // Given the known and unknown bits, compute a range that the LHS could be
4222 if (KnownOne | KnownZero) {
4223 if (Ty->isUnsigned()) { // Unsigned comparison.
4225 uint64_t RHSVal = CI->getZExtValue();
4226 ComputeUnsignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
4228 switch (I.getOpcode()) { // LE/GE have been folded already.
4229 default: assert(0 && "Unknown setcc opcode!");
4230 case Instruction::SetEQ:
4231 if (Max < RHSVal || Min > RHSVal)
4232 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4234 case Instruction::SetNE:
4235 if (Max < RHSVal || Min > RHSVal)
4236 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4238 case Instruction::SetLT:
4240 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4242 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4244 case Instruction::SetGT:
4246 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4248 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4251 } else { // Signed comparison.
4253 int64_t RHSVal = CI->getSExtValue();
4254 ComputeSignedMinMaxValuesFromKnownBits(Ty, KnownZero, KnownOne,
4256 switch (I.getOpcode()) { // LE/GE have been folded already.
4257 default: assert(0 && "Unknown setcc opcode!");
4258 case Instruction::SetEQ:
4259 if (Max < RHSVal || Min > RHSVal)
4260 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4262 case Instruction::SetNE:
4263 if (Max < RHSVal || Min > RHSVal)
4264 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4266 case Instruction::SetLT:
4268 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4270 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4272 case Instruction::SetGT:
4274 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4276 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4282 // Since the RHS is a constantInt (CI), if the left hand side is an
4283 // instruction, see if that instruction also has constants so that the
4284 // instruction can be folded into the setcc
4285 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4286 switch (LHSI->getOpcode()) {
4287 case Instruction::And:
4288 if (LHSI->hasOneUse() && isa<ConstantInt>(LHSI->getOperand(1)) &&
4289 LHSI->getOperand(0)->hasOneUse()) {
4290 ConstantInt *AndCST = cast<ConstantInt>(LHSI->getOperand(1));
4292 // If an operand is an AND of a truncating cast, we can widen the
4293 // and/compare to be the input width without changing the value
4294 // produced, eliminating a cast.
4295 if (CastInst *Cast = dyn_cast<CastInst>(LHSI->getOperand(0))) {
4296 // We can do this transformation if either the AND constant does not
4297 // have its sign bit set or if it is an equality comparison.
4298 // Extending a relational comparison when we're checking the sign
4299 // bit would not work.
4300 if (Cast->hasOneUse() && isa<TruncInst>(Cast) &&
4302 (AndCST->getZExtValue() == (uint64_t)AndCST->getSExtValue()) &&
4303 (CI->getZExtValue() == (uint64_t)CI->getSExtValue()))) {
4304 ConstantInt *NewCST;
4306 if (Cast->getOperand(0)->getType()->isSigned()) {
4307 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4308 AndCST->getZExtValue());
4309 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4310 CI->getZExtValue());
4312 NewCST = ConstantInt::get(Cast->getOperand(0)->getType(),
4313 AndCST->getZExtValue());
4314 NewCI = ConstantInt::get(Cast->getOperand(0)->getType(),
4315 CI->getZExtValue());
4317 Instruction *NewAnd =
4318 BinaryOperator::createAnd(Cast->getOperand(0), NewCST,
4320 InsertNewInstBefore(NewAnd, I);
4321 return new SetCondInst(I.getOpcode(), NewAnd, NewCI);
4325 // If this is: (X >> C1) & C2 != C3 (where any shift and any compare
4326 // could exist), turn it into (X & (C2 << C1)) != (C3 << C1). This
4327 // happens a LOT in code produced by the C front-end, for bitfield
4329 ShiftInst *Shift = dyn_cast<ShiftInst>(LHSI->getOperand(0));
4331 // Check to see if there is a noop-cast between the shift and the and.
4333 if (CastInst *CI = dyn_cast<CastInst>(LHSI->getOperand(0)))
4334 if (CI->getOpcode() == Instruction::BitCast)
4335 Shift = dyn_cast<ShiftInst>(CI->getOperand(0));
4339 ShAmt = Shift ? dyn_cast<ConstantInt>(Shift->getOperand(1)) : 0;
4340 const Type *Ty = Shift ? Shift->getType() : 0; // Type of the shift.
4341 const Type *AndTy = AndCST->getType(); // Type of the and.
4343 // We can fold this as long as we can't shift unknown bits
4344 // into the mask. This can only happen with signed shift
4345 // rights, as they sign-extend.
4347 bool CanFold = Shift->isLogicalShift();
4349 // To test for the bad case of the signed shr, see if any
4350 // of the bits shifted in could be tested after the mask.
4351 int ShAmtVal = Ty->getPrimitiveSizeInBits()-ShAmt->getZExtValue();
4352 if (ShAmtVal < 0) ShAmtVal = 0; // Out of range shift.
4354 Constant *OShAmt = ConstantInt::get(Type::UByteTy, ShAmtVal);
4356 ConstantExpr::getShl(ConstantInt::getAllOnesValue(AndTy),
4358 if (ConstantExpr::getAnd(ShVal, AndCST)->isNullValue())
4364 if (Shift->getOpcode() == Instruction::Shl)
4365 NewCst = ConstantExpr::getLShr(CI, ShAmt);
4367 NewCst = ConstantExpr::getShl(CI, ShAmt);
4369 // Check to see if we are shifting out any of the bits being
4371 if (ConstantExpr::get(Shift->getOpcode(), NewCst, ShAmt) != CI){
4372 // If we shifted bits out, the fold is not going to work out.
4373 // As a special case, check to see if this means that the
4374 // result is always true or false now.
4375 if (I.getOpcode() == Instruction::SetEQ)
4376 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4377 if (I.getOpcode() == Instruction::SetNE)
4378 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4380 I.setOperand(1, NewCst);
4381 Constant *NewAndCST;
4382 if (Shift->getOpcode() == Instruction::Shl)
4383 NewAndCST = ConstantExpr::getLShr(AndCST, ShAmt);
4385 NewAndCST = ConstantExpr::getShl(AndCST, ShAmt);
4386 LHSI->setOperand(1, NewAndCST);
4388 LHSI->setOperand(0, Shift->getOperand(0));
4390 Value *NewCast = InsertCastBefore(Instruction::BitCast,
4391 Shift->getOperand(0), AndTy,
4393 LHSI->setOperand(0, NewCast);
4395 WorkList.push_back(Shift); // Shift is dead.
4396 AddUsesToWorkList(I);
4402 // Turn ((X >> Y) & C) == 0 into (X & (C << Y)) == 0. The later is
4403 // preferable because it allows the C<<Y expression to be hoisted out
4404 // of a loop if Y is invariant and X is not.
4405 if (Shift && Shift->hasOneUse() && CI->isNullValue() &&
4406 I.isEquality() && !Shift->isArithmeticShift() &&
4407 isa<Instruction>(Shift->getOperand(0))) {
4410 if (Shift->getOpcode() == Instruction::LShr) {
4411 NS = new ShiftInst(Instruction::Shl, AndCST, Shift->getOperand(1),
4414 // Insert a logical shift.
4415 NS = new ShiftInst(Instruction::LShr, AndCST,
4416 Shift->getOperand(1), "tmp");
4418 InsertNewInstBefore(cast<Instruction>(NS), I);
4420 // If C's sign doesn't agree with the and, insert a cast now.
4421 if (NS->getType() != LHSI->getType())
4422 NS = InsertCastBefore(Instruction::BitCast, NS, LHSI->getType(),
4425 Value *ShiftOp = Shift->getOperand(0);
4426 if (ShiftOp->getType() != LHSI->getType())
4427 ShiftOp = InsertCastBefore(Instruction::BitCast, ShiftOp,
4428 LHSI->getType(), I);
4430 // Compute X & (C << Y).
4431 Instruction *NewAnd =
4432 BinaryOperator::createAnd(ShiftOp, NS, LHSI->getName());
4433 InsertNewInstBefore(NewAnd, I);
4435 I.setOperand(0, NewAnd);
4441 case Instruction::Shl: // (setcc (shl X, ShAmt), CI)
4442 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4443 if (I.isEquality()) {
4444 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4446 // Check that the shift amount is in range. If not, don't perform
4447 // undefined shifts. When the shift is visited it will be
4449 if (ShAmt->getZExtValue() >= TypeBits)
4452 // If we are comparing against bits always shifted out, the
4453 // comparison cannot succeed.
4455 ConstantExpr::getShl(ConstantExpr::getLShr(CI, ShAmt), ShAmt);
4456 if (Comp != CI) {// Comparing against a bit that we know is zero.
4457 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4458 Constant *Cst = ConstantBool::get(IsSetNE);
4459 return ReplaceInstUsesWith(I, Cst);
4462 if (LHSI->hasOneUse()) {
4463 // Otherwise strength reduce the shift into an and.
4464 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4465 uint64_t Val = (1ULL << (TypeBits-ShAmtVal))-1;
4468 if (CI->getType()->isUnsigned()) {
4469 Mask = ConstantInt::get(CI->getType(), Val);
4470 } else if (ShAmtVal != 0) {
4471 Mask = ConstantInt::get(CI->getType(), Val);
4473 Mask = ConstantInt::getAllOnesValue(CI->getType());
4477 BinaryOperator::createAnd(LHSI->getOperand(0),
4478 Mask, LHSI->getName()+".mask");
4479 Value *And = InsertNewInstBefore(AndI, I);
4480 return new SetCondInst(I.getOpcode(), And,
4481 ConstantExpr::getLShr(CI, ShAmt));
4487 case Instruction::LShr: // (setcc (shr X, ShAmt), CI)
4488 case Instruction::AShr:
4489 if (ConstantInt *ShAmt = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4490 if (I.isEquality()) {
4491 // Check that the shift amount is in range. If not, don't perform
4492 // undefined shifts. When the shift is visited it will be
4494 unsigned TypeBits = CI->getType()->getPrimitiveSizeInBits();
4495 if (ShAmt->getZExtValue() >= TypeBits)
4498 // If we are comparing against bits always shifted out, the
4499 // comparison cannot succeed.
4501 if (CI->getType()->isUnsigned())
4502 Comp = ConstantExpr::getLShr(ConstantExpr::getShl(CI, ShAmt),
4505 Comp = ConstantExpr::getAShr(ConstantExpr::getShl(CI, ShAmt),
4508 if (Comp != CI) {// Comparing against a bit that we know is zero.
4509 bool IsSetNE = I.getOpcode() == Instruction::SetNE;
4510 Constant *Cst = ConstantBool::get(IsSetNE);
4511 return ReplaceInstUsesWith(I, Cst);
4514 if (LHSI->hasOneUse() || CI->isNullValue()) {
4515 unsigned ShAmtVal = (unsigned)ShAmt->getZExtValue();
4517 // Otherwise strength reduce the shift into an and.
4518 uint64_t Val = ~0ULL; // All ones.
4519 Val <<= ShAmtVal; // Shift over to the right spot.
4522 if (CI->getType()->isUnsigned()) {
4523 Val &= ~0ULL >> (64-TypeBits);
4524 Mask = ConstantInt::get(CI->getType(), Val);
4526 Mask = ConstantInt::get(CI->getType(), Val);
4530 BinaryOperator::createAnd(LHSI->getOperand(0),
4531 Mask, LHSI->getName()+".mask");
4532 Value *And = InsertNewInstBefore(AndI, I);
4533 return new SetCondInst(I.getOpcode(), And,
4534 ConstantExpr::getShl(CI, ShAmt));
4540 case Instruction::SDiv:
4541 case Instruction::UDiv:
4542 // Fold: setcc ([us]div X, C1), C2 -> range test
4543 // Fold this div into the comparison, producing a range check.
4544 // Determine, based on the divide type, what the range is being
4545 // checked. If there is an overflow on the low or high side, remember
4546 // it, otherwise compute the range [low, hi) bounding the new value.
4547 // See: InsertRangeTest above for the kinds of replacements possible.
4548 if (ConstantInt *DivRHS = dyn_cast<ConstantInt>(LHSI->getOperand(1))) {
4549 // FIXME: If the operand types don't match the type of the divide
4550 // then don't attempt this transform. The code below doesn't have the
4551 // logic to deal with a signed divide and an unsigned compare (and
4552 // vice versa). This is because (x /s C1) <s C2 produces different
4553 // results than (x /s C1) <u C2 or (x /u C1) <s C2 or even
4554 // (x /u C1) <u C2. Simply casting the operands and result won't
4555 // work. :( The if statement below tests that condition and bails
4557 const Type *DivRHSTy = DivRHS->getType();
4558 unsigned DivOpCode = LHSI->getOpcode();
4559 if (I.isEquality() &&
4560 ((DivOpCode == Instruction::SDiv && DivRHSTy->isUnsigned()) ||
4561 (DivOpCode == Instruction::UDiv && DivRHSTy->isSigned())))
4564 // Initialize the variables that will indicate the nature of the
4566 bool LoOverflow = false, HiOverflow = false;
4567 ConstantInt *LoBound = 0, *HiBound = 0;
4569 // Compute Prod = CI * DivRHS. We are essentially solving an equation
4570 // of form X/C1=C2. We solve for X by multiplying C1 (DivRHS) and
4571 // C2 (CI). By solving for X we can turn this into a range check
4572 // instead of computing a divide.
4574 cast<ConstantInt>(ConstantExpr::getMul(CI, DivRHS));
4576 // Determine if the product overflows by seeing if the product is
4577 // not equal to the divide. Make sure we do the same kind of divide
4578 // as in the LHS instruction that we're folding.
4579 bool ProdOV = !DivRHS->isNullValue() &&
4580 (DivOpCode == Instruction::SDiv ?
4581 ConstantExpr::getSDiv(Prod, DivRHS) :
4582 ConstantExpr::getUDiv(Prod, DivRHS)) != CI;
4584 // Get the SetCC opcode
4585 Instruction::BinaryOps Opcode = I.getOpcode();
4587 if (DivRHS->isNullValue()) {
4588 // Don't hack on divide by zeros!
4589 } else if (DivOpCode == Instruction::UDiv) { // udiv
4591 LoOverflow = ProdOV;
4592 HiOverflow = ProdOV || AddWithOverflow(HiBound, LoBound, DivRHS);
4593 } else if (isPositive(DivRHS)) { // Divisor is > 0.
4594 if (CI->isNullValue()) { // (X / pos) op 0
4596 LoBound = cast<ConstantInt>(ConstantExpr::getNeg(SubOne(DivRHS)));
4598 } else if (isPositive(CI)) { // (X / pos) op pos
4600 LoOverflow = ProdOV;
4601 HiOverflow = ProdOV || AddWithOverflow(HiBound, Prod, DivRHS);
4602 } else { // (X / pos) op neg
4603 Constant *DivRHSH = ConstantExpr::getNeg(SubOne(DivRHS));
4604 LoOverflow = AddWithOverflow(LoBound, Prod,
4605 cast<ConstantInt>(DivRHSH));
4607 HiOverflow = ProdOV;
4609 } else { // Divisor is < 0.
4610 if (CI->isNullValue()) { // (X / neg) op 0
4611 LoBound = AddOne(DivRHS);
4612 HiBound = cast<ConstantInt>(ConstantExpr::getNeg(DivRHS));
4613 if (HiBound == DivRHS)
4614 LoBound = 0; // - INTMIN = INTMIN
4615 } else if (isPositive(CI)) { // (X / neg) op pos
4616 HiOverflow = LoOverflow = ProdOV;
4618 LoOverflow = AddWithOverflow(LoBound, Prod, AddOne(DivRHS));
4619 HiBound = AddOne(Prod);
4620 } else { // (X / neg) op neg
4622 LoOverflow = HiOverflow = ProdOV;
4623 HiBound = cast<ConstantInt>(ConstantExpr::getSub(Prod, DivRHS));
4626 // Dividing by a negate swaps the condition.
4627 Opcode = SetCondInst::getSwappedCondition(Opcode);
4631 Value *X = LHSI->getOperand(0);
4633 default: assert(0 && "Unhandled setcc opcode!");
4634 case Instruction::SetEQ:
4635 if (LoOverflow && HiOverflow)
4636 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4637 else if (HiOverflow)
4638 return new SetCondInst(Instruction::SetGE, X, LoBound);
4639 else if (LoOverflow)
4640 return new SetCondInst(Instruction::SetLT, X, HiBound);
4642 return InsertRangeTest(X, LoBound, HiBound, true, I);
4643 case Instruction::SetNE:
4644 if (LoOverflow && HiOverflow)
4645 return ReplaceInstUsesWith(I, ConstantBool::getTrue());
4646 else if (HiOverflow)
4647 return new SetCondInst(Instruction::SetLT, X, LoBound);
4648 else if (LoOverflow)
4649 return new SetCondInst(Instruction::SetGE, X, HiBound);
4651 return InsertRangeTest(X, LoBound, HiBound, false, I);
4652 case Instruction::SetLT:
4654 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4655 return new SetCondInst(Instruction::SetLT, X, LoBound);
4656 case Instruction::SetGT:
4658 return ReplaceInstUsesWith(I, ConstantBool::getFalse());
4659 return new SetCondInst(Instruction::SetGE, X, HiBound);
4666 // Simplify seteq and setne instructions with integer constant RHS.
4667 if (I.isEquality()) {
4668 bool isSetNE = I.getOpcode() == Instruction::SetNE;
4670 // If the first operand is (add|sub|and|or|xor|rem) with a constant, and
4671 // the second operand is a constant, simplify a bit.
4672 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0)) {
4673 switch (BO->getOpcode()) {
4674 case Instruction::SRem:
4675 // If we have a signed (X % (2^c)) == 0, turn it into an unsigned one.
4676 if (CI->isNullValue() && isa<ConstantInt>(BO->getOperand(1)) &&
4678 int64_t V = cast<ConstantInt>(BO->getOperand(1))->getSExtValue();
4679 if (V > 1 && isPowerOf2_64(V)) {
4680 Value *NewRem = InsertNewInstBefore(BinaryOperator::createURem(
4681 BO->getOperand(0), BO->getOperand(1), BO->getName()), I);
4682 return BinaryOperator::create(I.getOpcode(), NewRem,
4683 Constant::getNullValue(BO->getType()));
4687 case Instruction::Add:
4688 // Replace ((add A, B) != C) with (A != C-B) if B & C are constants.
4689 if (ConstantInt *BOp1C = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4690 if (BO->hasOneUse())
4691 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4692 ConstantExpr::getSub(CI, BOp1C));
4693 } else if (CI->isNullValue()) {
4694 // Replace ((add A, B) != 0) with (A != -B) if A or B is
4695 // efficiently invertible, or if the add has just this one use.
4696 Value *BOp0 = BO->getOperand(0), *BOp1 = BO->getOperand(1);
4698 if (Value *NegVal = dyn_castNegVal(BOp1))
4699 return new SetCondInst(I.getOpcode(), BOp0, NegVal);
4700 else if (Value *NegVal = dyn_castNegVal(BOp0))
4701 return new SetCondInst(I.getOpcode(), NegVal, BOp1);
4702 else if (BO->hasOneUse()) {
4703 Instruction *Neg = BinaryOperator::createNeg(BOp1, BO->getName());
4705 InsertNewInstBefore(Neg, I);
4706 return new SetCondInst(I.getOpcode(), BOp0, Neg);
4710 case Instruction::Xor:
4711 // For the xor case, we can xor two constants together, eliminating
4712 // the explicit xor.
4713 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1)))
4714 return BinaryOperator::create(I.getOpcode(), BO->getOperand(0),
4715 ConstantExpr::getXor(CI, BOC));
4718 case Instruction::Sub:
4719 // Replace (([sub|xor] A, B) != 0) with (A != B)
4720 if (CI->isNullValue())
4721 return new SetCondInst(I.getOpcode(), BO->getOperand(0),
4725 case Instruction::Or:
4726 // If bits are being or'd in that are not present in the constant we
4727 // are comparing against, then the comparison could never succeed!
4728 if (Constant *BOC = dyn_cast<Constant>(BO->getOperand(1))) {
4729 Constant *NotCI = ConstantExpr::getNot(CI);
4730 if (!ConstantExpr::getAnd(BOC, NotCI)->isNullValue())
4731 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4735 case Instruction::And:
4736 if (ConstantInt *BOC = dyn_cast<ConstantInt>(BO->getOperand(1))) {
4737 // If bits are being compared against that are and'd out, then the
4738 // comparison can never succeed!
4739 if (!ConstantExpr::getAnd(CI,
4740 ConstantExpr::getNot(BOC))->isNullValue())
4741 return ReplaceInstUsesWith(I, ConstantBool::get(isSetNE));
4743 // If we have ((X & C) == C), turn it into ((X & C) != 0).
4744 if (CI == BOC && isOneBitSet(CI))
4745 return new SetCondInst(isSetNE ? Instruction::SetEQ :
4746 Instruction::SetNE, Op0,
4747 Constant::getNullValue(CI->getType()));
4749 // Replace (and X, (1 << size(X)-1) != 0) with x < 0, converting X
4750 // to be a signed value as appropriate.
4751 if (isSignBit(BOC)) {
4752 Value *X = BO->getOperand(0);
4753 // If 'X' is not signed, insert a cast now...
4754 if (!BOC->getType()->isSigned()) {
4755 const Type *DestTy = BOC->getType()->getSignedVersion();
4756 X = InsertCastBefore(Instruction::BitCast, X, DestTy, I);
4758 return new SetCondInst(isSetNE ? Instruction::SetLT :
4759 Instruction::SetGE, X,
4760 Constant::getNullValue(X->getType()));
4763 // ((X & ~7) == 0) --> X < 8
4764 if (CI->isNullValue() && isHighOnes(BOC)) {
4765 Value *X = BO->getOperand(0);
4766 Constant *NegX = ConstantExpr::getNeg(BOC);
4768 // If 'X' is signed, insert a cast now.
4769 if (NegX->getType()->isSigned()) {
4770 const Type *DestTy = NegX->getType()->getUnsignedVersion();
4771 X = InsertCastBefore(Instruction::BitCast, X, DestTy, I);
4772 NegX = ConstantExpr::getBitCast(NegX, DestTy);
4775 return new SetCondInst(isSetNE ? Instruction::SetGE :
4776 Instruction::SetLT, X, NegX);
4782 } else if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(Op0)) {
4783 // Handle set{eq|ne} <intrinsic>, intcst.
4784 switch (II->getIntrinsicID()) {
4786 case Intrinsic::bswap_i16: // seteq (bswap(x)), c -> seteq(x,bswap(c))
4787 WorkList.push_back(II); // Dead?
4788 I.setOperand(0, II->getOperand(1));
4789 I.setOperand(1, ConstantInt::get(Type::UShortTy,
4790 ByteSwap_16(CI->getZExtValue())));
4792 case Intrinsic::bswap_i32: // seteq (bswap(x)), c -> seteq(x,bswap(c))
4793 WorkList.push_back(II); // Dead?
4794 I.setOperand(0, II->getOperand(1));
4795 I.setOperand(1, ConstantInt::get(Type::UIntTy,
4796 ByteSwap_32(CI->getZExtValue())));
4798 case Intrinsic::bswap_i64: // seteq (bswap(x)), c -> seteq(x,bswap(c))
4799 WorkList.push_back(II); // Dead?
4800 I.setOperand(0, II->getOperand(1));
4801 I.setOperand(1, ConstantInt::get(Type::ULongTy,
4802 ByteSwap_64(CI->getZExtValue())));
4806 } else { // Not a SetEQ/SetNE
4807 // If the LHS is a cast from an integral value of the same size,
4808 if (CastInst *Cast = dyn_cast<CastInst>(Op0)) {
4809 Value *CastOp = Cast->getOperand(0);
4810 const Type *SrcTy = CastOp->getType();
4811 unsigned SrcTySize = SrcTy->getPrimitiveSizeInBits();
4812 if (SrcTy != Cast->getType() && SrcTy->isInteger() &&
4813 SrcTySize == Cast->getType()->getPrimitiveSizeInBits()) {
4814 assert((SrcTy->isSigned() ^ Cast->getType()->isSigned()) &&
4815 "Source and destination signednesses should differ!");
4816 if (Cast->getType()->isSigned()) {
4817 // If this is a signed comparison, check for comparisons in the
4818 // vicinity of zero.
4819 if (I.getOpcode() == Instruction::SetLT && CI->isNullValue())
4821 return BinaryOperator::createSetGT(CastOp,
4822 ConstantInt::get(SrcTy, (1ULL << (SrcTySize-1))-1));
4823 else if (I.getOpcode() == Instruction::SetGT &&
4824 cast<ConstantInt>(CI)->getSExtValue() == -1)
4825 // X > -1 => x < 128
4826 return BinaryOperator::createSetLT(CastOp,
4827 ConstantInt::get(SrcTy, 1ULL << (SrcTySize-1)));
4829 ConstantInt *CUI = cast<ConstantInt>(CI);
4830 if (I.getOpcode() == Instruction::SetLT &&
4831 CUI->getZExtValue() == 1ULL << (SrcTySize-1))
4832 // X < 128 => X > -1
4833 return BinaryOperator::createSetGT(CastOp,
4834 ConstantInt::get(SrcTy, -1));
4835 else if (I.getOpcode() == Instruction::SetGT &&
4836 CUI->getZExtValue() == (1ULL << (SrcTySize-1))-1)
4838 return BinaryOperator::createSetLT(CastOp,
4839 Constant::getNullValue(SrcTy));
4846 // Handle setcc with constant RHS's that can be integer, FP or pointer.
4847 if (Constant *RHSC = dyn_cast<Constant>(Op1)) {
4848 if (Instruction *LHSI = dyn_cast<Instruction>(Op0))
4849 switch (LHSI->getOpcode()) {
4850 case Instruction::GetElementPtr:
4851 if (RHSC->isNullValue()) {
4852 // Transform setcc GEP P, int 0, int 0, int 0, null -> setcc P, null
4853 bool isAllZeros = true;
4854 for (unsigned i = 1, e = LHSI->getNumOperands(); i != e; ++i)
4855 if (!isa<Constant>(LHSI->getOperand(i)) ||
4856 !cast<Constant>(LHSI->getOperand(i))->isNullValue()) {
4861 return new SetCondInst(I.getOpcode(), LHSI->getOperand(0),
4862 Constant::getNullValue(LHSI->getOperand(0)->getType()));
4866 case Instruction::PHI:
4867 if (Instruction *NV = FoldOpIntoPhi(I))
4870 case Instruction::Select:
4871 // If either operand of the select is a constant, we can fold the
4872 // comparison into the select arms, which will cause one to be
4873 // constant folded and the select turned into a bitwise or.
4874 Value *Op1 = 0, *Op2 = 0;
4875 if (LHSI->hasOneUse()) {
4876 if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(1))) {
4877 // Fold the known value into the constant operand.
4878 Op1 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4879 // Insert a new SetCC of the other select operand.
4880 Op2 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4881 LHSI->getOperand(2), RHSC,
4883 } else if (Constant *C = dyn_cast<Constant>(LHSI->getOperand(2))) {
4884 // Fold the known value into the constant operand.
4885 Op2 = ConstantExpr::get(I.getOpcode(), C, RHSC);
4886 // Insert a new SetCC of the other select operand.
4887 Op1 = InsertNewInstBefore(new SetCondInst(I.getOpcode(),
4888 LHSI->getOperand(1), RHSC,
4894 return new SelectInst(LHSI->getOperand(0), Op1, Op2);
4899 // If we can optimize a 'setcc GEP, P' or 'setcc P, GEP', do so now.
4900 if (User *GEP = dyn_castGetElementPtr(Op0))
4901 if (Instruction *NI = FoldGEPSetCC(GEP, Op1, I.getOpcode(), I))
4903 if (User *GEP = dyn_castGetElementPtr(Op1))
4904 if (Instruction *NI = FoldGEPSetCC(GEP, Op0,
4905 SetCondInst::getSwappedCondition(I.getOpcode()), I))
4908 // Test to see if the operands of the setcc are casted versions of other
4909 // values. If the cast can be stripped off both arguments, we do so now.
4910 if (CastInst *CI = dyn_cast<CastInst>(Op0)) {
4911 Value *CastOp0 = CI->getOperand(0);
4912 if (CI->isLosslessCast() && I.isEquality() &&
4913 (isa<Constant>(Op1) || isa<CastInst>(Op1))) {
4914 // We keep moving the cast from the left operand over to the right
4915 // operand, where it can often be eliminated completely.
4918 // If operand #1 is a cast instruction, see if we can eliminate it as
4920 if (CastInst *CI2 = dyn_cast<CastInst>(Op1)) {
4921 Value *CI2Op0 = CI2->getOperand(0);
4922 if (CI2Op0->getType()->canLosslesslyBitCastTo(Op0->getType()))
4926 // If Op1 is a constant, we can fold the cast into the constant.
4927 if (Op1->getType() != Op0->getType())
4928 if (Constant *Op1C = dyn_cast<Constant>(Op1)) {
4929 Op1 = ConstantExpr::getBitCast(Op1C, Op0->getType());
4931 // Otherwise, cast the RHS right before the setcc
4932 Op1 = InsertCastBefore(Instruction::BitCast, Op1, Op0->getType(), I);
4934 return BinaryOperator::create(I.getOpcode(), Op0, Op1);
4937 // Handle the special case of: setcc (cast bool to X), <cst>
4938 // This comes up when you have code like
4941 // For generality, we handle any zero-extension of any operand comparison
4942 // with a constant or another cast from the same type.
4943 if (isa<ConstantInt>(Op1) || isa<CastInst>(Op1))
4944 if (Instruction *R = visitSetCondInstWithCastAndCast(I))
4948 if (I.isEquality()) {
4950 if (match(Op0, m_Xor(m_Value(A), m_Value(B))) &&
4951 (A == Op1 || B == Op1)) {
4952 // (A^B) == A -> B == 0
4953 Value *OtherVal = A == Op1 ? B : A;
4954 return BinaryOperator::create(I.getOpcode(), OtherVal,
4955 Constant::getNullValue(A->getType()));
4956 } else if (match(Op1, m_Xor(m_Value(A), m_Value(B))) &&
4957 (A == Op0 || B == Op0)) {
4958 // A == (A^B) -> B == 0
4959 Value *OtherVal = A == Op0 ? B : A;
4960 return BinaryOperator::create(I.getOpcode(), OtherVal,
4961 Constant::getNullValue(A->getType()));
4962 } else if (match(Op0, m_Sub(m_Value(A), m_Value(B))) && A == Op1) {
4963 // (A-B) == A -> B == 0
4964 return BinaryOperator::create(I.getOpcode(), B,
4965 Constant::getNullValue(B->getType()));
4966 } else if (match(Op1, m_Sub(m_Value(A), m_Value(B))) && A == Op0) {
4967 // A == (A-B) -> B == 0
4968 return BinaryOperator::create(I.getOpcode(), B,
4969 Constant::getNullValue(B->getType()));
4973 // (X&Z) == (Y&Z) -> (X^Y) & Z == 0
4974 if (Op0->hasOneUse() && Op1->hasOneUse() &&
4975 match(Op0, m_And(m_Value(A), m_Value(B))) &&
4976 match(Op1, m_And(m_Value(C), m_Value(D)))) {
4977 Value *X = 0, *Y = 0, *Z = 0;
4980 X = B; Y = D; Z = A;
4981 } else if (A == D) {
4982 X = B; Y = C; Z = A;
4983 } else if (B == C) {
4984 X = A; Y = D; Z = B;
4985 } else if (B == D) {
4986 X = A; Y = C; Z = B;
4989 if (X) { // Build (X^Y) & Z
4990 Op1 = InsertNewInstBefore(BinaryOperator::createXor(X, Y, "tmp"), I);
4991 Op1 = InsertNewInstBefore(BinaryOperator::createAnd(Op1, Z, "tmp"), I);
4992 I.setOperand(0, Op1);
4993 I.setOperand(1, Constant::getNullValue(Op1->getType()));
4998 return Changed ? &I : 0;
5001 // visitSetCondInstWithCastAndCast - Handle setcond (cast x to y), (cast/cst).
5002 // We only handle extending casts so far.
5004 Instruction *InstCombiner::visitSetCondInstWithCastAndCast(SetCondInst &SCI) {
5005 const CastInst *LHSCI = cast<CastInst>(SCI.getOperand(0));
5006 Value *LHSCIOp = LHSCI->getOperand(0);
5007 const Type *SrcTy = LHSCIOp->getType();
5008 const Type *DestTy = SCI.getOperand(0)->getType();
5011 if (!DestTy->isIntegral() || !SrcTy->isIntegral())
5014 unsigned SrcBits = SrcTy->getPrimitiveSizeInBits();
5015 unsigned DestBits = DestTy->getPrimitiveSizeInBits();
5016 if (SrcBits >= DestBits) return 0; // Only handle extending cast.
5018 // Is this a sign or zero extension?
5019 bool isSignSrc = SrcTy->isSigned();
5020 bool isSignDest = DestTy->isSigned();
5022 if (CastInst *CI = dyn_cast<CastInst>(SCI.getOperand(1))) {
5023 // Not an extension from the same type?
5024 RHSCIOp = CI->getOperand(0);
5025 if (RHSCIOp->getType() != LHSCIOp->getType()) return 0;
5026 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(SCI.getOperand(1))) {
5027 // Compute the constant that would happen if we truncated to SrcTy then
5028 // reextended to DestTy.
5029 Constant *Res1 = ConstantExpr::getTrunc(CI, SrcTy);
5030 Constant *Res2 = ConstantExpr::getCast(LHSCI->getOpcode(), Res1, DestTy);
5033 // Make sure that src sign and dest sign match. For example,
5035 // %A = cast short %X to uint
5036 // %B = setgt uint %A, 1330
5038 // It is incorrect to transform this into
5040 // %B = setgt short %X, 1330
5042 // because %A may have negative value.
5043 // However, it is OK if SrcTy is bool (See cast-set.ll testcase)
5044 // OR operation is EQ/NE.
5045 if (isSignSrc == isSignDest || SrcTy == Type::BoolTy || SCI.isEquality())
5050 // If the value cannot be represented in the shorter type, we cannot emit
5051 // a simple comparison.
5052 if (SCI.getOpcode() == Instruction::SetEQ)
5053 return ReplaceInstUsesWith(SCI, ConstantBool::getFalse());
5054 if (SCI.getOpcode() == Instruction::SetNE)
5055 return ReplaceInstUsesWith(SCI, ConstantBool::getTrue());
5057 // Evaluate the comparison for LT.
5059 if (DestTy->isSigned()) {
5060 // We're performing a signed comparison.
5062 // Signed extend and signed comparison.
5063 if (cast<ConstantInt>(CI)->getSExtValue() < 0)// X < (small) --> false
5064 Result = ConstantBool::getFalse();
5066 Result = ConstantBool::getTrue(); // X < (large) --> true
5068 // Unsigned extend and signed comparison.
5069 if (cast<ConstantInt>(CI)->getSExtValue() < 0)
5070 Result = ConstantBool::getFalse();
5072 Result = ConstantBool::getTrue();
5075 // We're performing an unsigned comparison.
5077 // Unsigned extend & compare -> always true.
5078 Result = ConstantBool::getTrue();
5080 // We're performing an unsigned comp with a sign extended value.
5081 // This is true if the input is >= 0. [aka >s -1]
5082 Constant *NegOne = ConstantIntegral::getAllOnesValue(SrcTy);
5083 Result = InsertNewInstBefore(BinaryOperator::createSetGT(LHSCIOp,
5084 NegOne, SCI.getName()), SCI);
5088 // Finally, return the value computed.
5089 if (SCI.getOpcode() == Instruction::SetLT) {
5090 return ReplaceInstUsesWith(SCI, Result);
5092 assert(SCI.getOpcode()==Instruction::SetGT &&"SetCC should be folded!");
5093 if (Constant *CI = dyn_cast<Constant>(Result))
5094 return ReplaceInstUsesWith(SCI, ConstantExpr::getNot(CI));
5096 return BinaryOperator::createNot(Result);
5103 // Okay, just insert a compare of the reduced operands now!
5104 return BinaryOperator::create(SCI.getOpcode(), LHSCIOp, RHSCIOp);
5107 Instruction *InstCombiner::visitShiftInst(ShiftInst &I) {
5108 assert(I.getOperand(1)->getType() == Type::UByteTy);
5109 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
5110 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5112 // shl X, 0 == X and shr X, 0 == X
5113 // shl 0, X == 0 and shr 0, X == 0
5114 if (Op1 == Constant::getNullValue(Type::UByteTy) ||
5115 Op0 == Constant::getNullValue(Op0->getType()))
5116 return ReplaceInstUsesWith(I, Op0);
5118 if (isa<UndefValue>(Op0)) { // undef >>s X -> undef
5119 if (!isLeftShift && I.getType()->isSigned())
5120 return ReplaceInstUsesWith(I, Op0);
5121 else // undef << X -> 0 AND undef >>u X -> 0
5122 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5124 if (isa<UndefValue>(Op1)) {
5125 if (isLeftShift || I.getType()->isUnsigned())// X << undef, X >>u undef -> 0
5126 return ReplaceInstUsesWith(I, Constant::getNullValue(I.getType()));
5128 return ReplaceInstUsesWith(I, Op0); // X >>s undef -> X
5131 // ashr int -1, X = -1 (for any arithmetic shift rights of ~0)
5132 if (I.getOpcode() == Instruction::AShr)
5133 if (ConstantInt *CSI = dyn_cast<ConstantInt>(Op0))
5134 if (CSI->isAllOnesValue())
5135 return ReplaceInstUsesWith(I, CSI);
5137 // Try to fold constant and into select arguments.
5138 if (isa<Constant>(Op0))
5139 if (SelectInst *SI = dyn_cast<SelectInst>(Op1))
5140 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5143 // See if we can turn a signed shr into an unsigned shr.
5144 if (I.isArithmeticShift()) {
5145 if (MaskedValueIsZero(Op0,
5146 1ULL << (I.getType()->getPrimitiveSizeInBits()-1))) {
5147 return new ShiftInst(Instruction::LShr, Op0, Op1, I.getName());
5151 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op1))
5152 if (CUI->getType()->isUnsigned())
5153 if (Instruction *Res = FoldShiftByConstant(Op0, CUI, I))
5158 Instruction *InstCombiner::FoldShiftByConstant(Value *Op0, ConstantInt *Op1,
5160 bool isLeftShift = I.getOpcode() == Instruction::Shl;
5161 bool isSignedShift = isLeftShift ? Op0->getType()->isSigned() :
5162 I.getOpcode() == Instruction::AShr;
5163 bool isUnsignedShift = !isSignedShift;
5165 // See if we can simplify any instructions used by the instruction whose sole
5166 // purpose is to compute bits we don't care about.
5167 uint64_t KnownZero, KnownOne;
5168 if (SimplifyDemandedBits(&I, I.getType()->getIntegralTypeMask(),
5169 KnownZero, KnownOne))
5172 // shl uint X, 32 = 0 and shr ubyte Y, 9 = 0, ... just don't eliminate shr
5173 // of a signed value.
5175 unsigned TypeBits = Op0->getType()->getPrimitiveSizeInBits();
5176 if (Op1->getZExtValue() >= TypeBits) {
5177 if (isUnsignedShift || isLeftShift)
5178 return ReplaceInstUsesWith(I, Constant::getNullValue(Op0->getType()));
5180 I.setOperand(1, ConstantInt::get(Type::UByteTy, TypeBits-1));
5185 // ((X*C1) << C2) == (X * (C1 << C2))
5186 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(Op0))
5187 if (BO->getOpcode() == Instruction::Mul && isLeftShift)
5188 if (Constant *BOOp = dyn_cast<Constant>(BO->getOperand(1)))
5189 return BinaryOperator::createMul(BO->getOperand(0),
5190 ConstantExpr::getShl(BOOp, Op1));
5192 // Try to fold constant and into select arguments.
5193 if (SelectInst *SI = dyn_cast<SelectInst>(Op0))
5194 if (Instruction *R = FoldOpIntoSelect(I, SI, this))
5196 if (isa<PHINode>(Op0))
5197 if (Instruction *NV = FoldOpIntoPhi(I))
5200 if (Op0->hasOneUse()) {
5201 if (BinaryOperator *Op0BO = dyn_cast<BinaryOperator>(Op0)) {
5202 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5205 switch (Op0BO->getOpcode()) {
5207 case Instruction::Add:
5208 case Instruction::And:
5209 case Instruction::Or:
5210 case Instruction::Xor:
5211 // These operators commute.
5212 // Turn (Y + (X >> C)) << C -> (X + (Y << C)) & (~0 << C)
5213 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5214 match(Op0BO->getOperand(1),
5215 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5216 Instruction *YS = new ShiftInst(Instruction::Shl,
5217 Op0BO->getOperand(0), Op1,
5219 InsertNewInstBefore(YS, I); // (Y << C)
5221 BinaryOperator::create(Op0BO->getOpcode(), YS, V1,
5222 Op0BO->getOperand(1)->getName());
5223 InsertNewInstBefore(X, I); // (X + (Y << C))
5224 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5225 C2 = ConstantExpr::getShl(C2, Op1);
5226 return BinaryOperator::createAnd(X, C2);
5229 // Turn (Y + ((X >> C) & CC)) << C -> ((X & (CC << C)) + (Y << C))
5230 if (isLeftShift && Op0BO->getOperand(1)->hasOneUse() &&
5231 match(Op0BO->getOperand(1),
5232 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5233 m_ConstantInt(CC))) && V2 == Op1 &&
5234 cast<BinaryOperator>(Op0BO->getOperand(1))->getOperand(0)->hasOneUse()) {
5235 Instruction *YS = new ShiftInst(Instruction::Shl,
5236 Op0BO->getOperand(0), Op1,
5238 InsertNewInstBefore(YS, I); // (Y << C)
5240 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5241 V1->getName()+".mask");
5242 InsertNewInstBefore(XM, I); // X & (CC << C)
5244 return BinaryOperator::create(Op0BO->getOpcode(), YS, XM);
5248 case Instruction::Sub:
5249 // Turn ((X >> C) + Y) << C -> (X + (Y << C)) & (~0 << C)
5250 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5251 match(Op0BO->getOperand(0),
5252 m_Shr(m_Value(V1), m_ConstantInt(CC))) && CC == Op1) {
5253 Instruction *YS = new ShiftInst(Instruction::Shl,
5254 Op0BO->getOperand(1), Op1,
5256 InsertNewInstBefore(YS, I); // (Y << C)
5258 BinaryOperator::create(Op0BO->getOpcode(), V1, YS,
5259 Op0BO->getOperand(0)->getName());
5260 InsertNewInstBefore(X, I); // (X + (Y << C))
5261 Constant *C2 = ConstantInt::getAllOnesValue(X->getType());
5262 C2 = ConstantExpr::getShl(C2, Op1);
5263 return BinaryOperator::createAnd(X, C2);
5266 // Turn (((X >> C)&CC) + Y) << C -> (X + (Y << C)) & (CC << C)
5267 if (isLeftShift && Op0BO->getOperand(0)->hasOneUse() &&
5268 match(Op0BO->getOperand(0),
5269 m_And(m_Shr(m_Value(V1), m_Value(V2)),
5270 m_ConstantInt(CC))) && V2 == Op1 &&
5271 cast<BinaryOperator>(Op0BO->getOperand(0))
5272 ->getOperand(0)->hasOneUse()) {
5273 Instruction *YS = new ShiftInst(Instruction::Shl,
5274 Op0BO->getOperand(1), Op1,
5276 InsertNewInstBefore(YS, I); // (Y << C)
5278 BinaryOperator::createAnd(V1, ConstantExpr::getShl(CC, Op1),
5279 V1->getName()+".mask");
5280 InsertNewInstBefore(XM, I); // X & (CC << C)
5282 return BinaryOperator::create(Op0BO->getOpcode(), XM, YS);
5289 // If the operand is an bitwise operator with a constant RHS, and the
5290 // shift is the only use, we can pull it out of the shift.
5291 if (ConstantInt *Op0C = dyn_cast<ConstantInt>(Op0BO->getOperand(1))) {
5292 bool isValid = true; // Valid only for And, Or, Xor
5293 bool highBitSet = false; // Transform if high bit of constant set?
5295 switch (Op0BO->getOpcode()) {
5296 default: isValid = false; break; // Do not perform transform!
5297 case Instruction::Add:
5298 isValid = isLeftShift;
5300 case Instruction::Or:
5301 case Instruction::Xor:
5304 case Instruction::And:
5309 // If this is a signed shift right, and the high bit is modified
5310 // by the logical operation, do not perform the transformation.
5311 // The highBitSet boolean indicates the value of the high bit of
5312 // the constant which would cause it to be modified for this
5315 if (isValid && !isLeftShift && isSignedShift) {
5316 uint64_t Val = Op0C->getZExtValue();
5317 isValid = ((Val & (1 << (TypeBits-1))) != 0) == highBitSet;
5321 Constant *NewRHS = ConstantExpr::get(I.getOpcode(), Op0C, Op1);
5323 Instruction *NewShift =
5324 new ShiftInst(I.getOpcode(), Op0BO->getOperand(0), Op1,
5327 InsertNewInstBefore(NewShift, I);
5329 return BinaryOperator::create(Op0BO->getOpcode(), NewShift,
5336 // Find out if this is a shift of a shift by a constant.
5337 ShiftInst *ShiftOp = 0;
5338 if (ShiftInst *Op0SI = dyn_cast<ShiftInst>(Op0))
5340 else if (BitCastInst *CI = dyn_cast<BitCastInst>(Op0)) {
5341 // If this is a noop-integer cast of a shift instruction, use the shift.
5342 if (isa<ShiftInst>(CI->getOperand(0))) {
5343 ShiftOp = cast<ShiftInst>(CI->getOperand(0));
5347 if (ShiftOp && isa<ConstantInt>(ShiftOp->getOperand(1))) {
5348 // Find the operands and properties of the input shift. Note that the
5349 // signedness of the input shift may differ from the current shift if there
5350 // is a noop cast between the two.
5351 bool isShiftOfLeftShift = ShiftOp->getOpcode() == Instruction::Shl;
5352 bool isShiftOfSignedShift = isShiftOfLeftShift ?
5353 ShiftOp->getType()->isSigned() :
5354 ShiftOp->getOpcode() == Instruction::AShr;
5355 bool isShiftOfUnsignedShift = !isShiftOfSignedShift;
5357 ConstantInt *ShiftAmt1C = cast<ConstantInt>(ShiftOp->getOperand(1));
5359 unsigned ShiftAmt1 = (unsigned)ShiftAmt1C->getZExtValue();
5360 unsigned ShiftAmt2 = (unsigned)Op1->getZExtValue();
5362 // Check for (A << c1) << c2 and (A >> c1) >> c2.
5363 if (isLeftShift == isShiftOfLeftShift) {
5364 // Do not fold these shifts if the first one is signed and the second one
5365 // is unsigned and this is a right shift. Further, don't do any folding
5367 if (isShiftOfSignedShift && isUnsignedShift && !isLeftShift)
5370 unsigned Amt = ShiftAmt1+ShiftAmt2; // Fold into one big shift.
5371 if (Amt > Op0->getType()->getPrimitiveSizeInBits())
5372 Amt = Op0->getType()->getPrimitiveSizeInBits();
5374 Value *Op = ShiftOp->getOperand(0);
5375 ShiftInst *ShiftResult = new ShiftInst(I.getOpcode(), Op,
5376 ConstantInt::get(Type::UByteTy, Amt));
5377 if (I.getType() == ShiftResult->getType())
5379 InsertNewInstBefore(ShiftResult, I);
5380 return CastInst::create(Instruction::BitCast, ShiftResult, I.getType());
5383 // Check for (A << c1) >> c2 or (A >> c1) << c2. If we are dealing with
5384 // signed types, we can only support the (A >> c1) << c2 configuration,
5385 // because it can not turn an arbitrary bit of A into a sign bit.
5386 if (isUnsignedShift || isLeftShift) {
5387 // Calculate bitmask for what gets shifted off the edge.
5388 Constant *C = ConstantIntegral::getAllOnesValue(I.getType());
5390 C = ConstantExpr::getShl(C, ShiftAmt1C);
5392 C = ConstantExpr::getLShr(C, ShiftAmt1C);
5394 Value *Op = ShiftOp->getOperand(0);
5395 if (Op->getType() != C->getType())
5396 Op = InsertCastBefore(Instruction::BitCast, Op, I.getType(), I);
5399 BinaryOperator::createAnd(Op, C, Op->getName()+".mask");
5400 InsertNewInstBefore(Mask, I);
5402 // Figure out what flavor of shift we should use...
5403 if (ShiftAmt1 == ShiftAmt2) {
5404 return ReplaceInstUsesWith(I, Mask); // (A << c) >> c === A & c2
5405 } else if (ShiftAmt1 < ShiftAmt2) {
5406 return new ShiftInst(I.getOpcode(), Mask,
5407 ConstantInt::get(Type::UByteTy, ShiftAmt2-ShiftAmt1));
5408 } else if (isShiftOfUnsignedShift || isShiftOfLeftShift) {
5409 if (isShiftOfUnsignedShift && !isShiftOfLeftShift && isSignedShift) {
5410 return new ShiftInst(Instruction::LShr, Mask,
5411 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5413 return new ShiftInst(ShiftOp->getOpcode(), Mask,
5414 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5417 // (X >>s C1) << C2 where C1 > C2 === (X >>s (C1-C2)) & mask
5418 Instruction *Shift =
5419 new ShiftInst(ShiftOp->getOpcode(), Mask,
5420 ConstantInt::get(Type::UByteTy, ShiftAmt1-ShiftAmt2));
5421 InsertNewInstBefore(Shift, I);
5423 C = ConstantIntegral::getAllOnesValue(Shift->getType());
5424 C = ConstantExpr::getShl(C, Op1);
5425 return BinaryOperator::createAnd(Shift, C, Op->getName()+".mask");
5428 // We can handle signed (X << C1) >>s C2 if it's a sign extend. In
5429 // this case, C1 == C2 and C1 is 8, 16, or 32.
5430 if (ShiftAmt1 == ShiftAmt2) {
5431 const Type *SExtType = 0;
5432 switch (Op0->getType()->getPrimitiveSizeInBits() - ShiftAmt1) {
5433 case 8 : SExtType = Type::SByteTy; break;
5434 case 16: SExtType = Type::ShortTy; break;
5435 case 32: SExtType = Type::IntTy; break;
5439 Instruction *NewTrunc =
5440 new TruncInst(ShiftOp->getOperand(0), SExtType, "sext");
5441 InsertNewInstBefore(NewTrunc, I);
5442 return new SExtInst(NewTrunc, I.getType());
5451 /// DecomposeSimpleLinearExpr - Analyze 'Val', seeing if it is a simple linear
5452 /// expression. If so, decompose it, returning some value X, such that Val is
5455 static Value *DecomposeSimpleLinearExpr(Value *Val, unsigned &Scale,
5457 assert(Val->getType() == Type::UIntTy && "Unexpected allocation size type!");
5458 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
5459 if (CI->getType()->isUnsigned()) {
5460 Offset = CI->getZExtValue();
5462 return ConstantInt::get(Type::UIntTy, 0);
5464 } else if (Instruction *I = dyn_cast<Instruction>(Val)) {
5465 if (I->getNumOperands() == 2) {
5466 if (ConstantInt *CUI = dyn_cast<ConstantInt>(I->getOperand(1))) {
5467 if (CUI->getType()->isUnsigned()) {
5468 if (I->getOpcode() == Instruction::Shl) {
5469 // This is a value scaled by '1 << the shift amt'.
5470 Scale = 1U << CUI->getZExtValue();
5472 return I->getOperand(0);
5473 } else if (I->getOpcode() == Instruction::Mul) {
5474 // This value is scaled by 'CUI'.
5475 Scale = CUI->getZExtValue();
5477 return I->getOperand(0);
5478 } else if (I->getOpcode() == Instruction::Add) {
5479 // We have X+C. Check to see if we really have (X*C2)+C1,
5480 // where C1 is divisible by C2.
5483 DecomposeSimpleLinearExpr(I->getOperand(0), SubScale, Offset);
5484 Offset += CUI->getZExtValue();
5485 if (SubScale > 1 && (Offset % SubScale == 0)) {
5495 // Otherwise, we can't look past this.
5502 /// PromoteCastOfAllocation - If we find a cast of an allocation instruction,
5503 /// try to eliminate the cast by moving the type information into the alloc.
5504 Instruction *InstCombiner::PromoteCastOfAllocation(CastInst &CI,
5505 AllocationInst &AI) {
5506 const PointerType *PTy = dyn_cast<PointerType>(CI.getType());
5507 if (!PTy) return 0; // Not casting the allocation to a pointer type.
5509 // Remove any uses of AI that are dead.
5510 assert(!CI.use_empty() && "Dead instructions should be removed earlier!");
5511 std::vector<Instruction*> DeadUsers;
5512 for (Value::use_iterator UI = AI.use_begin(), E = AI.use_end(); UI != E; ) {
5513 Instruction *User = cast<Instruction>(*UI++);
5514 if (isInstructionTriviallyDead(User)) {
5515 while (UI != E && *UI == User)
5516 ++UI; // If this instruction uses AI more than once, don't break UI.
5518 // Add operands to the worklist.
5519 AddUsesToWorkList(*User);
5521 DOUT << "IC: DCE: " << *User;
5523 User->eraseFromParent();
5524 removeFromWorkList(User);
5528 // Get the type really allocated and the type casted to.
5529 const Type *AllocElTy = AI.getAllocatedType();
5530 const Type *CastElTy = PTy->getElementType();
5531 if (!AllocElTy->isSized() || !CastElTy->isSized()) return 0;
5533 unsigned AllocElTyAlign = TD->getTypeAlignment(AllocElTy);
5534 unsigned CastElTyAlign = TD->getTypeAlignment(CastElTy);
5535 if (CastElTyAlign < AllocElTyAlign) return 0;
5537 // If the allocation has multiple uses, only promote it if we are strictly
5538 // increasing the alignment of the resultant allocation. If we keep it the
5539 // same, we open the door to infinite loops of various kinds.
5540 if (!AI.hasOneUse() && CastElTyAlign == AllocElTyAlign) return 0;
5542 uint64_t AllocElTySize = TD->getTypeSize(AllocElTy);
5543 uint64_t CastElTySize = TD->getTypeSize(CastElTy);
5544 if (CastElTySize == 0 || AllocElTySize == 0) return 0;
5546 // See if we can satisfy the modulus by pulling a scale out of the array
5548 unsigned ArraySizeScale, ArrayOffset;
5549 Value *NumElements = // See if the array size is a decomposable linear expr.
5550 DecomposeSimpleLinearExpr(AI.getOperand(0), ArraySizeScale, ArrayOffset);
5552 // If we can now satisfy the modulus, by using a non-1 scale, we really can
5554 if ((AllocElTySize*ArraySizeScale) % CastElTySize != 0 ||
5555 (AllocElTySize*ArrayOffset ) % CastElTySize != 0) return 0;
5557 unsigned Scale = (AllocElTySize*ArraySizeScale)/CastElTySize;
5562 // If the allocation size is constant, form a constant mul expression
5563 Amt = ConstantInt::get(Type::UIntTy, Scale);
5564 if (isa<ConstantInt>(NumElements) && NumElements->getType()->isUnsigned())
5565 Amt = ConstantExpr::getMul(
5566 cast<ConstantInt>(NumElements), cast<ConstantInt>(Amt));
5567 // otherwise multiply the amount and the number of elements
5568 else if (Scale != 1) {
5569 Instruction *Tmp = BinaryOperator::createMul(Amt, NumElements, "tmp");
5570 Amt = InsertNewInstBefore(Tmp, AI);
5574 if (unsigned Offset = (AllocElTySize*ArrayOffset)/CastElTySize) {
5575 Value *Off = ConstantInt::get(Type::UIntTy, Offset);
5576 Instruction *Tmp = BinaryOperator::createAdd(Amt, Off, "tmp");
5577 Amt = InsertNewInstBefore(Tmp, AI);
5580 std::string Name = AI.getName(); AI.setName("");
5581 AllocationInst *New;
5582 if (isa<MallocInst>(AI))
5583 New = new MallocInst(CastElTy, Amt, AI.getAlignment(), Name);
5585 New = new AllocaInst(CastElTy, Amt, AI.getAlignment(), Name);
5586 InsertNewInstBefore(New, AI);
5588 // If the allocation has multiple uses, insert a cast and change all things
5589 // that used it to use the new cast. This will also hack on CI, but it will
5591 if (!AI.hasOneUse()) {
5592 AddUsesToWorkList(AI);
5593 // New is the allocation instruction, pointer typed. AI is the original
5594 // allocation instruction, also pointer typed. Thus, cast to use is BitCast.
5595 CastInst *NewCast = new BitCastInst(New, AI.getType(), "tmpcast");
5596 InsertNewInstBefore(NewCast, AI);
5597 AI.replaceAllUsesWith(NewCast);
5599 return ReplaceInstUsesWith(CI, New);
5602 /// CanEvaluateInDifferentType - Return true if we can take the specified value
5603 /// and return it without inserting any new casts. This is used by code that
5604 /// tries to decide whether promoting or shrinking integer operations to wider
5605 /// or smaller types will allow us to eliminate a truncate or extend.
5606 static bool CanEvaluateInDifferentType(Value *V, const Type *Ty,
5607 int &NumCastsRemoved) {
5608 if (isa<Constant>(V)) return true;
5610 Instruction *I = dyn_cast<Instruction>(V);
5611 if (!I || !I->hasOneUse()) return false;
5613 switch (I->getOpcode()) {
5614 case Instruction::And:
5615 case Instruction::Or:
5616 case Instruction::Xor:
5617 // These operators can all arbitrarily be extended or truncated.
5618 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved) &&
5619 CanEvaluateInDifferentType(I->getOperand(1), Ty, NumCastsRemoved);
5620 case Instruction::AShr:
5621 case Instruction::LShr:
5622 case Instruction::Shl:
5623 // If this is just a bitcast changing the sign of the operation, we can
5624 // convert if the operand can be converted.
5625 if (V->getType()->getPrimitiveSizeInBits() == Ty->getPrimitiveSizeInBits())
5626 return CanEvaluateInDifferentType(I->getOperand(0), Ty, NumCastsRemoved);
5628 case Instruction::Trunc:
5629 case Instruction::ZExt:
5630 case Instruction::SExt:
5631 case Instruction::BitCast:
5632 // If this is a cast from the destination type, we can trivially eliminate
5633 // it, and this will remove a cast overall.
5634 if (I->getOperand(0)->getType() == Ty) {
5635 // If the first operand is itself a cast, and is eliminable, do not count
5636 // this as an eliminable cast. We would prefer to eliminate those two
5638 if (isa<CastInst>(I->getOperand(0)))
5646 // TODO: Can handle more cases here.
5653 /// EvaluateInDifferentType - Given an expression that
5654 /// CanEvaluateInDifferentType returns true for, actually insert the code to
5655 /// evaluate the expression.
5656 Value *InstCombiner::EvaluateInDifferentType(Value *V, const Type *Ty,
5658 if (Constant *C = dyn_cast<Constant>(V))
5659 return ConstantExpr::getIntegerCast(C, Ty, isSigned /*Sext or ZExt*/);
5661 // Otherwise, it must be an instruction.
5662 Instruction *I = cast<Instruction>(V);
5663 Instruction *Res = 0;
5664 switch (I->getOpcode()) {
5665 case Instruction::And:
5666 case Instruction::Or:
5667 case Instruction::Xor: {
5668 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5669 Value *RHS = EvaluateInDifferentType(I->getOperand(1), Ty, isSigned);
5670 Res = BinaryOperator::create((Instruction::BinaryOps)I->getOpcode(),
5671 LHS, RHS, I->getName());
5674 case Instruction::AShr:
5675 case Instruction::LShr:
5676 case Instruction::Shl: {
5677 Value *LHS = EvaluateInDifferentType(I->getOperand(0), Ty, isSigned);
5678 Res = new ShiftInst((Instruction::OtherOps)I->getOpcode(), LHS,
5679 I->getOperand(1), I->getName());
5682 case Instruction::Trunc:
5683 case Instruction::ZExt:
5684 case Instruction::SExt:
5685 case Instruction::BitCast:
5686 // If the source type of the cast is the type we're trying for then we can
5687 // just return the source. There's no need to insert it because its not new.
5688 if (I->getOperand(0)->getType() == Ty)
5689 return I->getOperand(0);
5691 // Some other kind of cast, which shouldn't happen, so just ..
5694 // TODO: Can handle more cases here.
5695 assert(0 && "Unreachable!");
5699 return InsertNewInstBefore(Res, *I);
5702 /// @brief Implement the transforms common to all CastInst visitors.
5703 Instruction *InstCombiner::commonCastTransforms(CastInst &CI) {
5704 Value *Src = CI.getOperand(0);
5706 // Casting undef to anything results in undef so might as just replace it and
5707 // get rid of the cast.
5708 if (isa<UndefValue>(Src)) // cast undef -> undef
5709 return ReplaceInstUsesWith(CI, UndefValue::get(CI.getType()));
5711 // Many cases of "cast of a cast" are eliminable. If its eliminable we just
5712 // eliminate it now.
5713 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
5714 if (Instruction::CastOps opc =
5715 isEliminableCastPair(CSrc, CI.getOpcode(), CI.getType(), TD)) {
5716 // The first cast (CSrc) is eliminable so we need to fix up or replace
5717 // the second cast (CI). CSrc will then have a good chance of being dead.
5718 return CastInst::create(opc, CSrc->getOperand(0), CI.getType());
5722 // If casting the result of a getelementptr instruction with no offset, turn
5723 // this into a cast of the original pointer!
5725 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Src)) {
5726 bool AllZeroOperands = true;
5727 for (unsigned i = 1, e = GEP->getNumOperands(); i != e; ++i)
5728 if (!isa<Constant>(GEP->getOperand(i)) ||
5729 !cast<Constant>(GEP->getOperand(i))->isNullValue()) {
5730 AllZeroOperands = false;
5733 if (AllZeroOperands) {
5734 // Changing the cast operand is usually not a good idea but it is safe
5735 // here because the pointer operand is being replaced with another
5736 // pointer operand so the opcode doesn't need to change.
5737 CI.setOperand(0, GEP->getOperand(0));
5742 // If we are casting a malloc or alloca to a pointer to a type of the same
5743 // size, rewrite the allocation instruction to allocate the "right" type.
5744 if (AllocationInst *AI = dyn_cast<AllocationInst>(Src))
5745 if (Instruction *V = PromoteCastOfAllocation(CI, *AI))
5748 // If we are casting a select then fold the cast into the select
5749 if (SelectInst *SI = dyn_cast<SelectInst>(Src))
5750 if (Instruction *NV = FoldOpIntoSelect(CI, SI, this))
5753 // If we are casting a PHI then fold the cast into the PHI
5754 if (isa<PHINode>(Src))
5755 if (Instruction *NV = FoldOpIntoPhi(CI))
5761 /// Only the TRUNC, ZEXT, SEXT, and BITCONVERT can have both operands as
5762 /// integers. This function implements the common transforms for all those
5764 /// @brief Implement the transforms common to CastInst with integer operands
5765 Instruction *InstCombiner::commonIntCastTransforms(CastInst &CI) {
5766 if (Instruction *Result = commonCastTransforms(CI))
5769 Value *Src = CI.getOperand(0);
5770 const Type *SrcTy = Src->getType();
5771 const Type *DestTy = CI.getType();
5772 unsigned SrcBitSize = SrcTy->getPrimitiveSizeInBits();
5773 unsigned DestBitSize = DestTy->getPrimitiveSizeInBits();
5775 // See if we can simplify any instructions used by the LHS whose sole
5776 // purpose is to compute bits we don't care about.
5777 uint64_t KnownZero = 0, KnownOne = 0;
5778 if (SimplifyDemandedBits(&CI, DestTy->getIntegralTypeMask(),
5779 KnownZero, KnownOne))
5782 // If the source isn't an instruction or has more than one use then we
5783 // can't do anything more.
5784 if (!isa<Instruction>(Src) || !Src->hasOneUse())
5787 // Attempt to propagate the cast into the instruction.
5788 Instruction *SrcI = cast<Instruction>(Src);
5789 int NumCastsRemoved = 0;
5790 if (CanEvaluateInDifferentType(SrcI, DestTy, NumCastsRemoved)) {
5791 // If this cast is a truncate, evaluting in a different type always
5792 // eliminates the cast, so it is always a win. If this is a noop-cast
5793 // this just removes a noop cast which isn't pointful, but simplifies
5794 // the code. If this is a zero-extension, we need to do an AND to
5795 // maintain the clear top-part of the computation, so we require that
5796 // the input have eliminated at least one cast. If this is a sign
5797 // extension, we insert two new casts (to do the extension) so we
5798 // require that two casts have been eliminated.
5799 bool DoXForm = CI.isNoopCast(TD->getIntPtrType());
5801 switch (CI.getOpcode()) {
5802 case Instruction::Trunc:
5805 case Instruction::ZExt:
5806 DoXForm = NumCastsRemoved >= 1;
5808 case Instruction::SExt:
5809 DoXForm = NumCastsRemoved >= 2;
5811 case Instruction::BitCast:
5815 // All the others use floating point so we shouldn't actually
5816 // get here because of the check above.
5817 assert(!"Unknown cast type .. unreachable");
5823 Value *Res = EvaluateInDifferentType(SrcI, DestTy,
5824 CI.getOpcode() == Instruction::SExt);
5825 assert(Res->getType() == DestTy);
5826 switch (CI.getOpcode()) {
5827 default: assert(0 && "Unknown cast type!");
5828 case Instruction::Trunc:
5829 case Instruction::BitCast:
5830 // Just replace this cast with the result.
5831 return ReplaceInstUsesWith(CI, Res);
5832 case Instruction::ZExt: {
5833 // We need to emit an AND to clear the high bits.
5834 assert(SrcBitSize < DestBitSize && "Not a zext?");
5836 ConstantInt::get(Type::ULongTy, (1ULL << SrcBitSize)-1);
5837 if (DestBitSize < 64)
5838 C = ConstantExpr::getTrunc(C, DestTy);
5840 assert(DestBitSize == 64);
5841 C = ConstantExpr::getBitCast(C, DestTy);
5843 return BinaryOperator::createAnd(Res, C);
5845 case Instruction::SExt:
5846 // We need to emit a cast to truncate, then a cast to sext.
5847 return CastInst::create(Instruction::SExt,
5848 InsertCastBefore(Instruction::Trunc, Res, Src->getType(),
5854 Value *Op0 = SrcI->getNumOperands() > 0 ? SrcI->getOperand(0) : 0;
5855 Value *Op1 = SrcI->getNumOperands() > 1 ? SrcI->getOperand(1) : 0;
5857 switch (SrcI->getOpcode()) {
5858 case Instruction::Add:
5859 case Instruction::Mul:
5860 case Instruction::And:
5861 case Instruction::Or:
5862 case Instruction::Xor:
5863 // If we are discarding information, or just changing the sign,
5865 if (DestBitSize <= SrcBitSize && DestBitSize != 1) {
5866 // Don't insert two casts if they cannot be eliminated. We allow
5867 // two casts to be inserted if the sizes are the same. This could
5868 // only be converting signedness, which is a noop.
5869 if (DestBitSize == SrcBitSize ||
5870 !ValueRequiresCast(Op1, DestTy,TD) ||
5871 !ValueRequiresCast(Op0, DestTy, TD)) {
5872 Instruction::CastOps opcode = CI.getOpcode();
5873 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
5874 Value *Op1c = InsertOperandCastBefore(opcode, Op1, DestTy, SrcI);
5875 return BinaryOperator::create(
5876 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
5880 // cast (xor bool X, true) to int --> xor (cast bool X to int), 1
5881 if (isa<ZExtInst>(CI) && SrcBitSize == 1 &&
5882 SrcI->getOpcode() == Instruction::Xor &&
5883 Op1 == ConstantBool::getTrue() &&
5884 (!Op0->hasOneUse() || !isa<SetCondInst>(Op0))) {
5885 Value *New = InsertOperandCastBefore(Instruction::ZExt, Op0, DestTy, &CI);
5886 return BinaryOperator::createXor(New, ConstantInt::get(CI.getType(), 1));
5889 case Instruction::SDiv:
5890 case Instruction::UDiv:
5891 case Instruction::SRem:
5892 case Instruction::URem:
5893 // If we are just changing the sign, rewrite.
5894 if (DestBitSize == SrcBitSize) {
5895 // Don't insert two casts if they cannot be eliminated. We allow
5896 // two casts to be inserted if the sizes are the same. This could
5897 // only be converting signedness, which is a noop.
5898 if (!ValueRequiresCast(Op1, DestTy,TD) ||
5899 !ValueRequiresCast(Op0, DestTy, TD)) {
5900 Value *Op0c = InsertOperandCastBefore(Instruction::BitCast,
5902 Value *Op1c = InsertOperandCastBefore(Instruction::BitCast,
5904 return BinaryOperator::create(
5905 cast<BinaryOperator>(SrcI)->getOpcode(), Op0c, Op1c);
5910 case Instruction::Shl:
5911 // Allow changing the sign of the source operand. Do not allow
5912 // changing the size of the shift, UNLESS the shift amount is a
5913 // constant. We must not change variable sized shifts to a smaller
5914 // size, because it is undefined to shift more bits out than exist
5916 if (DestBitSize == SrcBitSize ||
5917 (DestBitSize < SrcBitSize && isa<Constant>(Op1))) {
5918 Instruction::CastOps opcode = (DestBitSize == SrcBitSize ?
5919 Instruction::BitCast : Instruction::Trunc);
5920 Value *Op0c = InsertOperandCastBefore(opcode, Op0, DestTy, SrcI);
5921 return new ShiftInst(Instruction::Shl, Op0c, Op1);
5924 case Instruction::AShr:
5925 // If this is a signed shr, and if all bits shifted in are about to be
5926 // truncated off, turn it into an unsigned shr to allow greater
5928 if (DestBitSize < SrcBitSize &&
5929 isa<ConstantInt>(Op1)) {
5930 unsigned ShiftAmt = cast<ConstantInt>(Op1)->getZExtValue();
5931 if (SrcBitSize > ShiftAmt && SrcBitSize-ShiftAmt >= DestBitSize) {
5932 // Insert the new logical shift right.
5933 return new ShiftInst(Instruction::LShr, Op0, Op1);
5938 case Instruction::SetEQ:
5939 case Instruction::SetNE:
5940 // If we are just checking for a seteq of a single bit and casting it
5941 // to an integer. If so, shift the bit to the appropriate place then
5942 // cast to integer to avoid the comparison.
5943 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
5944 uint64_t Op1CV = Op1C->getZExtValue();
5945 // cast (X == 0) to int --> X^1 iff X has only the low bit set.
5946 // cast (X == 0) to int --> (X>>1)^1 iff X has only the 2nd bit set.
5947 // cast (X == 1) to int --> X iff X has only the low bit set.
5948 // cast (X == 2) to int --> X>>1 iff X has only the 2nd bit set.
5949 // cast (X != 0) to int --> X iff X has only the low bit set.
5950 // cast (X != 0) to int --> X>>1 iff X has only the 2nd bit set.
5951 // cast (X != 1) to int --> X^1 iff X has only the low bit set.
5952 // cast (X != 2) to int --> (X>>1)^1 iff X has only the 2nd bit set.
5953 if (Op1CV == 0 || isPowerOf2_64(Op1CV)) {
5954 // If Op1C some other power of two, convert:
5955 uint64_t KnownZero, KnownOne;
5956 uint64_t TypeMask = Op1->getType()->getIntegralTypeMask();
5957 ComputeMaskedBits(Op0, TypeMask, KnownZero, KnownOne);
5959 if (isPowerOf2_64(KnownZero^TypeMask)) { // Exactly 1 possible 1?
5960 bool isSetNE = SrcI->getOpcode() == Instruction::SetNE;
5961 if (Op1CV && (Op1CV != (KnownZero^TypeMask))) {
5962 // (X&4) == 2 --> false
5963 // (X&4) != 2 --> true
5964 Constant *Res = ConstantBool::get(isSetNE);
5965 Res = ConstantExpr::getZExt(Res, CI.getType());
5966 return ReplaceInstUsesWith(CI, Res);
5969 unsigned ShiftAmt = Log2_64(KnownZero^TypeMask);
5972 // Perform a logical shr by shiftamt.
5973 // Insert the shift to put the result in the low bit.
5974 In = InsertNewInstBefore(
5975 new ShiftInst(Instruction::LShr, In,
5976 ConstantInt::get(Type::UByteTy, ShiftAmt),
5977 In->getName()+".lobit"), CI);
5980 if ((Op1CV != 0) == isSetNE) { // Toggle the low bit.
5981 Constant *One = ConstantInt::get(In->getType(), 1);
5982 In = BinaryOperator::createXor(In, One, "tmp");
5983 InsertNewInstBefore(cast<Instruction>(In), CI);
5986 if (CI.getType() == In->getType())
5987 return ReplaceInstUsesWith(CI, In);
5989 return CastInst::createIntegerCast(In, CI.getType(), false/*ZExt*/);
5998 Instruction *InstCombiner::visitTrunc(CastInst &CI) {
5999 if (Instruction *Result = commonIntCastTransforms(CI))
6002 Value *Src = CI.getOperand(0);
6003 const Type *Ty = CI.getType();
6004 unsigned DestBitWidth = Ty->getPrimitiveSizeInBits();
6006 if (Instruction *SrcI = dyn_cast<Instruction>(Src)) {
6007 switch (SrcI->getOpcode()) {
6009 case Instruction::LShr:
6010 // We can shrink lshr to something smaller if we know the bits shifted in
6011 // are already zeros.
6012 if (ConstantInt *ShAmtV = dyn_cast<ConstantInt>(SrcI->getOperand(1))) {
6013 unsigned ShAmt = ShAmtV->getZExtValue();
6015 // Get a mask for the bits shifting in.
6016 uint64_t Mask = (~0ULL >> (64-ShAmt)) << DestBitWidth;
6017 Value* SrcIOp0 = SrcI->getOperand(0);
6018 if (SrcI->hasOneUse() && MaskedValueIsZero(SrcIOp0, Mask)) {
6019 if (ShAmt >= DestBitWidth) // All zeros.
6020 return ReplaceInstUsesWith(CI, Constant::getNullValue(Ty));
6022 // Okay, we can shrink this. Truncate the input, then return a new
6024 Value *V = InsertCastBefore(Instruction::Trunc, SrcIOp0, Ty, CI);
6025 return new ShiftInst(Instruction::LShr, V, SrcI->getOperand(1));
6027 } else { // This is a variable shr.
6029 // Turn 'trunc (lshr X, Y) to bool' into '(X & (1 << Y)) != 0'. This is
6030 // more LLVM instructions, but allows '1 << Y' to be hoisted if
6031 // loop-invariant and CSE'd.
6032 if (CI.getType() == Type::BoolTy && SrcI->hasOneUse()) {
6033 Value *One = ConstantInt::get(SrcI->getType(), 1);
6035 Value *V = InsertNewInstBefore(new ShiftInst(Instruction::Shl, One,
6036 SrcI->getOperand(1),
6038 V = InsertNewInstBefore(BinaryOperator::createAnd(V,
6039 SrcI->getOperand(0),
6041 Value *Zero = Constant::getNullValue(V->getType());
6042 return BinaryOperator::createSetNE(V, Zero);
6052 Instruction *InstCombiner::visitZExt(CastInst &CI) {
6053 // If one of the common conversion will work ..
6054 if (Instruction *Result = commonIntCastTransforms(CI))
6057 Value *Src = CI.getOperand(0);
6059 // If this is a cast of a cast
6060 if (CastInst *CSrc = dyn_cast<CastInst>(Src)) { // A->B->C cast
6061 // If this is a TRUNC followed by a ZEXT then we are dealing with integral
6062 // types and if the sizes are just right we can convert this into a logical
6063 // 'and' which will be much cheaper than the pair of casts.
6064 if (isa<TruncInst>(CSrc)) {
6065 // Get the sizes of the types involved
6066 Value *A = CSrc->getOperand(0);
6067 unsigned SrcSize = A->getType()->getPrimitiveSizeInBits();
6068 unsigned MidSize = CSrc->getType()->getPrimitiveSizeInBits();
6069 unsigned DstSize = CI.getType()->getPrimitiveSizeInBits();
6070 // If we're actually extending zero bits and the trunc is a no-op
6071 if (MidSize < DstSize && SrcSize == DstSize) {
6072 // Replace both of the casts with an And of the type mask.
6073 uint64_t AndValue = CSrc->getType()->getIntegralTypeMask();
6074 Constant *AndConst = ConstantInt::get(A->getType(), AndValue);
6076 BinaryOperator::createAnd(CSrc->getOperand(0), AndConst);
6077 // Unfortunately, if the type changed, we need to cast it back.
6078 if (And->getType() != CI.getType()) {
6079 And->setName(CSrc->getName()+".mask");
6080 InsertNewInstBefore(And, CI);
6081 And = CastInst::createIntegerCast(And, CI.getType(), false/*ZExt*/);
6091 Instruction *InstCombiner::visitSExt(CastInst &CI) {
6092 return commonIntCastTransforms(CI);
6095 Instruction *InstCombiner::visitFPTrunc(CastInst &CI) {
6096 return commonCastTransforms(CI);
6099 Instruction *InstCombiner::visitFPExt(CastInst &CI) {
6100 return commonCastTransforms(CI);
6103 Instruction *InstCombiner::visitFPToUI(CastInst &CI) {
6104 return commonCastTransforms(CI);
6107 Instruction *InstCombiner::visitFPToSI(CastInst &CI) {
6108 return commonCastTransforms(CI);
6111 Instruction *InstCombiner::visitUIToFP(CastInst &CI) {
6112 return commonCastTransforms(CI);
6115 Instruction *InstCombiner::visitSIToFP(CastInst &CI) {
6116 return commonCastTransforms(CI);
6119 Instruction *InstCombiner::visitPtrToInt(CastInst &CI) {
6120 return commonCastTransforms(CI);
6123 Instruction *InstCombiner::visitIntToPtr(CastInst &CI) {
6124 return commonCastTransforms(CI);
6127 Instruction *InstCombiner::visitBitCast(CastInst &CI) {
6129 // If the operands are integer typed then apply the integer transforms,
6130 // otherwise just apply the common ones.
6131 Value *Src = CI.getOperand(0);
6132 const Type *SrcTy = Src->getType();
6133 const Type *DestTy = CI.getType();
6135 if (SrcTy->isInteger() && DestTy->isInteger()) {
6136 if (Instruction *Result = commonIntCastTransforms(CI))
6139 if (Instruction *Result = commonCastTransforms(CI))
6144 // Get rid of casts from one type to the same type. These are useless and can
6145 // be replaced by the operand.
6146 if (DestTy == Src->getType())
6147 return ReplaceInstUsesWith(CI, Src);
6149 // If the source and destination are pointers, and this cast is equivalent to
6150 // a getelementptr X, 0, 0, 0... turn it into the appropriate getelementptr.
6151 // This can enhance SROA and other transforms that want type-safe pointers.
6152 if (const PointerType *DstPTy = dyn_cast<PointerType>(DestTy)) {
6153 if (const PointerType *SrcPTy = dyn_cast<PointerType>(SrcTy)) {
6154 const Type *DstElTy = DstPTy->getElementType();
6155 const Type *SrcElTy = SrcPTy->getElementType();
6157 Constant *ZeroUInt = Constant::getNullValue(Type::UIntTy);
6158 unsigned NumZeros = 0;
6159 while (SrcElTy != DstElTy &&
6160 isa<CompositeType>(SrcElTy) && !isa<PointerType>(SrcElTy) &&
6161 SrcElTy->getNumContainedTypes() /* not "{}" */) {
6162 SrcElTy = cast<CompositeType>(SrcElTy)->getTypeAtIndex(ZeroUInt);
6166 // If we found a path from the src to dest, create the getelementptr now.
6167 if (SrcElTy == DstElTy) {
6168 std::vector<Value*> Idxs(NumZeros+1, ZeroUInt);
6169 return new GetElementPtrInst(Src, Idxs);
6174 if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(Src)) {
6175 if (SVI->hasOneUse()) {
6176 // Okay, we have (bitconvert (shuffle ..)). Check to see if this is
6177 // a bitconvert to a vector with the same # elts.
6178 if (isa<PackedType>(DestTy) &&
6179 cast<PackedType>(DestTy)->getNumElements() ==
6180 SVI->getType()->getNumElements()) {
6182 // If either of the operands is a cast from CI.getType(), then
6183 // evaluating the shuffle in the casted destination's type will allow
6184 // us to eliminate at least one cast.
6185 if (((Tmp = dyn_cast<CastInst>(SVI->getOperand(0))) &&
6186 Tmp->getOperand(0)->getType() == DestTy) ||
6187 ((Tmp = dyn_cast<CastInst>(SVI->getOperand(1))) &&
6188 Tmp->getOperand(0)->getType() == DestTy)) {
6189 Value *LHS = InsertOperandCastBefore(Instruction::BitCast,
6190 SVI->getOperand(0), DestTy, &CI);
6191 Value *RHS = InsertOperandCastBefore(Instruction::BitCast,
6192 SVI->getOperand(1), DestTy, &CI);
6193 // Return a new shuffle vector. Use the same element ID's, as we
6194 // know the vector types match #elts.
6195 return new ShuffleVectorInst(LHS, RHS, SVI->getOperand(2));
6203 /// GetSelectFoldableOperands - We want to turn code that looks like this:
6205 /// %D = select %cond, %C, %A
6207 /// %C = select %cond, %B, 0
6210 /// Assuming that the specified instruction is an operand to the select, return
6211 /// a bitmask indicating which operands of this instruction are foldable if they
6212 /// equal the other incoming value of the select.
6214 static unsigned GetSelectFoldableOperands(Instruction *I) {
6215 switch (I->getOpcode()) {
6216 case Instruction::Add:
6217 case Instruction::Mul:
6218 case Instruction::And:
6219 case Instruction::Or:
6220 case Instruction::Xor:
6221 return 3; // Can fold through either operand.
6222 case Instruction::Sub: // Can only fold on the amount subtracted.
6223 case Instruction::Shl: // Can only fold on the shift amount.
6224 case Instruction::LShr:
6225 case Instruction::AShr:
6228 return 0; // Cannot fold
6232 /// GetSelectFoldableConstant - For the same transformation as the previous
6233 /// function, return the identity constant that goes into the select.
6234 static Constant *GetSelectFoldableConstant(Instruction *I) {
6235 switch (I->getOpcode()) {
6236 default: assert(0 && "This cannot happen!"); abort();
6237 case Instruction::Add:
6238 case Instruction::Sub:
6239 case Instruction::Or:
6240 case Instruction::Xor:
6241 return Constant::getNullValue(I->getType());
6242 case Instruction::Shl:
6243 case Instruction::LShr:
6244 case Instruction::AShr:
6245 return Constant::getNullValue(Type::UByteTy);
6246 case Instruction::And:
6247 return ConstantInt::getAllOnesValue(I->getType());
6248 case Instruction::Mul:
6249 return ConstantInt::get(I->getType(), 1);
6253 /// FoldSelectOpOp - Here we have (select c, TI, FI), and we know that TI and FI
6254 /// have the same opcode and only one use each. Try to simplify this.
6255 Instruction *InstCombiner::FoldSelectOpOp(SelectInst &SI, Instruction *TI,
6257 if (TI->getNumOperands() == 1) {
6258 // If this is a non-volatile load or a cast from the same type,
6261 if (TI->getOperand(0)->getType() != FI->getOperand(0)->getType())
6264 return 0; // unknown unary op.
6267 // Fold this by inserting a select from the input values.
6268 SelectInst *NewSI = new SelectInst(SI.getCondition(), TI->getOperand(0),
6269 FI->getOperand(0), SI.getName()+".v");
6270 InsertNewInstBefore(NewSI, SI);
6271 return CastInst::create(Instruction::CastOps(TI->getOpcode()), NewSI,
6275 // Only handle binary operators here.
6276 if (!isa<ShiftInst>(TI) && !isa<BinaryOperator>(TI))
6279 // Figure out if the operations have any operands in common.
6280 Value *MatchOp, *OtherOpT, *OtherOpF;
6282 if (TI->getOperand(0) == FI->getOperand(0)) {
6283 MatchOp = TI->getOperand(0);
6284 OtherOpT = TI->getOperand(1);
6285 OtherOpF = FI->getOperand(1);
6286 MatchIsOpZero = true;
6287 } else if (TI->getOperand(1) == FI->getOperand(1)) {
6288 MatchOp = TI->getOperand(1);
6289 OtherOpT = TI->getOperand(0);
6290 OtherOpF = FI->getOperand(0);
6291 MatchIsOpZero = false;
6292 } else if (!TI->isCommutative()) {
6294 } else if (TI->getOperand(0) == FI->getOperand(1)) {
6295 MatchOp = TI->getOperand(0);
6296 OtherOpT = TI->getOperand(1);
6297 OtherOpF = FI->getOperand(0);
6298 MatchIsOpZero = true;
6299 } else if (TI->getOperand(1) == FI->getOperand(0)) {
6300 MatchOp = TI->getOperand(1);
6301 OtherOpT = TI->getOperand(0);
6302 OtherOpF = FI->getOperand(1);
6303 MatchIsOpZero = true;
6308 // If we reach here, they do have operations in common.
6309 SelectInst *NewSI = new SelectInst(SI.getCondition(), OtherOpT,
6310 OtherOpF, SI.getName()+".v");
6311 InsertNewInstBefore(NewSI, SI);
6313 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TI)) {
6315 return BinaryOperator::create(BO->getOpcode(), MatchOp, NewSI);
6317 return BinaryOperator::create(BO->getOpcode(), NewSI, MatchOp);
6320 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), MatchOp, NewSI);
6322 return new ShiftInst(cast<ShiftInst>(TI)->getOpcode(), NewSI, MatchOp);
6326 Instruction *InstCombiner::visitSelectInst(SelectInst &SI) {
6327 Value *CondVal = SI.getCondition();
6328 Value *TrueVal = SI.getTrueValue();
6329 Value *FalseVal = SI.getFalseValue();
6331 // select true, X, Y -> X
6332 // select false, X, Y -> Y
6333 if (ConstantBool *C = dyn_cast<ConstantBool>(CondVal))
6334 return ReplaceInstUsesWith(SI, C->getValue() ? TrueVal : FalseVal);
6336 // select C, X, X -> X
6337 if (TrueVal == FalseVal)
6338 return ReplaceInstUsesWith(SI, TrueVal);
6340 if (isa<UndefValue>(TrueVal)) // select C, undef, X -> X
6341 return ReplaceInstUsesWith(SI, FalseVal);
6342 if (isa<UndefValue>(FalseVal)) // select C, X, undef -> X
6343 return ReplaceInstUsesWith(SI, TrueVal);
6344 if (isa<UndefValue>(CondVal)) { // select undef, X, Y -> X or Y
6345 if (isa<Constant>(TrueVal))
6346 return ReplaceInstUsesWith(SI, TrueVal);
6348 return ReplaceInstUsesWith(SI, FalseVal);
6351 if (SI.getType() == Type::BoolTy)
6352 if (ConstantBool *C = dyn_cast<ConstantBool>(TrueVal)) {
6353 if (C->getValue()) {
6354 // Change: A = select B, true, C --> A = or B, C
6355 return BinaryOperator::createOr(CondVal, FalseVal);
6357 // Change: A = select B, false, C --> A = and !B, C
6359 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6360 "not."+CondVal->getName()), SI);
6361 return BinaryOperator::createAnd(NotCond, FalseVal);
6363 } else if (ConstantBool *C = dyn_cast<ConstantBool>(FalseVal)) {
6364 if (C->getValue() == false) {
6365 // Change: A = select B, C, false --> A = and B, C
6366 return BinaryOperator::createAnd(CondVal, TrueVal);
6368 // Change: A = select B, C, true --> A = or !B, C
6370 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6371 "not."+CondVal->getName()), SI);
6372 return BinaryOperator::createOr(NotCond, TrueVal);
6376 // Selecting between two integer constants?
6377 if (ConstantInt *TrueValC = dyn_cast<ConstantInt>(TrueVal))
6378 if (ConstantInt *FalseValC = dyn_cast<ConstantInt>(FalseVal)) {
6379 // select C, 1, 0 -> cast C to int
6380 if (FalseValC->isNullValue() && TrueValC->getZExtValue() == 1) {
6381 return CastInst::create(Instruction::ZExt, CondVal, SI.getType());
6382 } else if (TrueValC->isNullValue() && FalseValC->getZExtValue() == 1) {
6383 // select C, 0, 1 -> cast !C to int
6385 InsertNewInstBefore(BinaryOperator::createNot(CondVal,
6386 "not."+CondVal->getName()), SI);
6387 return CastInst::create(Instruction::ZExt, NotCond, SI.getType());
6390 if (SetCondInst *IC = dyn_cast<SetCondInst>(SI.getCondition())) {
6392 // (x <s 0) ? -1 : 0 -> sra x, 31
6393 // (x >u 2147483647) ? -1 : 0 -> sra x, 31
6394 if (TrueValC->isAllOnesValue() && FalseValC->isNullValue())
6395 if (ConstantInt *CmpCst = dyn_cast<ConstantInt>(IC->getOperand(1))) {
6396 bool CanXForm = false;
6397 if (CmpCst->getType()->isSigned())
6398 CanXForm = CmpCst->isNullValue() &&
6399 IC->getOpcode() == Instruction::SetLT;
6401 unsigned Bits = CmpCst->getType()->getPrimitiveSizeInBits();
6402 CanXForm = (CmpCst->getZExtValue() == ~0ULL >> (64-Bits+1)) &&
6403 IC->getOpcode() == Instruction::SetGT;
6407 // The comparison constant and the result are not neccessarily the
6408 // same width. Make an all-ones value by inserting a AShr.
6409 Value *X = IC->getOperand(0);
6410 unsigned Bits = X->getType()->getPrimitiveSizeInBits();
6411 Constant *ShAmt = ConstantInt::get(Type::UByteTy, Bits-1);
6412 Instruction *SRA = new ShiftInst(Instruction::AShr, X,
6414 InsertNewInstBefore(SRA, SI);
6416 // Finally, convert to the type of the select RHS. We figure out
6417 // if this requires a SExt, Trunc or BitCast based on the sizes.
6418 Instruction::CastOps opc = Instruction::BitCast;
6419 unsigned SRASize = SRA->getType()->getPrimitiveSizeInBits();
6420 unsigned SISize = SI.getType()->getPrimitiveSizeInBits();
6421 if (SRASize < SISize)
6422 opc = Instruction::SExt;
6423 else if (SRASize > SISize)
6424 opc = Instruction::Trunc;
6425 return CastInst::create(opc, SRA, SI.getType());
6430 // If one of the constants is zero (we know they can't both be) and we
6431 // have a setcc instruction with zero, and we have an 'and' with the
6432 // non-constant value, eliminate this whole mess. This corresponds to
6433 // cases like this: ((X & 27) ? 27 : 0)
6434 if (TrueValC->isNullValue() || FalseValC->isNullValue())
6435 if (IC->isEquality() && isa<ConstantInt>(IC->getOperand(1)) &&
6436 cast<Constant>(IC->getOperand(1))->isNullValue())
6437 if (Instruction *ICA = dyn_cast<Instruction>(IC->getOperand(0)))
6438 if (ICA->getOpcode() == Instruction::And &&
6439 isa<ConstantInt>(ICA->getOperand(1)) &&
6440 (ICA->getOperand(1) == TrueValC ||
6441 ICA->getOperand(1) == FalseValC) &&
6442 isOneBitSet(cast<ConstantInt>(ICA->getOperand(1)))) {
6443 // Okay, now we know that everything is set up, we just don't
6444 // know whether we have a setne or seteq and whether the true or
6445 // false val is the zero.
6446 bool ShouldNotVal = !TrueValC->isNullValue();
6447 ShouldNotVal ^= IC->getOpcode() == Instruction::SetNE;
6450 V = InsertNewInstBefore(BinaryOperator::create(
6451 Instruction::Xor, V, ICA->getOperand(1)), SI);
6452 return ReplaceInstUsesWith(SI, V);
6457 // See if we are selecting two values based on a comparison of the two values.
6458 if (SetCondInst *SCI = dyn_cast<SetCondInst>(CondVal)) {
6459 if (SCI->getOperand(0) == TrueVal && SCI->getOperand(1) == FalseVal) {
6460 // Transform (X == Y) ? X : Y -> Y
6461 if (SCI->getOpcode() == Instruction::SetEQ)
6462 return ReplaceInstUsesWith(SI, FalseVal);
6463 // Transform (X != Y) ? X : Y -> X
6464 if (SCI->getOpcode() == Instruction::SetNE)
6465 return ReplaceInstUsesWith(SI, TrueVal);
6466 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6468 } else if (SCI->getOperand(0) == FalseVal && SCI->getOperand(1) == TrueVal){
6469 // Transform (X == Y) ? Y : X -> X
6470 if (SCI->getOpcode() == Instruction::SetEQ)
6471 return ReplaceInstUsesWith(SI, FalseVal);
6472 // Transform (X != Y) ? Y : X -> Y
6473 if (SCI->getOpcode() == Instruction::SetNE)
6474 return ReplaceInstUsesWith(SI, TrueVal);
6475 // NOTE: if we wanted to, this is where to detect MIN/MAX/ABS/etc.
6479 if (Instruction *TI = dyn_cast<Instruction>(TrueVal))
6480 if (Instruction *FI = dyn_cast<Instruction>(FalseVal))
6481 if (TI->hasOneUse() && FI->hasOneUse()) {
6482 Instruction *AddOp = 0, *SubOp = 0;
6484 // Turn (select C, (op X, Y), (op X, Z)) -> (op X, (select C, Y, Z))
6485 if (TI->getOpcode() == FI->getOpcode())
6486 if (Instruction *IV = FoldSelectOpOp(SI, TI, FI))
6489 // Turn select C, (X+Y), (X-Y) --> (X+(select C, Y, (-Y))). This is
6490 // even legal for FP.
6491 if (TI->getOpcode() == Instruction::Sub &&
6492 FI->getOpcode() == Instruction::Add) {
6493 AddOp = FI; SubOp = TI;
6494 } else if (FI->getOpcode() == Instruction::Sub &&
6495 TI->getOpcode() == Instruction::Add) {
6496 AddOp = TI; SubOp = FI;
6500 Value *OtherAddOp = 0;
6501 if (SubOp->getOperand(0) == AddOp->getOperand(0)) {
6502 OtherAddOp = AddOp->getOperand(1);
6503 } else if (SubOp->getOperand(0) == AddOp->getOperand(1)) {
6504 OtherAddOp = AddOp->getOperand(0);
6508 // So at this point we know we have (Y -> OtherAddOp):
6509 // select C, (add X, Y), (sub X, Z)
6510 Value *NegVal; // Compute -Z
6511 if (Constant *C = dyn_cast<Constant>(SubOp->getOperand(1))) {
6512 NegVal = ConstantExpr::getNeg(C);
6514 NegVal = InsertNewInstBefore(
6515 BinaryOperator::createNeg(SubOp->getOperand(1), "tmp"), SI);
6518 Value *NewTrueOp = OtherAddOp;
6519 Value *NewFalseOp = NegVal;
6521 std::swap(NewTrueOp, NewFalseOp);
6522 Instruction *NewSel =
6523 new SelectInst(CondVal, NewTrueOp,NewFalseOp,SI.getName()+".p");
6525 NewSel = InsertNewInstBefore(NewSel, SI);
6526 return BinaryOperator::createAdd(SubOp->getOperand(0), NewSel);
6531 // See if we can fold the select into one of our operands.
6532 if (SI.getType()->isInteger()) {
6533 // See the comment above GetSelectFoldableOperands for a description of the
6534 // transformation we are doing here.
6535 if (Instruction *TVI = dyn_cast<Instruction>(TrueVal))
6536 if (TVI->hasOneUse() && TVI->getNumOperands() == 2 &&
6537 !isa<Constant>(FalseVal))
6538 if (unsigned SFO = GetSelectFoldableOperands(TVI)) {
6539 unsigned OpToFold = 0;
6540 if ((SFO & 1) && FalseVal == TVI->getOperand(0)) {
6542 } else if ((SFO & 2) && FalseVal == TVI->getOperand(1)) {
6547 Constant *C = GetSelectFoldableConstant(TVI);
6548 std::string Name = TVI->getName(); TVI->setName("");
6549 Instruction *NewSel =
6550 new SelectInst(SI.getCondition(), TVI->getOperand(2-OpToFold), C,
6552 InsertNewInstBefore(NewSel, SI);
6553 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(TVI))
6554 return BinaryOperator::create(BO->getOpcode(), FalseVal, NewSel);
6555 else if (ShiftInst *SI = dyn_cast<ShiftInst>(TVI))
6556 return new ShiftInst(SI->getOpcode(), FalseVal, NewSel);
6558 assert(0 && "Unknown instruction!!");
6563 if (Instruction *FVI = dyn_cast<Instruction>(FalseVal))
6564 if (FVI->hasOneUse() && FVI->getNumOperands() == 2 &&
6565 !isa<Constant>(TrueVal))
6566 if (unsigned SFO = GetSelectFoldableOperands(FVI)) {
6567 unsigned OpToFold = 0;
6568 if ((SFO & 1) && TrueVal == FVI->getOperand(0)) {
6570 } else if ((SFO & 2) && TrueVal == FVI->getOperand(1)) {
6575 Constant *C = GetSelectFoldableConstant(FVI);
6576 std::string Name = FVI->getName(); FVI->setName("");
6577 Instruction *NewSel =
6578 new SelectInst(SI.getCondition(), C, FVI->getOperand(2-OpToFold),
6580 InsertNewInstBefore(NewSel, SI);
6581 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(FVI))
6582 return BinaryOperator::create(BO->getOpcode(), TrueVal, NewSel);
6583 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FVI))
6584 return new ShiftInst(SI->getOpcode(), TrueVal, NewSel);
6586 assert(0 && "Unknown instruction!!");
6592 if (BinaryOperator::isNot(CondVal)) {
6593 SI.setOperand(0, BinaryOperator::getNotArgument(CondVal));
6594 SI.setOperand(1, FalseVal);
6595 SI.setOperand(2, TrueVal);
6602 /// GetKnownAlignment - If the specified pointer has an alignment that we can
6603 /// determine, return it, otherwise return 0.
6604 static unsigned GetKnownAlignment(Value *V, TargetData *TD) {
6605 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(V)) {
6606 unsigned Align = GV->getAlignment();
6607 if (Align == 0 && TD)
6608 Align = TD->getTypeAlignment(GV->getType()->getElementType());
6610 } else if (AllocationInst *AI = dyn_cast<AllocationInst>(V)) {
6611 unsigned Align = AI->getAlignment();
6612 if (Align == 0 && TD) {
6613 if (isa<AllocaInst>(AI))
6614 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6615 else if (isa<MallocInst>(AI)) {
6616 // Malloc returns maximally aligned memory.
6617 Align = TD->getTypeAlignment(AI->getType()->getElementType());
6618 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::DoubleTy));
6619 Align = std::max(Align, (unsigned)TD->getTypeAlignment(Type::LongTy));
6623 } else if (isa<BitCastInst>(V) ||
6624 (isa<ConstantExpr>(V) &&
6625 cast<ConstantExpr>(V)->getOpcode() == Instruction::BitCast)) {
6626 User *CI = cast<User>(V);
6627 if (isa<PointerType>(CI->getOperand(0)->getType()))
6628 return GetKnownAlignment(CI->getOperand(0), TD);
6630 } else if (isa<GetElementPtrInst>(V) ||
6631 (isa<ConstantExpr>(V) &&
6632 cast<ConstantExpr>(V)->getOpcode()==Instruction::GetElementPtr)) {
6633 User *GEPI = cast<User>(V);
6634 unsigned BaseAlignment = GetKnownAlignment(GEPI->getOperand(0), TD);
6635 if (BaseAlignment == 0) return 0;
6637 // If all indexes are zero, it is just the alignment of the base pointer.
6638 bool AllZeroOperands = true;
6639 for (unsigned i = 1, e = GEPI->getNumOperands(); i != e; ++i)
6640 if (!isa<Constant>(GEPI->getOperand(i)) ||
6641 !cast<Constant>(GEPI->getOperand(i))->isNullValue()) {
6642 AllZeroOperands = false;
6645 if (AllZeroOperands)
6646 return BaseAlignment;
6648 // Otherwise, if the base alignment is >= the alignment we expect for the
6649 // base pointer type, then we know that the resultant pointer is aligned at
6650 // least as much as its type requires.
6653 const Type *BasePtrTy = GEPI->getOperand(0)->getType();
6654 if (TD->getTypeAlignment(cast<PointerType>(BasePtrTy)->getElementType())
6656 const Type *GEPTy = GEPI->getType();
6657 return TD->getTypeAlignment(cast<PointerType>(GEPTy)->getElementType());
6665 /// visitCallInst - CallInst simplification. This mostly only handles folding
6666 /// of intrinsic instructions. For normal calls, it allows visitCallSite to do
6667 /// the heavy lifting.
6669 Instruction *InstCombiner::visitCallInst(CallInst &CI) {
6670 IntrinsicInst *II = dyn_cast<IntrinsicInst>(&CI);
6671 if (!II) return visitCallSite(&CI);
6673 // Intrinsics cannot occur in an invoke, so handle them here instead of in
6675 if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(II)) {
6676 bool Changed = false;
6678 // memmove/cpy/set of zero bytes is a noop.
6679 if (Constant *NumBytes = dyn_cast<Constant>(MI->getLength())) {
6680 if (NumBytes->isNullValue()) return EraseInstFromFunction(CI);
6682 if (ConstantInt *CI = dyn_cast<ConstantInt>(NumBytes))
6683 if (CI->getZExtValue() == 1) {
6684 // Replace the instruction with just byte operations. We would
6685 // transform other cases to loads/stores, but we don't know if
6686 // alignment is sufficient.
6690 // If we have a memmove and the source operation is a constant global,
6691 // then the source and dest pointers can't alias, so we can change this
6692 // into a call to memcpy.
6693 if (MemMoveInst *MMI = dyn_cast<MemMoveInst>(II)) {
6694 if (GlobalVariable *GVSrc = dyn_cast<GlobalVariable>(MMI->getSource()))
6695 if (GVSrc->isConstant()) {
6696 Module *M = CI.getParent()->getParent()->getParent();
6698 if (CI.getCalledFunction()->getFunctionType()->getParamType(2) ==
6700 Name = "llvm.memcpy.i32";
6702 Name = "llvm.memcpy.i64";
6703 Function *MemCpy = M->getOrInsertFunction(Name,
6704 CI.getCalledFunction()->getFunctionType());
6705 CI.setOperand(0, MemCpy);
6710 // If we can determine a pointer alignment that is bigger than currently
6711 // set, update the alignment.
6712 if (isa<MemCpyInst>(MI) || isa<MemMoveInst>(MI)) {
6713 unsigned Alignment1 = GetKnownAlignment(MI->getOperand(1), TD);
6714 unsigned Alignment2 = GetKnownAlignment(MI->getOperand(2), TD);
6715 unsigned Align = std::min(Alignment1, Alignment2);
6716 if (MI->getAlignment()->getZExtValue() < Align) {
6717 MI->setAlignment(ConstantInt::get(Type::UIntTy, Align));
6720 } else if (isa<MemSetInst>(MI)) {
6721 unsigned Alignment = GetKnownAlignment(MI->getDest(), TD);
6722 if (MI->getAlignment()->getZExtValue() < Alignment) {
6723 MI->setAlignment(ConstantInt::get(Type::UIntTy, Alignment));
6728 if (Changed) return II;
6730 switch (II->getIntrinsicID()) {
6732 case Intrinsic::ppc_altivec_lvx:
6733 case Intrinsic::ppc_altivec_lvxl:
6734 case Intrinsic::x86_sse_loadu_ps:
6735 case Intrinsic::x86_sse2_loadu_pd:
6736 case Intrinsic::x86_sse2_loadu_dq:
6737 // Turn PPC lvx -> load if the pointer is known aligned.
6738 // Turn X86 loadups -> load if the pointer is known aligned.
6739 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6740 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
6741 PointerType::get(II->getType()), CI);
6742 return new LoadInst(Ptr);
6745 case Intrinsic::ppc_altivec_stvx:
6746 case Intrinsic::ppc_altivec_stvxl:
6747 // Turn stvx -> store if the pointer is known aligned.
6748 if (GetKnownAlignment(II->getOperand(2), TD) >= 16) {
6749 const Type *OpPtrTy = PointerType::get(II->getOperand(1)->getType());
6750 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(2),
6752 return new StoreInst(II->getOperand(1), Ptr);
6755 case Intrinsic::x86_sse_storeu_ps:
6756 case Intrinsic::x86_sse2_storeu_pd:
6757 case Intrinsic::x86_sse2_storeu_dq:
6758 case Intrinsic::x86_sse2_storel_dq:
6759 // Turn X86 storeu -> store if the pointer is known aligned.
6760 if (GetKnownAlignment(II->getOperand(1), TD) >= 16) {
6761 const Type *OpPtrTy = PointerType::get(II->getOperand(2)->getType());
6762 Value *Ptr = InsertCastBefore(Instruction::BitCast, II->getOperand(1),
6764 return new StoreInst(II->getOperand(2), Ptr);
6768 case Intrinsic::x86_sse_cvttss2si: {
6769 // These intrinsics only demands the 0th element of its input vector. If
6770 // we can simplify the input based on that, do so now.
6772 if (Value *V = SimplifyDemandedVectorElts(II->getOperand(1), 1,
6774 II->setOperand(1, V);
6780 case Intrinsic::ppc_altivec_vperm:
6781 // Turn vperm(V1,V2,mask) -> shuffle(V1,V2,mask) if mask is a constant.
6782 if (ConstantPacked *Mask = dyn_cast<ConstantPacked>(II->getOperand(3))) {
6783 assert(Mask->getNumOperands() == 16 && "Bad type for intrinsic!");
6785 // Check that all of the elements are integer constants or undefs.
6786 bool AllEltsOk = true;
6787 for (unsigned i = 0; i != 16; ++i) {
6788 if (!isa<ConstantInt>(Mask->getOperand(i)) &&
6789 !isa<UndefValue>(Mask->getOperand(i))) {
6796 // Cast the input vectors to byte vectors.
6797 Value *Op0 = InsertCastBefore(Instruction::BitCast,
6798 II->getOperand(1), Mask->getType(), CI);
6799 Value *Op1 = InsertCastBefore(Instruction::BitCast,
6800 II->getOperand(2), Mask->getType(), CI);
6801 Value *Result = UndefValue::get(Op0->getType());
6803 // Only extract each element once.
6804 Value *ExtractedElts[32];
6805 memset(ExtractedElts, 0, sizeof(ExtractedElts));
6807 for (unsigned i = 0; i != 16; ++i) {
6808 if (isa<UndefValue>(Mask->getOperand(i)))
6810 unsigned Idx =cast<ConstantInt>(Mask->getOperand(i))->getZExtValue();
6811 Idx &= 31; // Match the hardware behavior.
6813 if (ExtractedElts[Idx] == 0) {
6815 new ExtractElementInst(Idx < 16 ? Op0 : Op1, Idx&15, "tmp");
6816 InsertNewInstBefore(Elt, CI);
6817 ExtractedElts[Idx] = Elt;
6820 // Insert this value into the result vector.
6821 Result = new InsertElementInst(Result, ExtractedElts[Idx], i,"tmp");
6822 InsertNewInstBefore(cast<Instruction>(Result), CI);
6824 return CastInst::create(Instruction::BitCast, Result, CI.getType());
6829 case Intrinsic::stackrestore: {
6830 // If the save is right next to the restore, remove the restore. This can
6831 // happen when variable allocas are DCE'd.
6832 if (IntrinsicInst *SS = dyn_cast<IntrinsicInst>(II->getOperand(1))) {
6833 if (SS->getIntrinsicID() == Intrinsic::stacksave) {
6834 BasicBlock::iterator BI = SS;
6836 return EraseInstFromFunction(CI);
6840 // If the stack restore is in a return/unwind block and if there are no
6841 // allocas or calls between the restore and the return, nuke the restore.
6842 TerminatorInst *TI = II->getParent()->getTerminator();
6843 if (isa<ReturnInst>(TI) || isa<UnwindInst>(TI)) {
6844 BasicBlock::iterator BI = II;
6845 bool CannotRemove = false;
6846 for (++BI; &*BI != TI; ++BI) {
6847 if (isa<AllocaInst>(BI) ||
6848 (isa<CallInst>(BI) && !isa<IntrinsicInst>(BI))) {
6849 CannotRemove = true;
6854 return EraseInstFromFunction(CI);
6861 return visitCallSite(II);
6864 // InvokeInst simplification
6866 Instruction *InstCombiner::visitInvokeInst(InvokeInst &II) {
6867 return visitCallSite(&II);
6870 // visitCallSite - Improvements for call and invoke instructions.
6872 Instruction *InstCombiner::visitCallSite(CallSite CS) {
6873 bool Changed = false;
6875 // If the callee is a constexpr cast of a function, attempt to move the cast
6876 // to the arguments of the call/invoke.
6877 if (transformConstExprCastCall(CS)) return 0;
6879 Value *Callee = CS.getCalledValue();
6881 if (Function *CalleeF = dyn_cast<Function>(Callee))
6882 if (CalleeF->getCallingConv() != CS.getCallingConv()) {
6883 Instruction *OldCall = CS.getInstruction();
6884 // If the call and callee calling conventions don't match, this call must
6885 // be unreachable, as the call is undefined.
6886 new StoreInst(ConstantBool::getTrue(),
6887 UndefValue::get(PointerType::get(Type::BoolTy)), OldCall);
6888 if (!OldCall->use_empty())
6889 OldCall->replaceAllUsesWith(UndefValue::get(OldCall->getType()));
6890 if (isa<CallInst>(OldCall)) // Not worth removing an invoke here.
6891 return EraseInstFromFunction(*OldCall);
6895 if (isa<ConstantPointerNull>(Callee) || isa<UndefValue>(Callee)) {
6896 // This instruction is not reachable, just remove it. We insert a store to
6897 // undef so that we know that this code is not reachable, despite the fact
6898 // that we can't modify the CFG here.
6899 new StoreInst(ConstantBool::getTrue(),
6900 UndefValue::get(PointerType::get(Type::BoolTy)),
6901 CS.getInstruction());
6903 if (!CS.getInstruction()->use_empty())
6904 CS.getInstruction()->
6905 replaceAllUsesWith(UndefValue::get(CS.getInstruction()->getType()));
6907 if (InvokeInst *II = dyn_cast<InvokeInst>(CS.getInstruction())) {
6908 // Don't break the CFG, insert a dummy cond branch.
6909 new BranchInst(II->getNormalDest(), II->getUnwindDest(),
6910 ConstantBool::getTrue(), II);
6912 return EraseInstFromFunction(*CS.getInstruction());
6915 const PointerType *PTy = cast<PointerType>(Callee->getType());
6916 const FunctionType *FTy = cast<FunctionType>(PTy->getElementType());
6917 if (FTy->isVarArg()) {
6918 // See if we can optimize any arguments passed through the varargs area of
6920 for (CallSite::arg_iterator I = CS.arg_begin()+FTy->getNumParams(),
6921 E = CS.arg_end(); I != E; ++I)
6922 if (CastInst *CI = dyn_cast<CastInst>(*I)) {
6923 // If this cast does not effect the value passed through the varargs
6924 // area, we can eliminate the use of the cast.
6925 Value *Op = CI->getOperand(0);
6926 if (CI->isLosslessCast()) {
6933 return Changed ? CS.getInstruction() : 0;
6936 // transformConstExprCastCall - If the callee is a constexpr cast of a function,
6937 // attempt to move the cast to the arguments of the call/invoke.
6939 bool InstCombiner::transformConstExprCastCall(CallSite CS) {
6940 if (!isa<ConstantExpr>(CS.getCalledValue())) return false;
6941 ConstantExpr *CE = cast<ConstantExpr>(CS.getCalledValue());
6942 if (CE->getOpcode() != Instruction::BitCast ||
6943 !isa<Function>(CE->getOperand(0)))
6945 Function *Callee = cast<Function>(CE->getOperand(0));
6946 Instruction *Caller = CS.getInstruction();
6948 // Okay, this is a cast from a function to a different type. Unless doing so
6949 // would cause a type conversion of one of our arguments, change this call to
6950 // be a direct call with arguments casted to the appropriate types.
6952 const FunctionType *FT = Callee->getFunctionType();
6953 const Type *OldRetTy = Caller->getType();
6955 // Check to see if we are changing the return type...
6956 if (OldRetTy != FT->getReturnType()) {
6957 if (Callee->isExternal() &&
6958 !Caller->use_empty() &&
6959 !(OldRetTy->canLosslesslyBitCastTo(FT->getReturnType()) ||
6960 (isa<PointerType>(FT->getReturnType()) &&
6961 TD->getIntPtrType()->canLosslesslyBitCastTo(OldRetTy)))
6963 return false; // Cannot transform this return value...
6965 // If the callsite is an invoke instruction, and the return value is used by
6966 // a PHI node in a successor, we cannot change the return type of the call
6967 // because there is no place to put the cast instruction (without breaking
6968 // the critical edge). Bail out in this case.
6969 if (!Caller->use_empty())
6970 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller))
6971 for (Value::use_iterator UI = II->use_begin(), E = II->use_end();
6973 if (PHINode *PN = dyn_cast<PHINode>(*UI))
6974 if (PN->getParent() == II->getNormalDest() ||
6975 PN->getParent() == II->getUnwindDest())
6979 unsigned NumActualArgs = unsigned(CS.arg_end()-CS.arg_begin());
6980 unsigned NumCommonArgs = std::min(FT->getNumParams(), NumActualArgs);
6982 CallSite::arg_iterator AI = CS.arg_begin();
6983 for (unsigned i = 0, e = NumCommonArgs; i != e; ++i, ++AI) {
6984 const Type *ParamTy = FT->getParamType(i);
6985 const Type *ActTy = (*AI)->getType();
6986 ConstantInt *c = dyn_cast<ConstantInt>(*AI);
6987 //Either we can cast directly, or we can upconvert the argument
6988 bool isConvertible = ActTy->canLosslesslyBitCastTo(ParamTy) ||
6989 (ParamTy->isIntegral() && ActTy->isIntegral() &&
6990 ParamTy->isSigned() == ActTy->isSigned() &&
6991 ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize()) ||
6992 (c && ParamTy->getPrimitiveSize() >= ActTy->getPrimitiveSize() &&
6993 c->getSExtValue() > 0);
6994 if (Callee->isExternal() && !isConvertible) return false;
6997 if (FT->getNumParams() < NumActualArgs && !FT->isVarArg() &&
6998 Callee->isExternal())
6999 return false; // Do not delete arguments unless we have a function body...
7001 // Okay, we decided that this is a safe thing to do: go ahead and start
7002 // inserting cast instructions as necessary...
7003 std::vector<Value*> Args;
7004 Args.reserve(NumActualArgs);
7006 AI = CS.arg_begin();
7007 for (unsigned i = 0; i != NumCommonArgs; ++i, ++AI) {
7008 const Type *ParamTy = FT->getParamType(i);
7009 if ((*AI)->getType() == ParamTy) {
7010 Args.push_back(*AI);
7012 Instruction::CastOps opcode = CastInst::getCastOpcode(*AI,
7013 (*AI)->getType()->isSigned(), ParamTy, ParamTy->isSigned());
7014 CastInst *NewCast = CastInst::create(opcode, *AI, ParamTy, "tmp");
7015 Args.push_back(InsertNewInstBefore(NewCast, *Caller));
7019 // If the function takes more arguments than the call was taking, add them
7021 for (unsigned i = NumCommonArgs; i != FT->getNumParams(); ++i)
7022 Args.push_back(Constant::getNullValue(FT->getParamType(i)));
7024 // If we are removing arguments to the function, emit an obnoxious warning...
7025 if (FT->getNumParams() < NumActualArgs)
7026 if (!FT->isVarArg()) {
7027 cerr << "WARNING: While resolving call to function '"
7028 << Callee->getName() << "' arguments were dropped!\n";
7030 // Add all of the arguments in their promoted form to the arg list...
7031 for (unsigned i = FT->getNumParams(); i != NumActualArgs; ++i, ++AI) {
7032 const Type *PTy = getPromotedType((*AI)->getType());
7033 if (PTy != (*AI)->getType()) {
7034 // Must promote to pass through va_arg area!
7035 Instruction::CastOps opcode = CastInst::getCastOpcode(
7036 *AI, (*AI)->getType()->isSigned(), PTy, PTy->isSigned());
7037 Instruction *Cast = CastInst::create(opcode, *AI, PTy, "tmp");
7038 InsertNewInstBefore(Cast, *Caller);
7039 Args.push_back(Cast);
7041 Args.push_back(*AI);
7046 if (FT->getReturnType() == Type::VoidTy)
7047 Caller->setName(""); // Void type should not have a name...
7050 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7051 NC = new InvokeInst(Callee, II->getNormalDest(), II->getUnwindDest(),
7052 Args, Caller->getName(), Caller);
7053 cast<InvokeInst>(II)->setCallingConv(II->getCallingConv());
7055 NC = new CallInst(Callee, Args, Caller->getName(), Caller);
7056 if (cast<CallInst>(Caller)->isTailCall())
7057 cast<CallInst>(NC)->setTailCall();
7058 cast<CallInst>(NC)->setCallingConv(cast<CallInst>(Caller)->getCallingConv());
7061 // Insert a cast of the return type as necessary...
7063 if (Caller->getType() != NV->getType() && !Caller->use_empty()) {
7064 if (NV->getType() != Type::VoidTy) {
7065 const Type *CallerTy = Caller->getType();
7066 Instruction::CastOps opcode = CastInst::getCastOpcode(
7067 NC, NC->getType()->isSigned(), CallerTy, CallerTy->isSigned());
7068 NV = NC = CastInst::create(opcode, NC, CallerTy, "tmp");
7070 // If this is an invoke instruction, we should insert it after the first
7071 // non-phi, instruction in the normal successor block.
7072 if (InvokeInst *II = dyn_cast<InvokeInst>(Caller)) {
7073 BasicBlock::iterator I = II->getNormalDest()->begin();
7074 while (isa<PHINode>(I)) ++I;
7075 InsertNewInstBefore(NC, *I);
7077 // Otherwise, it's a call, just insert cast right after the call instr
7078 InsertNewInstBefore(NC, *Caller);
7080 AddUsersToWorkList(*Caller);
7082 NV = UndefValue::get(Caller->getType());
7086 if (Caller->getType() != Type::VoidTy && !Caller->use_empty())
7087 Caller->replaceAllUsesWith(NV);
7088 Caller->getParent()->getInstList().erase(Caller);
7089 removeFromWorkList(Caller);
7093 /// FoldPHIArgBinOpIntoPHI - If we have something like phi [add (a,b), add(c,d)]
7094 /// and if a/b/c/d and the add's all have a single use, turn this into two phi's
7095 /// and a single binop.
7096 Instruction *InstCombiner::FoldPHIArgBinOpIntoPHI(PHINode &PN) {
7097 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7098 assert(isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst) ||
7099 isa<GetElementPtrInst>(FirstInst));
7100 unsigned Opc = FirstInst->getOpcode();
7101 Value *LHSVal = FirstInst->getOperand(0);
7102 Value *RHSVal = FirstInst->getOperand(1);
7104 const Type *LHSType = LHSVal->getType();
7105 const Type *RHSType = RHSVal->getType();
7107 // Scan to see if all operands are the same opcode, all have one use, and all
7108 // kill their operands (i.e. the operands have one use).
7109 for (unsigned i = 0; i != PN.getNumIncomingValues(); ++i) {
7110 Instruction *I = dyn_cast<Instruction>(PN.getIncomingValue(i));
7111 if (!I || I->getOpcode() != Opc || !I->hasOneUse() ||
7112 // Verify type of the LHS matches so we don't fold setcc's of different
7113 // types or GEP's with different index types.
7114 I->getOperand(0)->getType() != LHSType ||
7115 I->getOperand(1)->getType() != RHSType)
7118 // Keep track of which operand needs a phi node.
7119 if (I->getOperand(0) != LHSVal) LHSVal = 0;
7120 if (I->getOperand(1) != RHSVal) RHSVal = 0;
7123 // Otherwise, this is safe to transform, determine if it is profitable.
7125 // If this is a GEP, and if the index (not the pointer) needs a PHI, bail out.
7126 // Indexes are often folded into load/store instructions, so we don't want to
7127 // hide them behind a phi.
7128 if (isa<GetElementPtrInst>(FirstInst) && RHSVal == 0)
7131 Value *InLHS = FirstInst->getOperand(0);
7132 Value *InRHS = FirstInst->getOperand(1);
7133 PHINode *NewLHS = 0, *NewRHS = 0;
7135 NewLHS = new PHINode(LHSType, FirstInst->getOperand(0)->getName()+".pn");
7136 NewLHS->reserveOperandSpace(PN.getNumOperands()/2);
7137 NewLHS->addIncoming(InLHS, PN.getIncomingBlock(0));
7138 InsertNewInstBefore(NewLHS, PN);
7143 NewRHS = new PHINode(RHSType, FirstInst->getOperand(1)->getName()+".pn");
7144 NewRHS->reserveOperandSpace(PN.getNumOperands()/2);
7145 NewRHS->addIncoming(InRHS, PN.getIncomingBlock(0));
7146 InsertNewInstBefore(NewRHS, PN);
7150 // Add all operands to the new PHIs.
7151 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7153 Value *NewInLHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7154 NewLHS->addIncoming(NewInLHS, PN.getIncomingBlock(i));
7157 Value *NewInRHS =cast<Instruction>(PN.getIncomingValue(i))->getOperand(1);
7158 NewRHS->addIncoming(NewInRHS, PN.getIncomingBlock(i));
7162 if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7163 return BinaryOperator::create(BinOp->getOpcode(), LHSVal, RHSVal);
7164 else if (ShiftInst *SI = dyn_cast<ShiftInst>(FirstInst))
7165 return new ShiftInst(SI->getOpcode(), LHSVal, RHSVal);
7167 assert(isa<GetElementPtrInst>(FirstInst));
7168 return new GetElementPtrInst(LHSVal, RHSVal);
7172 /// isSafeToSinkLoad - Return true if we know that it is safe sink the load out
7173 /// of the block that defines it. This means that it must be obvious the value
7174 /// of the load is not changed from the point of the load to the end of the
7176 static bool isSafeToSinkLoad(LoadInst *L) {
7177 BasicBlock::iterator BBI = L, E = L->getParent()->end();
7179 for (++BBI; BBI != E; ++BBI)
7180 if (BBI->mayWriteToMemory())
7186 // FoldPHIArgOpIntoPHI - If all operands to a PHI node are the same "unary"
7187 // operator and they all are only used by the PHI, PHI together their
7188 // inputs, and do the operation once, to the result of the PHI.
7189 Instruction *InstCombiner::FoldPHIArgOpIntoPHI(PHINode &PN) {
7190 Instruction *FirstInst = cast<Instruction>(PN.getIncomingValue(0));
7192 // Scan the instruction, looking for input operations that can be folded away.
7193 // If all input operands to the phi are the same instruction (e.g. a cast from
7194 // the same type or "+42") we can pull the operation through the PHI, reducing
7195 // code size and simplifying code.
7196 Constant *ConstantOp = 0;
7197 const Type *CastSrcTy = 0;
7198 bool isVolatile = false;
7199 if (isa<CastInst>(FirstInst)) {
7200 CastSrcTy = FirstInst->getOperand(0)->getType();
7201 } else if (isa<BinaryOperator>(FirstInst) || isa<ShiftInst>(FirstInst)) {
7202 // Can fold binop or shift here if the RHS is a constant, otherwise call
7203 // FoldPHIArgBinOpIntoPHI.
7204 ConstantOp = dyn_cast<Constant>(FirstInst->getOperand(1));
7205 if (ConstantOp == 0)
7206 return FoldPHIArgBinOpIntoPHI(PN);
7207 } else if (LoadInst *LI = dyn_cast<LoadInst>(FirstInst)) {
7208 isVolatile = LI->isVolatile();
7209 // We can't sink the load if the loaded value could be modified between the
7210 // load and the PHI.
7211 if (LI->getParent() != PN.getIncomingBlock(0) ||
7212 !isSafeToSinkLoad(LI))
7214 } else if (isa<GetElementPtrInst>(FirstInst)) {
7215 if (FirstInst->getNumOperands() == 2)
7216 return FoldPHIArgBinOpIntoPHI(PN);
7217 // Can't handle general GEPs yet.
7220 return 0; // Cannot fold this operation.
7223 // Check to see if all arguments are the same operation.
7224 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7225 if (!isa<Instruction>(PN.getIncomingValue(i))) return 0;
7226 Instruction *I = cast<Instruction>(PN.getIncomingValue(i));
7227 if (!I->hasOneUse() || I->getOpcode() != FirstInst->getOpcode())
7230 if (I->getOperand(0)->getType() != CastSrcTy)
7231 return 0; // Cast operation must match.
7232 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
7233 // We can't sink the load if the loaded value could be modified between the
7234 // load and the PHI.
7235 if (LI->isVolatile() != isVolatile ||
7236 LI->getParent() != PN.getIncomingBlock(i) ||
7237 !isSafeToSinkLoad(LI))
7239 } else if (I->getOperand(1) != ConstantOp) {
7244 // Okay, they are all the same operation. Create a new PHI node of the
7245 // correct type, and PHI together all of the LHS's of the instructions.
7246 PHINode *NewPN = new PHINode(FirstInst->getOperand(0)->getType(),
7247 PN.getName()+".in");
7248 NewPN->reserveOperandSpace(PN.getNumOperands()/2);
7250 Value *InVal = FirstInst->getOperand(0);
7251 NewPN->addIncoming(InVal, PN.getIncomingBlock(0));
7253 // Add all operands to the new PHI.
7254 for (unsigned i = 1, e = PN.getNumIncomingValues(); i != e; ++i) {
7255 Value *NewInVal = cast<Instruction>(PN.getIncomingValue(i))->getOperand(0);
7256 if (NewInVal != InVal)
7258 NewPN->addIncoming(NewInVal, PN.getIncomingBlock(i));
7263 // The new PHI unions all of the same values together. This is really
7264 // common, so we handle it intelligently here for compile-time speed.
7268 InsertNewInstBefore(NewPN, PN);
7272 // Insert and return the new operation.
7273 if (CastInst* FirstCI = dyn_cast<CastInst>(FirstInst))
7274 return CastInst::create(FirstCI->getOpcode(), PhiVal, PN.getType());
7275 else if (isa<LoadInst>(FirstInst))
7276 return new LoadInst(PhiVal, "", isVolatile);
7277 else if (BinaryOperator *BinOp = dyn_cast<BinaryOperator>(FirstInst))
7278 return BinaryOperator::create(BinOp->getOpcode(), PhiVal, ConstantOp);
7280 return new ShiftInst(cast<ShiftInst>(FirstInst)->getOpcode(),
7281 PhiVal, ConstantOp);
7284 /// DeadPHICycle - Return true if this PHI node is only used by a PHI node cycle
7286 static bool DeadPHICycle(PHINode *PN, std::set<PHINode*> &PotentiallyDeadPHIs) {
7287 if (PN->use_empty()) return true;
7288 if (!PN->hasOneUse()) return false;
7290 // Remember this node, and if we find the cycle, return.
7291 if (!PotentiallyDeadPHIs.insert(PN).second)
7294 if (PHINode *PU = dyn_cast<PHINode>(PN->use_back()))
7295 return DeadPHICycle(PU, PotentiallyDeadPHIs);
7300 // PHINode simplification
7302 Instruction *InstCombiner::visitPHINode(PHINode &PN) {
7303 // If LCSSA is around, don't mess with Phi nodes
7304 if (mustPreserveAnalysisID(LCSSAID)) return 0;
7306 if (Value *V = PN.hasConstantValue())
7307 return ReplaceInstUsesWith(PN, V);
7309 // If all PHI operands are the same operation, pull them through the PHI,
7310 // reducing code size.
7311 if (isa<Instruction>(PN.getIncomingValue(0)) &&
7312 PN.getIncomingValue(0)->hasOneUse())
7313 if (Instruction *Result = FoldPHIArgOpIntoPHI(PN))
7316 // If this is a trivial cycle in the PHI node graph, remove it. Basically, if
7317 // this PHI only has a single use (a PHI), and if that PHI only has one use (a
7318 // PHI)... break the cycle.
7320 if (PHINode *PU = dyn_cast<PHINode>(PN.use_back())) {
7321 std::set<PHINode*> PotentiallyDeadPHIs;
7322 PotentiallyDeadPHIs.insert(&PN);
7323 if (DeadPHICycle(PU, PotentiallyDeadPHIs))
7324 return ReplaceInstUsesWith(PN, UndefValue::get(PN.getType()));
7330 static Value *InsertCastToIntPtrTy(Value *V, const Type *DTy,
7331 Instruction *InsertPoint,
7333 unsigned PtrSize = DTy->getPrimitiveSize();
7334 unsigned VTySize = V->getType()->getPrimitiveSize();
7335 // We must cast correctly to the pointer type. Ensure that we
7336 // sign extend the integer value if it is smaller as this is
7337 // used for address computation.
7338 Instruction::CastOps opcode =
7339 (VTySize < PtrSize ? Instruction::SExt :
7340 (VTySize == PtrSize ? Instruction::BitCast : Instruction::Trunc));
7341 return IC->InsertCastBefore(opcode, V, DTy, *InsertPoint);
7345 Instruction *InstCombiner::visitGetElementPtrInst(GetElementPtrInst &GEP) {
7346 Value *PtrOp = GEP.getOperand(0);
7347 // Is it 'getelementptr %P, long 0' or 'getelementptr %P'
7348 // If so, eliminate the noop.
7349 if (GEP.getNumOperands() == 1)
7350 return ReplaceInstUsesWith(GEP, PtrOp);
7352 if (isa<UndefValue>(GEP.getOperand(0)))
7353 return ReplaceInstUsesWith(GEP, UndefValue::get(GEP.getType()));
7355 bool HasZeroPointerIndex = false;
7356 if (Constant *C = dyn_cast<Constant>(GEP.getOperand(1)))
7357 HasZeroPointerIndex = C->isNullValue();
7359 if (GEP.getNumOperands() == 2 && HasZeroPointerIndex)
7360 return ReplaceInstUsesWith(GEP, PtrOp);
7362 // Eliminate unneeded casts for indices.
7363 bool MadeChange = false;
7364 gep_type_iterator GTI = gep_type_begin(GEP);
7365 for (unsigned i = 1, e = GEP.getNumOperands(); i != e; ++i, ++GTI)
7366 if (isa<SequentialType>(*GTI)) {
7367 if (CastInst *CI = dyn_cast<CastInst>(GEP.getOperand(i))) {
7368 Value *Src = CI->getOperand(0);
7369 const Type *SrcTy = Src->getType();
7370 const Type *DestTy = CI->getType();
7371 if (Src->getType()->isInteger()) {
7372 if (SrcTy->getPrimitiveSizeInBits() ==
7373 DestTy->getPrimitiveSizeInBits()) {
7374 // We can always eliminate a cast from ulong or long to the other.
7375 // We can always eliminate a cast from uint to int or the other on
7376 // 32-bit pointer platforms.
7377 if (DestTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()){
7379 GEP.setOperand(i, Src);
7381 } else if (SrcTy->getPrimitiveSize() < DestTy->getPrimitiveSize() &&
7382 SrcTy->getPrimitiveSize() == 4) {
7383 // We can always eliminate a cast from int to [u]long. We can
7384 // eliminate a cast from uint to [u]long iff the target is a 32-bit
7386 if (SrcTy->isSigned() ||
7387 SrcTy->getPrimitiveSizeInBits() >= TD->getPointerSizeInBits()) {
7389 GEP.setOperand(i, Src);
7394 // If we are using a wider index than needed for this platform, shrink it
7395 // to what we need. If the incoming value needs a cast instruction,
7396 // insert it. This explicit cast can make subsequent optimizations more
7398 Value *Op = GEP.getOperand(i);
7399 if (Op->getType()->getPrimitiveSize() > TD->getPointerSize())
7400 if (Constant *C = dyn_cast<Constant>(Op)) {
7401 GEP.setOperand(i, ConstantExpr::getTrunc(C,
7402 TD->getIntPtrType()->getSignedVersion()));
7405 Op = InsertCastBefore(Instruction::Trunc, Op, TD->getIntPtrType(),
7407 GEP.setOperand(i, Op);
7411 // If this is a constant idx, make sure to canonicalize it to be a signed
7412 // operand, otherwise CSE and other optimizations are pessimized.
7413 if (ConstantInt *CUI = dyn_cast<ConstantInt>(Op))
7414 if (CUI->getType()->isUnsigned()) {
7416 ConstantExpr::getBitCast(CUI, CUI->getType()->getSignedVersion()));
7420 if (MadeChange) return &GEP;
7422 // Combine Indices - If the source pointer to this getelementptr instruction
7423 // is a getelementptr instruction, combine the indices of the two
7424 // getelementptr instructions into a single instruction.
7426 std::vector<Value*> SrcGEPOperands;
7427 if (User *Src = dyn_castGetElementPtr(PtrOp))
7428 SrcGEPOperands.assign(Src->op_begin(), Src->op_end());
7430 if (!SrcGEPOperands.empty()) {
7431 // Note that if our source is a gep chain itself that we wait for that
7432 // chain to be resolved before we perform this transformation. This
7433 // avoids us creating a TON of code in some cases.
7435 if (isa<GetElementPtrInst>(SrcGEPOperands[0]) &&
7436 cast<Instruction>(SrcGEPOperands[0])->getNumOperands() == 2)
7437 return 0; // Wait until our source is folded to completion.
7439 std::vector<Value *> Indices;
7441 // Find out whether the last index in the source GEP is a sequential idx.
7442 bool EndsWithSequential = false;
7443 for (gep_type_iterator I = gep_type_begin(*cast<User>(PtrOp)),
7444 E = gep_type_end(*cast<User>(PtrOp)); I != E; ++I)
7445 EndsWithSequential = !isa<StructType>(*I);
7447 // Can we combine the two pointer arithmetics offsets?
7448 if (EndsWithSequential) {
7449 // Replace: gep (gep %P, long B), long A, ...
7450 // With: T = long A+B; gep %P, T, ...
7452 Value *Sum, *SO1 = SrcGEPOperands.back(), *GO1 = GEP.getOperand(1);
7453 if (SO1 == Constant::getNullValue(SO1->getType())) {
7455 } else if (GO1 == Constant::getNullValue(GO1->getType())) {
7458 // If they aren't the same type, convert both to an integer of the
7459 // target's pointer size.
7460 if (SO1->getType() != GO1->getType()) {
7461 if (Constant *SO1C = dyn_cast<Constant>(SO1)) {
7462 SO1 = ConstantExpr::getIntegerCast(SO1C, GO1->getType(), true);
7463 } else if (Constant *GO1C = dyn_cast<Constant>(GO1)) {
7464 GO1 = ConstantExpr::getIntegerCast(GO1C, SO1->getType(), true);
7466 unsigned PS = TD->getPointerSize();
7467 if (SO1->getType()->getPrimitiveSize() == PS) {
7468 // Convert GO1 to SO1's type.
7469 GO1 = InsertCastToIntPtrTy(GO1, SO1->getType(), &GEP, this);
7471 } else if (GO1->getType()->getPrimitiveSize() == PS) {
7472 // Convert SO1 to GO1's type.
7473 SO1 = InsertCastToIntPtrTy(SO1, GO1->getType(), &GEP, this);
7475 const Type *PT = TD->getIntPtrType();
7476 SO1 = InsertCastToIntPtrTy(SO1, PT, &GEP, this);
7477 GO1 = InsertCastToIntPtrTy(GO1, PT, &GEP, this);
7481 if (isa<Constant>(SO1) && isa<Constant>(GO1))
7482 Sum = ConstantExpr::getAdd(cast<Constant>(SO1), cast<Constant>(GO1));
7484 Sum = BinaryOperator::createAdd(SO1, GO1, PtrOp->getName()+".sum");
7485 InsertNewInstBefore(cast<Instruction>(Sum), GEP);
7489 // Recycle the GEP we already have if possible.
7490 if (SrcGEPOperands.size() == 2) {
7491 GEP.setOperand(0, SrcGEPOperands[0]);
7492 GEP.setOperand(1, Sum);
7495 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7496 SrcGEPOperands.end()-1);
7497 Indices.push_back(Sum);
7498 Indices.insert(Indices.end(), GEP.op_begin()+2, GEP.op_end());
7500 } else if (isa<Constant>(*GEP.idx_begin()) &&
7501 cast<Constant>(*GEP.idx_begin())->isNullValue() &&
7502 SrcGEPOperands.size() != 1) {
7503 // Otherwise we can do the fold if the first index of the GEP is a zero
7504 Indices.insert(Indices.end(), SrcGEPOperands.begin()+1,
7505 SrcGEPOperands.end());
7506 Indices.insert(Indices.end(), GEP.idx_begin()+1, GEP.idx_end());
7509 if (!Indices.empty())
7510 return new GetElementPtrInst(SrcGEPOperands[0], Indices, GEP.getName());
7512 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(PtrOp)) {
7513 // GEP of global variable. If all of the indices for this GEP are
7514 // constants, we can promote this to a constexpr instead of an instruction.
7516 // Scan for nonconstants...
7517 std::vector<Constant*> Indices;
7518 User::op_iterator I = GEP.idx_begin(), E = GEP.idx_end();
7519 for (; I != E && isa<Constant>(*I); ++I)
7520 Indices.push_back(cast<Constant>(*I));
7522 if (I == E) { // If they are all constants...
7523 Constant *CE = ConstantExpr::getGetElementPtr(GV, Indices);
7525 // Replace all uses of the GEP with the new constexpr...
7526 return ReplaceInstUsesWith(GEP, CE);
7528 } else if (Value *X = getBitCastOperand(PtrOp)) { // Is the operand a cast?
7529 if (!isa<PointerType>(X->getType())) {
7530 // Not interesting. Source pointer must be a cast from pointer.
7531 } else if (HasZeroPointerIndex) {
7532 // transform: GEP (cast [10 x ubyte]* X to [0 x ubyte]*), long 0, ...
7533 // into : GEP [10 x ubyte]* X, long 0, ...
7535 // This occurs when the program declares an array extern like "int X[];"
7537 const PointerType *CPTy = cast<PointerType>(PtrOp->getType());
7538 const PointerType *XTy = cast<PointerType>(X->getType());
7539 if (const ArrayType *XATy =
7540 dyn_cast<ArrayType>(XTy->getElementType()))
7541 if (const ArrayType *CATy =
7542 dyn_cast<ArrayType>(CPTy->getElementType()))
7543 if (CATy->getElementType() == XATy->getElementType()) {
7544 // At this point, we know that the cast source type is a pointer
7545 // to an array of the same type as the destination pointer
7546 // array. Because the array type is never stepped over (there
7547 // is a leading zero) we can fold the cast into this GEP.
7548 GEP.setOperand(0, X);
7551 } else if (GEP.getNumOperands() == 2) {
7552 // Transform things like:
7553 // %t = getelementptr ubyte* cast ([2 x int]* %str to uint*), uint %V
7554 // into: %t1 = getelementptr [2 x int*]* %str, int 0, uint %V; cast
7555 const Type *SrcElTy = cast<PointerType>(X->getType())->getElementType();
7556 const Type *ResElTy=cast<PointerType>(PtrOp->getType())->getElementType();
7557 if (isa<ArrayType>(SrcElTy) &&
7558 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType()) ==
7559 TD->getTypeSize(ResElTy)) {
7560 Value *V = InsertNewInstBefore(
7561 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
7562 GEP.getOperand(1), GEP.getName()), GEP);
7563 // V and GEP are both pointer types --> BitCast
7564 return new BitCastInst(V, GEP.getType());
7567 // Transform things like:
7568 // getelementptr sbyte* cast ([100 x double]* X to sbyte*), int %tmp
7569 // (where tmp = 8*tmp2) into:
7570 // getelementptr [100 x double]* %arr, int 0, int %tmp.2
7572 if (isa<ArrayType>(SrcElTy) &&
7573 (ResElTy == Type::SByteTy || ResElTy == Type::UByteTy)) {
7574 uint64_t ArrayEltSize =
7575 TD->getTypeSize(cast<ArrayType>(SrcElTy)->getElementType());
7577 // Check to see if "tmp" is a scale by a multiple of ArrayEltSize. We
7578 // allow either a mul, shift, or constant here.
7580 ConstantInt *Scale = 0;
7581 if (ArrayEltSize == 1) {
7582 NewIdx = GEP.getOperand(1);
7583 Scale = ConstantInt::get(NewIdx->getType(), 1);
7584 } else if (ConstantInt *CI = dyn_cast<ConstantInt>(GEP.getOperand(1))) {
7585 NewIdx = ConstantInt::get(CI->getType(), 1);
7587 } else if (Instruction *Inst =dyn_cast<Instruction>(GEP.getOperand(1))){
7588 if (Inst->getOpcode() == Instruction::Shl &&
7589 isa<ConstantInt>(Inst->getOperand(1))) {
7591 cast<ConstantInt>(Inst->getOperand(1))->getZExtValue();
7592 if (Inst->getType()->isSigned())
7593 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7595 Scale = ConstantInt::get(Inst->getType(), 1ULL << ShAmt);
7596 NewIdx = Inst->getOperand(0);
7597 } else if (Inst->getOpcode() == Instruction::Mul &&
7598 isa<ConstantInt>(Inst->getOperand(1))) {
7599 Scale = cast<ConstantInt>(Inst->getOperand(1));
7600 NewIdx = Inst->getOperand(0);
7604 // If the index will be to exactly the right offset with the scale taken
7605 // out, perform the transformation.
7606 if (Scale && Scale->getZExtValue() % ArrayEltSize == 0) {
7607 if (isa<ConstantInt>(Scale))
7608 Scale = ConstantInt::get(Scale->getType(),
7609 Scale->getZExtValue() / ArrayEltSize);
7610 if (Scale->getZExtValue() != 1) {
7611 Constant *C = ConstantExpr::getIntegerCast(Scale, NewIdx->getType(),
7613 Instruction *Sc = BinaryOperator::createMul(NewIdx, C, "idxscale");
7614 NewIdx = InsertNewInstBefore(Sc, GEP);
7617 // Insert the new GEP instruction.
7618 Instruction *NewGEP =
7619 new GetElementPtrInst(X, Constant::getNullValue(Type::IntTy),
7620 NewIdx, GEP.getName());
7621 NewGEP = InsertNewInstBefore(NewGEP, GEP);
7622 // The NewGEP must be pointer typed, so must the old one -> BitCast
7623 return new BitCastInst(NewGEP, GEP.getType());
7632 Instruction *InstCombiner::visitAllocationInst(AllocationInst &AI) {
7633 // Convert: malloc Ty, C - where C is a constant != 1 into: malloc [C x Ty], 1
7634 if (AI.isArrayAllocation()) // Check C != 1
7635 if (const ConstantInt *C = dyn_cast<ConstantInt>(AI.getArraySize())) {
7637 ArrayType::get(AI.getAllocatedType(), C->getZExtValue());
7638 AllocationInst *New = 0;
7640 // Create and insert the replacement instruction...
7641 if (isa<MallocInst>(AI))
7642 New = new MallocInst(NewTy, 0, AI.getAlignment(), AI.getName());
7644 assert(isa<AllocaInst>(AI) && "Unknown type of allocation inst!");
7645 New = new AllocaInst(NewTy, 0, AI.getAlignment(), AI.getName());
7648 InsertNewInstBefore(New, AI);
7650 // Scan to the end of the allocation instructions, to skip over a block of
7651 // allocas if possible...
7653 BasicBlock::iterator It = New;
7654 while (isa<AllocationInst>(*It)) ++It;
7656 // Now that I is pointing to the first non-allocation-inst in the block,
7657 // insert our getelementptr instruction...
7659 Value *NullIdx = Constant::getNullValue(Type::IntTy);
7660 Value *V = new GetElementPtrInst(New, NullIdx, NullIdx,
7661 New->getName()+".sub", It);
7663 // Now make everything use the getelementptr instead of the original
7665 return ReplaceInstUsesWith(AI, V);
7666 } else if (isa<UndefValue>(AI.getArraySize())) {
7667 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7670 // If alloca'ing a zero byte object, replace the alloca with a null pointer.
7671 // Note that we only do this for alloca's, because malloc should allocate and
7672 // return a unique pointer, even for a zero byte allocation.
7673 if (isa<AllocaInst>(AI) && AI.getAllocatedType()->isSized() &&
7674 TD->getTypeSize(AI.getAllocatedType()) == 0)
7675 return ReplaceInstUsesWith(AI, Constant::getNullValue(AI.getType()));
7680 Instruction *InstCombiner::visitFreeInst(FreeInst &FI) {
7681 Value *Op = FI.getOperand(0);
7683 // Change free <ty>* (cast <ty2>* X to <ty>*) into free <ty2>* X
7684 if (CastInst *CI = dyn_cast<CastInst>(Op))
7685 if (isa<PointerType>(CI->getOperand(0)->getType())) {
7686 FI.setOperand(0, CI->getOperand(0));
7690 // free undef -> unreachable.
7691 if (isa<UndefValue>(Op)) {
7692 // Insert a new store to null because we cannot modify the CFG here.
7693 new StoreInst(ConstantBool::getTrue(),
7694 UndefValue::get(PointerType::get(Type::BoolTy)), &FI);
7695 return EraseInstFromFunction(FI);
7698 // If we have 'free null' delete the instruction. This can happen in stl code
7699 // when lots of inlining happens.
7700 if (isa<ConstantPointerNull>(Op))
7701 return EraseInstFromFunction(FI);
7707 /// InstCombineLoadCast - Fold 'load (cast P)' -> cast (load P)' when possible.
7708 static Instruction *InstCombineLoadCast(InstCombiner &IC, LoadInst &LI) {
7709 User *CI = cast<User>(LI.getOperand(0));
7710 Value *CastOp = CI->getOperand(0);
7712 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7713 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7714 const Type *SrcPTy = SrcTy->getElementType();
7716 if (DestPTy->isInteger() || isa<PointerType>(DestPTy) ||
7717 isa<PackedType>(DestPTy)) {
7718 // If the source is an array, the code below will not succeed. Check to
7719 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7721 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7722 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7723 if (ASrcTy->getNumElements() != 0) {
7724 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
7725 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7726 SrcTy = cast<PointerType>(CastOp->getType());
7727 SrcPTy = SrcTy->getElementType();
7730 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy) ||
7731 isa<PackedType>(SrcPTy)) &&
7732 // Do not allow turning this into a load of an integer, which is then
7733 // casted to a pointer, this pessimizes pointer analysis a lot.
7734 (isa<PointerType>(SrcPTy) == isa<PointerType>(LI.getType())) &&
7735 IC.getTargetData().getTypeSize(SrcPTy) ==
7736 IC.getTargetData().getTypeSize(DestPTy)) {
7738 // Okay, we are casting from one integer or pointer type to another of
7739 // the same size. Instead of casting the pointer before the load, cast
7740 // the result of the loaded value.
7741 Value *NewLoad = IC.InsertNewInstBefore(new LoadInst(CastOp,
7743 LI.isVolatile()),LI);
7744 // Now cast the result of the load.
7745 return new BitCastInst(NewLoad, LI.getType());
7752 /// isSafeToLoadUnconditionally - Return true if we know that executing a load
7753 /// from this value cannot trap. If it is not obviously safe to load from the
7754 /// specified pointer, we do a quick local scan of the basic block containing
7755 /// ScanFrom, to determine if the address is already accessed.
7756 static bool isSafeToLoadUnconditionally(Value *V, Instruction *ScanFrom) {
7757 // If it is an alloca or global variable, it is always safe to load from.
7758 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
7760 // Otherwise, be a little bit agressive by scanning the local block where we
7761 // want to check to see if the pointer is already being loaded or stored
7762 // from/to. If so, the previous load or store would have already trapped,
7763 // so there is no harm doing an extra load (also, CSE will later eliminate
7764 // the load entirely).
7765 BasicBlock::iterator BBI = ScanFrom, E = ScanFrom->getParent()->begin();
7770 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7771 if (LI->getOperand(0) == V) return true;
7772 } else if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
7773 if (SI->getOperand(1) == V) return true;
7779 Instruction *InstCombiner::visitLoadInst(LoadInst &LI) {
7780 Value *Op = LI.getOperand(0);
7782 // load (cast X) --> cast (load X) iff safe
7783 if (isa<CastInst>(Op))
7784 if (Instruction *Res = InstCombineLoadCast(*this, LI))
7787 // None of the following transforms are legal for volatile loads.
7788 if (LI.isVolatile()) return 0;
7790 if (&LI.getParent()->front() != &LI) {
7791 BasicBlock::iterator BBI = &LI; --BBI;
7792 // If the instruction immediately before this is a store to the same
7793 // address, do a simple form of store->load forwarding.
7794 if (StoreInst *SI = dyn_cast<StoreInst>(BBI))
7795 if (SI->getOperand(1) == LI.getOperand(0))
7796 return ReplaceInstUsesWith(LI, SI->getOperand(0));
7797 if (LoadInst *LIB = dyn_cast<LoadInst>(BBI))
7798 if (LIB->getOperand(0) == LI.getOperand(0))
7799 return ReplaceInstUsesWith(LI, LIB);
7802 if (GetElementPtrInst *GEPI = dyn_cast<GetElementPtrInst>(Op))
7803 if (isa<ConstantPointerNull>(GEPI->getOperand(0)) ||
7804 isa<UndefValue>(GEPI->getOperand(0))) {
7805 // Insert a new store to null instruction before the load to indicate
7806 // that this code is not reachable. We do this instead of inserting
7807 // an unreachable instruction directly because we cannot modify the
7809 new StoreInst(UndefValue::get(LI.getType()),
7810 Constant::getNullValue(Op->getType()), &LI);
7811 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7814 if (Constant *C = dyn_cast<Constant>(Op)) {
7815 // load null/undef -> undef
7816 if ((C->isNullValue() || isa<UndefValue>(C))) {
7817 // Insert a new store to null instruction before the load to indicate that
7818 // this code is not reachable. We do this instead of inserting an
7819 // unreachable instruction directly because we cannot modify the CFG.
7820 new StoreInst(UndefValue::get(LI.getType()),
7821 Constant::getNullValue(Op->getType()), &LI);
7822 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7825 // Instcombine load (constant global) into the value loaded.
7826 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(Op))
7827 if (GV->isConstant() && !GV->isExternal())
7828 return ReplaceInstUsesWith(LI, GV->getInitializer());
7830 // Instcombine load (constantexpr_GEP global, 0, ...) into the value loaded.
7831 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Op))
7832 if (CE->getOpcode() == Instruction::GetElementPtr) {
7833 if (GlobalVariable *GV = dyn_cast<GlobalVariable>(CE->getOperand(0)))
7834 if (GV->isConstant() && !GV->isExternal())
7836 ConstantFoldLoadThroughGEPConstantExpr(GV->getInitializer(), CE))
7837 return ReplaceInstUsesWith(LI, V);
7838 if (CE->getOperand(0)->isNullValue()) {
7839 // Insert a new store to null instruction before the load to indicate
7840 // that this code is not reachable. We do this instead of inserting
7841 // an unreachable instruction directly because we cannot modify the
7843 new StoreInst(UndefValue::get(LI.getType()),
7844 Constant::getNullValue(Op->getType()), &LI);
7845 return ReplaceInstUsesWith(LI, UndefValue::get(LI.getType()));
7848 } else if (CE->isCast()) {
7849 if (Instruction *Res = InstCombineLoadCast(*this, LI))
7854 if (Op->hasOneUse()) {
7855 // Change select and PHI nodes to select values instead of addresses: this
7856 // helps alias analysis out a lot, allows many others simplifications, and
7857 // exposes redundancy in the code.
7859 // Note that we cannot do the transformation unless we know that the
7860 // introduced loads cannot trap! Something like this is valid as long as
7861 // the condition is always false: load (select bool %C, int* null, int* %G),
7862 // but it would not be valid if we transformed it to load from null
7865 if (SelectInst *SI = dyn_cast<SelectInst>(Op)) {
7866 // load (select (Cond, &V1, &V2)) --> select(Cond, load &V1, load &V2).
7867 if (isSafeToLoadUnconditionally(SI->getOperand(1), SI) &&
7868 isSafeToLoadUnconditionally(SI->getOperand(2), SI)) {
7869 Value *V1 = InsertNewInstBefore(new LoadInst(SI->getOperand(1),
7870 SI->getOperand(1)->getName()+".val"), LI);
7871 Value *V2 = InsertNewInstBefore(new LoadInst(SI->getOperand(2),
7872 SI->getOperand(2)->getName()+".val"), LI);
7873 return new SelectInst(SI->getCondition(), V1, V2);
7876 // load (select (cond, null, P)) -> load P
7877 if (Constant *C = dyn_cast<Constant>(SI->getOperand(1)))
7878 if (C->isNullValue()) {
7879 LI.setOperand(0, SI->getOperand(2));
7883 // load (select (cond, P, null)) -> load P
7884 if (Constant *C = dyn_cast<Constant>(SI->getOperand(2)))
7885 if (C->isNullValue()) {
7886 LI.setOperand(0, SI->getOperand(1));
7894 /// InstCombineStoreToCast - Fold 'store V, (cast P)' -> store (cast V), P'
7896 static Instruction *InstCombineStoreToCast(InstCombiner &IC, StoreInst &SI) {
7897 User *CI = cast<User>(SI.getOperand(1));
7898 Value *CastOp = CI->getOperand(0);
7900 const Type *DestPTy = cast<PointerType>(CI->getType())->getElementType();
7901 if (const PointerType *SrcTy = dyn_cast<PointerType>(CastOp->getType())) {
7902 const Type *SrcPTy = SrcTy->getElementType();
7904 if (DestPTy->isInteger() || isa<PointerType>(DestPTy)) {
7905 // If the source is an array, the code below will not succeed. Check to
7906 // see if a trivial 'gep P, 0, 0' will help matters. Only do this for
7908 if (const ArrayType *ASrcTy = dyn_cast<ArrayType>(SrcPTy))
7909 if (Constant *CSrc = dyn_cast<Constant>(CastOp))
7910 if (ASrcTy->getNumElements() != 0) {
7911 std::vector<Value*> Idxs(2, Constant::getNullValue(Type::IntTy));
7912 CastOp = ConstantExpr::getGetElementPtr(CSrc, Idxs);
7913 SrcTy = cast<PointerType>(CastOp->getType());
7914 SrcPTy = SrcTy->getElementType();
7917 if ((SrcPTy->isInteger() || isa<PointerType>(SrcPTy)) &&
7918 IC.getTargetData().getTypeSize(SrcPTy) ==
7919 IC.getTargetData().getTypeSize(DestPTy)) {
7921 // Okay, we are casting from one integer or pointer type to another of
7922 // the same size. Instead of casting the pointer before the store, cast
7923 // the value to be stored.
7925 Instruction::CastOps opcode = Instruction::BitCast;
7926 Value *SIOp0 = SI.getOperand(0);
7927 if (isa<PointerType>(SrcPTy)) {
7928 if (SIOp0->getType()->isIntegral())
7929 opcode = Instruction::IntToPtr;
7930 } else if (SrcPTy->isIntegral()) {
7931 if (isa<PointerType>(SIOp0->getType()))
7932 opcode = Instruction::PtrToInt;
7934 if (Constant *C = dyn_cast<Constant>(SIOp0))
7935 NewCast = ConstantExpr::getCast(opcode, C, SrcPTy);
7937 NewCast = IC.InsertNewInstBefore(
7938 CastInst::create(opcode, SIOp0, SrcPTy, SIOp0->getName()+".c"), SI);
7939 return new StoreInst(NewCast, CastOp);
7946 Instruction *InstCombiner::visitStoreInst(StoreInst &SI) {
7947 Value *Val = SI.getOperand(0);
7948 Value *Ptr = SI.getOperand(1);
7950 if (isa<UndefValue>(Ptr)) { // store X, undef -> noop (even if volatile)
7951 EraseInstFromFunction(SI);
7956 // Do really simple DSE, to catch cases where there are several consequtive
7957 // stores to the same location, separated by a few arithmetic operations. This
7958 // situation often occurs with bitfield accesses.
7959 BasicBlock::iterator BBI = &SI;
7960 for (unsigned ScanInsts = 6; BBI != SI.getParent()->begin() && ScanInsts;
7964 if (StoreInst *PrevSI = dyn_cast<StoreInst>(BBI)) {
7965 // Prev store isn't volatile, and stores to the same location?
7966 if (!PrevSI->isVolatile() && PrevSI->getOperand(1) == SI.getOperand(1)) {
7969 EraseInstFromFunction(*PrevSI);
7975 // If this is a load, we have to stop. However, if the loaded value is from
7976 // the pointer we're loading and is producing the pointer we're storing,
7977 // then *this* store is dead (X = load P; store X -> P).
7978 if (LoadInst *LI = dyn_cast<LoadInst>(BBI)) {
7979 if (LI == Val && LI->getOperand(0) == Ptr) {
7980 EraseInstFromFunction(SI);
7984 // Otherwise, this is a load from some other location. Stores before it
7989 // Don't skip over loads or things that can modify memory.
7990 if (BBI->mayWriteToMemory())
7995 if (SI.isVolatile()) return 0; // Don't hack volatile stores.
7997 // store X, null -> turns into 'unreachable' in SimplifyCFG
7998 if (isa<ConstantPointerNull>(Ptr)) {
7999 if (!isa<UndefValue>(Val)) {
8000 SI.setOperand(0, UndefValue::get(Val->getType()));
8001 if (Instruction *U = dyn_cast<Instruction>(Val))
8002 WorkList.push_back(U); // Dropped a use.
8005 return 0; // Do not modify these!
8008 // store undef, Ptr -> noop
8009 if (isa<UndefValue>(Val)) {
8010 EraseInstFromFunction(SI);
8015 // If the pointer destination is a cast, see if we can fold the cast into the
8017 if (isa<CastInst>(Ptr))
8018 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8020 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Ptr))
8022 if (Instruction *Res = InstCombineStoreToCast(*this, SI))
8026 // If this store is the last instruction in the basic block, and if the block
8027 // ends with an unconditional branch, try to move it to the successor block.
8029 if (BranchInst *BI = dyn_cast<BranchInst>(BBI))
8030 if (BI->isUnconditional()) {
8031 // Check to see if the successor block has exactly two incoming edges. If
8032 // so, see if the other predecessor contains a store to the same location.
8033 // if so, insert a PHI node (if needed) and move the stores down.
8034 BasicBlock *Dest = BI->getSuccessor(0);
8036 pred_iterator PI = pred_begin(Dest);
8037 BasicBlock *Other = 0;
8038 if (*PI != BI->getParent())
8041 if (PI != pred_end(Dest)) {
8042 if (*PI != BI->getParent())
8047 if (++PI != pred_end(Dest))
8050 if (Other) { // If only one other pred...
8051 BBI = Other->getTerminator();
8052 // Make sure this other block ends in an unconditional branch and that
8053 // there is an instruction before the branch.
8054 if (isa<BranchInst>(BBI) && cast<BranchInst>(BBI)->isUnconditional() &&
8055 BBI != Other->begin()) {
8057 StoreInst *OtherStore = dyn_cast<StoreInst>(BBI);
8059 // If this instruction is a store to the same location.
8060 if (OtherStore && OtherStore->getOperand(1) == SI.getOperand(1)) {
8061 // Okay, we know we can perform this transformation. Insert a PHI
8062 // node now if we need it.
8063 Value *MergedVal = OtherStore->getOperand(0);
8064 if (MergedVal != SI.getOperand(0)) {
8065 PHINode *PN = new PHINode(MergedVal->getType(), "storemerge");
8066 PN->reserveOperandSpace(2);
8067 PN->addIncoming(SI.getOperand(0), SI.getParent());
8068 PN->addIncoming(OtherStore->getOperand(0), Other);
8069 MergedVal = InsertNewInstBefore(PN, Dest->front());
8072 // Advance to a place where it is safe to insert the new store and
8074 BBI = Dest->begin();
8075 while (isa<PHINode>(BBI)) ++BBI;
8076 InsertNewInstBefore(new StoreInst(MergedVal, SI.getOperand(1),
8077 OtherStore->isVolatile()), *BBI);
8079 // Nuke the old stores.
8080 EraseInstFromFunction(SI);
8081 EraseInstFromFunction(*OtherStore);
8093 Instruction *InstCombiner::visitBranchInst(BranchInst &BI) {
8094 // Change br (not X), label True, label False to: br X, label False, True
8096 BasicBlock *TrueDest;
8097 BasicBlock *FalseDest;
8098 if (match(&BI, m_Br(m_Not(m_Value(X)), TrueDest, FalseDest)) &&
8099 !isa<Constant>(X)) {
8100 // Swap Destinations and condition...
8102 BI.setSuccessor(0, FalseDest);
8103 BI.setSuccessor(1, TrueDest);
8107 // Cannonicalize setne -> seteq
8108 Instruction::BinaryOps Op; Value *Y;
8109 if (match(&BI, m_Br(m_SetCond(Op, m_Value(X), m_Value(Y)),
8110 TrueDest, FalseDest)))
8111 if ((Op == Instruction::SetNE || Op == Instruction::SetLE ||
8112 Op == Instruction::SetGE) && BI.getCondition()->hasOneUse()) {
8113 SetCondInst *I = cast<SetCondInst>(BI.getCondition());
8114 std::string Name = I->getName(); I->setName("");
8115 Instruction::BinaryOps NewOpcode = SetCondInst::getInverseCondition(Op);
8116 Value *NewSCC = BinaryOperator::create(NewOpcode, X, Y, Name, I);
8117 // Swap Destinations and condition...
8118 BI.setCondition(NewSCC);
8119 BI.setSuccessor(0, FalseDest);
8120 BI.setSuccessor(1, TrueDest);
8121 removeFromWorkList(I);
8122 I->getParent()->getInstList().erase(I);
8123 WorkList.push_back(cast<Instruction>(NewSCC));
8130 Instruction *InstCombiner::visitSwitchInst(SwitchInst &SI) {
8131 Value *Cond = SI.getCondition();
8132 if (Instruction *I = dyn_cast<Instruction>(Cond)) {
8133 if (I->getOpcode() == Instruction::Add)
8134 if (ConstantInt *AddRHS = dyn_cast<ConstantInt>(I->getOperand(1))) {
8135 // change 'switch (X+4) case 1:' into 'switch (X) case -3'
8136 for (unsigned i = 2, e = SI.getNumOperands(); i != e; i += 2)
8137 SI.setOperand(i,ConstantExpr::getSub(cast<Constant>(SI.getOperand(i)),
8139 SI.setOperand(0, I->getOperand(0));
8140 WorkList.push_back(I);
8147 /// CheapToScalarize - Return true if the value is cheaper to scalarize than it
8148 /// is to leave as a vector operation.
8149 static bool CheapToScalarize(Value *V, bool isConstant) {
8150 if (isa<ConstantAggregateZero>(V))
8152 if (ConstantPacked *C = dyn_cast<ConstantPacked>(V)) {
8153 if (isConstant) return true;
8154 // If all elts are the same, we can extract.
8155 Constant *Op0 = C->getOperand(0);
8156 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8157 if (C->getOperand(i) != Op0)
8161 Instruction *I = dyn_cast<Instruction>(V);
8162 if (!I) return false;
8164 // Insert element gets simplified to the inserted element or is deleted if
8165 // this is constant idx extract element and its a constant idx insertelt.
8166 if (I->getOpcode() == Instruction::InsertElement && isConstant &&
8167 isa<ConstantInt>(I->getOperand(2)))
8169 if (I->getOpcode() == Instruction::Load && I->hasOneUse())
8171 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I))
8172 if (BO->hasOneUse() &&
8173 (CheapToScalarize(BO->getOperand(0), isConstant) ||
8174 CheapToScalarize(BO->getOperand(1), isConstant)))
8180 /// getShuffleMask - Read and decode a shufflevector mask. It turns undef
8181 /// elements into values that are larger than the #elts in the input.
8182 static std::vector<unsigned> getShuffleMask(const ShuffleVectorInst *SVI) {
8183 unsigned NElts = SVI->getType()->getNumElements();
8184 if (isa<ConstantAggregateZero>(SVI->getOperand(2)))
8185 return std::vector<unsigned>(NElts, 0);
8186 if (isa<UndefValue>(SVI->getOperand(2)))
8187 return std::vector<unsigned>(NElts, 2*NElts);
8189 std::vector<unsigned> Result;
8190 const ConstantPacked *CP = cast<ConstantPacked>(SVI->getOperand(2));
8191 for (unsigned i = 0, e = CP->getNumOperands(); i != e; ++i)
8192 if (isa<UndefValue>(CP->getOperand(i)))
8193 Result.push_back(NElts*2); // undef -> 8
8195 Result.push_back(cast<ConstantInt>(CP->getOperand(i))->getZExtValue());
8199 /// FindScalarElement - Given a vector and an element number, see if the scalar
8200 /// value is already around as a register, for example if it were inserted then
8201 /// extracted from the vector.
8202 static Value *FindScalarElement(Value *V, unsigned EltNo) {
8203 assert(isa<PackedType>(V->getType()) && "Not looking at a vector?");
8204 const PackedType *PTy = cast<PackedType>(V->getType());
8205 unsigned Width = PTy->getNumElements();
8206 if (EltNo >= Width) // Out of range access.
8207 return UndefValue::get(PTy->getElementType());
8209 if (isa<UndefValue>(V))
8210 return UndefValue::get(PTy->getElementType());
8211 else if (isa<ConstantAggregateZero>(V))
8212 return Constant::getNullValue(PTy->getElementType());
8213 else if (ConstantPacked *CP = dyn_cast<ConstantPacked>(V))
8214 return CP->getOperand(EltNo);
8215 else if (InsertElementInst *III = dyn_cast<InsertElementInst>(V)) {
8216 // If this is an insert to a variable element, we don't know what it is.
8217 if (!isa<ConstantInt>(III->getOperand(2)))
8219 unsigned IIElt = cast<ConstantInt>(III->getOperand(2))->getZExtValue();
8221 // If this is an insert to the element we are looking for, return the
8224 return III->getOperand(1);
8226 // Otherwise, the insertelement doesn't modify the value, recurse on its
8228 return FindScalarElement(III->getOperand(0), EltNo);
8229 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(V)) {
8230 unsigned InEl = getShuffleMask(SVI)[EltNo];
8232 return FindScalarElement(SVI->getOperand(0), InEl);
8233 else if (InEl < Width*2)
8234 return FindScalarElement(SVI->getOperand(1), InEl - Width);
8236 return UndefValue::get(PTy->getElementType());
8239 // Otherwise, we don't know.
8243 Instruction *InstCombiner::visitExtractElementInst(ExtractElementInst &EI) {
8245 // If packed val is undef, replace extract with scalar undef.
8246 if (isa<UndefValue>(EI.getOperand(0)))
8247 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8249 // If packed val is constant 0, replace extract with scalar 0.
8250 if (isa<ConstantAggregateZero>(EI.getOperand(0)))
8251 return ReplaceInstUsesWith(EI, Constant::getNullValue(EI.getType()));
8253 if (ConstantPacked *C = dyn_cast<ConstantPacked>(EI.getOperand(0))) {
8254 // If packed val is constant with uniform operands, replace EI
8255 // with that operand
8256 Constant *op0 = C->getOperand(0);
8257 for (unsigned i = 1; i < C->getNumOperands(); ++i)
8258 if (C->getOperand(i) != op0) {
8263 return ReplaceInstUsesWith(EI, op0);
8266 // If extracting a specified index from the vector, see if we can recursively
8267 // find a previously computed scalar that was inserted into the vector.
8268 if (ConstantInt *IdxC = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8269 // This instruction only demands the single element from the input vector.
8270 // If the input vector has a single use, simplify it based on this use
8272 uint64_t IndexVal = IdxC->getZExtValue();
8273 if (EI.getOperand(0)->hasOneUse()) {
8275 if (Value *V = SimplifyDemandedVectorElts(EI.getOperand(0),
8278 EI.setOperand(0, V);
8283 if (Value *Elt = FindScalarElement(EI.getOperand(0), IndexVal))
8284 return ReplaceInstUsesWith(EI, Elt);
8287 if (Instruction *I = dyn_cast<Instruction>(EI.getOperand(0))) {
8288 if (I->hasOneUse()) {
8289 // Push extractelement into predecessor operation if legal and
8290 // profitable to do so
8291 if (BinaryOperator *BO = dyn_cast<BinaryOperator>(I)) {
8292 bool isConstantElt = isa<ConstantInt>(EI.getOperand(1));
8293 if (CheapToScalarize(BO, isConstantElt)) {
8294 ExtractElementInst *newEI0 =
8295 new ExtractElementInst(BO->getOperand(0), EI.getOperand(1),
8296 EI.getName()+".lhs");
8297 ExtractElementInst *newEI1 =
8298 new ExtractElementInst(BO->getOperand(1), EI.getOperand(1),
8299 EI.getName()+".rhs");
8300 InsertNewInstBefore(newEI0, EI);
8301 InsertNewInstBefore(newEI1, EI);
8302 return BinaryOperator::create(BO->getOpcode(), newEI0, newEI1);
8304 } else if (isa<LoadInst>(I)) {
8305 Value *Ptr = InsertCastBefore(Instruction::BitCast, I->getOperand(0),
8306 PointerType::get(EI.getType()), EI);
8307 GetElementPtrInst *GEP =
8308 new GetElementPtrInst(Ptr, EI.getOperand(1), I->getName() + ".gep");
8309 InsertNewInstBefore(GEP, EI);
8310 return new LoadInst(GEP);
8313 if (InsertElementInst *IE = dyn_cast<InsertElementInst>(I)) {
8314 // Extracting the inserted element?
8315 if (IE->getOperand(2) == EI.getOperand(1))
8316 return ReplaceInstUsesWith(EI, IE->getOperand(1));
8317 // If the inserted and extracted elements are constants, they must not
8318 // be the same value, extract from the pre-inserted value instead.
8319 if (isa<Constant>(IE->getOperand(2)) &&
8320 isa<Constant>(EI.getOperand(1))) {
8321 AddUsesToWorkList(EI);
8322 EI.setOperand(0, IE->getOperand(0));
8325 } else if (ShuffleVectorInst *SVI = dyn_cast<ShuffleVectorInst>(I)) {
8326 // If this is extracting an element from a shufflevector, figure out where
8327 // it came from and extract from the appropriate input element instead.
8328 if (ConstantInt *Elt = dyn_cast<ConstantInt>(EI.getOperand(1))) {
8329 unsigned SrcIdx = getShuffleMask(SVI)[Elt->getZExtValue()];
8331 if (SrcIdx < SVI->getType()->getNumElements())
8332 Src = SVI->getOperand(0);
8333 else if (SrcIdx < SVI->getType()->getNumElements()*2) {
8334 SrcIdx -= SVI->getType()->getNumElements();
8335 Src = SVI->getOperand(1);
8337 return ReplaceInstUsesWith(EI, UndefValue::get(EI.getType()));
8339 return new ExtractElementInst(Src, SrcIdx);
8346 /// CollectSingleShuffleElements - If V is a shuffle of values that ONLY returns
8347 /// elements from either LHS or RHS, return the shuffle mask and true.
8348 /// Otherwise, return false.
8349 static bool CollectSingleShuffleElements(Value *V, Value *LHS, Value *RHS,
8350 std::vector<Constant*> &Mask) {
8351 assert(V->getType() == LHS->getType() && V->getType() == RHS->getType() &&
8352 "Invalid CollectSingleShuffleElements");
8353 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8355 if (isa<UndefValue>(V)) {
8356 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
8358 } else if (V == LHS) {
8359 for (unsigned i = 0; i != NumElts; ++i)
8360 Mask.push_back(ConstantInt::get(Type::UIntTy, i));
8362 } else if (V == RHS) {
8363 for (unsigned i = 0; i != NumElts; ++i)
8364 Mask.push_back(ConstantInt::get(Type::UIntTy, i+NumElts));
8366 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8367 // If this is an insert of an extract from some other vector, include it.
8368 Value *VecOp = IEI->getOperand(0);
8369 Value *ScalarOp = IEI->getOperand(1);
8370 Value *IdxOp = IEI->getOperand(2);
8372 if (!isa<ConstantInt>(IdxOp))
8374 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8376 if (isa<UndefValue>(ScalarOp)) { // inserting undef into vector.
8377 // Okay, we can handle this if the vector we are insertinting into is
8379 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8380 // If so, update the mask to reflect the inserted undef.
8381 Mask[InsertedIdx] = UndefValue::get(Type::UIntTy);
8384 } else if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)){
8385 if (isa<ConstantInt>(EI->getOperand(1)) &&
8386 EI->getOperand(0)->getType() == V->getType()) {
8387 unsigned ExtractedIdx =
8388 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8390 // This must be extracting from either LHS or RHS.
8391 if (EI->getOperand(0) == LHS || EI->getOperand(0) == RHS) {
8392 // Okay, we can handle this if the vector we are insertinting into is
8394 if (CollectSingleShuffleElements(VecOp, LHS, RHS, Mask)) {
8395 // If so, update the mask to reflect the inserted value.
8396 if (EI->getOperand(0) == LHS) {
8397 Mask[InsertedIdx & (NumElts-1)] =
8398 ConstantInt::get(Type::UIntTy, ExtractedIdx);
8400 assert(EI->getOperand(0) == RHS);
8401 Mask[InsertedIdx & (NumElts-1)] =
8402 ConstantInt::get(Type::UIntTy, ExtractedIdx+NumElts);
8411 // TODO: Handle shufflevector here!
8416 /// CollectShuffleElements - We are building a shuffle of V, using RHS as the
8417 /// RHS of the shuffle instruction, if it is not null. Return a shuffle mask
8418 /// that computes V and the LHS value of the shuffle.
8419 static Value *CollectShuffleElements(Value *V, std::vector<Constant*> &Mask,
8421 assert(isa<PackedType>(V->getType()) &&
8422 (RHS == 0 || V->getType() == RHS->getType()) &&
8423 "Invalid shuffle!");
8424 unsigned NumElts = cast<PackedType>(V->getType())->getNumElements();
8426 if (isa<UndefValue>(V)) {
8427 Mask.assign(NumElts, UndefValue::get(Type::UIntTy));
8429 } else if (isa<ConstantAggregateZero>(V)) {
8430 Mask.assign(NumElts, ConstantInt::get(Type::UIntTy, 0));
8432 } else if (InsertElementInst *IEI = dyn_cast<InsertElementInst>(V)) {
8433 // If this is an insert of an extract from some other vector, include it.
8434 Value *VecOp = IEI->getOperand(0);
8435 Value *ScalarOp = IEI->getOperand(1);
8436 Value *IdxOp = IEI->getOperand(2);
8438 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8439 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8440 EI->getOperand(0)->getType() == V->getType()) {
8441 unsigned ExtractedIdx =
8442 cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8443 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8445 // Either the extracted from or inserted into vector must be RHSVec,
8446 // otherwise we'd end up with a shuffle of three inputs.
8447 if (EI->getOperand(0) == RHS || RHS == 0) {
8448 RHS = EI->getOperand(0);
8449 Value *V = CollectShuffleElements(VecOp, Mask, RHS);
8450 Mask[InsertedIdx & (NumElts-1)] =
8451 ConstantInt::get(Type::UIntTy, NumElts+ExtractedIdx);
8456 Value *V = CollectShuffleElements(EI->getOperand(0), Mask, RHS);
8457 // Everything but the extracted element is replaced with the RHS.
8458 for (unsigned i = 0; i != NumElts; ++i) {
8459 if (i != InsertedIdx)
8460 Mask[i] = ConstantInt::get(Type::UIntTy, NumElts+i);
8465 // If this insertelement is a chain that comes from exactly these two
8466 // vectors, return the vector and the effective shuffle.
8467 if (CollectSingleShuffleElements(IEI, EI->getOperand(0), RHS, Mask))
8468 return EI->getOperand(0);
8473 // TODO: Handle shufflevector here!
8475 // Otherwise, can't do anything fancy. Return an identity vector.
8476 for (unsigned i = 0; i != NumElts; ++i)
8477 Mask.push_back(ConstantInt::get(Type::UIntTy, i));
8481 Instruction *InstCombiner::visitInsertElementInst(InsertElementInst &IE) {
8482 Value *VecOp = IE.getOperand(0);
8483 Value *ScalarOp = IE.getOperand(1);
8484 Value *IdxOp = IE.getOperand(2);
8486 // If the inserted element was extracted from some other vector, and if the
8487 // indexes are constant, try to turn this into a shufflevector operation.
8488 if (ExtractElementInst *EI = dyn_cast<ExtractElementInst>(ScalarOp)) {
8489 if (isa<ConstantInt>(EI->getOperand(1)) && isa<ConstantInt>(IdxOp) &&
8490 EI->getOperand(0)->getType() == IE.getType()) {
8491 unsigned NumVectorElts = IE.getType()->getNumElements();
8492 unsigned ExtractedIdx=cast<ConstantInt>(EI->getOperand(1))->getZExtValue();
8493 unsigned InsertedIdx = cast<ConstantInt>(IdxOp)->getZExtValue();
8495 if (ExtractedIdx >= NumVectorElts) // Out of range extract.
8496 return ReplaceInstUsesWith(IE, VecOp);
8498 if (InsertedIdx >= NumVectorElts) // Out of range insert.
8499 return ReplaceInstUsesWith(IE, UndefValue::get(IE.getType()));
8501 // If we are extracting a value from a vector, then inserting it right
8502 // back into the same place, just use the input vector.
8503 if (EI->getOperand(0) == VecOp && ExtractedIdx == InsertedIdx)
8504 return ReplaceInstUsesWith(IE, VecOp);
8506 // We could theoretically do this for ANY input. However, doing so could
8507 // turn chains of insertelement instructions into a chain of shufflevector
8508 // instructions, and right now we do not merge shufflevectors. As such,
8509 // only do this in a situation where it is clear that there is benefit.
8510 if (isa<UndefValue>(VecOp) || isa<ConstantAggregateZero>(VecOp)) {
8511 // Turn this into shuffle(EIOp0, VecOp, Mask). The result has all of
8512 // the values of VecOp, except then one read from EIOp0.
8513 // Build a new shuffle mask.
8514 std::vector<Constant*> Mask;
8515 if (isa<UndefValue>(VecOp))
8516 Mask.assign(NumVectorElts, UndefValue::get(Type::UIntTy));
8518 assert(isa<ConstantAggregateZero>(VecOp) && "Unknown thing");
8519 Mask.assign(NumVectorElts, ConstantInt::get(Type::UIntTy,
8522 Mask[InsertedIdx] = ConstantInt::get(Type::UIntTy, ExtractedIdx);
8523 return new ShuffleVectorInst(EI->getOperand(0), VecOp,
8524 ConstantPacked::get(Mask));
8527 // If this insertelement isn't used by some other insertelement, turn it
8528 // (and any insertelements it points to), into one big shuffle.
8529 if (!IE.hasOneUse() || !isa<InsertElementInst>(IE.use_back())) {
8530 std::vector<Constant*> Mask;
8532 Value *LHS = CollectShuffleElements(&IE, Mask, RHS);
8533 if (RHS == 0) RHS = UndefValue::get(LHS->getType());
8534 // We now have a shuffle of LHS, RHS, Mask.
8535 return new ShuffleVectorInst(LHS, RHS, ConstantPacked::get(Mask));
8544 Instruction *InstCombiner::visitShuffleVectorInst(ShuffleVectorInst &SVI) {
8545 Value *LHS = SVI.getOperand(0);
8546 Value *RHS = SVI.getOperand(1);
8547 std::vector<unsigned> Mask = getShuffleMask(&SVI);
8549 bool MadeChange = false;
8551 // Undefined shuffle mask -> undefined value.
8552 if (isa<UndefValue>(SVI.getOperand(2)))
8553 return ReplaceInstUsesWith(SVI, UndefValue::get(SVI.getType()));
8555 // TODO: If we have shuffle(x, undef, mask) and any elements of mask refer to
8556 // the undef, change them to undefs.
8558 // Canonicalize shuffle(x ,x,mask) -> shuffle(x, undef,mask')
8559 // Canonicalize shuffle(undef,x,mask) -> shuffle(x, undef,mask').
8560 if (LHS == RHS || isa<UndefValue>(LHS)) {
8561 if (isa<UndefValue>(LHS) && LHS == RHS) {
8562 // shuffle(undef,undef,mask) -> undef.
8563 return ReplaceInstUsesWith(SVI, LHS);
8566 // Remap any references to RHS to use LHS.
8567 std::vector<Constant*> Elts;
8568 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8570 Elts.push_back(UndefValue::get(Type::UIntTy));
8572 if ((Mask[i] >= e && isa<UndefValue>(RHS)) ||
8573 (Mask[i] < e && isa<UndefValue>(LHS)))
8574 Mask[i] = 2*e; // Turn into undef.
8576 Mask[i] &= (e-1); // Force to LHS.
8577 Elts.push_back(ConstantInt::get(Type::UIntTy, Mask[i]));
8580 SVI.setOperand(0, SVI.getOperand(1));
8581 SVI.setOperand(1, UndefValue::get(RHS->getType()));
8582 SVI.setOperand(2, ConstantPacked::get(Elts));
8583 LHS = SVI.getOperand(0);
8584 RHS = SVI.getOperand(1);
8588 // Analyze the shuffle, are the LHS or RHS and identity shuffles?
8589 bool isLHSID = true, isRHSID = true;
8591 for (unsigned i = 0, e = Mask.size(); i != e; ++i) {
8592 if (Mask[i] >= e*2) continue; // Ignore undef values.
8593 // Is this an identity shuffle of the LHS value?
8594 isLHSID &= (Mask[i] == i);
8596 // Is this an identity shuffle of the RHS value?
8597 isRHSID &= (Mask[i]-e == i);
8600 // Eliminate identity shuffles.
8601 if (isLHSID) return ReplaceInstUsesWith(SVI, LHS);
8602 if (isRHSID) return ReplaceInstUsesWith(SVI, RHS);
8604 // If the LHS is a shufflevector itself, see if we can combine it with this
8605 // one without producing an unusual shuffle. Here we are really conservative:
8606 // we are absolutely afraid of producing a shuffle mask not in the input
8607 // program, because the code gen may not be smart enough to turn a merged
8608 // shuffle into two specific shuffles: it may produce worse code. As such,
8609 // we only merge two shuffles if the result is one of the two input shuffle
8610 // masks. In this case, merging the shuffles just removes one instruction,
8611 // which we know is safe. This is good for things like turning:
8612 // (splat(splat)) -> splat.
8613 if (ShuffleVectorInst *LHSSVI = dyn_cast<ShuffleVectorInst>(LHS)) {
8614 if (isa<UndefValue>(RHS)) {
8615 std::vector<unsigned> LHSMask = getShuffleMask(LHSSVI);
8617 std::vector<unsigned> NewMask;
8618 for (unsigned i = 0, e = Mask.size(); i != e; ++i)
8620 NewMask.push_back(2*e);
8622 NewMask.push_back(LHSMask[Mask[i]]);
8624 // If the result mask is equal to the src shuffle or this shuffle mask, do
8626 if (NewMask == LHSMask || NewMask == Mask) {
8627 std::vector<Constant*> Elts;
8628 for (unsigned i = 0, e = NewMask.size(); i != e; ++i) {
8629 if (NewMask[i] >= e*2) {
8630 Elts.push_back(UndefValue::get(Type::UIntTy));
8632 Elts.push_back(ConstantInt::get(Type::UIntTy, NewMask[i]));
8635 return new ShuffleVectorInst(LHSSVI->getOperand(0),
8636 LHSSVI->getOperand(1),
8637 ConstantPacked::get(Elts));
8642 return MadeChange ? &SVI : 0;
8647 void InstCombiner::removeFromWorkList(Instruction *I) {
8648 WorkList.erase(std::remove(WorkList.begin(), WorkList.end(), I),
8653 /// TryToSinkInstruction - Try to move the specified instruction from its
8654 /// current block into the beginning of DestBlock, which can only happen if it's
8655 /// safe to move the instruction past all of the instructions between it and the
8656 /// end of its block.
8657 static bool TryToSinkInstruction(Instruction *I, BasicBlock *DestBlock) {
8658 assert(I->hasOneUse() && "Invariants didn't hold!");
8660 // Cannot move control-flow-involving, volatile loads, vaarg, etc.
8661 if (isa<PHINode>(I) || I->mayWriteToMemory()) return false;
8663 // Do not sink alloca instructions out of the entry block.
8664 if (isa<AllocaInst>(I) && I->getParent() == &DestBlock->getParent()->front())
8667 // We can only sink load instructions if there is nothing between the load and
8668 // the end of block that could change the value.
8669 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
8670 for (BasicBlock::iterator Scan = LI, E = LI->getParent()->end();
8672 if (Scan->mayWriteToMemory())
8676 BasicBlock::iterator InsertPos = DestBlock->begin();
8677 while (isa<PHINode>(InsertPos)) ++InsertPos;
8679 I->moveBefore(InsertPos);
8684 /// OptimizeConstantExpr - Given a constant expression and target data layout
8685 /// information, symbolically evaluate the constant expr to something simpler
8687 static Constant *OptimizeConstantExpr(ConstantExpr *CE, const TargetData *TD) {
8690 Constant *Ptr = CE->getOperand(0);
8691 if (CE->getOpcode() == Instruction::GetElementPtr && Ptr->isNullValue() &&
8692 cast<PointerType>(Ptr->getType())->getElementType()->isSized()) {
8693 // If this is a constant expr gep that is effectively computing an
8694 // "offsetof", fold it into 'cast int Size to T*' instead of 'gep 0, 0, 12'
8695 bool isFoldableGEP = true;
8696 for (unsigned i = 1, e = CE->getNumOperands(); i != e; ++i)
8697 if (!isa<ConstantInt>(CE->getOperand(i)))
8698 isFoldableGEP = false;
8699 if (isFoldableGEP) {
8700 std::vector<Value*> Ops(CE->op_begin()+1, CE->op_end());
8701 uint64_t Offset = TD->getIndexedOffset(Ptr->getType(), Ops);
8702 Constant *C = ConstantInt::get(TD->getIntPtrType(), Offset);
8703 return ConstantExpr::getIntToPtr(C, CE->getType());
8711 /// AddReachableCodeToWorklist - Walk the function in depth-first order, adding
8712 /// all reachable code to the worklist.
8714 /// This has a couple of tricks to make the code faster and more powerful. In
8715 /// particular, we constant fold and DCE instructions as we go, to avoid adding
8716 /// them to the worklist (this significantly speeds up instcombine on code where
8717 /// many instructions are dead or constant). Additionally, if we find a branch
8718 /// whose condition is a known constant, we only visit the reachable successors.
8720 static void AddReachableCodeToWorklist(BasicBlock *BB,
8721 std::set<BasicBlock*> &Visited,
8722 std::vector<Instruction*> &WorkList,
8723 const TargetData *TD) {
8724 // We have now visited this block! If we've already been here, bail out.
8725 if (!Visited.insert(BB).second) return;
8727 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) {
8728 Instruction *Inst = BBI++;
8730 // DCE instruction if trivially dead.
8731 if (isInstructionTriviallyDead(Inst)) {
8733 DOUT << "IC: DCE: " << *Inst;
8734 Inst->eraseFromParent();
8738 // ConstantProp instruction if trivially constant.
8739 if (Constant *C = ConstantFoldInstruction(Inst)) {
8740 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
8741 C = OptimizeConstantExpr(CE, TD);
8742 DOUT << "IC: ConstFold to: " << *C << " from: " << *Inst;
8743 Inst->replaceAllUsesWith(C);
8745 Inst->eraseFromParent();
8749 WorkList.push_back(Inst);
8752 // Recursively visit successors. If this is a branch or switch on a constant,
8753 // only visit the reachable successor.
8754 TerminatorInst *TI = BB->getTerminator();
8755 if (BranchInst *BI = dyn_cast<BranchInst>(TI)) {
8756 if (BI->isConditional() && isa<ConstantBool>(BI->getCondition())) {
8757 bool CondVal = cast<ConstantBool>(BI->getCondition())->getValue();
8758 AddReachableCodeToWorklist(BI->getSuccessor(!CondVal), Visited, WorkList,
8762 } else if (SwitchInst *SI = dyn_cast<SwitchInst>(TI)) {
8763 if (ConstantInt *Cond = dyn_cast<ConstantInt>(SI->getCondition())) {
8764 // See if this is an explicit destination.
8765 for (unsigned i = 1, e = SI->getNumSuccessors(); i != e; ++i)
8766 if (SI->getCaseValue(i) == Cond) {
8767 AddReachableCodeToWorklist(SI->getSuccessor(i), Visited, WorkList,TD);
8771 // Otherwise it is the default destination.
8772 AddReachableCodeToWorklist(SI->getSuccessor(0), Visited, WorkList, TD);
8777 for (unsigned i = 0, e = TI->getNumSuccessors(); i != e; ++i)
8778 AddReachableCodeToWorklist(TI->getSuccessor(i), Visited, WorkList, TD);
8781 bool InstCombiner::runOnFunction(Function &F) {
8782 bool Changed = false;
8783 TD = &getAnalysis<TargetData>();
8786 // Do a depth-first traversal of the function, populate the worklist with
8787 // the reachable instructions. Ignore blocks that are not reachable. Keep
8788 // track of which blocks we visit.
8789 std::set<BasicBlock*> Visited;
8790 AddReachableCodeToWorklist(F.begin(), Visited, WorkList, TD);
8792 // Do a quick scan over the function. If we find any blocks that are
8793 // unreachable, remove any instructions inside of them. This prevents
8794 // the instcombine code from having to deal with some bad special cases.
8795 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
8796 if (!Visited.count(BB)) {
8797 Instruction *Term = BB->getTerminator();
8798 while (Term != BB->begin()) { // Remove instrs bottom-up
8799 BasicBlock::iterator I = Term; --I;
8801 DOUT << "IC: DCE: " << *I;
8804 if (!I->use_empty())
8805 I->replaceAllUsesWith(UndefValue::get(I->getType()));
8806 I->eraseFromParent();
8811 while (!WorkList.empty()) {
8812 Instruction *I = WorkList.back(); // Get an instruction from the worklist
8813 WorkList.pop_back();
8815 // Check to see if we can DCE the instruction.
8816 if (isInstructionTriviallyDead(I)) {
8817 // Add operands to the worklist.
8818 if (I->getNumOperands() < 4)
8819 AddUsesToWorkList(*I);
8822 DOUT << "IC: DCE: " << *I;
8824 I->eraseFromParent();
8825 removeFromWorkList(I);
8829 // Instruction isn't dead, see if we can constant propagate it.
8830 if (Constant *C = ConstantFoldInstruction(I)) {
8831 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C))
8832 C = OptimizeConstantExpr(CE, TD);
8833 DOUT << "IC: ConstFold to: " << *C << " from: " << *I;
8835 // Add operands to the worklist.
8836 AddUsesToWorkList(*I);
8837 ReplaceInstUsesWith(*I, C);
8840 I->eraseFromParent();
8841 removeFromWorkList(I);
8845 // See if we can trivially sink this instruction to a successor basic block.
8846 if (I->hasOneUse()) {
8847 BasicBlock *BB = I->getParent();
8848 BasicBlock *UserParent = cast<Instruction>(I->use_back())->getParent();
8849 if (UserParent != BB) {
8850 bool UserIsSuccessor = false;
8851 // See if the user is one of our successors.
8852 for (succ_iterator SI = succ_begin(BB), E = succ_end(BB); SI != E; ++SI)
8853 if (*SI == UserParent) {
8854 UserIsSuccessor = true;
8858 // If the user is one of our immediate successors, and if that successor
8859 // only has us as a predecessors (we'd have to split the critical edge
8860 // otherwise), we can keep going.
8861 if (UserIsSuccessor && !isa<PHINode>(I->use_back()) &&
8862 next(pred_begin(UserParent)) == pred_end(UserParent))
8863 // Okay, the CFG is simple enough, try to sink this instruction.
8864 Changed |= TryToSinkInstruction(I, UserParent);
8868 // Now that we have an instruction, try combining it to simplify it...
8869 if (Instruction *Result = visit(*I)) {
8871 // Should we replace the old instruction with a new one?
8873 DOUT << "IC: Old = " << *I
8874 << " New = " << *Result;
8876 // Everything uses the new instruction now.
8877 I->replaceAllUsesWith(Result);
8879 // Push the new instruction and any users onto the worklist.
8880 WorkList.push_back(Result);
8881 AddUsersToWorkList(*Result);
8883 // Move the name to the new instruction first...
8884 std::string OldName = I->getName(); I->setName("");
8885 Result->setName(OldName);
8887 // Insert the new instruction into the basic block...
8888 BasicBlock *InstParent = I->getParent();
8889 BasicBlock::iterator InsertPos = I;
8891 if (!isa<PHINode>(Result)) // If combining a PHI, don't insert
8892 while (isa<PHINode>(InsertPos)) // middle of a block of PHIs.
8895 InstParent->getInstList().insert(InsertPos, Result);
8897 // Make sure that we reprocess all operands now that we reduced their
8899 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8900 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8901 WorkList.push_back(OpI);
8903 // Instructions can end up on the worklist more than once. Make sure
8904 // we do not process an instruction that has been deleted.
8905 removeFromWorkList(I);
8907 // Erase the old instruction.
8908 InstParent->getInstList().erase(I);
8910 DOUT << "IC: MOD = " << *I;
8912 // If the instruction was modified, it's possible that it is now dead.
8913 // if so, remove it.
8914 if (isInstructionTriviallyDead(I)) {
8915 // Make sure we process all operands now that we are reducing their
8917 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i)
8918 if (Instruction *OpI = dyn_cast<Instruction>(I->getOperand(i)))
8919 WorkList.push_back(OpI);
8921 // Instructions may end up in the worklist more than once. Erase all
8922 // occurrences of this instruction.
8923 removeFromWorkList(I);
8924 I->eraseFromParent();
8926 WorkList.push_back(Result);
8927 AddUsersToWorkList(*Result);
8937 FunctionPass *llvm::createInstructionCombiningPass() {
8938 return new InstCombiner();